title
stringlengths 14
154
| paper_url
stringlengths 42
42
| authors
listlengths 1
21
| type
stringclasses 3
values | abstract
stringlengths 413
2.52k
| keywords
stringlengths 4
397
| TL;DR
stringlengths 5
250
⌀ | submission_number
int64 2
14.3k
| arxiv_id
stringlengths 10
10
⌀ | embedding
listlengths 384
384
| github_url
stringlengths 0
126
| github_stars
int64 0
55k
| num_models
int64 0
82
| num_datasets
int64 0
7
| num_spaces
int64 0
100
|
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
TexTailor: Customized Text-aligned Texturing via Effective Resampling
|
https://openreview.net/forum?id=1NprT9Kz0d
|
[
"Suin Lee",
"Daeshik Kim"
] |
Poster
|
We present TexTailor, a novel method for generating consistent object textures from textual descriptions. Existing text-to-texture synthesis approaches utilize depth-aware diffusion models to progressively generate images and synthesize textures across predefined multiple viewpoints. However, these approaches lead to a gradual shift in texture properties across viewpoints due to (1) insufficient integration of previously synthesized textures at each viewpoint during the diffusion process and (2) the autoregressive nature of the texture synthesis process. Moreover, the predefined selection of camera positions, which does not account for the object's geometry, limits the effective use of texture information synthesized from different viewpoints, ultimately degrading overall texture consistency. In TexTailor, we address these issues by (1) applying a resampling scheme that repeatedly integrates information from previously synthesized textures within the diffusion process, and (2) fine-tuning a depth-aware diffusion model on these resampled textures. During this process, we observed that using only a few training images restricts the model's original ability to generate high-fidelity images aligned with the conditioning, and therefore propose an performance preservation loss to mitigate this issue. Additionally, we improve the synthesis of view-consistent textures by adaptively adjusting camera positions based on the object's geometry. Experiments on a subset of the Objaverse dataset and the ShapeNet car dataset demonstrate that TexTailor outperforms state-of-the-art methods in synthesizing view-consistent textures.
|
3D texture synthesis, diffusion model, resampling
| null | 11,533
| null |
[
-0.034125447273254395,
-0.05042264983057976,
0.07580787688493729,
-0.002096068812534213,
0.06239906698465347,
-0.042740341275930405,
-0.07124777138233185,
-0.020383892580866814,
0.043135881423950195,
-0.02181437611579895,
-0.055608708411455154,
-0.025385675951838493,
0.009795957244932652,
0.013639932498335838,
0.0030419575050473213,
-0.036508411169052124,
0.08643069863319397,
0.12000416964292526,
-0.08054924756288528,
-0.010478982701897621,
-0.022744476795196533,
0.014974272809922695,
0.013144698925316334,
-0.0595010444521904,
0.054342348128557205,
0.07579350471496582,
0.03169623389840126,
-0.07080356031656265,
0.049041472375392914,
-0.06323445588350296,
0.0018714952748268843,
-0.02517721988260746,
-0.043948885053396225,
0.046131622046232224,
-0.04363395646214485,
0.07552513480186462,
-0.04800589010119438,
0.005558056756854057,
-0.011521494947373867,
-0.017683569341897964,
-0.0016278241528198123,
0.07638686150312424,
-0.024398846551775932,
0.018179118633270264,
0.10108442604541779,
0.017469678074121475,
0.031977903097867966,
0.019795255735516548,
-0.018131062388420105,
0.016611730679869652,
-0.056568726897239685,
-0.04611704871058464,
-0.09069044888019562,
0.009037281386554241,
-0.029915954917669296,
0.09979566931724548,
0.0580565482378006,
-0.0013541162479668856,
0.03289167582988739,
0.05115847289562225,
0.008154913783073425,
0.0021097867283970118,
0.041220348328351974,
0.006654589902609587,
-0.056957609951496124,
-0.009473398327827454,
0.09355863183736801,
0.015672173351049423,
-0.010936228558421135,
0.028485897928476334,
-0.0698828399181366,
0.04392635449767113,
-0.05212275683879852,
-0.013548370450735092,
-0.00961785577237606,
0.040452294051647186,
0.05219266936182976,
-0.05073542147874832,
-0.01979925110936165,
-0.10427465289831161,
0.021761145442724228,
0.010865632444620132,
0.054475001990795135,
-0.012253358028829098,
-0.0481223501265049,
-0.0033476389944553375,
0.011928698979318142,
0.005293248221278191,
0.01610511727631092,
-0.02996397577226162,
-0.013340145349502563,
-0.030206356197595596,
-0.07977088540792465,
-0.017332671210169792,
-0.03706272691488266,
-0.020967228338122368,
-0.0022661476396024227,
-0.06322108209133148,
0.022992722690105438,
0.06820118427276611,
0.02473868988454342,
-0.0684642493724823,
0.0030127819627523422,
0.005296770483255386,
0.0432952381670475,
-0.08596931397914886,
0.09673324972391129,
0.05550654977560043,
-0.04939645156264305,
0.061232443898916245,
0.04710535705089569,
0.021132225170731544,
-0.07718788832426071,
0.01864022947847843,
-0.010930491611361504,
-0.021268771961331367,
-0.03288973867893219,
-0.03364323079586029,
-0.015008291229605675,
0.022854171693325043,
-0.018393542617559433,
0.02356506697833538,
-0.013983295299112797,
0.013599984347820282,
0.015363240614533424,
-0.10526047646999359,
0.00682962266728282,
2.0376074178451226e-33,
-0.0017368346452713013,
-0.028599070385098457,
0.059585168957710266,
0.054877690970897675,
0.08497665077447891,
0.0014818118652328849,
-0.03594933822751045,
-0.01663053035736084,
0.012244818732142448,
-0.03950175642967224,
-0.03281858563423157,
-0.043031658977270126,
-0.07579658180475235,
0.1366308480501175,
-0.019732624292373657,
-0.020867493003606796,
-0.032596655189991,
0.10505874454975128,
0.042384449392557144,
0.03826214373111725,
-0.04221673682332039,
0.00367905106395483,
0.018255217000842094,
-0.04252820089459419,
-0.08019030839204788,
0.08058960735797882,
0.06070184335112572,
-0.006212123204022646,
-0.05424728989601135,
-0.012392440810799599,
-0.11801411211490631,
-0.05128660425543785,
0.03525887057185173,
0.06156078726053238,
0.003062823787331581,
0.024897968396544456,
-0.00397307425737381,
-0.012112571857869625,
-0.0028199178632348776,
0.08242110162973404,
0.006288517732173204,
0.09775464981794357,
0.01198305282741785,
-0.0456506721675396,
-0.07713346928358078,
-0.01836545206606388,
0.04018786922097206,
0.03140505775809288,
-0.13671666383743286,
-0.014586263336241245,
0.03972553461790085,
0.051899224519729614,
0.0007452687714248896,
-0.018721332773566246,
-0.09669773280620575,
-0.08989837765693665,
0.06984442472457886,
-0.05424944683909416,
0.05888304114341736,
-0.0051015829667449,
-0.022340914234519005,
0.05862582102417946,
-0.050855156034231186,
0.05399531498551369,
0.022326674312353134,
-0.01652233488857746,
0.05826306343078613,
-0.00555694242939353,
-0.01795823872089386,
-0.026580624282360077,
-0.05311685428023338,
0.04055878520011902,
-0.01898435316979885,
-0.039043840020895004,
0.040422141551971436,
0.0021720046643167734,
0.011264617554843426,
-0.011600570753216743,
0.024492835626006126,
0.011274168267846107,
-0.08034012466669083,
0.03653410077095032,
0.020439349114894867,
-0.10631633549928665,
-0.09253881126642227,
-0.042387522757053375,
-0.000062556478951592,
-0.06559840589761734,
-0.005197751801460981,
-0.013322784565389156,
-0.0863446295261383,
-0.0401148647069931,
-0.012096135877072811,
-0.0906277522444725,
0.11418846994638443,
-1.710651011003596e-33,
0.05100540816783905,
-0.012289813719689846,
-0.014843638055026531,
0.0894077718257904,
-0.00015427243488375098,
-0.016093438491225243,
0.01663004793226719,
0.03082050196826458,
0.01136308629065752,
-0.16017495095729828,
-0.03669523075222969,
0.028114257380366325,
0.011649081483483315,
-0.08952312171459198,
-0.08099625259637833,
-0.06660369783639908,
0.038571979850530624,
-0.006765733007341623,
-0.019201522693037987,
0.008331045508384705,
0.0422617644071579,
0.04680303484201431,
-0.009957105852663517,
-0.0057028355076909065,
-0.0753941610455513,
0.1165243536233902,
-0.04107549414038658,
0.05891059711575508,
0.023402292281389236,
-0.0434906966984272,
0.053165748715400696,
-0.08074870705604553,
0.08000710606575012,
-0.08188488334417343,
-0.057899992913007736,
0.07260069996118546,
0.028329944238066673,
0.014478395693004131,
-0.08005248010158539,
0.0849071815609932,
0.07455862313508987,
-0.05360414460301399,
-0.027594413608312607,
0.027324674651026726,
-0.027389679104089737,
-0.015480312518775463,
-0.05290970578789711,
-0.09369416534900665,
0.05441867560148239,
0.09724780172109604,
0.0017875756602734327,
0.03159358724951744,
-0.0211966373026371,
-0.015716303139925003,
0.018936118111014366,
-0.06741918623447418,
-0.16240213811397552,
-0.0052282242104411125,
0.08490458130836487,
0.037439145147800446,
-0.0005293184658512473,
-0.08888658881187439,
-0.0002613211690913886,
-0.10195521265268326,
0.0135525306686759,
0.006499094422906637,
-0.021711289882659912,
-0.024079570546746254,
0.083804652094841,
0.06460502743721008,
0.09365762770175934,
-0.029824327677488327,
0.01174839772284031,
0.08020321279764175,
0.07236611843109131,
-0.04729108139872551,
0.052747052162885666,
0.018272938206791878,
0.027656305581331253,
-0.0446489118039608,
0.02285861037671566,
-0.01560871209949255,
-0.020052550360560417,
0.1353418081998825,
0.06380730867385864,
0.03448992595076561,
-0.00963119138032198,
0.003630406456068158,
0.07965674251317978,
0.04136205092072487,
0.011049903929233551,
-0.030298655852675438,
0.008000069297850132,
0.10171627998352051,
-0.03480753302574158,
-4.7850821971451296e-8,
-0.021454067900776863,
0.06546910107135773,
0.007981440983712673,
0.0455697700381279,
-0.07775076478719711,
-0.03885602578520775,
0.048283305019140244,
0.06298021227121353,
0.01769244857132435,
-0.09832793474197388,
0.0731896460056305,
-0.027248702943325043,
-0.00681197177618742,
-0.0526399090886116,
0.04484926909208298,
0.09896320849657059,
-0.03840316832065582,
-0.0070303697139024734,
-0.014143706299364567,
0.003819749690592289,
-0.03846849128603935,
0.031078152358531952,
-0.012544430792331696,
-0.08734925091266632,
0.05094151571393013,
0.028055677190423012,
-0.03222710266709328,
0.053469087928533554,
0.06570915877819061,
-0.047832049429416656,
0.005607051774859428,
0.01393338292837143,
-0.004276372492313385,
0.04037713631987572,
0.008564398624002934,
-0.005650671198964119,
-0.021273173391819,
-0.03971976414322853,
0.004613661672919989,
-0.05898067355155945,
-0.015889715403318405,
0.058532100170850754,
0.0008938760729506612,
0.04932979866862297,
0.014361455105245113,
0.003096642903983593,
0.10215684026479721,
-0.02459518425166607,
-0.049535732716321945,
-0.04407888278365135,
-0.04772547259926796,
0.008403867483139038,
0.000469663820695132,
0.07163810729980469,
0.07826295495033264,
-0.014843628741800785,
0.05376994237303734,
0.01906256005167961,
0.018657157197594643,
0.05012543499469757,
-0.010748500935733318,
0.020335469394922256,
-0.02440454624593258,
-0.007488432805985212
] | 0
| 0
| 0
| 0
|
|
Generalization and Distributed Learning of GFlowNets
|
https://openreview.net/forum?id=PJNhZoCjLh
|
[
"Tiago Silva",
"Amauri H Souza",
"Omar Rivasplata",
"Vikas Garg",
"Samuel Kaski",
"Diego Mesquita"
] |
Poster
|
Conventional wisdom attributes the success of Generative Flow Networks (GFlowNets) to their ability to exploit the compositional structure of the sample space for learning generalizable flow functions (Bengio et al., 2021). Despite the abundance of empirical evidence, formalizing this belief with verifiable non-vacuous statistical guarantees has remained elusive. We address this issue with the first data-dependent generalization bounds for GFlowNets. We also elucidate the negative impact of the state space size on the generalization performance of these models via Azuma-Hoeffding-type oracle PAC-Bayesian inequalities. We leverage our theoretical insights to design a novel distributed learning algorithm for GFlowNets, which we call *Subgraph Asynchronous Learning* (SAL). In a nutshell, SAL utilizes a divide-and-conquer strategy: multiple GFlowNets are trained in parallel on smaller subnetworks of the flow network, and then aggregated with an additional GFlowNet that allocates appropriate flow to each subnetwork. Our experiments with synthetic and real-world problems demonstrate the benefits of SAL over centralized training in terms of mode coverage and distribution matching.
|
GFlowNets
|
We derive non-vacuous statistical guarantees and introduce the first distributed learning method for GFlowNets with network-level parallelization.
| 11,527
| null |
[
-0.08640678972005844,
-0.06512480974197388,
0.09326230734586716,
0.004733944311738014,
0.05976906418800354,
-0.019467711448669434,
-0.03181052953004837,
-0.03232596814632416,
0.0038762351032346487,
-0.0041024587117135525,
-0.10705870389938354,
0.010935583151876926,
0.01885918155312538,
0.006962432060390711,
-0.009135689586400986,
0.023453889414668083,
0.07850863039493561,
0.030502036213874817,
-0.06934786587953568,
-0.06865621358156204,
0.0007463939837180078,
0.02528545632958412,
0.011642860248684883,
-0.007602235302329063,
0.05235967040061951,
-0.0002682847552932799,
-0.002416463103145361,
-0.0686315968632698,
0.03681355342268944,
0.029218971729278564,
0.047784119844436646,
0.05869617685675621,
-0.014176574535667896,
-0.004631251096725464,
-0.050153013318777084,
0.03066752851009369,
-0.05094372108578682,
0.037488628178834915,
0.054218314588069916,
0.07703317701816559,
0.08612900972366333,
0.013574538752436638,
0.0472392737865448,
0.028457947075366974,
-0.002002389868721366,
0.02106183022260666,
0.0019298469414934516,
-0.05204469710588455,
-0.0016777331475168467,
-0.0008242442854680121,
0.02568654902279377,
-0.029885338619351387,
-0.07475432008504868,
0.12152906507253647,
0.02539202757179737,
-0.019858552142977715,
0.021916240453720093,
-0.03957774117588997,
-0.02561170794069767,
0.019311927258968353,
-0.012440433725714684,
-0.1558208018541336,
-0.05591234192252159,
-0.012441464699804783,
-0.07375901937484741,
0.007483975030481815,
0.035708703100681305,
0.1279848963022232,
-0.01989552192389965,
0.002016848186030984,
0.03942794352769852,
0.06955544650554657,
-0.12073986977338791,
0.0549064539372921,
0.025315430015325546,
0.05712570995092392,
0.12292704731225967,
-0.01045638881623745,
0.047893453389406204,
-0.06606890261173248,
0.07421977818012238,
0.12987126410007477,
0.034310586750507355,
-0.03888338431715965,
0.06307270377874374,
-0.017121972516179085,
-0.013178405351936817,
0.05273370072245598,
0.04297957569360733,
-0.02364700473845005,
-0.08594194054603577,
0.05044662579894066,
-0.037697821855545044,
-0.04161245748400688,
0.03846127167344093,
-0.002630728529766202,
-0.04600291699171066,
-0.11625474691390991,
0.00006453325477195904,
0.11484319716691971,
-0.020819777622818947,
0.0687505230307579,
0.0987808108329773,
-0.06757550686597824,
0.045140139758586884,
0.012583070434629917,
-0.020475249737501144,
0.052874088287353516,
0.04491816833615303,
-0.07048224657773972,
0.012987022288143635,
0.08418601006269455,
0.08836062252521515,
-0.04608489200472832,
-0.07744552940130234,
0.006204866338521242,
-0.031134558841586113,
-0.015977531671524048,
-0.03584638237953186,
0.0787205696105957,
-0.06148320063948631,
-0.014807314611971378,
0.015992343425750732,
-0.002655779244378209,
0.04957103729248047,
-0.01663542352616787,
-0.06168026849627495,
2.8722060192060467e-33,
0.0431942380964756,
-0.0582832433283329,
0.08071797341108322,
0.013920162804424763,
0.10251478105783463,
0.05858616903424263,
0.0417347252368927,
0.004846549592912197,
0.007020157761871815,
0.015670550987124443,
-0.0227802786976099,
0.027970433235168457,
-0.09476488828659058,
0.04920518770813942,
0.05689125508069992,
0.03277112916111946,
-0.053135763853788376,
0.018487675115466118,
-0.0032549554016441107,
-0.029955687001347542,
0.08516456186771393,
-0.11660604178905487,
-0.045413896441459656,
0.012670621275901794,
-0.021311718970537186,
-0.05653691291809082,
0.027541017159819603,
0.03648560494184494,
0.017815204337239265,
0.021828055381774902,
-0.09513363987207413,
-0.04293959587812424,
0.010046235285699368,
0.05636196956038475,
-0.034024566411972046,
-0.039171259850263596,
-0.104732446372509,
-0.03385964781045914,
0.018824361264705658,
-0.05902186781167984,
-0.039066705852746964,
0.00973218772560358,
0.021839533001184464,
-0.013617847114801407,
-0.09055901318788528,
-0.08121096342802048,
-0.008930634707212448,
-0.016346318647265434,
-0.017970949411392212,
-0.019708527252078056,
0.08289489895105362,
-0.02429618500173092,
-0.0345824733376503,
-0.025550255551934242,
0.03230854496359825,
0.029291238635778427,
0.02898498997092247,
0.08580220490694046,
0.043239474296569824,
0.06326865404844284,
-0.026600191369652748,
-0.005388640332967043,
-0.04158527031540871,
0.06611964851617813,
0.038618575781583786,
0.07667867839336395,
-0.0723092332482338,
0.03867427259683609,
0.12460145354270935,
-0.06536713987588882,
0.001650037243962288,
-0.0010458540637046099,
-0.0924203023314476,
-0.0005984805175103247,
0.05770111456513405,
-0.024920539930462837,
0.04566636681556702,
-0.04196572303771973,
-0.026577621698379517,
0.09082464128732681,
-0.09704076498746872,
0.02465556003153324,
-0.03659506142139435,
0.039910271763801575,
-0.0020800463389605284,
-0.015119897201657295,
0.03565916046500206,
-0.053182538598775864,
-0.047854457050561905,
0.02940547466278076,
-0.0980193242430687,
-0.07112199068069458,
0.009487581439316273,
0.006601906381547451,
0.004856168758124113,
-2.7521406588710736e-33,
-0.08305931091308594,
0.07936859875917435,
-0.07685121893882751,
0.03962063044309616,
0.008317353203892708,
-0.0772533044219017,
0.030659697949886322,
-0.006730394903570414,
-0.05908786132931709,
-0.0444803349673748,
-0.0227082259953022,
0.08589830994606018,
-0.011409606784582138,
-0.045848190784454346,
0.02448035590350628,
-0.020731613039970398,
0.013755813241004944,
-0.0039035710506141186,
0.043040309101343155,
0.0043992879800498486,
0.027993736788630486,
0.050070032477378845,
-0.06640215963125229,
0.005663443356752396,
0.02309386059641838,
-0.0038899460341781378,
-0.03661441430449486,
0.0966666042804718,
0.01958124339580536,
0.059694308787584305,
-0.039501406252384186,
-0.08415443450212479,
-0.11089285463094711,
-0.000982913887128234,
-0.011751137673854828,
-0.009379536844789982,
0.09191147983074188,
0.06531468778848648,
-0.08048207312822342,
0.04159284010529518,
0.05380173772573471,
-0.08608842641115189,
-0.094795361161232,
0.02422473393380642,
0.007184796966612339,
0.027681993320584297,
-0.006819108035415411,
-0.003443018766120076,
-0.0017011308809742332,
0.006300013978034258,
-0.08632410317659378,
0.08594255894422531,
-0.07513529807329178,
0.05377349629998207,
-0.03435605764389038,
-0.012873978354036808,
0.07082215696573257,
0.013460787944495678,
0.036747366189956665,
-0.024793731048703194,
-0.05648932233452797,
-0.0725490003824234,
-0.019580915570259094,
-0.005150954704731703,
-0.016466107219457626,
-0.01765538565814495,
0.00632455013692379,
0.04101186245679855,
-0.010974016040563583,
0.07342268526554108,
-0.0005444438429549336,
0.05967039614915848,
0.026399068534374237,
0.013583328574895859,
0.04570915922522545,
0.0461740680038929,
0.07963664084672928,
-0.023010458797216415,
-0.019486840814352036,
0.02011161856353283,
-0.038237281143665314,
0.02401069737970829,
0.008510608226060867,
0.07637617737054825,
0.07508324831724167,
0.036484986543655396,
0.07350058108568192,
-0.023229330778121948,
0.04872506111860275,
-0.00467836856842041,
-0.06310679763555527,
0.013212707825005054,
-0.06252602487802505,
-0.006084558088332415,
-0.05981838330626488,
-4.3533631810532825e-8,
-0.05978250876069069,
0.048735763877630234,
-0.00587795116007328,
-0.04306110367178917,
0.10279898345470428,
0.014658208005130291,
-0.05054953694343567,
0.0423755906522274,
0.0050399717874825,
0.035615187138319016,
0.056724149733781815,
0.047739386558532715,
0.016490669921040535,
-0.06528016179800034,
-0.0001462821091990918,
0.045885052531957626,
-0.030431345105171204,
-0.0515659861266613,
-0.036094892770051956,
-0.05737495794892311,
0.035830166190862656,
-0.030365169048309326,
-0.012338652275502682,
-0.004041813313961029,
-0.006421731784939766,
-0.11455129832029343,
0.0502873919904232,
0.002808227203786373,
0.016661586239933968,
0.027801720425486565,
-0.027336623519659042,
0.021390600129961967,
0.000232971491641365,
-0.01977446861565113,
0.04921342432498932,
0.06705658882856369,
0.006398526020348072,
0.007996099069714546,
0.04761354625225067,
-0.08021761476993561,
-0.016806764528155327,
0.005074788816273212,
0.0011800839565694332,
-0.024009939283132553,
-0.022506091743707657,
-0.002030102303251624,
0.008171027526259422,
0.003939833026379347,
0.05137856304645538,
0.04486898332834244,
0.06784779578447342,
0.042951010167598724,
-0.002725595375522971,
0.05777629464864731,
0.045109380036592484,
-0.023688463494181633,
-0.015557407401502132,
-0.14407822489738464,
0.02412935346364975,
-0.0027774556074291468,
-0.06068985164165497,
-0.0024456894025206566,
-0.06551840901374817,
-0.02468433789908886
] | 0
| 0
| 0
| 0
|
|
MA-RLHF: Reinforcement Learning from Human Feedback with Macro Actions
|
https://openreview.net/forum?id=WWXjMYZxfH
|
[
"Yekun Chai",
"Haoran Sun",
"Huang Fang",
"Shuohuan Wang",
"Yu Sun",
"Hua Wu"
] |
Poster
|
Reinforcement learning from human feedback (RLHF) has demonstrated effectiveness in aligning large language models (LLMs) with human preferences. However, token-level RLHF suffers from the credit assignment problem over long sequences, where delayed rewards make it challenging for the model to discern which actions contributed to preferred outcomes. This hinders learning efficiency and slows convergence.In this paper, we propose MA-RLHF, a simple yet effective RLHF framework that incorporates macro actions --- sequences of tokens or higher-level language constructs --- into the learning process. By operating at higher level of abstraction, our approach reduces the temporal distance between actions and rewards, facilitating faster and more accurate credit assignment. This results in more stable policy gradient estimates and enhances learning efficiency within each episode, all without increasing computational complexity during training or inference. We validate our approach through extensive experiments across various model sizes and tasks, including text summarization, dialogue generation, question answering, and program synthesis. Our method achieves substantial performance improvements over standard RLHF, with performance gains of up to 30\% in text summarization and code generation, 18\% in dialogue, and 8\% in question answering tasks. Notably, our approach reaches parity with vanilla RLHF $1.7 \sim 2$ times faster in terms of training time and continues to outperform it with further training. We make our code and data publicly available at \url{https://github.com/ernie-research/MA-RLHF}.
|
Human Alignment, Large Language Models, Reinforcement Learning
|
This paper introduces MA-RLHF, a framework that incorporates macro actions into RLHF for large language models, addressing the credit assignment problem and significantly improving learning efficiency and performance across various tasks.
| 11,526
| null |
[
-0.07231111079454422,
-0.06117536127567291,
0.010294086299836636,
0.009132987819612026,
0.05644643306732178,
0.08504873514175415,
0.0050904639065265656,
0.04147570952773094,
0.05321996659040451,
0.004986716900020838,
-0.04560254141688347,
-0.09060239046812057,
0.02696182206273079,
0.04652482271194458,
-0.0007228106842376292,
-0.007285622414201498,
0.0658402144908905,
0.04073745757341385,
-0.046299755573272705,
-0.04157283157110214,
0.025580864399671555,
0.07134637981653214,
0.08929252624511719,
0.030633728951215744,
0.024139054119586945,
-0.013812066987156868,
-0.04887441173195839,
-0.06366785615682602,
0.007462768349796534,
-0.014958679676055908,
0.08055409789085388,
-0.02785676345229149,
0.13581280410289764,
0.002360885264351964,
-0.09422153979539871,
0.06809215247631073,
-0.12170601636171341,
-0.016364557668566704,
-0.043886635452508926,
-0.07425718754529953,
-0.06323718279600143,
-0.046821314841508865,
0.04313567653298378,
-0.009460465982556343,
0.10362007468938828,
-0.005216529592871666,
0.030671492218971252,
-0.07273197174072266,
0.025995243340730667,
0.018982458859682083,
-0.05409387871623039,
0.013781728222966194,
0.0277620367705822,
0.029629506170749664,
-0.06054845079779625,
0.042521487921476364,
0.02451324090361595,
0.05194362252950668,
0.02197893336415291,
-0.10013138502836227,
-0.05357866361737251,
-0.1506241410970688,
-0.02007305435836315,
-0.014717167243361473,
-0.056368324905633926,
-0.06499522179365158,
-0.01062434446066618,
0.07116920500993729,
-0.03329594433307648,
0.07776277512311935,
-0.04149022698402405,
0.07934650033712387,
-0.00789655838161707,
0.03828577324748039,
-0.015086301602423191,
0.06013515964150429,
0.003146137110888958,
-0.049643922597169876,
0.04892902076244354,
0.006445207167416811,
-0.03190384432673454,
-0.0347975455224514,
0.07935506850481033,
-0.010207301005721092,
0.06284229457378387,
-0.01456656027585268,
0.04482702538371086,
0.041810642927885056,
0.022322991862893105,
-0.013561575673520565,
0.022852452471852303,
0.000088156433776021,
0.09041421115398407,
-0.043656572699546814,
-0.014111220836639404,
0.0627877414226532,
-0.07845187932252884,
-0.08099888265132904,
-0.058584701269865036,
0.09652087837457657,
0.03184957057237625,
0.13240286707878113,
-0.03578479215502739,
-0.08029304444789886,
-0.003056707326322794,
0.03324789181351662,
-0.0022568691056221724,
0.01898220367729664,
0.021954573690891266,
-0.10111059993505478,
0.05047125741839409,
0.033701661974191666,
-0.03720754384994507,
0.04253140464425087,
0.023840641602873802,
0.015559680759906769,
0.04614568501710892,
-0.07712069898843765,
0.02150656282901764,
0.09622937440872192,
0.039455775171518326,
0.006143786478787661,
-0.003675176063552499,
0.004645267501473427,
-0.018135089427232742,
-0.024021275341510773,
-0.008569218218326569,
1.3866641967658342e-33,
0.0666825920343399,
0.012035459280014038,
0.007925188168883324,
0.06207330897450447,
0.022292988374829292,
0.022556088864803314,
0.03597912937402725,
0.047643110156059265,
-0.0360657274723053,
-0.04589056968688965,
0.04293863847851753,
-0.01980498433113098,
-0.040741465985774994,
0.07933681458234787,
0.013768270611763,
-0.03434528037905693,
-0.062135692685842514,
-0.0013765828916803002,
-0.01299293152987957,
-0.00478522852063179,
0.03793244808912277,
-0.008154252544045448,
0.023251377046108246,
-0.04682554304599762,
0.06404867768287659,
-0.0029253161046653986,
0.08544030040502548,
-0.0743064358830452,
0.00079794303746894,
-0.0019835394341498613,
-0.09039905667304993,
-0.03436165302991867,
-0.012079884298145771,
0.006195363588631153,
-0.009541459381580353,
0.017266450449824333,
-0.037271659821271896,
-0.029901666566729546,
0.035777997225522995,
-0.03158683329820633,
-0.03877948224544525,
-0.030977841466665268,
0.019830673933029175,
0.015186690725386143,
-0.100651316344738,
-0.0765351727604866,
-0.022001279518008232,
0.02410586178302765,
-0.05290793627500534,
0.08009690791368484,
0.0230534877628088,
-0.0035083277616649866,
-0.03626685217022896,
-0.04904339835047722,
-0.004870731849223375,
0.03476487472653389,
-0.041402947157621384,
0.07558535784482956,
-0.0384628102183342,
0.025864191353321075,
-0.012351895682513714,
0.021165834739804268,
0.015949690714478493,
0.09625505656003952,
-0.034569431096315384,
0.07698459178209305,
-0.051391907036304474,
0.056517813354730606,
0.06786716729402542,
-0.012092672288417816,
-0.00905209593474865,
0.046589627861976624,
0.030153634026646614,
-0.023755993694067,
0.06574174761772156,
-0.004685469437390566,
0.04244881123304367,
-0.10836244374513626,
0.013207506388425827,
0.03730436787009239,
-0.012979567050933838,
-0.03634857386350632,
-0.006177528761327267,
-0.024054860696196556,
-0.016011524945497513,
-0.0659794732928276,
0.06701772660017014,
-0.07727178186178207,
0.03493506461381912,
-0.05711674317717552,
-0.022491898387670517,
-0.06588386744260788,
0.02929779142141342,
0.030628401786088943,
0.030690398067235947,
-1.8402966502977707e-33,
0.033661309629678726,
-0.024158425629138947,
-0.04273884370923042,
0.009443002752959728,
-0.014273152686655521,
0.0026808998081833124,
-0.02900940738618374,
0.007652345113456249,
-0.021590927615761757,
-0.04844876006245613,
-0.08028602600097656,
0.0365731418132782,
0.04648667946457863,
0.009768849238753319,
0.08458135277032852,
-0.0412495993077755,
0.0088662039488554,
-0.008225059136748314,
0.030132681131362915,
0.05620179325342178,
0.012997372075915337,
0.0896725282073021,
-0.1033729836344719,
0.033868174999952316,
-0.03864807263016701,
0.007315850351005793,
-0.04794636741280556,
0.07743577659130096,
0.0020506547298282385,
0.0170957800000906,
-0.08767759799957275,
0.003456237493082881,
-0.041712675243616104,
-0.016530167311429977,
-0.04528291895985603,
0.040940385311841965,
0.057792846113443375,
0.01748053915798664,
-0.1023322269320488,
0.0931849256157875,
0.10918381065130234,
-0.13088007271289825,
-0.0700898990035057,
-0.010273286141455173,
-0.06720589846372604,
-0.01592588797211647,
-0.07182737439870834,
-0.027222154662013054,
-0.018375035375356674,
0.009360624477267265,
-0.09970366954803467,
0.012074615806341171,
-0.06695293635129929,
0.023520322516560555,
-0.09617320448160172,
-0.0517176017165184,
-0.011322857812047005,
0.000598377431742847,
0.06323196738958359,
0.013086141087114811,
-0.07264818996191025,
0.05539184808731079,
0.024874551221728325,
-0.010765775106847286,
0.04113801196217537,
-0.04699856415390968,
0.018695147708058357,
-0.032189901918172836,
0.009408952668309212,
0.023853126913309097,
0.06743531674146652,
0.004252045881003141,
0.02966618537902832,
0.01659706048667431,
0.09537763893604279,
0.011689082719385624,
-0.005803599487990141,
-0.059898070991039276,
-0.09803348034620285,
-0.07390555739402771,
0.0355471670627594,
0.054899685084819794,
0.04887695237994194,
0.045778341591358185,
0.028023069724440575,
0.04289020225405693,
0.043878789991140366,
0.08503837883472443,
0.011199959553778172,
0.020497964695096016,
-0.008801769465208054,
0.009709775447845459,
0.03091532550752163,
0.0244255643337965,
-0.07614251971244812,
-4.2722891890889514e-8,
-0.03318653255701065,
0.04693492874503136,
0.00045152025995776057,
0.049828797578811646,
0.032636743038892746,
-0.01452610082924366,
-0.05974279344081879,
0.01793428137898445,
-0.01285692397505045,
-0.025755489245057106,
0.06005942076444626,
0.07250797003507614,
0.03206067532300949,
-0.1527850180864334,
-0.059883199632167816,
0.08251432329416275,
-0.03739932179450989,
-0.04311097785830498,
0.0013578924117609859,
0.03154800087213516,
0.08587057888507843,
0.02014794759452343,
-0.00397532619535923,
-0.01972268894314766,
-0.052608802914619446,
-0.04829099029302597,
0.012081771157681942,
0.15107327699661255,
0.0488886721432209,
-0.036766741424798965,
0.045612093061208725,
0.016767317429184914,
-0.04025799036026001,
0.0057633123360574245,
0.05380295217037201,
-0.0020727466326206923,
0.09885117411613464,
-0.05940406024456024,
-0.026127777993679047,
-0.011754723265767097,
-0.0033837163355201483,
0.05027331784367561,
-0.022437648847699165,
-0.006757756695151329,
-0.0009244512766599655,
0.007020408287644386,
-0.10572392493486404,
-0.12763331830501556,
0.10471587628126144,
-0.0908452495932579,
-0.005755324382334948,
-0.012605953961610794,
-0.055001262575387955,
0.02737981267273426,
0.03843965008854866,
0.03420821577310562,
-0.006665391847491264,
-0.03140662983059883,
0.06067509204149246,
0.0177009180188179,
0.048713039606809616,
-0.004470019601285458,
-0.048217855393886566,
-0.013512823730707169
] | 0
| 0
| 0
| 0
|
|
Bonsai: Gradient-free Graph Condensation for Node Classification
|
https://openreview.net/forum?id=5x88lQ2MsH
|
[
"Mridul Gupta",
"Samyak Jain",
"Vansh Ramani",
"HARIPRASAD KODAMANA",
"Sayan Ranu"
] |
Poster
|
Graph condensation has emerged as a promising avenue to enable scalable training of GNNs by compressing the training dataset while preserving essential graph characteristics. Our study uncovers significant shortcomings in current graph condensation techniques. First, the majority of the algorithms paradoxically require training on the full dataset to perform condensation. Second, due to their gradient-emulating approach, these methods require fresh condensation for any change in hyperparameters or GNN architecture, limiting their flexibility and reusability. To address these challenges, we present Bonsai, a novel graph condensation method empowered by the observation that *computation trees* form the fundamental processing units of message-passing GNNs. Bonsai condenses datasets by encoding a careful selection of *exemplar* trees that maximize the representation of all computation trees in the training set. This unique approach imparts Bonsai as the first linear-time, model-agnostic graph condensation algorithm for node classification that outperforms existing baselines across $7$ real-world datasets on accuracy, while being $22$ times faster on average. Bonsai is grounded in rigorous mathematical guarantees on the adopted approximation strategies, making it robust to GNN architectures, datasets, and parameters.
|
Graph Neural Networks, Machine Learning, Data Distillation, Graph Distillation, Dataset Distillation, Sustainable AI, Graph Condensation, Data Condensation, Dataset Condensation
|
An unsupervised and model/hyper-parameter agnostic graph condensation algorithm for Node CLassification
| 11,523
|
2410.17579
|
[
-0.08454012125730515,
-0.01161501556634903,
0.05645575374364853,
0.06786339730024338,
0.06966899335384369,
-0.012012793682515621,
-0.04570664465427399,
-0.051829319447278976,
-0.040885213762521744,
-0.0019243821734562516,
-0.02475360594689846,
-0.056942135095596313,
0.018619973212480545,
0.03663848340511322,
-0.0606018602848053,
0.021549224853515625,
0.07440270483493805,
0.0561332181096077,
-0.06166689097881317,
-0.02216115966439247,
-0.013254144228994846,
0.049655236303806305,
-0.03390982002019882,
0.030793575569987297,
0.10898249596357346,
-0.04564165323972702,
-0.05779071897268295,
-0.08280360698699951,
0.06955603510141373,
0.02514658123254776,
0.04848535358905792,
-0.008510233834385872,
-0.06791342049837112,
0.10514488816261292,
-0.07224603742361069,
0.04723063483834267,
0.008229002356529236,
-0.01086625549942255,
0.006989439949393272,
-0.028588619083166122,
0.04916166886687279,
0.022023389115929604,
-0.019317099824547768,
0.07476093620061874,
0.08556969463825226,
0.0797140970826149,
-0.012979931198060513,
0.018690116703510284,
0.0074744257144629955,
0.02636420540511608,
-0.053921282291412354,
-0.036083053797483444,
-0.0413968600332737,
0.08475963026285172,
-0.01158859208226204,
-0.040782783180475235,
-0.005705411545932293,
-0.04016900807619095,
-0.0045855119824409485,
0.05297880619764328,
-0.02467280440032482,
-0.08978327363729477,
0.01770598813891411,
-0.04895323887467384,
-0.027398234233260155,
0.0035339719615876675,
0.02165563404560089,
0.07476908713579178,
0.022689707577228546,
0.06696716696023941,
0.06293955445289612,
0.09654609858989716,
-0.10990692675113678,
0.0871984139084816,
-0.016727253794670105,
0.0630066841840744,
0.046180225908756256,
0.020856179296970367,
-0.02251480706036091,
-0.019608598202466965,
0.0631839707493782,
0.048655182123184204,
0.015715915709733963,
0.008333990350365639,
0.06491059064865112,
-0.022128505632281303,
-0.026870280504226685,
0.001730829244479537,
-0.030566127970814705,
0.021031059324741364,
-0.09824930876493454,
-0.008296599611639977,
0.03427311033010483,
-0.03890058025717735,
-0.05555809661746025,
0.009896326810121536,
0.015019217506051064,
-0.028379740193486214,
-0.06912527978420258,
0.10866867750883102,
-0.05507310479879379,
0.01155372615903616,
0.06208600103855133,
-0.07576140016317368,
-0.006690269336104393,
0.029848087579011917,
-0.00841098465025425,
0.04017774015665054,
0.04521487280726433,
-0.06134882941842079,
-0.007890855893492699,
0.08007404953241348,
-0.008244441822171211,
-0.050451334565877914,
0.026382988318800926,
0.009447109885513783,
-0.02860403247177601,
-0.002812376944348216,
0.01337075512856245,
0.06965404748916626,
-0.038608308881521225,
0.04851954057812691,
-0.01305988896638155,
0.01652573235332966,
-0.019540494307875633,
-0.01502619031816721,
-0.09262603521347046,
3.3717174662376e-33,
0.0872722715139389,
0.05400240421295166,
0.08036699146032333,
-0.021909423172473907,
0.09584349393844604,
0.014363217167556286,
-0.0021377471275627613,
0.006701932288706303,
-0.011661645956337452,
0.01280176267027855,
-0.08237876743078232,
-0.0550936758518219,
-0.012443659827113152,
0.047153424471616745,
0.09030161052942276,
-0.017757248133420944,
-0.029382409527897835,
-0.0013449782272800803,
0.000264269852777943,
-0.056418705731630325,
0.04967343807220459,
-0.032741617411375046,
0.02673427201807499,
0.07874433696269989,
-0.01642727665603161,
0.007258579134941101,
0.05373090133070946,
0.021509559825062752,
0.024716826155781746,
-0.0019957993645220995,
-0.06102624535560608,
0.040207017213106155,
0.04505492001771927,
0.03678689897060394,
0.021438023075461388,
-0.047919899225234985,
-0.04802507907152176,
-0.03584785759449005,
0.027456989511847496,
-0.05374812334775925,
-0.051349032670259476,
0.04947109892964363,
0.01529216393828392,
-0.044464416801929474,
-0.04082462191581726,
-0.09738475829362869,
-0.06901244819164276,
-0.008098854683339596,
0.022743666544556618,
-0.04495709389448166,
0.00707218935713172,
-0.021328933537006378,
-0.043084219098091125,
0.07619073241949081,
0.002104364335536957,
0.001610052539035678,
0.12070370465517044,
0.08197163790464401,
0.017950057983398438,
0.03560330718755722,
0.061507128179073334,
-0.03736021742224693,
-0.0452042780816555,
0.05788908153772354,
-0.036626219749450684,
0.0201487448066473,
-0.10206985473632812,
0.1017908826470375,
0.0008480289834551513,
-0.03906818851828575,
-0.049706459045410156,
0.03957301005721092,
-0.027846483513712883,
-0.02929377183318138,
0.05119180679321289,
-0.0009731067693792284,
-0.033471208065748215,
-0.14972172677516937,
-0.03779701516032219,
0.015779873356223106,
-0.0404694527387619,
-0.022540856152772903,
-0.008003033697605133,
0.010254417546093464,
-0.05398717150092125,
-0.09416235983371735,
0.05022421106696129,
0.04360543191432953,
0.020957836881279945,
-0.03609595075249672,
-0.11928941309452057,
-0.017338262870907784,
0.020869046449661255,
-0.03527450934052467,
-0.02088107168674469,
-2.6672561914172807e-33,
-0.008763141930103302,
0.11508236825466156,
-0.06683588773012161,
0.07643883675336838,
0.02217569574713707,
-0.04147107154130936,
-0.019212249666452408,
-0.06119924038648605,
-0.09214414656162262,
0.01728048175573349,
-0.010001030750572681,
-0.006154570262879133,
0.0197443924844265,
0.014027715660631657,
0.0271940715610981,
0.006446146871894598,
-0.0682525485754013,
-0.01197552029043436,
-0.025230534374713898,
0.021332956850528717,
-0.013658789917826653,
0.08510572463274002,
-0.09234742075204849,
0.003418954089283943,
0.04205594211816788,
-0.03480580076575279,
0.03593083471059799,
0.07478775084018707,
-0.00006154338188935071,
0.0100276293233037,
-0.08188599348068237,
0.013762527145445347,
-0.03904356434941292,
-0.09721797704696655,
0.029430100694298744,
-0.0018829089822247624,
0.08256705850362778,
-0.05480131879448891,
-0.03853148594498634,
0.025991195812821388,
0.03804115578532219,
-0.04329067841172218,
-0.07373335212469101,
0.05753975734114647,
-0.003861077129840851,
0.04052668437361717,
-0.06725133955478668,
0.006430600304156542,
-0.07264432311058044,
0.06674276292324066,
-0.039806317538022995,
0.029537634924054146,
-0.024759965017437935,
0.057990580797195435,
-0.005901089403778315,
-0.12257256358861923,
-0.02153708040714264,
0.06454528123140335,
0.0197147224098444,
-0.04111034423112869,
-0.03526962175965309,
-0.05528241768479347,
-0.015307693742215633,
-0.030696338042616844,
0.010364384390413761,
-0.0788595899939537,
0.015878500416874886,
0.044894736260175705,
0.028609629720449448,
0.0037747134920209646,
0.0396282896399498,
0.02998383343219757,
-0.05434831976890564,
-0.0029242150485515594,
-0.003332410240545869,
0.004696886986494064,
0.03134538605809212,
-0.053646646440029144,
-0.05493564158678055,
-0.07051271945238113,
-0.018848497420549393,
-0.004205979872494936,
0.021757347509264946,
0.03990964591503143,
0.08926746249198914,
0.0441160723567009,
0.04012560471892357,
0.07860350608825684,
0.024150747805833817,
0.06824962794780731,
-0.06043738126754761,
0.030009105801582336,
-0.03205002099275589,
0.05833172798156738,
-0.05813874304294586,
-3.8407470981383085e-8,
-0.20400981605052948,
0.07600434124469757,
0.003565230406820774,
0.02827862650156021,
0.06292182952165604,
-0.03898400068283081,
0.08706096559762955,
0.17781271040439606,
-0.01777089387178421,
0.03528769686818123,
0.046618394553661346,
0.010469458997249603,
-0.02743861824274063,
-0.05581120029091835,
0.016014263033866882,
-0.005147217772901058,
-0.00906672514975071,
-0.025006655603647232,
0.011055877432227135,
0.026765596121549606,
-0.0012082790490239859,
-0.0006301789544522762,
-0.026827974244952202,
0.04762531816959381,
0.024636734277009964,
-0.16696010529994965,
-0.001960678957402706,
0.08931992202997208,
0.00965830311179161,
0.0032633922528475523,
-0.06564266979694366,
0.03594702109694481,
0.056355953216552734,
-0.059919051826000214,
-0.025359775871038437,
0.11979936808347702,
-0.003480682848021388,
0.015499233268201351,
-0.013232888653874397,
-0.027555886656045914,
0.005429841578006744,
0.10587213933467865,
-0.0719277560710907,
-0.010002000257372856,
-0.00656656501814723,
-0.006526729557663202,
0.04829159751534462,
0.000389524910133332,
0.03582144156098366,
0.0018166870577260852,
-0.007047618273645639,
-0.027157459408044815,
-0.024212811142206192,
0.0012448105262592435,
-0.00954337790608406,
0.03125832602381706,
-0.0780535563826561,
-0.03967123478651047,
0.031261224299669266,
0.051673609763383865,
0.025910180062055588,
-0.04230516776442528,
-0.03413820266723633,
-0.03589654713869095
] | 0
| 0
| 0
| 0
|
|
Logically Consistent Language Models via Neuro-Symbolic Integration
|
https://openreview.net/forum?id=7PGluppo4k
|
[
"Diego Calanzone",
"Stefano Teso",
"Antonio Vergari"
] |
Poster
|
Large language models (LLMs) are a promising venue for natural language understanding and generation tasks. However, current LLMs are far from reliable: they are prone to generate non-factual information and, more crucially, to contradict themselves when prompted to reason about relations between real entities of the world. These problems are currently addressed with large scale fine-tuning or by delegating consistent reasoning to external tools. In this work, we strive for a middle ground and leverage a training objective based on a principled neuro-symbolic loss that teaches a LLM to be consistent with external knowledge in the form of a set of facts and rules. Fine-tuning with such a loss on a limited set of facts enables our LLMs to be more logically consistent than previous baselines for a given constraint. Our approach also allows to easily combine multiple logical constraints at once in a principled way, delivering LLMs that are more consistent w.r.t. all the selected rules. Moreover, our method allows LLMs to extrapolate to unseen but semantically similar factual knowledge, represented in unseen datasets, more systematically.
|
probabilistic reasoning, logical consistency, LLMs, neuro-symbolic, semantic loss
|
We show that principled probabilistic reasoning can teach an LLM to be logically consistent with a set of external facts and rules (and itself), allowing to extrapolate to unseen but semantically similar factual knowledge.
| 11,510
|
2409.13724
|
[
0.026438770815730095,
-0.07181569933891296,
0.02677215449512005,
0.04239460825920105,
0.04408111795783043,
0.015776917338371277,
-0.017619755119085312,
-0.03242071717977524,
-0.0432015024125576,
-0.02155800350010395,
-0.028416601940989494,
-0.1023833155632019,
0.09086000174283981,
0.04916829243302345,
0.030674591660499573,
0.09229466319084167,
0.05868208780884743,
0.08715476840734482,
-0.12063288688659668,
-0.046912919729948044,
0.056147500872612,
0.03522205352783203,
0.03209628537297249,
0.01247729267925024,
-0.010963983833789825,
0.015447665005922318,
0.07169006019830704,
-0.006472141016274691,
0.05821170285344124,
-0.014849692583084106,
0.08349191397428513,
0.0008674146956764162,
-0.11532524228096008,
-0.0039058069232851267,
0.07917196303606033,
0.07554873824119568,
-0.07770276814699173,
0.042809244245290756,
0.024685919284820557,
-0.012163570150732994,
0.005787622183561325,
0.000804383831564337,
0.004880350083112717,
-0.014403202570974827,
0.08820830285549164,
0.0633544921875,
-0.008594944141805172,
0.010230455547571182,
-0.04013259708881378,
0.014195442199707031,
-0.0439017079770565,
-0.05956372618675232,
-0.03143440559506416,
0.03713999688625336,
0.002350397640839219,
0.059891942888498306,
-0.004102844279259443,
0.022598912939429283,
-0.054148998111486435,
-0.046421751379966736,
0.0046206992119550705,
-0.08775997906923294,
-0.06435870379209518,
-0.0587579645216465,
-0.025964226573705673,
-0.028438154608011246,
-0.02861882746219635,
0.05745421722531319,
-0.043658360838890076,
0.10772145539522171,
-0.016459379345178604,
0.044683776795864105,
-0.04512918367981911,
0.08190637081861496,
-0.07287435978651047,
0.11816104501485825,
0.060652658343315125,
-0.055474549531936646,
0.060434699058532715,
-0.02669973485171795,
-0.05950137972831726,
0.06636577844619751,
0.0007252099458128214,
0.045314889401197433,
0.14982804656028748,
-0.015305633656680584,
0.060013946145772934,
0.05111514404416084,
0.010253350250422955,
0.09122412651777267,
-0.02091274969279766,
-0.05976012349128723,
0.13239261507987976,
0.029249288141727448,
0.06511850655078888,
0.047782931476831436,
-0.0028105510864406824,
-0.08586020767688751,
-0.04422207921743393,
0.09568800032138824,
0.007116954308003187,
0.18333886563777924,
-0.019413262605667114,
-0.08334911614656448,
-0.01605139672756195,
-0.03406079486012459,
-0.00015519699081778526,
0.01774343103170395,
0.0016298460541293025,
-0.06598259508609772,
0.038722243160009384,
0.07260145246982574,
0.03352919593453407,
-0.015614517033100128,
0.0345771424472332,
0.0018489487702026963,
0.02554147131741047,
-0.013897661119699478,
0.05477241426706314,
-0.02384660765528679,
-0.06898635625839233,
0.01491547841578722,
-0.001613207277841866,
0.11170310527086258,
0.07492341101169586,
-0.09398665279150009,
-0.02015974000096321,
3.1263179135643296e-34,
0.021106697618961334,
0.021218687295913696,
0.01771099865436554,
0.019232463091611862,
0.05774328112602234,
0.0629386156797409,
-0.03450893610715866,
0.040111154317855835,
-0.03332354500889778,
0.085113525390625,
-0.04056078568100929,
-0.00497883977368474,
-0.0629325658082962,
0.0389145202934742,
-0.006248035002499819,
0.016923172399401665,
-0.054334819316864014,
0.03785413131117821,
-0.045522574335336685,
-0.02382916957139969,
-0.012805148027837276,
-0.00616874173283577,
-0.03959246352314949,
-0.03874976560473442,
0.01727711409330368,
0.006942057516425848,
0.07863364368677139,
-0.027356985956430435,
0.0033722997177392244,
-0.026857515797019005,
-0.06309640407562256,
-0.06613688170909882,
0.06827370077371597,
0.0508953258395195,
0.03297033905982971,
-0.07443369179964066,
-0.0038380862679332495,
-0.06050053983926773,
0.01934385672211647,
-0.0012508425861597061,
0.006551176775246859,
0.05961684137582779,
0.012996903620660305,
0.031165875494480133,
-0.04528002068400383,
-0.03651642054319382,
-0.006682578939944506,
0.0033289254643023014,
0.0013985141413286328,
-0.09635564684867859,
0.03532520309090614,
0.04583129659295082,
-0.009970035403966904,
-0.06619469821453094,
-0.0334964245557785,
-0.05394952371716499,
-0.04667931795120239,
0.011016425676643848,
0.0015998625895008445,
0.019095227122306824,
0.03271748498082161,
0.017725927755236626,
-0.0007486263639293611,
0.10831186175346375,
0.038099322468042374,
0.06702066212892532,
-0.03284895420074463,
0.04082893580198288,
0.0918908342719078,
-0.09569703042507172,
-0.06821413338184357,
-0.008002457208931446,
0.004643469583243132,
-0.0892089381814003,
-0.008393012918531895,
-0.014050101861357689,
0.04614018276333809,
-0.062336139380931854,
0.004505041521042585,
0.006185169331729412,
0.025517040863633156,
0.006425630301237106,
-0.01598171889781952,
-0.028954781591892242,
0.012562408111989498,
-0.06912372261285782,
0.03830946609377861,
-0.05289636179804802,
-0.030213404446840286,
-0.053057167679071426,
-0.07853102684020996,
-0.09463869780302048,
-0.03902645781636238,
-0.037587687373161316,
-0.014780889265239239,
-4.7643810004609474e-34,
-0.04283453896641731,
-0.0025618455838412046,
-0.005611340049654245,
0.034135643392801285,
-0.030807247385382652,
-0.06499477475881577,
-0.034727517515420914,
0.018290020525455475,
-0.013184324838221073,
-0.07691633701324463,
-0.07362041622400284,
-0.06694722175598145,
0.05306902900338173,
-0.06224231794476509,
-0.0495569184422493,
-0.0008982021827250719,
-0.05379154905676842,
0.012976573780179024,
0.008739009499549866,
0.08342471718788147,
-0.01769651658833027,
0.11153794825077057,
-0.12010542303323746,
0.031360577791929245,
-0.010389662347733974,
0.06112758442759514,
0.009188667871057987,
0.0643218532204628,
-0.06955000758171082,
0.06496486067771912,
-0.006631080061197281,
-0.024065280333161354,
-0.003613104345276952,
-0.0859467089176178,
-0.0810784325003624,
-0.0031189932487905025,
0.048938002437353134,
-0.04211394861340523,
-0.0656730979681015,
0.06100637838244438,
0.03774966672062874,
-0.00996977649629116,
-0.05663955584168434,
-0.018403327092528343,
-0.05381254479289055,
-0.010650192387402058,
-0.10059268772602081,
0.0109599269926548,
0.008828222751617432,
0.07635156065225601,
-0.00952010229229927,
-0.04852915182709694,
-0.06683122366666794,
-0.0023483936674892902,
0.0006668346468359232,
-0.076667919754982,
-0.04903016239404678,
-0.05861624330282211,
0.009512308053672314,
-0.0445823073387146,
-0.08025118708610535,
-0.0074595920741558075,
0.02769024670124054,
-0.04537307098507881,
0.025573451071977615,
0.01936064474284649,
-0.01734732650220394,
0.028897520154714584,
0.0345320962369442,
-0.003387352218851447,
0.025252513587474823,
0.021632198244333267,
-0.03179923817515373,
0.013059427961707115,
0.038469262421131134,
0.010623947717249393,
0.04909951612353325,
-0.029740571975708008,
0.00896387454122305,
-0.07165344059467316,
0.04792629927396774,
-0.017224781215190887,
0.053496360778808594,
0.09676612168550491,
0.02341104857623577,
0.031020352616906166,
0.05268358439207077,
0.09440457075834274,
0.029120296239852905,
0.07271549105644226,
-0.04441060498356819,
0.018328502774238586,
-0.008683277294039726,
0.06462627649307251,
-0.10102511197328568,
-3.941778814464669e-8,
-0.08339820802211761,
0.05293358117341995,
-0.040592603385448456,
0.10984431952238083,
0.04567079246044159,
-0.08261097967624664,
-0.03987232595682144,
0.011417494155466557,
-0.032439131289720535,
-0.012366599403321743,
0.0437740683555603,
0.026767488569021225,
-0.05083361268043518,
-0.04727364331483841,
-0.023375602439045906,
0.04427768662571907,
0.003538920311257243,
0.012718995101749897,
-0.027662357315421104,
-0.00974275078624487,
0.07402109354734421,
0.05103699862957001,
-0.02049398608505726,
0.03544905409216881,
0.1150054931640625,
-0.05894199758768082,
-0.0506821908056736,
0.09884876012802124,
0.052056919783353806,
0.03561338037252426,
0.047361936420202255,
-0.056173160672187805,
0.05275674909353256,
-0.05848497524857521,
0.006209332495927811,
0.08608346432447433,
-0.017757074907422066,
-0.028598343953490257,
-0.015050175599753857,
-0.08541350066661835,
0.005493655800819397,
0.050223831087350845,
0.007597723044455051,
0.03678014501929283,
0.002020105253905058,
-0.04423696547746658,
-0.05548865348100662,
-0.07544087618589401,
0.0027645626105368137,
-0.01676459237933159,
-0.04506496340036392,
0.026384569704532623,
0.01584501937031746,
0.007894436828792095,
0.05538863688707352,
-0.01515733078122139,
-0.042196113616228104,
0.01972966268658638,
-0.011466652154922485,
0.01909826137125492,
0.0070502799935638905,
0.03351159393787384,
-0.014098044484853745,
0.02015649899840355
] | 0
| 0
| 0
| 0
|
|
Narrowing Information Bottleneck Theory for Multimodal Image-Text Representations Interpretability
|
https://openreview.net/forum?id=INqLJwqUmc
|
[
"Zhiyu Zhu",
"Zhibo Jin",
"Jiayu Zhang",
"NAN YANG",
"Jiahao Huang",
"Jianlong Zhou",
"Fang Chen"
] |
Poster
|
The task of identifying multimodal image-text representations has garnered increasing attention, particularly with models such as CLIP (Contrastive Language-Image Pretraining), which demonstrate exceptional performance in learning complex associations between images and text. Despite these advancements, ensuring the interpretability of such models is paramount for their safe deployment in real-world applications, such as healthcare. While numerous interpretability methods have been developed for unimodal tasks, these approaches often fail to transfer effectively to multimodal contexts due to inherent differences in the representation structures. Bottleneck methods, well-established in information theory, have been applied to enhance CLIP's interpretability. However, they are often hindered by strong assumptions or intrinsic randomness. To overcome these challenges, we propose the Narrowing Information Bottleneck Theory, a novel framework that fundamentally redefines the traditional bottleneck approach. This theory is specifically designed to satisfy contemporary attribution axioms, providing a more robust and reliable solution for improving the interpretability of multimodal models. In our experiments, compared to state-of-the-art methods, our approach enhances image interpretability by an average of 9\%, text interpretability by an average of 58.83\%, and accelerates processing speed by 63.95\%. Our code is publicly accessible at https://github.com/LMBTough/NIB.
|
Interpretability, CLIP
| null | 11,505
|
2502.14889
|
[
0.029358038678765297,
-0.07435255497694016,
-0.006018271204084158,
0.024677129462361336,
0.13819848001003265,
-0.017906440421938896,
0.014898017048835754,
-0.007804029621183872,
0.056424349546432495,
-0.11082948744297028,
-0.03456651046872139,
-0.008510058745741844,
-0.0071216365322470665,
0.072704017162323,
-0.04070547595620155,
-0.024785393849015236,
0.05946724861860275,
0.07642872631549835,
-0.16595196723937988,
-0.009481006301939487,
0.08468025177717209,
-0.02282565087080002,
0.07791800796985626,
-0.0215565487742424,
0.07001140713691711,
0.049489084631204605,
-0.07937083393335342,
-0.05915190279483795,
-0.005806736182421446,
-0.03802451863884926,
0.00749139953404665,
-0.012830393388867378,
0.056368160992860794,
0.012465830892324448,
-0.07188113033771515,
0.08576465398073196,
-0.029253367334604263,
0.084529347717762,
-0.013540704734623432,
0.021814856678247452,
-0.020670214667916298,
0.041562970727682114,
-0.013637570664286613,
0.06438900530338287,
0.08592799305915833,
-0.008920753374695778,
-0.01899552345275879,
-0.024868039414286613,
-0.016112040728330612,
0.0057774921879172325,
-0.12729912996292114,
0.0017551735509186983,
-0.008039131760597229,
0.04609939828515053,
-0.024621350690722466,
-0.028438353911042213,
0.03030487708747387,
0.012019139714539051,
-0.050540946424007416,
-0.01223567221313715,
-0.010825269855558872,
-0.07171057909727097,
-0.01945965364575386,
0.05497215688228607,
0.07129817456007004,
0.046638645231723785,
0.06613775342702866,
-0.03497019782662392,
-0.06398238241672516,
0.0025406647473573685,
-0.047838080674409866,
0.05326424166560173,
-0.0696144700050354,
0.07475680857896805,
0.009584806859493256,
0.016089385375380516,
0.10827366262674332,
0.025157175958156586,
0.03758964687585831,
-0.09567597508430481,
0.04909275472164154,
-0.005673993844538927,
0.07646119594573975,
-0.0034717784728854895,
0.12297932058572769,
0.012170558795332909,
-0.028914235532283783,
0.036603838205337524,
-0.04264073818922043,
0.06667158752679825,
-0.08060349524021149,
-0.06437066197395325,
-0.012063803151249886,
0.02723751775920391,
0.08544418215751648,
0.04471326246857643,
-0.009973394684493542,
-0.11172494292259216,
0.003157978877425194,
0.07171744853258133,
-0.0018122991314157844,
-0.026987873017787933,
0.00401461124420166,
0.0010292950319126248,
0.02011258341372013,
-0.07246806472539902,
0.06797458231449127,
-0.05157098174095154,
-0.008310425095260143,
-0.012806260026991367,
0.06920786201953888,
0.01118355244398117,
-0.04058433324098587,
0.015142418444156647,
0.01742761768400669,
-0.03989759832620621,
0.017270680516958237,
-0.018072880804538727,
0.059092748910188675,
0.03317195922136307,
-0.029512230306863785,
-0.005107279401272535,
-0.05712816119194031,
-0.03978422284126282,
0.005930415820330381,
-0.02347223274409771,
-0.0684792622923851,
1.3336054123410893e-33,
-0.0017337797908112407,
-0.0028372188098728657,
0.08131925016641617,
0.056808289140462875,
0.09943042695522308,
-0.026356130838394165,
-0.1352468878030777,
-0.018946003168821335,
-0.032104816287755966,
0.025340382009744644,
0.004110866226255894,
0.0371260829269886,
-0.03908064588904381,
0.07422299683094025,
0.032007649540901184,
-0.008868237026035786,
-0.07947881519794464,
0.0521017350256443,
0.006528304424136877,
-0.0014403461245819926,
0.03102927841246128,
-0.044937677681446075,
-0.07564286142587662,
-0.008809493854641914,
-0.013842500746250153,
-0.02854432910680771,
0.05782143026590347,
-0.0622556135058403,
-0.004590017255395651,
-0.0024449636694043875,
-0.20420007407665253,
0.032060831785202026,
0.07456627488136292,
-0.0675194263458252,
0.0496913343667984,
-0.11598940193653107,
-0.028829170390963554,
-0.018958471715450287,
-0.015400622971355915,
-0.0495455302298069,
0.006130828056484461,
0.000035400636988924816,
-0.025902869179844856,
-0.02254626713693142,
-0.05674654245376587,
0.008605340495705605,
0.016728568822145462,
-0.006751922890543938,
-0.12200643867254257,
0.000018842798453988507,
0.014239550568163395,
0.03142431005835533,
0.023789074271917343,
-0.05991044268012047,
-0.012403015047311783,
-0.03122125007212162,
0.03397664427757263,
0.029078058898448944,
0.025420721620321274,
0.03102564997971058,
-0.002839780645444989,
0.07627496868371964,
0.06559067219495773,
0.026002299040555954,
0.027838407084345818,
0.03791369870305061,
-0.07257421314716339,
0.04154561087489128,
0.05302101746201515,
-0.029716946184635162,
-0.06700793653726578,
0.07114469259977341,
-0.04195212200284004,
-0.04495885223150253,
-0.036240000277757645,
0.03453954681754112,
0.020983127877116203,
-0.04283292964100838,
0.02462705224752426,
0.07551445066928864,
-0.05103781819343567,
-0.026141420006752014,
0.0525752529501915,
-0.05347583070397377,
-0.0853566974401474,
0.02150963991880417,
0.022869527339935303,
-0.10170862078666687,
-0.06990426033735275,
-0.035894252359867096,
-0.02692706324160099,
0.05072351172566414,
-0.016978364437818527,
-0.027012240141630173,
0.04643932357430458,
-2.4267015580970693e-33,
-0.026444273069500923,
0.0189769696444273,
0.018538087606430054,
0.0521530918776989,
0.06527047604322433,
-0.02820652350783348,
0.014351269230246544,
0.0654023140668869,
0.0647759884595871,
-0.08918450027704239,
-0.010977773927152157,
-0.008188365027308464,
-0.003166948677971959,
-0.039079662412405014,
0.02888570725917816,
0.001685412134975195,
0.05066906660795212,
0.05142316594719887,
0.010713671334087849,
0.07725951820611954,
0.041436005383729935,
0.038648732006549835,
-0.04529574513435364,
0.02755347453057766,
-0.06116368621587753,
0.10881185531616211,
-0.07966653257608414,
-0.031165264546871185,
-0.019207382574677467,
-0.0669020265340805,
-0.06936593353748322,
0.010205244645476341,
-0.0408296063542366,
-0.07523025572299957,
-0.06112317368388176,
0.043425604701042175,
0.03236769884824753,
-0.004385712556540966,
-0.04977332428097725,
0.07454987615346909,
0.1431623101234436,
-0.011798086576163769,
-0.08539916574954987,
-0.04083036631345749,
0.028451498597860336,
-0.09367315471172333,
-0.018954476341605186,
-0.03570520505309105,
0.0177288968116045,
0.1181114986538887,
-0.003017899813130498,
-0.02162865363061428,
-0.07802107185125351,
0.05281421169638634,
-0.0538945347070694,
-0.0820830836892128,
-0.03698761388659477,
-0.04741086810827255,
0.06066112592816353,
0.02141997404396534,
-0.02127816714346409,
-0.014251339249312878,
-0.11838105320930481,
-0.03243501856923103,
0.00420808931812644,
-0.059272535145282745,
0.01714611053466797,
-0.04904940724372864,
0.033470652997493744,
0.0314786322414875,
-0.016154520213603973,
-0.015989847481250763,
-0.016262218356132507,
0.04812069609761238,
0.028221912682056427,
0.021340245380997658,
0.040797699242830276,
-0.021872229874134064,
-0.04947437345981598,
-0.021739276126027107,
0.069084033370018,
-0.032205767929553986,
0.06371352821588516,
0.02065892145037651,
0.013998756185173988,
0.07433610409498215,
-0.01706678234040737,
-0.04609452560544014,
0.0884518027305603,
0.019694969058036804,
-0.008733310736715794,
0.03269992768764496,
-0.052906692028045654,
0.10677596926689148,
-0.08897136151790619,
-4.2175997805316e-8,
0.02051353082060814,
-0.04267154261469841,
-0.010965160094201565,
-0.02560807578265667,
0.0127344885841012,
-0.025471869856119156,
-0.06066735088825226,
0.03297049552202225,
0.010613362304866314,
-0.04675024002790451,
0.09638162702322006,
0.018133847042918205,
-0.03304203972220421,
-0.05043938755989075,
-0.03710424527525902,
0.036023274064064026,
-0.042458780109882355,
0.04226944223046303,
0.01629544235765934,
0.0006246048142202199,
0.027946682646870613,
-0.018227364867925644,
-0.03044421784579754,
-0.038419779390096664,
0.040628910064697266,
-0.04571089521050453,
-0.04376012831926346,
0.07528704404830933,
0.04115566238760948,
-0.06376360356807709,
-0.008211316540837288,
0.02926735393702984,
-0.0010225848527625203,
-0.016905292868614197,
0.004604230169206858,
0.009128854610025883,
0.0003348645113874227,
-0.05438351631164551,
-0.02921614609658718,
-0.018177680671215057,
0.07354572415351868,
0.03398854658007622,
0.0056901369243860245,
0.04721172899007797,
0.009367452934384346,
-0.01097275223582983,
0.022948089987039566,
-0.028122033923864365,
0.01493185292929411,
-0.0452808141708374,
0.027329519391059875,
0.01779460348188877,
0.03615706413984299,
0.11627127975225449,
0.03424578160047531,
-0.040438055992126465,
0.07000234723091125,
0.0665149837732315,
0.008107322268188,
0.08773353695869446,
0.03679581359028816,
0.03814077749848366,
-0.02568201720714569,
-0.016393279656767845
] | 0
| 0
| 0
| 0
|
|
Efficient Source-Free Time-Series Adaptation via Parameter Subspace Disentanglement
|
https://openreview.net/forum?id=Q5Sawm0nqo
|
[
"Gaurav Patel",
"Christopher Michael Sandino",
"Behrooz Mahasseni",
"Ellen L. Zippi",
"Erdrin Azemi",
"Ali Moin",
"Juri Minxha"
] |
Poster
|
In this paper, we propose a framework for efficient Source-Free Domain Adaptation (SFDA) in the context of time-series, focusing on enhancing both parameter efficiency and data-sample utilization. Our approach introduces an improved paradigm for source-model preparation and target-side adaptation, aiming to enhance training efficiency during target adaptation. Specifically, we reparameterize the source model's weights in a Tucker-style decomposed manner, factorizing the model into a compact form during the source model preparation phase. During target-side adaptation, only a subset of these decomposed factors is fine-tuned, leading to significant improvements in training efficiency. We demonstrate using PAC Bayesian analysis that this selective fine-tuning strategy implicitly regularizes the adaptation process by constraining the model's learning capacity. Furthermore, this re-parameterization reduces the overall model size and enhances inference efficiency, making the approach particularly well suited for resource-constrained devices. Additionally, we demonstrate that our framework is compatible with various SFDA methods and achieves significant computational efficiency, reducing the number of fine-tuned parameters and inference overhead in terms of MACs by over 90\% while maintaining model performance.
|
Time Series, Source-Free Domain Adaptation, Efficiency
|
Efficient source-free domain adaptation framework for time series that improves parameter and sample efficiency of the adaptation process.
| 11,503
|
2410.02147
|
[
-0.050359904766082764,
-0.09704559296369553,
0.02396605722606182,
0.09037159383296967,
0.11357036232948303,
0.005091714207082987,
-0.026438137516379356,
-0.07132682204246521,
0.011400841176509857,
-0.06255324929952621,
-0.10003674775362015,
-0.04828621447086334,
-0.0019738194532692432,
-0.021842636168003082,
-0.004888779483735561,
-0.047351110726594925,
0.07540352642536163,
0.052589356899261475,
-0.01768874190747738,
0.007547943387180567,
0.044861022382974625,
0.00959734246134758,
0.02346220426261425,
-0.022293418645858765,
0.0446026511490345,
-0.018007392063736916,
0.06894005835056305,
-0.0036062600556761026,
0.03921986371278763,
0.002530414378270507,
-0.007095237262547016,
0.04711274802684784,
-0.056153178215026855,
-0.04445107653737068,
-0.03509018197655678,
0.012619059532880783,
-0.013990587554872036,
0.0605599544942379,
-0.019313232973217964,
0.051693812012672424,
-0.006889850366860628,
0.049040116369724274,
0.04038648307323456,
0.054328519850969315,
-0.05575845018029213,
-0.01894582062959671,
0.02773020602762699,
-0.052590418606996536,
0.04941721633076668,
0.03794421628117561,
-0.007037607952952385,
0.0014573248336091638,
-0.11347916722297668,
0.07614130526781082,
-0.025179298594594002,
0.05317715182900429,
0.03882664814591408,
0.04113221913576126,
0.013384915888309479,
-0.02402457781136036,
-0.09402712434530258,
-0.05772002041339874,
-0.02027607150375843,
-0.0033410757314413786,
-0.020782869309186935,
0.018662454560399055,
0.0015892343362793326,
0.11698393523693085,
-0.00680667394772172,
0.00807282142341137,
-0.09344310313463211,
0.08658672869205475,
-0.038608282804489136,
-0.044651880860328674,
0.02943192981183529,
0.05448475852608681,
0.04129408299922943,
-0.005080097820609808,
0.09954971820116043,
-0.07923581451177597,
0.03738608956336975,
0.006766516249626875,
-0.04019163176417351,
-0.028030965477228165,
0.1084437295794487,
0.03394177556037903,
0.004429804626852274,
0.03435099124908447,
0.06489984691143036,
-0.013206196017563343,
-0.026729464530944824,
0.008843590505421162,
0.05884752795100212,
-0.008624402806162834,
-0.004383638966828585,
0.04209846258163452,
-0.012471826747059822,
-0.014510281383991241,
-0.005286491010338068,
0.08439678698778152,
0.016750937327742577,
0.07946114242076874,
0.033675748854875565,
0.0012105999048799276,
-0.030195172876119614,
-0.060775794088840485,
0.04385261610150337,
0.111082062125206,
0.03321357071399689,
-0.005099378991872072,
0.07247941941022873,
0.0988466814160347,
0.03063725307583809,
-0.02456745132803917,
0.004511318635195494,
-0.012282397598028183,
-0.0078586395829916,
-0.11333674937486649,
-0.04210015758872032,
0.04430003836750984,
-0.018964096903800964,
-0.00831613503396511,
-0.016486866399645805,
-0.03736918419599533,
0.026276204735040665,
-0.031032510101795197,
-0.09502571076154709,
2.1978208693774694e-33,
0.05466551333665848,
0.05531332269310951,
0.05362705886363983,
-0.008321551606059074,
-0.018089409917593002,
0.02782765030860901,
0.035779207944869995,
0.010129374451935291,
-0.005136954132467508,
0.03102201782166958,
-0.023723118007183075,
-0.001507176784798503,
-0.16969257593154907,
0.02518031932413578,
0.09214635193347931,
-0.001217294717207551,
-0.03024420514702797,
0.13368141651153564,
0.02298385463654995,
0.010555585846304893,
0.021093476563692093,
0.007799320854246616,
-0.06451940536499023,
-0.030190996825695038,
0.016359122470021248,
0.019005540758371353,
0.026851192116737366,
-0.005304047372192144,
-0.09210895001888275,
0.008634759113192558,
-0.08476734906435013,
-0.010622844099998474,
-0.0397847518324852,
-0.01338907703757286,
0.0474795401096344,
-0.11130152642726898,
0.009989782236516476,
0.012304777279496193,
0.010541430674493313,
0.003565455786883831,
-0.01650075986981392,
-0.00197099638171494,
0.10243841260671616,
-0.08830065280199051,
-0.07332616299390793,
-0.10136664658784866,
0.019316859543323517,
-0.07587917149066925,
-0.026194626465439796,
0.016537772491574287,
0.056120265275239944,
-0.004548982717096806,
0.04992029815912247,
-0.0992409884929657,
-0.08143822103738785,
-0.03829224035143852,
0.012921187095344067,
-0.006867663469165564,
0.04237136244773865,
0.09910966455936432,
-0.055380888283252716,
0.052353356033563614,
-0.06527696549892426,
0.0688776895403862,
0.0021276436746120453,
0.030542083084583282,
0.009879743680357933,
0.04727661609649658,
0.011017994955182076,
0.02890031971037388,
-0.02754274755716324,
-0.016564426943659782,
-0.05133356526494026,
0.005269884597510099,
0.07768578827381134,
-0.06269805878400803,
0.04168137535452843,
0.0074361469596624374,
-0.04259994998574257,
0.028271356597542763,
0.0031486458610743284,
0.057485468685626984,
0.010776866227388382,
-0.023458383977413177,
-0.013649269007146358,
-0.021944884210824966,
0.038992904126644135,
-0.08317035436630249,
-0.04112550616264343,
-0.11496146023273468,
-0.0076094018295407295,
-0.0007071991567499936,
-0.06615662574768066,
-0.03611297160387039,
0.007629490457475185,
-2.0449312986637224e-33,
-0.018266763538122177,
-0.06301707774400711,
0.0192099679261446,
-0.0031832626555114985,
0.005761002656072378,
-0.004644245840609074,
-0.00906821433454752,
0.05059860646724701,
0.054792676120996475,
-0.06327236443758011,
-0.058913033455610275,
-0.09769273549318314,
-0.03157268092036247,
-0.07512051612138748,
0.014805988408625126,
-0.014759652316570282,
-0.11739827692508698,
-0.00008489681931678206,
-0.08248714357614517,
-0.06781776249408722,
0.015816466882824898,
0.0333312526345253,
0.009768457151949406,
-0.011401718482375145,
-0.008710147812962532,
0.010153747163712978,
0.032510243356227875,
0.141037255525589,
0.037190042436122894,
0.008717754855751991,
-0.12474824488162994,
0.014882095158100128,
-0.01767178252339363,
0.008425221778452396,
-0.10650067776441574,
0.0692477822303772,
0.04024739935994148,
0.03797711431980133,
-0.022389691323041916,
0.09020265191793442,
0.022025421261787415,
0.040822990238666534,
-0.047727059572935104,
0.03605835512280464,
-0.021221652626991272,
0.0560956746339798,
-0.03961581736803055,
0.012993301264941692,
0.05288224294781685,
0.01909114606678486,
0.020073609426617622,
0.032157205045223236,
-0.038932107388973236,
-0.0030211370903998613,
-0.027983298525214195,
-0.053556378930807114,
0.007895718328654766,
-0.018450742587447166,
0.07915444672107697,
-0.047273483127355576,
-0.05729914456605911,
-0.04286320134997368,
-0.007234846707433462,
-0.07094908505678177,
0.014258824288845062,
-0.015504501760005951,
0.023960936814546585,
-0.009921407327055931,
0.04891550913453102,
0.023381048813462257,
0.030975477769970894,
0.007992499507963657,
0.029277438297867775,
-0.013780266977846622,
-0.04376543313264847,
-0.025805780664086342,
0.06788888573646545,
-0.026694405823946,
-0.027264626696705818,
0.009852468967437744,
-0.00004662959327106364,
0.0032196545507758856,
0.0006791915511712432,
0.0636773332953453,
0.03620107099413872,
0.146889790892601,
0.045848071575164795,
0.06238621100783348,
-0.024989839643239975,
-0.0016147978603839874,
-0.11157314479351044,
0.010955280624330044,
0.0361383818089962,
0.12440390139818192,
-0.05963060259819031,
-4.806823028502549e-8,
-0.060918185859918594,
0.12489617615938187,
0.012834873050451279,
0.025269249454140663,
-0.03456582874059677,
-0.01989465206861496,
-0.03310789540410042,
0.04984735697507858,
0.08541946858167648,
0.02120397426187992,
0.0940907895565033,
-0.041115276515483856,
-0.007833842188119888,
0.03483077138662338,
-0.04868847131729126,
0.10771147161722183,
0.013663156889379025,
0.01909126155078411,
-0.04509969428181648,
-0.08778735250234604,
-0.018828071653842926,
0.0621064156293869,
0.011521187610924244,
0.004453770350664854,
0.0936017706990242,
-0.103396937251091,
-0.00952051766216755,
0.13961954414844513,
0.09381671249866486,
0.013058566488325596,
-0.044598449021577835,
0.00458949152380228,
-0.10725966840982437,
-0.044064171612262726,
-0.047746092081069946,
-0.008926586247980595,
0.0009119769092649221,
-0.01863810233771801,
0.0041011725552380085,
0.007274195551872253,
-0.04022824764251709,
0.06812282651662827,
-0.08740159124135971,
0.02697073668241501,
-0.0008659386658109725,
-0.03033636324107647,
0.01079435646533966,
-0.030401291325688362,
0.01446618977934122,
-0.016862675547599792,
0.06363851577043533,
-0.014141729101538658,
-0.016239911317825317,
-0.04361732304096222,
0.042811885476112366,
0.07512952387332916,
0.011720936745405197,
-0.04083898290991783,
0.03500277176499367,
0.014859032817184925,
-0.02388574555516243,
-0.004402459599077702,
-0.02151823602616787,
0.028882386162877083
] | 0
| 0
| 0
| 0
|
|
WardropNet: Traffic Flow Predictions via Equilibrium-Augmented Learning
|
https://openreview.net/forum?id=7FHSPd3SRE
|
[
"Kai Jungel",
"Dario Paccagnan",
"Axel Parmentier",
"Maximilian Schiffer"
] |
Poster
|
When optimizing transportation systems, anticipating traffic flows is a central element. Yet, computing such traffic equilibria remains computationally expensive. Against this background, we introduce a novel combinatorial optimization augmented neural network pipeline that allows for fast and accurate traffic flow predictions. We propose WardropNet, a neural network that combines classical layers with a subsequent equilibrium layer: the first ones inform the latter by predicting the parameterization of the equilibrium problem's latency functions. Using supervised learning we minimize the difference between the actual traffic flow and the predicted output. We show how to leverage a Bregman divergence fitting the geometry of the equilibria, which allows for end-to-end learning. WardropNet outperforms pure learning-based approaches in predicting traffic equilibria for realistic and stylized traffic scenarios. On realistic scenarios, WardropNet improves on average for time-invariant predictions by up to 72\% and for time-variant predictions by up to 23\% over pure learning-based approaches.
|
structured learning, combinatorial optimization augmented machine learning, traffic equilibrium prediction
| null | 11,497
|
2410.06656
|
[
-0.04985470324754715,
-0.1172865629196167,
0.010404312051832676,
-0.01217765174806118,
0.018862266093492508,
0.02781079337000847,
0.03398493304848671,
-0.02964610792696476,
-0.01327521726489067,
-0.01574736274778843,
-0.09752827137708664,
-0.05448073521256447,
-0.007477364037185907,
0.036298226565122604,
-0.09796717762947083,
0.025709856301546097,
0.08061980456113815,
0.05782899260520935,
-0.04887933284044266,
-0.06918147951364517,
0.012930437922477722,
0.045039862394332886,
-0.02449829876422882,
-0.0011310280533507466,
0.01919671706855297,
0.0342424102127552,
0.029086194932460785,
-0.003903243923559785,
-0.06818614155054092,
-0.021387504413723946,
0.031607456505298615,
-0.028620272874832153,
-0.07063563168048859,
-0.022143423557281494,
-0.10080043226480484,
-0.0019351027440279722,
-0.1195257157087326,
0.02091788873076439,
0.0042514558881521225,
-0.0021894958335906267,
0.03338877856731415,
-0.010212099179625511,
-0.029452268034219742,
0.04039029777050018,
0.05412405729293823,
0.005711722187697887,
0.06704983860254288,
-0.04336380958557129,
0.03572128713130951,
-0.04656871408224106,
0.02477591298520565,
0.039064496755599976,
0.022677358239889145,
0.05784967541694641,
-0.05462704226374626,
0.009421808645129204,
-0.012212559580802917,
0.010905795730650425,
-0.0579029843211174,
0.012779110111296177,
-0.03549261763691902,
-0.13686107099056244,
0.0019744927994906902,
0.004627761896699667,
-0.04249601066112518,
-0.007879235781729221,
0.03801063448190689,
0.0760100707411766,
0.002777773654088378,
0.019656933844089508,
0.05074299871921539,
0.09023450314998627,
-0.0847744271159172,
0.010717449709773064,
0.11899806559085846,
0.026265043765306473,
0.09855739772319794,
0.02374301105737686,
0.024349119514226913,
-0.07046062499284744,
0.020721551030874252,
-0.010347675532102585,
0.010880131274461746,
0.04835379496216774,
0.10907087475061417,
-0.026886222884058952,
-0.029246129095554352,
0.07616914063692093,
0.05897773802280426,
-0.04376360401511192,
-0.05421227216720581,
0.005109291058033705,
0.025039171800017357,
-0.07854565978050232,
0.05058775469660759,
0.05035994574427605,
-0.016845108941197395,
-0.1051681712269783,
-0.019753307104110718,
0.12945342063903809,
-0.08648119121789932,
0.05780430883169174,
0.02222607284784317,
-0.00677080312743783,
0.11228425800800323,
-0.012739384546875954,
0.07700159400701523,
0.04177912324666977,
0.03178500756621361,
-0.061064690351486206,
0.072642020881176,
-0.04475238174200058,
0.07273367047309875,
-0.05548489838838577,
0.020049581304192543,
0.0525854155421257,
-0.027212711051106453,
-0.0641108974814415,
-0.02868514321744442,
0.1195133626461029,
-0.13044853508472443,
-0.0032687699422240257,
-0.039208896458148956,
0.0881751999258995,
0.012854677625000477,
0.005142997018992901,
-0.030674608424305916,
-4.744199231825267e-34,
-0.04891215264797211,
-0.030865414068102837,
0.06696691364049911,
0.01270165853202343,
0.0320814847946167,
0.025166882202029228,
0.007916499860584736,
0.010060534812510014,
0.05396692082285881,
-0.05720876529812813,
-0.05733606591820717,
-0.011883159168064594,
-0.07022450119256973,
0.04448601230978966,
-0.011563844978809357,
-0.008718561381101608,
-0.11518455296754837,
0.026609590277075768,
0.006834875792264938,
-0.05540895089507103,
0.022774649783968925,
-0.060193996876478195,
-0.07822833955287933,
-0.045130655169487,
0.0054990146309137344,
0.016338521614670753,
0.05564573407173157,
0.00608627637848258,
0.0856984332203865,
0.01007646881043911,
-0.06635090708732605,
0.08346644788980484,
-0.001144086942076683,
0.04268387705087662,
0.021810121834278107,
0.04587888345122337,
-0.01711464859545231,
0.00794232077896595,
0.07854081690311432,
0.037645235657691956,
-0.02848249487578869,
-0.015608788467943668,
-0.028195256367325783,
-0.025220561772584915,
-0.13212108612060547,
-0.12514421343803406,
0.04411562904715538,
-0.014429021626710892,
-0.06482969224452972,
-0.03183167427778244,
0.03072194568812847,
0.01451236754655838,
-0.018224123865365982,
-0.09859656542539597,
-0.08194434642791748,
-0.02697732299566269,
0.06662130355834961,
0.056164372712373734,
-0.003310045227408409,
0.06279293447732925,
0.012700704857707024,
0.0059260232374072075,
-0.012392185628414154,
-0.020051583647727966,
-0.0051717571914196014,
0.039750974625349045,
0.047008395195007324,
0.09983986616134644,
0.07422506809234619,
-0.04166785627603531,
0.02467062883079052,
0.014424609951674938,
-0.009615452028810978,
-0.015795530751347542,
0.07880251109600067,
0.03528590500354767,
-0.010865445248782635,
-0.10050829499959946,
-0.012925629504024982,
0.06851842999458313,
-0.0037034358829259872,
0.044834040105342865,
-0.019388336688280106,
-0.007053579203784466,
0.028809072449803352,
-0.07501965016126633,
0.06814642250537872,
-0.07899440079927444,
-0.008934922516345978,
-0.01804313063621521,
-0.08503395318984985,
0.014050777070224285,
0.048241037875413895,
0.06734110414981842,
0.020958248525857925,
-9.328387050209735e-34,
-0.009763404726982117,
0.0017272343393415213,
-0.07653031498193741,
0.0521165169775486,
-0.04292026907205582,
0.011794229969382286,
-0.03697612136602402,
0.019141769036650658,
0.027117429301142693,
0.04079857096076012,
0.03126043826341629,
-0.0032773232087492943,
0.049750976264476776,
-0.05528249591588974,
0.048928629606962204,
-0.055998045951128006,
0.0005336038302630186,
-0.03349972143769264,
0.01226731762290001,
-0.04678528010845184,
-0.015142902731895447,
-0.023137198761105537,
-0.09514632076025009,
0.005231773015111685,
-0.05440697446465492,
-0.010595508851110935,
0.012133195996284485,
0.05316680669784546,
0.0031494356226176023,
0.054268449544906616,
-0.08882501721382141,
-0.09019040316343307,
-0.008600205183029175,
-0.02147817052900791,
-0.05518827214837074,
0.1283331960439682,
0.08143100887537003,
0.002169774379581213,
-0.06678536534309387,
-0.017484378069639206,
0.03078588843345642,
-0.11426053941249847,
0.019921516999602318,
-0.0033344896510243416,
-0.019969714805483818,
-0.03228689730167389,
0.041481055319309235,
-0.004673155955970287,
-0.005243645515292883,
0.0414062961935997,
0.052988551557064056,
0.05652967095375061,
-0.10947009176015854,
0.031069016084074974,
-0.033899445086717606,
0.02069402113556862,
0.021452998742461205,
0.005430395249277353,
0.11065356433391571,
0.007709946483373642,
0.00849226489663124,
-0.015127385035157204,
-0.06456194072961807,
0.019623393192887306,
-0.03230457380414009,
-0.108192078769207,
-0.07717996090650558,
-0.04481206089258194,
0.012050837278366089,
-0.0023844719398766756,
0.010397273115813732,
0.043630655854940414,
0.035729385912418365,
0.024146761745214462,
0.02635772153735161,
-0.03852805867791176,
0.09145776927471161,
0.015576029196381569,
-0.013581927865743637,
-0.018284503370523453,
0.006171183194965124,
0.04549568146467209,
0.06804665923118591,
0.07562461495399475,
0.035470400005578995,
-0.015773579478263855,
0.05223096162080765,
0.053146205842494965,
0.02370157465338707,
0.02349400892853737,
-0.04334968328475952,
0.03216252848505974,
-0.01812772825360298,
0.021482069045305252,
-0.09900609403848648,
-3.540954196523671e-8,
-0.06057250499725342,
0.052226949483156204,
0.023201240226626396,
-0.03652677312493324,
0.02128731645643711,
0.03811684623360634,
0.020317096263170242,
0.06811186671257019,
-0.03801226615905762,
-0.026659555733203888,
0.11288639158010483,
0.04304076358675957,
-0.03642013296484947,
-0.03717370331287384,
-0.024991344660520554,
0.03698736056685448,
-0.006184376310557127,
-0.02876078523695469,
0.018457725644111633,
-0.013944624923169613,
-0.028920689597725868,
0.05908714607357979,
-0.026257282122969627,
-0.04908403754234314,
0.043646980077028275,
-0.02404879219830036,
0.001557921408675611,
0.07622238248586655,
0.03351190686225891,
0.00008062136475928128,
-0.07724884152412415,
0.009070439264178276,
-0.01560267899185419,
-0.0754501149058342,
0.01621849648654461,
0.07443509995937347,
0.017955781891942024,
0.031894341111183167,
0.004773252177983522,
-0.03130537271499634,
0.00694277323782444,
0.042442742735147476,
-0.034086428582668304,
-0.03638791665434837,
0.024962671101093292,
0.011871105059981346,
0.03439248725771904,
0.0007476042956113815,
0.05905117839574814,
-0.04355489835143089,
0.06125709414482117,
-0.04573855921626091,
0.06003251671791077,
0.04260960593819618,
0.04636296257376671,
-0.06361113488674164,
-0.0428939126431942,
-0.133736252784729,
0.03250763192772865,
0.08893319964408875,
-0.05032264441251755,
0.08380769193172455,
-0.06106450408697128,
0.01798727922141552
] |
https://github.com/tumbais/ml-co-pipeline-trafficprediction
| 6
| 0
| 0
| 0
|
To Clip or not to Clip: the Dynamics of SGD with Gradient Clipping in High-Dimensions
|
https://openreview.net/forum?id=jmN1zXMq0O
|
[
"Noah Marshall",
"Ke Liang Xiao",
"Atish Agarwala",
"Elliot Paquette"
] |
Poster
|
The success of modern machine learning is due in part to the adaptive optimization methods that have been developed to deal with the difficulties of training large models over complex datasets. One such method is gradient clipping: a practical procedure with limited theoretical underpinnings. In this work, we study clipping in a least squares problem under streaming SGD. We develop a theoretical analysis of the learning dynamics in the limit of large intrinsic dimension—a model and dataset dependent notion of dimensionality. In this limit we find a deterministic equation that describes the evolution of the loss and demonstrate that this equation predicts the path of clipped SGD on synthetic, CIFAR10, and Wikitext2 data. We show that with Gaussian noise clipping cannot improve SGD performance. Yet, in other noisy settings, clipping can provide benefits with tuning of the clipping threshold. We propose a simple heuristic for near optimal scheduling of the clipping threshold which requires the tuning of only one hyperparameter. We conclude with a discussion about the links between high-dimensional clipping and neural network training.
|
gradient clipping, high-dimensional probability, stochastic optimization, deep learning theory
| null | 11,493
|
2406.11733
|
[
-0.11533354222774506,
-0.12049442529678345,
0.03254452720284462,
0.0334618054330349,
0.06589566916227341,
-0.022161955013871193,
-0.0027842402923852205,
-0.028923356905579567,
0.012454153038561344,
-0.1004832461476326,
0.0029325352516025305,
0.0372106097638607,
-0.057905226945877075,
-0.0044410028494894505,
-0.057974494993686676,
-0.048627231270074844,
0.0679856389760971,
0.10770391672849655,
-0.09966405481100082,
-0.06386327743530273,
0.03932039812207222,
-0.027249781414866447,
-0.0018547450890764594,
0.0187600739300251,
0.08306285738945007,
-0.0018549474189057946,
0.01428535021841526,
-0.04631199315190315,
0.018018824979662895,
0.03580060228705406,
0.10275012999773026,
-0.07186757773160934,
0.011997978202998638,
0.002761595882475376,
-0.07442120462656021,
0.00884623359888792,
-0.033567070960998535,
0.06465671956539154,
0.03880564495921135,
-0.004320813808590174,
-0.012991569004952908,
-0.059059225022792816,
-0.03491431474685669,
-0.010361983440816402,
0.019734904170036316,
-0.018183542415499687,
-0.00895990151911974,
-0.09838126599788666,
-0.03017807938158512,
0.024821799248456955,
-0.05554214492440224,
-0.09130492806434631,
-0.09504520893096924,
0.0873049944639206,
0.011995123699307442,
-0.018422627821564674,
0.016957562416791916,
0.02217155694961548,
0.04388218745589256,
0.037480928003787994,
0.0021113513503223658,
-0.15588697791099548,
0.033845603466033936,
-0.020085617899894714,
0.013873245567083359,
-0.0289747416973114,
0.040074992924928665,
0.05147415027022362,
-0.015304363332688808,
0.10682980716228485,
-0.021665727719664574,
0.11766631156206131,
-0.1384783238172531,
0.05069569870829582,
0.010595221072435379,
0.018941449001431465,
0.07936522364616394,
-0.015314090065658092,
-0.0014788673724979162,
-0.030307121574878693,
0.03041410818696022,
0.028180716559290886,
0.08254935592412949,
-0.014383479952812195,
0.05886096879839897,
-0.04432779923081398,
0.03143889456987381,
0.020917179062962532,
0.034743573516607285,
-0.02841184474527836,
-0.015549244359135628,
0.05151988938450813,
-0.04669167473912239,
0.015269219875335693,
-0.011020757257938385,
0.010237333364784718,
-0.025754369795322418,
-0.06831381469964981,
0.016650637611746788,
0.057070836424827576,
-0.031274545937776566,
0.06230076402425766,
0.03271113336086273,
0.07227002829313278,
0.05833359807729721,
-0.04133504629135132,
0.05840098112821579,
0.026122860610485077,
0.008869013749063015,
-0.0728834792971611,
0.08386272192001343,
0.11667963862419128,
0.0253238994628191,
-0.0012802163837477565,
0.022692423313856125,
0.022547880187630653,
-0.015864375978708267,
-0.03734761103987694,
-0.03145056590437889,
0.13931508362293243,
-0.07733660936355591,
0.0247522983700037,
-0.06589517742395401,
0.02822444587945938,
0.0495593398809433,
0.0947834774851799,
-0.030912814661860466,
2.4319675891177608e-33,
-0.020167436450719833,
0.035218942910432816,
-0.01666889153420925,
-0.07524782419204712,
0.04520529881119728,
-0.04468270763754845,
0.015247550792992115,
-0.027825459837913513,
-0.016375089064240456,
-0.028973281383514404,
-0.023755766451358795,
-0.058105189353227615,
-0.07993259280920029,
0.05262916907668114,
0.05969085916876793,
-0.028533536940813065,
-0.0036164680495858192,
0.055138006806373596,
0.032954562455415726,
-0.05492934212088585,
0.07829374819993973,
-0.08762889355421066,
-0.0252941083163023,
-0.0008153460221365094,
0.022088749334216118,
-0.048915572464466095,
0.04153706878423691,
0.004163051024079323,
-0.009787840768694878,
0.05110182613134384,
-0.08300939947366714,
0.014001237228512764,
-0.005893774796277285,
0.01220171432942152,
0.03808050975203514,
0.02497614361345768,
0.034245815128088,
-0.010101775638759136,
0.031872451305389404,
-0.015599728561937809,
0.0025503297802060843,
0.05440293997526169,
0.02057814598083496,
-0.0327727235853672,
-0.13518737256526947,
-0.123768150806427,
0.09288773685693741,
0.033098530024290085,
-0.08457895368337631,
-0.04573741555213928,
0.020551307126879692,
-0.0019380190642550588,
-0.040704481303691864,
-0.021992523223161697,
-0.02761991322040558,
-0.009773697704076767,
0.06081826239824295,
-0.05921101197600365,
0.0244782492518425,
0.07509277015924454,
-0.01080798264592886,
0.03688813000917435,
0.0134479571133852,
0.04814119264483452,
-0.03240274637937546,
-0.016338281333446503,
0.031404707580804825,
0.047202080488204956,
0.008721934631466866,
-0.001660242909565568,
-0.04359244555234909,
0.00650763139128685,
-0.04815218225121498,
-0.05557893216609955,
0.08088283985853195,
-0.02566528506577015,
0.08438655734062195,
-0.09415142983198166,
0.002419186057522893,
0.026193497702479362,
-0.07680811733007431,
0.03857981786131859,
0.004034075420349836,
-0.13652051985263824,
-0.07911673933267593,
-0.11179019510746002,
0.028722012415528297,
-0.09157819300889969,
-0.043984342366456985,
-0.05764198303222656,
-0.1637912541627884,
0.014411264099180698,
-0.016313573345541954,
-0.01270864624530077,
-0.001251386827789247,
-2.1883439971869416e-33,
-0.003638996509835124,
0.04127771407365799,
-0.007990756072103977,
0.11514217406511307,
-0.012048682197928429,
0.03831964358687401,
-0.04917089268565178,
0.03833432495594025,
-0.1069948747754097,
-0.06954675167798996,
-0.01013224571943283,
-0.030253615230321884,
-0.012330024503171444,
0.004276840481907129,
-0.009515166282653809,
-0.0026802450884133577,
-0.015156763605773449,
0.02975291758775711,
-0.005670945625752211,
-0.052067145705223083,
0.03084966354072094,
0.019480975344777107,
0.003914618398994207,
0.08733872324228287,
-0.05018240585923195,
0.019689464941620827,
-0.04367319867014885,
0.10514335334300995,
-0.0070059397257864475,
0.01638849824666977,
-0.06686584651470184,
-0.0278942808508873,
-0.0876745730638504,
0.00028747160104103386,
-0.069146066904068,
0.0482436902821064,
0.04451083764433861,
0.007721893023699522,
0.007146111223846674,
0.10474183410406113,
0.06327904015779495,
0.011522461660206318,
-0.045927874743938446,
0.022652072831988335,
0.00029345363145694137,
-0.02033202350139618,
0.004038311541080475,
0.04263315349817276,
0.014063971117138863,
-0.02555602788925171,
0.0024341654498130083,
0.06013582646846771,
-0.009579496458172798,
0.1159081906080246,
-0.019417276605963707,
0.005388327408581972,
-0.01981513947248459,
-0.0256754532456398,
0.017756236717104912,
0.013411351479589939,
-0.027404233813285828,
-0.018718469887971878,
-0.06346158683300018,
-0.032496288418769836,
0.01115624513477087,
-0.02915525808930397,
0.04495241492986679,
0.072441965341568,
0.020357679575681686,
0.009044216945767403,
-0.008314707316458225,
0.014788956381380558,
0.04643338918685913,
0.003217601217329502,
-0.042427703738212585,
0.018066540360450745,
0.01924179121851921,
0.0012558966409415007,
-0.034012243151664734,
0.0008217991562560201,
0.07432850450277328,
0.03154938668012619,
0.05569003149867058,
0.009975696913897991,
0.031669024378061295,
0.11579670011997223,
0.04805725812911987,
0.05811559781432152,
0.014948681928217411,
-0.04397660866379738,
-0.035109955817461014,
-0.010154844261705875,
-0.03802717849612236,
0.037214282900094986,
-0.04608554020524025,
-4.437561074155383e-8,
-0.06052713841199875,
0.025348301976919174,
0.03618858754634857,
-0.013243990018963814,
0.07590864598751068,
-0.03235214948654175,
0.0032274723052978516,
0.05214334651827812,
0.0029226280748844147,
0.028980571776628494,
0.09269203990697861,
-0.034046124666929245,
-0.04327322542667389,
0.009886194951832294,
-0.009842824190855026,
0.01585170067846775,
-0.04672195762395859,
0.02155441604554653,
0.0006228653946891427,
0.028538666665554047,
-0.022445444017648697,
0.043218277394771576,
0.06997285783290863,
-0.06412537395954132,
0.059271346777677536,
-0.08277709037065506,
0.0448065921664238,
0.0676332637667656,
0.13467779755592346,
0.020636750385165215,
-0.06143457069993019,
0.03401384502649307,
0.0026111092884093523,
-0.015144712291657925,
0.011690287850797176,
0.0564497672021389,
0.04791749268770218,
-0.04088140279054642,
-0.00838213600218296,
0.03518928587436676,
-0.00986082386225462,
0.015019756741821766,
0.009563982486724854,
-0.011437173932790756,
-0.0061071282252669334,
0.0185907743871212,
0.0640290230512619,
-0.05642646178603172,
0.0023196195252239704,
0.004090451635420322,
0.05668552219867706,
-0.05729227140545845,
0.02393115498125553,
0.10758552700281143,
0.10046935826539993,
-0.012991873547434807,
-0.04162367805838585,
-0.05093255639076233,
-0.030918482691049576,
0.02084176056087017,
0.010838735848665237,
-0.03485598415136337,
-0.10226830840110779,
0.029261047020554543
] |
https://github.com/nmarzz/clip
| 0
| 0
| 0
| 0
|
Discrete Diffusion Schrödinger Bridge Matching for Graph Transformation
|
https://openreview.net/forum?id=tQyh0gnfqW
|
[
"Jun Hyeong Kim",
"Seonghwan Kim",
"Seokhyun Moon",
"Hyeongwoo Kim",
"Jeheon Woo",
"Woo Youn Kim"
] |
Poster
|
Transporting between arbitrary distributions is a fundamental goal in generative modeling.
Recently proposed diffusion bridge models provide a potential solution, but they rely on a joint distribution that is difficult to obtain in practice.
Furthermore, formulations based on continuous domains limit their applicability to discrete domains such as graphs.
To overcome these limitations, we propose Discrete Diffusion Schrödinger Bridge Matching (DDSBM), a novel framework that utilizes continuous-time Markov chains to solve the SB problem in a high-dimensional discrete state space.
Our approach extends Iterative Markovian Fitting to discrete domains, and we have proved its convergence to the SB.
Furthermore, we adapt our framework for the graph transformation, and show that our design choice of underlying dynamics characterized by independent modifications of nodes and edges can be interpreted as the entropy-regularized version of optimal transport with a cost function described by the graph edit distance.
To demonstrate the effectiveness of our framework, we have applied DDSBM to molecular optimization in the field of chemistry.
Experimental results demonstrate that DDSBM effectively optimizes molecules' property-of-interest with minimal graph transformation, successfully retaining other features. Source code is available [here](https://github.com/junhkim1226/DDSBM).
|
Schrödinger Bridge, Discrete Diffusion Model, Molecular Optimization
|
We propose DDSBM, a framework for solving the Schrödinger Bridge problem in high-dimensional discrete spaces, applied to graph-based molecular optimization.
| 11,491
| null |
[
-0.062082432210445404,
-0.005450533237308264,
0.019510142505168915,
-0.03981953486800194,
-0.02829865738749504,
-0.010163577273488045,
-0.04844244197010994,
-0.04365608096122742,
0.052822817116975784,
0.01000817958265543,
-0.07010698318481445,
-0.02990332804620266,
-0.006276410538703203,
0.08043269068002701,
0.020860061049461365,
0.04689299315214157,
0.07006405293941498,
0.037753209471702576,
-0.03406735509634018,
-0.0059408219531178474,
0.00880359299480915,
0.052173253148794174,
0.01004977896809578,
-0.05244148150086403,
0.010388530790805817,
0.07687736302614212,
0.1120515912771225,
-0.027941057458519936,
0.013023574836552143,
-0.009721743874251842,
0.06402292847633362,
-0.0038082655519247055,
-0.09102072566747665,
-0.05198274552822113,
0.002244817093014717,
0.06428465992212296,
-0.0866832286119461,
0.030641358345746994,
0.05258273333311081,
-0.03717790171504021,
0.03352363035082817,
0.06243673339486122,
-0.018977655097842216,
0.0647578239440918,
0.003972028382122517,
-0.02720215730369091,
-0.032251641154289246,
-0.03366996347904205,
0.043104834854602814,
-0.04063115641474724,
0.05656382441520691,
0.01643720455467701,
-0.040036022663116455,
0.030476750805974007,
0.01502446923404932,
0.039828043431043625,
0.0687451958656311,
-0.04868127778172493,
0.010041493922472,
-0.015329811722040176,
0.016519607976078987,
-0.06057125702500343,
0.0009970952523872256,
0.03520076721906662,
-0.009841629303991795,
-0.009747548028826714,
0.0736163854598999,
0.03291518986225128,
-0.02684752084314823,
0.07424753904342651,
-0.04416622593998909,
0.04581932723522186,
-0.024588704109191895,
0.05361368507146835,
0.07999860495328903,
-0.024715259671211243,
0.046478573232889175,
0.01100117340683937,
0.005071282386779785,
-0.017109941691160202,
-0.035339176654815674,
0.015142741613090038,
0.06219223141670227,
0.04054796323180199,
0.024789147078990936,
0.07125183939933777,
-0.10877500474452972,
0.03313763067126274,
0.0390826016664505,
-0.00912471953779459,
-0.03337729722261429,
0.02306964062154293,
-0.037684839218854904,
-0.06570008397102356,
-0.04077352583408356,
0.014428508467972279,
0.024808676913380623,
-0.015431440435349941,
0.07575874775648117,
0.048764027655124664,
-0.0115488450974226,
0.003289812710136175,
0.017835944890975952,
-0.04743674024939537,
0.01818663626909256,
-0.04147515818476677,
-0.043138206005096436,
0.060912590473890305,
-0.009907822124660015,
0.025716817006468773,
0.057938795536756516,
0.03122570924460888,
0.07715239375829697,
0.06907261908054352,
-0.06598205119371414,
-0.054040368646383286,
0.025631101801991463,
-0.07022617757320404,
-0.012402980588376522,
0.045721109956502914,
-0.06914813816547394,
0.022078169509768486,
-0.06074368581175804,
-0.02009878121316433,
-0.08155903965234756,
0.03274304047226906,
0.0009781060507521033,
-2.8772490277243185e-34,
0.03700419142842293,
0.0017316265730187297,
0.0036803358234465122,
-0.002830841578543186,
0.06074037775397301,
-0.0019249471370130777,
-0.03756332024931908,
-0.023648036643862724,
-0.06771017611026764,
-0.013015467673540115,
-0.08715840429067612,
0.03086905926465988,
-0.07541736960411072,
-0.003108002245426178,
0.009799419902265072,
0.0133035434409976,
-0.02224365808069706,
0.04592328891158104,
0.08843228965997696,
-0.0824279934167862,
-0.0017936431104317307,
-0.03814311325550079,
-0.013159678317606449,
0.08759080618619919,
-0.018668007105588913,
0.012690121307969093,
-0.002915765391662717,
-0.02571704052388668,
0.07830096036195755,
-0.027141457423567772,
-0.1222376674413681,
0.07115913927555084,
-0.02813601866364479,
-0.012856571935117245,
0.06364808976650238,
-0.0010210819309577346,
-0.016480322927236557,
-0.047790806740522385,
-0.05243507772684097,
-0.1258215457201004,
-0.03321881592273712,
0.019942089915275574,
0.001131031196564436,
-0.037904463708400726,
-0.09813784807920456,
0.01852709800004959,
0.040566690266132355,
-0.05738500878214836,
-0.04979509487748146,
-0.03560878708958626,
-0.001798303215764463,
-0.0465557686984539,
-0.00036633858690038323,
0.036743856966495514,
-0.035687435418367386,
-0.038578614592552185,
0.05559731647372246,
-0.008476713672280312,
0.03251918405294418,
0.08867163956165314,
-0.022470682859420776,
0.05745328217744827,
0.04211444780230522,
0.04187062382698059,
0.08304860442876816,
0.0074865254573524,
-0.060064494609832764,
0.0106575982645154,
0.06567635387182236,
-0.006750686094164848,
0.00943940319120884,
0.0701400488615036,
0.007452339865267277,
-0.04269813746213913,
0.06746872514486313,
-0.03636222332715988,
-0.006063542794436216,
-0.06592243164777756,
-0.00003270886008976959,
0.05673963949084282,
-0.08446987718343735,
-0.003356687491759658,
-0.018894841894507408,
0.02919057011604309,
0.02799752913415432,
-0.15180890262126923,
-0.03975182771682739,
-0.10369797050952911,
-0.09254126250743866,
-0.024992994964122772,
-0.0053636799566447735,
-0.08882346004247665,
0.02557128481566906,
0.056378670036792755,
0.0697818174958229,
-2.2805892633287898e-33,
-0.025309255346655846,
0.06181066855788231,
0.1485314518213272,
0.05556928738951683,
0.019715305417776108,
-0.04745471850037575,
0.0036837037187069654,
0.0000343456631526351,
0.05281706526875496,
-0.018676774576306343,
-0.03468259423971176,
-0.009777500294148922,
-0.013294102624058723,
-0.07428069412708282,
-0.04346505180001259,
-0.024158064275979996,
-0.0016976858023554087,
-0.03086441569030285,
-0.02689392678439617,
0.03689046949148178,
0.017668111249804497,
0.10382585972547531,
-0.11103732138872147,
-0.01817600429058075,
0.023650353774428368,
0.07293681800365448,
0.021313393488526344,
0.01614939421415329,
0.018559303134679794,
0.021177420392632484,
-0.01566111110150814,
0.0003537468146532774,
-0.09169955551624298,
-0.00956740416586399,
-0.07975804060697556,
0.09638791531324387,
0.020260117948055267,
0.10280203819274902,
0.0215349979698658,
0.0028256026562303305,
-0.013708830811083317,
-0.06666651368141174,
-0.08921468257904053,
0.026905354112386703,
0.07014480233192444,
0.04424900561571121,
-0.05446038022637367,
-0.019013984128832817,
-0.004831509664654732,
-0.06775006651878357,
-0.0031087270472198725,
0.0975789800286293,
-0.002650077221915126,
-0.030986342579126358,
-0.050028908997774124,
-0.051394619047641754,
-0.045456983149051666,
0.05288514122366905,
0.09166401624679565,
-0.0038795347791165113,
-0.07464683055877686,
-0.04808693006634712,
0.05971909314393997,
0.032360974699258804,
-0.09642157703638077,
-0.037075795233249664,
-0.0009446885669603944,
-0.04941121116280556,
0.009770607575774193,
0.013120278716087341,
-0.029909895732998848,
-0.009595954790711403,
0.05954793840646744,
0.035403963178396225,
0.11042916774749756,
-0.1543549746274948,
0.09274914860725403,
-0.006941197440028191,
0.002965012099593878,
0.041991692036390305,
-0.014251811429858208,
0.09477423876523972,
-0.01840115711092949,
0.042158789932727814,
0.02454577200114727,
-0.02297898381948471,
0.009728953242301941,
-0.036773476749658585,
0.03479250892996788,
-0.0015660967910662293,
0.01377805881202221,
0.04340313747525215,
-0.03394981846213341,
0.018044363707304,
-0.07420219480991364,
-4.573648126893204e-8,
-0.06475283205509186,
-0.01275563146919012,
0.042576566338539124,
-0.04287685453891754,
0.017567945644259453,
0.027447625994682312,
0.1324775516986847,
0.09085715562105179,
-0.008547631092369556,
-0.00007967050623847172,
0.08144540339708328,
0.02650829777121544,
-0.01688344031572342,
-0.011000704020261765,
-0.09084735810756683,
0.007915680296719074,
-0.027841122820973396,
-0.04800809547305107,
0.0030118285212665796,
-0.027457155287265778,
-0.10926990956068039,
-0.006847247947007418,
-0.012276080437004566,
0.0695534348487854,
0.05573995038866997,
-0.10106927901506424,
-0.06233888491988182,
-0.08352172374725342,
0.052712954580783844,
0.010544140823185444,
-0.0729350671172142,
0.02028605155646801,
0.06419970095157623,
0.1397663652896881,
0.0003507625951897353,
-0.04372796788811684,
-0.02307841181755066,
-0.021163975819945335,
-0.04991009458899498,
0.025769155472517014,
0.0012231947621330619,
0.1059902161359787,
-0.059119660407304764,
0.01778276264667511,
-0.028089776635169983,
0.04567410424351692,
0.02909005805850029,
0.13106872141361237,
0.05454932898283005,
0.007791319862008095,
0.03315545246005058,
-0.025954624637961388,
-0.020248249173164368,
-0.09091918170452118,
-0.016372837126255035,
0.05455576255917549,
-0.01390076708048582,
-0.036987170577049255,
0.03776418790221214,
-0.04030800983309746,
-0.019190022721886635,
-0.03864169865846634,
-0.0835002139210701,
-0.03837665542960167
] | 0
| 0
| 0
| 0
|
|
Breaking Mental Set to Improve Reasoning through Diverse Multi-Agent Debate
|
https://openreview.net/forum?id=t6QHYUOQL7
|
[
"Yexiang Liu",
"Jie Cao",
"Zekun Li",
"Ran He",
"Tieniu Tan"
] |
Poster
|
Large Language Models (LLMs) have seen significant progress but continue to struggle with persistent reasoning mistakes.
Previous methods of *self-reflection* have been proven limited due to the models’ inherent fixed thinking patterns.
While Multi-Agent Debate (MAD) attempts to mitigate this by incorporating multiple agents, it often employs the same reasoning methods, even though assigning different personas to models. This leads to a "fixed mental set", where models rely on homogeneous thought processes without exploring alternative perspectives.
In this paper, we introduce Diverse Multi-Agent Debate (DMAD), a method that encourages agents to think with distinct reasoning approaches. By leveraging diverse problem-solving strategies, each agent can gain insights from different perspectives, refining its responses through discussion and collectively arriving at the optimal solution. DMAD effectively breaks the limitations of fixed mental sets. We evaluate DMAD against various prompting techniques, including *self-reflection* and traditional MAD, across multiple benchmarks using both LLMs and Multimodal LLMs. Our experiments show that DMAD consistently outperforms other methods, delivering better results than MAD in fewer rounds. Code is available at https://github.com/MraDonkey/DMAD.
|
Multi-Agent Debate, Large Language Models, Multimodal Large Language Models, Prompting, Self-Correction, Reasoning
| null | 11,487
| null |
[
0.024555865675210953,
-0.04884171858429909,
0.009592724964022636,
-0.055301595479249954,
0.02213858999311924,
-0.03547193855047226,
0.02234124206006527,
0.040530554950237274,
0.05179181694984436,
-0.005088122095912695,
-0.09623873233795166,
-0.08939872682094574,
0.01156159583479166,
0.040344949811697006,
0.10636016726493835,
0.045431673526763916,
0.09625159949064255,
-0.011773102916777134,
-0.0279021468013525,
-0.0386536680161953,
-0.032205283641815186,
-0.028250297531485558,
0.04344862326979637,
-0.03496256470680237,
-0.04138009250164032,
-0.026491308584809303,
-0.007623075973242521,
-0.09890305995941162,
0.01554433535784483,
0.022771088406443596,
0.08088979870080948,
0.05927983671426773,
-0.08398327976465225,
0.04883471131324768,
-0.01198774017393589,
0.08240463584661484,
-0.11317773908376694,
0.12873125076293945,
-0.016426431015133858,
-0.0404554083943367,
-0.07190494984388351,
-0.013380985707044601,
0.009072291664779186,
-0.06287863850593567,
0.03315047174692154,
-0.04223915562033653,
-0.025053508579730988,
0.04208046942949295,
0.0647406205534935,
-0.051630642265081406,
-0.15205618739128113,
-0.04561300203204155,
0.018128132447600365,
0.03220473229885101,
0.044954799115657806,
0.0022713292855769396,
0.09214794635772705,
0.1150471493601799,
0.021180303767323494,
-0.06987741589546204,
-0.03814883902668953,
-0.05509862303733826,
-0.0354321263730526,
0.013367523439228535,
-0.024489272385835648,
0.10557752847671509,
-0.04122529923915863,
0.04852273315191269,
-0.0704445168375969,
0.03773023188114166,
-0.01943337917327881,
0.0396292470395565,
0.03665309399366379,
-0.0004216174711473286,
0.047990188002586365,
0.054743919521570206,
-0.058392927050590515,
-0.05760689452290535,
0.0724305585026741,
-0.022811898961663246,
-0.061976995319128036,
0.01639845035970211,
-0.01405337080359459,
0.028517698869109154,
0.141856387257576,
-0.004792563151568174,
-0.006137808319181204,
0.06278020888566971,
0.052723877131938934,
0.03812900930643082,
-0.0218136478215456,
-0.02469979226589203,
0.08162323385477066,
0.0362212210893631,
0.0383761040866375,
0.09162771701812744,
0.008209173567593098,
-0.012335910461843014,
-0.012876213528215885,
0.0881626084446907,
0.041559621691703796,
0.07483229786157608,
-0.01572825200855732,
-0.0625007227063179,
-0.03990545868873596,
-0.03194640204310417,
0.061911553144454956,
0.003681091358885169,
0.0030198493041098118,
-0.08894094079732895,
-0.039127182215452194,
0.024100348353385925,
0.0175733994692564,
0.05423428863286972,
0.04410303756594658,
-0.038667239248752594,
0.10301754623651505,
0.04952163249254227,
0.016531141474843025,
0.009597518481314182,
0.05091976001858711,
0.012220809236168861,
0.10144559293985367,
0.02721291407942772,
0.18023626506328583,
0.04924995079636574,
0.027722608298063278,
7.249981713196425e-34,
0.08410828560590744,
0.03783512860536575,
0.018909744918346405,
0.08042077720165253,
0.03569542244076729,
0.02288755401968956,
-0.03362559899687767,
0.011418825015425682,
-0.0012079107109457254,
-0.012352872639894485,
0.06025124713778496,
-0.011928686872124672,
-0.021310895681381226,
0.028218330815434456,
-0.00026477721985429525,
0.005346654448658228,
-0.07397633790969849,
0.04677533730864525,
-0.053648192435503006,
-0.05527142062783241,
0.0382240004837513,
0.025818465277552605,
0.0544239804148674,
-0.0008748054387979209,
0.04462864622473717,
0.0033057546243071556,
0.038850195705890656,
-0.056923575699329376,
0.0021180626936256886,
0.015371199697256088,
-0.11065912246704102,
0.003939489834010601,
-0.0029969513416290283,
0.13984043896198273,
-0.06165008246898651,
-0.018063589930534363,
-0.04499099776148796,
-0.06888192147016525,
0.048077262938022614,
-0.022156711667776108,
-0.04587358608841896,
0.008019596338272095,
0.08689983189105988,
0.02635188028216362,
-0.033060573041439056,
-0.02151111327111721,
-0.10002008080482483,
-0.00853538978844881,
-0.03159411624073982,
0.005510298069566488,
0.021955549716949463,
0.04603300616145134,
0.026232801377773285,
-0.03985275700688362,
-0.04157369211316109,
-0.07351763546466827,
-0.030830757692456245,
-0.004384794272482395,
0.0173569917678833,
-0.03112681768834591,
0.00982397235929966,
0.016882941126823425,
-0.04083813726902008,
0.08782781660556793,
0.1046411544084549,
0.007085487712174654,
-0.04772104322910309,
0.0852091833949089,
0.08469407260417938,
0.00729463342577219,
-0.015557637438178062,
0.01615852303802967,
0.01656901091337204,
-0.055527959018945694,
-0.06540094316005707,
-0.06645796447992325,
-0.010307840071618557,
-0.041866809129714966,
0.004021971020847559,
-0.022199982777237892,
0.03829425200819969,
0.04004606604576111,
-0.11327420175075531,
-0.019736643880605698,
-0.03418610990047455,
-0.04208199307322502,
0.004558734130114317,
-0.06562691926956177,
0.005748124793171883,
-0.016546059399843216,
-0.04883774742484093,
-0.018540160730481148,
-0.06447196751832962,
-0.04117681086063385,
-0.03431932255625725,
-1.1447242249860794e-33,
-0.056900348514318466,
-0.06682419776916504,
-0.0592755489051342,
0.06358711421489716,
0.02297891490161419,
-0.06476859748363495,
0.007104127202183008,
-0.02144767716526985,
-0.002637987956404686,
-0.04403343424201012,
-0.07230570912361145,
-0.042408060282468796,
0.0003912279789801687,
0.02843409776687622,
-0.036716096103191376,
-0.08572462946176529,
0.023877067491412163,
0.00637446204200387,
0.026081372052431107,
-0.01485427562147379,
0.019236834719777107,
0.077996626496315,
-0.056290868669748306,
0.026909779757261276,
0.00553236436098814,
0.025912946090102196,
-0.0026981555856764317,
0.0017311618430539966,
0.03878620266914368,
0.018129415810108185,
-0.029764804989099503,
0.008248078636825085,
-0.015481073409318924,
-0.04547199234366417,
-0.013477080501616001,
0.11948332190513611,
-0.025021279230713844,
0.004101060330867767,
-0.05034082382917404,
0.0744171068072319,
-0.006659670267254114,
-0.11144167929887772,
-0.05734347552061081,
-0.023693261668086052,
-0.09178396314382553,
-0.0072641074657440186,
-0.09566503763198853,
0.008309543132781982,
-0.006265341304242611,
0.016829347237944603,
-0.04220256209373474,
-0.09446457773447037,
0.022493984550237656,
-0.028139589354395866,
-0.010666641406714916,
-0.08033167570829391,
0.025215597823262215,
-0.029327131807804108,
-0.015326675027608871,
-0.008705565705895424,
-0.07968167215585709,
0.05357752740383148,
0.047655798494815826,
-0.05412378907203674,
-0.001582476426847279,
-0.02569103054702282,
-0.05489969626069069,
-0.01152834389358759,
-0.025082431733608246,
0.0026021883822977543,
0.030902719125151634,
-0.01524396799504757,
-0.08034686744213104,
-0.012486284598708153,
0.1226666271686554,
0.05508356913924217,
-0.03276288881897926,
-0.02740469016134739,
-0.03974656015634537,
-0.0560968853533268,
-0.017136046662926674,
-0.006170467473566532,
0.02911464124917984,
0.06451071798801422,
-0.029769262298941612,
-0.0030465195886790752,
-0.029675057157874107,
0.15090472996234894,
0.017699550837278366,
0.03002130053937435,
0.005260335747152567,
0.021924791857600212,
0.036339886486530304,
0.0572935976088047,
-0.12389646470546722,
-3.924020930412553e-8,
-0.044741444289684296,
0.04389820992946625,
0.03421814739704132,
0.07833603024482727,
0.013642188161611557,
-0.037168849259614944,
-0.010904958471655846,
-0.012111077085137367,
-0.038279809057712555,
0.03503742814064026,
0.042802706360816956,
-0.025433801114559174,
0.0017495984211564064,
0.04209815710783005,
-0.00013399572344496846,
-0.017829157412052155,
-0.04806649312376976,
-0.008483144454658031,
-0.04167887941002846,
-0.07981951534748077,
0.04485037177801132,
-0.034600455313920975,
-0.07159359008073807,
-0.008436227217316628,
0.0260935015976429,
0.013047684915363789,
-0.09596602618694305,
0.05085629224777222,
-0.018499065190553665,
0.03866467624902725,
-0.006367741152644157,
0.015950195491313934,
0.0024139992892742157,
-0.017000950872898102,
0.03152456507086754,
0.018116513267159462,
-0.02277728170156479,
0.010709892958402634,
0.06734444200992584,
-0.03552190959453583,
-0.021199947223067284,
0.051654573529958725,
-0.014302062802016735,
0.04396475851535797,
0.04728512838482857,
-0.028503797948360443,
-0.11157658696174622,
-0.03446060046553612,
0.000059609825257211924,
-0.05646717548370361,
-0.03001805581152439,
-0.05021151900291443,
-0.014381542801856995,
0.08562144637107849,
0.07750951498746872,
0.03969921916723251,
0.04202937334775925,
-0.008534003980457783,
0.06209332123398781,
0.045237794518470764,
0.08522126078605652,
0.06411506980657578,
-0.050775282084941864,
0.0056428792886435986
] | 0
| 0
| 0
| 0
|
|
Resolution Attack: Exploiting Image Compression to Deceive Deep Neural Networks
|
https://openreview.net/forum?id=OFukl9Qg8P
|
[
"Wangjia Yu",
"Xiaomeng Fu",
"Qiao Li",
"Jizhong Han",
"Xiaodan Zhang"
] |
Poster
|
Model robustness is essential for ensuring the stability and reliability of machine learning systems. Despite extensive research on various aspects of model robustness, such as adversarial robustness and label noise robustness, the exploration of robustness towards different resolutions, remains less explored. To address this gap, we introduce a novel form of attack: the resolution attack. This attack aims to deceive both classifiers and human observers by generating images that exhibit different semantics across different resolutions. To implement the resolution attack, we propose an automated framework capable of generating dual-semantic images in a zero-shot manner. Specifically, we leverage large-scale diffusion models for their comprehensive ability to construct images and propose a staged denoising strategy to achieve a smoother transition across resolutions. Through the proposed framework, we conduct resolution attacks against various off-the-shelf classifiers. The experimental results exhibit high attack success rate, which not only validates the effectiveness of our proposed framework but also reveals the vulnerability of current classifiers towards different resolutions. Additionally, our framework, which incorporates features from two distinct objects, serves as a competitive tool for applications such as face swapping and facial camouflage. The code is available at https://github.com/ywj1/resolution-attack.
|
Resolution Attack、Image Generation、Deep Learning Robustness、Image Classification
|
This paper introduces the concept of resolution attacks, demonstrates how dual representation of images can lead to misclassification in deep learning classifiers, and how dual-stream generation denoising modules can generate dual-semantic images.
| 11,470
| null |
[
-0.07679097354412079,
-0.012910059653222561,
0.034020740538835526,
0.10828351974487305,
0.07648145407438278,
-0.03882671892642975,
-0.034843023866415024,
-0.1333850473165512,
-0.006685439497232437,
-0.028449634090065956,
-0.03149578720331192,
-0.04751165956258774,
0.037836380302906036,
0.009477868676185608,
-0.0401405431330204,
-0.030729835852980614,
0.034233272075653076,
0.1525353640317917,
-0.10076604038476944,
0.023172862827777863,
-0.015730254352092743,
-0.04681994020938873,
-0.031046263873577118,
-0.003495464799925685,
0.026859652251005173,
-0.019704904407262802,
-0.012653143145143986,
-0.055282942950725555,
0.0068139019422233105,
-0.04881327971816063,
0.07132457196712494,
-0.05533147230744362,
-0.07779868692159653,
0.035779938101768494,
-0.007892346009612083,
0.012043013237416744,
0.007686697877943516,
0.019645586609840393,
0.040172409266233444,
-0.044831786304712296,
0.025303540751338005,
0.003029357874765992,
0.012345216237008572,
0.022152867168188095,
0.03356153145432472,
0.03880968689918518,
0.029162148013710976,
0.03483067825436592,
0.009493785910308361,
-0.07854607701301575,
0.00444438261911273,
0.0005197460995987058,
-0.04790179058909416,
0.04722718521952629,
-0.09825676679611206,
-0.021754223853349686,
0.027003135532140732,
0.044357120990753174,
-0.011907552368938923,
0.08080654591321945,
-0.03704594075679779,
-0.04885225370526314,
0.014211113564670086,
0.04276035726070404,
-0.03878320753574371,
0.026242339983582497,
0.03829566016793251,
-0.0036455972585827112,
0.009324111044406891,
0.07930615544319153,
0.03290200233459473,
0.059878960251808167,
-0.012728340923786163,
0.03608693554997444,
0.0405636765062809,
0.014410428702831268,
-0.03321569412946701,
-0.04948422312736511,
0.06423655152320862,
-0.04000217840075493,
0.05167694762349129,
-0.03231799229979515,
0.059551410377025604,
0.02131003327667713,
0.09860177338123322,
-0.02858399972319603,
-0.012929468415677547,
0.017490804195404053,
0.05192402005195618,
-0.010340253822505474,
-0.1385323852300644,
0.00954838190227747,
0.01895960606634617,
0.005919117946177721,
0.03775404021143913,
0.005737696308642626,
-0.030407266691327095,
-0.058606430888175964,
-0.04352252930402756,
0.1168791726231575,
-0.04951730743050575,
-0.031586579978466034,
0.026602771133184433,
-0.07741688191890717,
0.09297458827495575,
0.024951636791229248,
0.08094686269760132,
-0.02222295291721821,
0.029574286192655563,
0.0019306916510686278,
0.07256605476140976,
0.04272199794650078,
0.011211936362087727,
-0.04633061960339546,
0.10688129812479019,
-0.019615529105067253,
-0.008669254370033741,
0.028272971510887146,
0.021883592009544373,
-0.03878781571984291,
0.06403322517871857,
0.024779099971055984,
-0.05043412744998932,
-0.08116433024406433,
-0.009345697239041328,
-0.008253321051597595,
-0.04460662230849266,
4.457760990271757e-35,
-0.0015806599985808134,
-0.018651578575372696,
0.05089348554611206,
0.06743517518043518,
0.10909081995487213,
-0.022442733868956566,
-0.041081544011831284,
0.03723287582397461,
-0.029238136485219002,
0.031239567324519157,
-0.025662733241915703,
-0.034426990896463394,
-0.061494871973991394,
0.11839602887630463,
0.03647927567362785,
0.07161381840705872,
0.002617478370666504,
0.046240322291851044,
-0.003225753316655755,
0.035453177988529205,
0.0708952397108078,
-0.06241366267204285,
-0.03766225650906563,
-0.014431729912757874,
-0.037004586309194565,
-0.06547760963439941,
0.043131276965141296,
-0.0225901547819376,
0.045705728232860565,
-0.002836828352883458,
-0.1453971266746521,
-0.013919309712946415,
0.060304611921310425,
0.06753668934106827,
0.017547445371747017,
-0.03348412364721298,
-0.010222483426332474,
0.030753690749406815,
-0.03340712934732437,
0.0354902483522892,
-0.0012822931166738272,
0.03148696944117546,
-0.06325893849134445,
-0.04214116558432579,
-0.04253166913986206,
0.016792917624115944,
-0.006893889978528023,
0.04443513974547386,
-0.09380968660116196,
-0.031738873571157455,
0.05815280228853226,
0.008399078622460365,
-0.021862048655748367,
-0.07225878536701202,
-0.035345833748579025,
-0.05163317173719406,
0.06231539696455002,
-0.04959990829229355,
-0.012160239741206169,
0.0025602243840694427,
0.020208612084388733,
-0.03426392003893852,
-0.07238362729549408,
0.028113188222050667,
-0.00669638067483902,
-0.011729173362255096,
0.008259954862296581,
0.03954165801405907,
-0.03797851502895355,
-0.013443995267152786,
-0.06093533709645271,
0.020245162770152092,
0.02251029573380947,
-0.05198213458061218,
0.034054674208164215,
0.022474821656942368,
0.11166156828403473,
0.047985486686229706,
0.014316162094473839,
0.053116969764232635,
-0.01002403162419796,
-0.06535419821739197,
0.02613653428852558,
-0.04740021377801895,
-0.07933797687292099,
0.013587841764092445,
-0.044861964881420135,
-0.04946316033601761,
-0.031831044703722,
-0.025213181972503662,
-0.033249009400606155,
-0.04377439618110657,
-0.004247589036822319,
0.0036868697497993708,
-0.007169549819082022,
-8.9477280002799e-34,
-0.02757779322564602,
0.06209057569503784,
-0.07305394113063812,
0.0168377086520195,
-0.06297964602708817,
-0.00433881301432848,
0.04314827173948288,
0.0800585150718689,
0.046108927577733994,
-0.06654105335474014,
-0.05635857954621315,
-0.022837277501821518,
0.02387799136340618,
-0.0595293864607811,
0.03199725225567818,
-0.0433502197265625,
0.04403412342071533,
-0.011515552178025246,
-0.07056659460067749,
0.03678836673498154,
0.03427531197667122,
0.16412733495235443,
-0.10243070125579834,
0.022427385672926903,
-0.03678083047270775,
0.03308466821908951,
-0.06736424565315247,
0.07985501736402512,
0.06753593683242798,
-0.01774028316140175,
-0.019186921417713165,
-0.033826809376478195,
0.03179657831788063,
-0.05325103923678398,
-0.07711484283208847,
0.08500395715236664,
0.04673776030540466,
-0.012540383264422417,
-0.03498624265193939,
0.09990578144788742,
0.02767634019255638,
-0.0859706848859787,
-0.0913805142045021,
0.09725236147642136,
0.04632554203271866,
-0.09705809503793716,
0.005632160697132349,
-0.03537483140826225,
0.011985204182565212,
0.013140193186700344,
-0.008317675441503525,
0.008998687379062176,
-0.10628382116556168,
0.05247557535767555,
-0.04346190392971039,
-0.037224579602479935,
-0.07244427502155304,
0.012085162103176117,
0.07244918495416641,
0.09253638237714767,
-0.01019079890102148,
-0.07293398678302765,
-0.09236988425254822,
-0.04314521700143814,
-0.041780099272727966,
0.0022598092909902334,
-0.08439082652330399,
0.002556098159402609,
0.028826497495174408,
0.0749073177576065,
0.039353758096694946,
-0.025531193241477013,
-0.04395848885178566,
0.0174536295235157,
-0.026954760774970055,
-0.0756010115146637,
0.016516851261258125,
-0.02911437302827835,
0.015450113452970982,
0.04946370795369148,
0.04951658099889755,
-0.06339190155267715,
0.09032724797725677,
0.041419293731451035,
0.08917989581823349,
0.011270983144640923,
0.03901983052492142,
0.013815993443131447,
0.04048541188240051,
0.05545998364686966,
-0.04347353056073189,
0.008200210519134998,
-0.029953425750136375,
0.038825348019599915,
0.020473288372159004,
-4.268024511588919e-8,
-0.06636613607406616,
0.09216982871294022,
-0.002729549538344145,
-0.02000567689538002,
-0.0387842059135437,
-0.04430514574050903,
-0.027536774054169655,
0.005820766091346741,
-0.0013866183580830693,
-0.02862839214503765,
0.1145893931388855,
0.008460317738354206,
0.00359740830026567,
-0.04879290238022804,
-0.06687415391206741,
0.07166237384080887,
-0.030977418646216393,
0.04589033126831055,
-0.037126392126083374,
-0.04663536697626114,
-0.07971888780593872,
0.02366139553487301,
-0.013502560555934906,
-0.054066434502601624,
0.03858800232410431,
-0.07874935120344162,
0.003801473183557391,
0.054208215326070786,
0.03915250673890114,
0.015873447060585022,
-0.07828706502914429,
0.013370331376791,
0.019563432782888412,
-0.015003102831542492,
0.021503327414393425,
0.12838785350322723,
0.012971293181180954,
0.046741023659706116,
0.035792358219623566,
-0.1189856007695198,
-0.05175234377384186,
0.0005574289825744927,
-0.026052694767713547,
0.020926496013998985,
-0.02917432226240635,
0.05215701833367348,
0.04279596731066704,
-0.06754684448242188,
0.039156362414360046,
0.06447733938694,
0.04567791894078255,
0.006437642499804497,
-0.002843892900273204,
0.09766904264688492,
0.004118968266993761,
-0.003627160796895623,
-0.008070130832493305,
-0.026953572407364845,
-0.020425714552402496,
0.15758633613586426,
0.055925846099853516,
0.03938879445195198,
0.02440938726067543,
0.0028297936078161
] | 0
| 0
| 0
| 0
|
|
SLMRec: Distilling Large Language Models into Small for Sequential Recommendation
|
https://openreview.net/forum?id=G4wARwjF8M
|
[
"Wujiang Xu",
"Qitian Wu",
"Zujie Liang",
"Jiaojiao Han",
"Xuying Ning",
"Yunxiao Shi",
"Wenfang Lin",
"Yongfeng Zhang"
] |
Poster
|
Sequential Recommendation (SR) task involves predicting the next item a user is likely to interact with, given their past interactions.
The SR models examine the sequence of a user's actions to discern more complex behavioral patterns and temporal dynamics.
Recent research demonstrates the great impact of LLMs on sequential recommendation systems, either viewing sequential recommendation as language modeling or serving as the backbone for user representation. Although these methods deliver outstanding performance, there is scant evidence of the necessity of a large language model and how large the language model is needed, especially in the sequential recommendation scene. Meanwhile, due to the huge size of LLMs, it is inefficient and impractical to apply a LLM-based model in real-world platforms that often need to process billions of traffic logs daily. In this paper, we explore the influence of LLMs' depth by conducting extensive experiments on large-scale industry datasets. Surprisingly, our motivational experiments reveal that most intermediate layers of LLMs are redundant, indicating that pruning the remaining layers can still maintain strong performance.
Motivated by this insight, we empower small language models for SR, namely SLMRec, which adopt a simple yet effective knowledge distillation method. Moreover, SLMRec is orthogonal to other post-training efficiency techniques, such as quantization and pruning, so that they can be leveraged in combination. Comprehensive experimental results illustrate that the proposed SLMRec model attains the best performance using only 13\% of the parameters found in LLM-based recommendation models while simultaneously achieving up to 6.6x and 8.0x speedups in training and inference time costs, respectively. Besides, we provide a theoretical justification for why small language models can perform comparably to large language models in SR.
|
Large Language Models; Knowledge Distillation
|
SLMs could perform better equipped with knowledge distillation than LLMs in recommendation task.
| 11,465
|
2405.17890
|
[
-0.06424777209758759,
-0.13339313864707947,
0.03449472784996033,
0.004757415037602186,
0.09378966689109802,
0.027446402236819267,
0.012278820388019085,
0.027047770097851753,
0.05937347561120987,
-0.11529937386512756,
-0.08118244260549545,
0.059669364243745804,
0.020878082141280174,
0.03714904561638832,
0.008521020412445068,
-0.05036946386098862,
0.14645378291606903,
-0.01206941343843937,
-0.04787714034318924,
-0.09288906306028366,
0.035721197724342346,
0.04409320652484894,
0.029383746907114983,
-0.033541832119226456,
0.011226840317249298,
0.01632562279701233,
-0.01356078777462244,
-0.004467932507395744,
-0.0011211257660761476,
0.00005515390512300655,
0.022708794102072716,
0.09486152976751328,
-0.01055929996073246,
0.018153436481952667,
-0.057656291872262955,
-0.0005088477046228945,
-0.08258222788572311,
-0.0485374853014946,
-0.045555874705314636,
-0.052108824253082275,
-0.028554653748869896,
0.004524039104580879,
0.00667561637237668,
0.030727583914995193,
0.012410457246005535,
-0.08464586734771729,
0.028788773342967033,
-0.04427693039178848,
-0.041967716068029404,
0.028323426842689514,
-0.07161020487546921,
-0.04049728810787201,
0.05953158065676689,
0.055625852197408676,
-0.05106353759765625,
-0.033081457018852234,
0.010831859894096851,
0.03200460225343704,
-0.024961454793810844,
-0.03237343207001686,
-0.05905885621905327,
-0.14359821379184723,
-0.05280964821577072,
-0.04187601059675217,
-0.02172362245619297,
-0.02777615748345852,
0.014398635365068913,
0.04428212717175484,
-0.04094097763299942,
0.0697740688920021,
-0.03112892061471939,
0.061720915138721466,
-0.05192974582314491,
0.05171891674399376,
0.002810451900586486,
-0.026600828394293785,
-0.003147493116557598,
-0.05949774011969566,
0.02889522910118103,
-0.09797292202711105,
-0.03681913763284683,
-0.044706910848617554,
0.011369244195520878,
0.03001503273844719,
0.0672571137547493,
-0.02988164685666561,
0.05112207308411598,
-0.020354686304926872,
-0.01148893777281046,
0.03276403620839119,
-0.04500877857208252,
-0.0031444167252629995,
0.05472611263394356,
-0.04657464101910591,
-0.011287327855825424,
0.09842761605978012,
-0.011777381412684917,
-0.014857133850455284,
-0.008102438412606716,
0.04450932890176773,
0.03568821772933006,
0.13083010911941528,
0.03981606289744377,
-0.018997741863131523,
-0.005969860590994358,
-0.02810402400791645,
-0.041206032037734985,
0.07885748893022537,
0.023359350860118866,
-0.00913798063993454,
-0.05244017392396927,
0.08509844541549683,
0.01479397527873516,
0.013351946137845516,
0.06122259795665741,
-0.06685996800661087,
0.04295505955815315,
-0.0550730861723423,
-0.051879413425922394,
0.10582330077886581,
-0.0022415427956730127,
0.10474123805761337,
0.06900084763765335,
-0.03870771825313568,
0.019378649070858955,
0.06515385210514069,
-0.007919196970760822,
8.3444046063654e-34,
0.08111076802015305,
-0.01721556857228279,
-0.02152383141219616,
0.05138842761516571,
0.04087081924080849,
-0.04382837563753128,
0.029020577669143677,
0.0737266018986702,
-0.03261281177401543,
-0.031000647693872452,
-0.03166288137435913,
0.05346354469656944,
-0.04207468405365944,
0.09299495071172714,
0.024483697488904,
0.02958548627793789,
0.015985814854502678,
0.045175086706876755,
-0.023864176124334335,
-0.07848141342401505,
0.047410398721694946,
0.010076411068439484,
0.055642083287239075,
-0.058990638703107834,
0.04278168827295303,
-0.0030352522153407335,
0.03848094865679741,
0.006239165551960468,
0.017481373623013496,
-0.005801775027066469,
-0.018983058631420135,
0.011707979254424572,
0.002182340482249856,
0.05233202502131462,
0.027527496218681335,
0.011290986090898514,
-0.03348951414227486,
-0.11428140103816986,
0.11617504060268402,
-0.027913767844438553,
-0.02789807878434658,
0.026623161509633064,
-0.0019778823480010033,
-0.020120320841670036,
-0.10572204738855362,
-0.03908234089612961,
-0.023244790732860565,
0.035827506333589554,
-0.0038527303840965033,
0.02283686213195324,
0.11780334264039993,
0.006829080171883106,
0.026000024750828743,
0.03312258794903755,
-0.03252387419342995,
-0.016998354345560074,
-0.012480252422392368,
-0.07134541869163513,
-0.005258571822196245,
0.03492151200771332,
0.0225691981613636,
0.04294665902853012,
0.02566193789243698,
0.033030956983566284,
0.05392925441265106,
0.010671625845134258,
0.009255537763237953,
0.11465147882699966,
0.018433082848787308,
-0.030500050634145737,
0.00045465404400601983,
0.024654999375343323,
0.001458813901990652,
-0.01781458593904972,
0.023568375036120415,
0.032374512404203415,
0.019888443872332573,
-0.11658284813165665,
-0.007272009737789631,
0.040580570697784424,
0.028935052454471588,
-0.01895788498222828,
0.006427079904824495,
-0.0639236643910408,
-0.03170803561806679,
-0.02846052311360836,
0.06355581432580948,
-0.1067158579826355,
-0.004924449138343334,
-0.049838703125715256,
-0.09640507400035858,
-0.011173038743436337,
-0.03581869602203369,
0.03913087397813797,
-0.0028859153389930725,
-1.5323231915250778e-33,
-0.014687726274132729,
-0.07769843935966492,
0.06594707816839218,
0.09509801119565964,
0.012686125934123993,
-0.08309061080217361,
-0.038853637874126434,
-0.011269915848970413,
-0.02923177368938923,
-0.015019461512565613,
-0.03247089311480522,
-0.055487290024757385,
0.12262649834156036,
-0.0048703947104513645,
0.0864291563630104,
-0.006581003312021494,
0.022416794672608376,
-0.003695941064506769,
0.021790506318211555,
0.013080734759569168,
0.03189832717180252,
0.042028069496154785,
-0.13703227043151855,
-0.01897260919213295,
-0.05011356994509697,
-0.03423667699098587,
-0.002371301408857107,
0.034406665712594986,
0.010148274712264538,
-0.02964096888899803,
-0.0017021532403305173,
-0.013903499580919743,
-0.03116179071366787,
-0.050055764615535736,
-0.03911437466740608,
0.04264683648943901,
0.03355550765991211,
0.00832382868975401,
-0.023754190653562546,
0.025590000674128532,
0.06529794633388519,
-0.03411436825990677,
-0.04031762480735779,
0.016000496223568916,
-0.060161929577589035,
-0.029792936518788338,
-0.1073908731341362,
0.014437077566981316,
-0.018557121977210045,
-0.04198968783020973,
-0.045755527913570404,
0.053827665746212006,
-0.00821696501225233,
0.017187045887112617,
-0.05146579071879387,
-0.027492325752973557,
-0.030426573008298874,
-0.01705312728881836,
0.03731490299105644,
-0.018804386258125305,
-0.017997656017541885,
0.02622036635875702,
0.035245299339294434,
-0.09725499898195267,
0.014305460266768932,
-0.06735038012266159,
0.02736154943704605,
-0.07666614651679993,
-0.04746043309569359,
-0.07584567368030548,
0.0951499491930008,
0.01891653798520565,
-0.0683349147439003,
0.04461543634533882,
-0.04057835415005684,
-0.06067657843232155,
-0.005752114579081535,
-0.06295141577720642,
-0.010320911183953285,
-0.04516143724322319,
0.022804008796811104,
0.05406750366091728,
-0.01401235070079565,
-0.027256770059466362,
0.08272871375083923,
0.056876908987760544,
0.0409943088889122,
0.13000330328941345,
-0.014987119473516941,
-0.02199583873152733,
0.0005342754302546382,
-0.007961414754390717,
-0.007730789482593536,
0.07597022503614426,
-0.03673670068383217,
-4.449170631914967e-8,
-0.10197661817073822,
0.028259798884391785,
0.03786313161253929,
0.10613059252500534,
0.04680447280406952,
-0.018164964392781258,
-0.01254959125071764,
0.1090490072965622,
0.0043786754831671715,
0.013698401860892773,
0.05563953518867493,
-0.0007312644156627357,
-0.053932614624500275,
-0.03228147700428963,
0.04221644625067711,
0.05383625999093056,
0.06241849064826965,
-0.060553163290023804,
-0.017353637143969536,
-0.018304618075489998,
0.06367073953151703,
0.060989730060100555,
0.038065072149038315,
0.015237768180668354,
0.07616829127073288,
-0.00931268185377121,
0.009093821980059147,
0.1392124742269516,
0.11502541601657867,
0.030766166746616364,
-0.008408516645431519,
-0.021075278520584106,
-0.016583161428570747,
-0.09193258732557297,
-0.011163303628563881,
-0.006989570334553719,
0.013169689103960991,
-0.07562313228845596,
-0.0015510455705225468,
-0.07494660466909409,
0.031484946608543396,
0.0037575806491076946,
0.023738078773021698,
0.06046894192695618,
-0.007886339910328388,
0.0011617491254583001,
-0.08380211144685745,
-0.13589893281459808,
0.03346768766641617,
-0.015435037203133106,
-0.021747790277004242,
-0.09166447073221207,
0.048855192959308624,
0.08098511397838593,
0.07459508627653122,
0.031432900577783585,
0.005426191259175539,
-0.03180762380361557,
0.051340412348508835,
0.04130486026406288,
0.058811500668525696,
-0.03415443375706673,
-0.08596988022327423,
0.07539987564086914
] |
https://github.com/wujiangxu/slmrec
| 20
| 0
| 0
| 0
|
How Do Large Language Models Understand Graph Patterns? A Benchmark for Graph Pattern Comprehension
|
https://openreview.net/forum?id=CkKEuLmRnr
|
[
"Xinnan Dai",
"Haohao Qu",
"Yifei Shen",
"Bohang Zhang",
"Qihao Wen",
"Wenqi Fan",
"Dongsheng Li",
"Jiliang Tang",
"Caihua Shan"
] |
Poster
|
Benchmarking the capabilities and limitations of large language models (LLMs) in graph-related tasks is becoming an increasingly popular and crucial area of research. Recent studies have shown that LLMs exhibit a preliminary ability to understand graph structures and node features. However, the potential of LLMs in graph pattern mining remains largely unexplored. This is a key component in fields such as computational chemistry, biology, and social network analysis. To bridge this gap, this work introduces a comprehensive benchmark to assess LLMs' capabilities in graph pattern tasks. We have developed a benchmark that evaluates whether LLMs can understand graph patterns based on either terminological or topological descriptions. Additionally, our benchmark tests the LLMs' capacity to autonomously discover graph patterns from data. The benchmark encompasses both synthetic and real datasets, and a variety of models, with a total of 11 tasks and 7 models. Our experimental framework is designed for easy expansion to accommodate new models and datasets. Our findings reveal that: (1) LLMs have preliminary abilities to understand graph patterns, with O1-mini outperforming in the majority of tasks; (2) Formatting input graph data to align with the knowledge acquired during pretraining can enhance performance; (3) LLMs employ diverse
potential algorithms to solve one task, with performance varying based on their execution capabilities.
|
Large language models, graph pattern, graph mining
|
We propose a comprehensive benchmark to assess LLMs' capabilities in graph pattern tasks.
| 11,464
|
2410.05298
|
[
-0.014963695779442787,
-0.07111605256795883,
0.0212677214294672,
0.01464071124792099,
0.09382178634405136,
-0.04640624672174454,
-0.07320234179496765,
-0.040183596312999725,
0.005287023261189461,
-0.011231680400669575,
-0.04040304198861122,
-0.0691809430718422,
0.05597564950585365,
0.05876634642481804,
0.05101817846298218,
0.05339447781443596,
0.0713479220867157,
0.002507939236238599,
-0.000384156039217487,
-0.05992767959833145,
0.05568917468190193,
0.01079531293362379,
0.02482011169195175,
-0.035271141678094864,
0.10831502079963684,
0.04134569317102432,
-0.05341539904475212,
-0.005336330272257328,
0.04124894738197327,
0.011096948757767677,
0.0796656608581543,
0.08842337876558304,
-0.01255420595407486,
0.08532978594303131,
0.005355995614081621,
0.03396293520927429,
-0.07452922314405441,
0.011699967086315155,
0.019533805549144745,
0.013949736952781677,
0.0052965376526117325,
-0.007072126492857933,
0.06348903477191925,
0.03265130892395973,
0.08034507185220718,
0.01924831047654152,
-0.04837881401181221,
-0.016571514308452606,
-0.02330849878489971,
-0.005790865980088711,
-0.12087234109640121,
-0.07424284517765045,
-0.04543953388929367,
0.017967643216252327,
0.03668585419654846,
0.010593296959996223,
0.057287413626909256,
0.01087601762264967,
0.003882860764861107,
-0.02982931397855282,
0.03667972609400749,
-0.14144277572631836,
-0.03047887608408928,
0.026390869170427322,
-0.010140922851860523,
0.019639745354652405,
0.012929011136293411,
0.08169816434383392,
0.04657719284296036,
0.029054904356598854,
-0.01603551022708416,
0.08560916781425476,
-0.03072689101099968,
0.07424882054328918,
0.061134278774261475,
0.049993377178907394,
-0.029826611280441284,
-0.015694987028837204,
-0.026030750945210457,
-0.12915495038032532,
-0.03571917861700058,
-0.002920052269473672,
0.021205363795161247,
0.031568899750709534,
0.039377305656671524,
0.007331511937081814,
0.0053385160863399506,
-0.04221426695585251,
-0.02658792957663536,
0.045978691428899765,
-0.0865759402513504,
-0.06938689947128296,
0.015039161778986454,
-0.05495558679103851,
0.010223698802292347,
0.07774276286363602,
0.06128096580505371,
-0.05189356207847595,
-0.0034166553523391485,
0.04044286906719208,
-0.010723710060119629,
0.040652841329574585,
0.03941415995359421,
-0.09418477863073349,
-0.029780084267258644,
0.01269766315817833,
-0.0458979532122612,
0.023878438398241997,
0.021553371101617813,
-0.02535771019756794,
-0.011603115126490593,
0.09295192360877991,
0.01937553659081459,
0.06220749020576477,
0.021034138277173042,
-0.07889249175786972,
0.07031463086605072,
-0.0827597975730896,
0.040650393813848495,
0.13361601531505585,
-0.029618628323078156,
0.09575658291578293,
0.02507806196808815,
-0.07735364139080048,
0.01901986449956894,
0.03824504464864731,
-0.0774364098906517,
9.677956111270244e-34,
0.11692433059215546,
0.0294412299990654,
0.014458582736551762,
0.03875017538666725,
0.10387259721755981,
0.015861565247178078,
-0.05903999134898186,
0.0005095030646771193,
-0.029236633330583572,
0.08778861165046692,
-0.08667388558387756,
0.05362609401345253,
-0.023795852437615395,
0.07196895778179169,
0.056054580956697464,
0.019823899492621422,
0.05400475859642029,
-0.013268721289932728,
-0.06317339092493057,
-0.05787033215165138,
0.008566237986087799,
-0.029723361134529114,
0.015443114563822746,
-0.006830974016338587,
0.014734961092472076,
0.08106432855129242,
0.029802465811371803,
-0.04201173037290573,
0.00850759632885456,
0.0033820432145148516,
-0.08141409605741501,
0.007072960492223501,
0.02856515161693096,
0.07707329839468002,
0.031205708160996437,
-0.031369488686323166,
-0.0026366577949374914,
-0.05626431480050087,
0.03201703354716301,
-0.07976090908050537,
-0.02237825095653534,
0.038263581693172455,
-0.007551409304141998,
0.018233291804790497,
-0.06137977913022041,
-0.036949124187231064,
-0.006888486910611391,
-0.027915552258491516,
0.020628657191991806,
-0.033798523247241974,
0.035423941910266876,
-0.04410422965884209,
-0.00034548970870673656,
0.07346203178167343,
0.01794878952205181,
0.010395967401564121,
-0.010026569478213787,
0.020339353010058403,
0.0004980749217793345,
0.1267213374376297,
0.07298790663480759,
0.020222200080752373,
0.0328415110707283,
0.06748411804437637,
0.05209249258041382,
0.007657898589968681,
-0.11823053658008575,
0.07134192436933517,
0.022850383073091507,
0.0002637777361087501,
-0.021608764305710793,
0.0906018614768982,
-0.045083507895469666,
-0.08513877540826797,
0.07078757137060165,
0.02279919758439064,
0.001945684663951397,
-0.13408498466014862,
-0.06493489444255829,
0.01650497317314148,
-0.015971746295690536,
-0.09804148972034454,
0.04059166461229324,
-0.047283586114645004,
-0.0824509784579277,
-0.040181200951337814,
0.09076936542987823,
-0.07349962741136551,
0.02515982650220394,
-0.03965816646814346,
-0.07228012382984161,
-0.054326560348272324,
-0.025463800877332687,
-0.030112765729427338,
0.024221044033765793,
-1.6616407944269762e-33,
-0.08527839183807373,
0.01649003103375435,
0.06331852078437805,
0.04224786534905434,
-0.007843255065381527,
-0.13711461424827576,
-0.04249601066112518,
-0.014031973667442799,
-0.04624798148870468,
0.010027342475950718,
-0.037081532180309296,
-0.06409259140491486,
0.028652193024754524,
-0.037466492503881454,
0.041349850594997406,
-0.06517782807350159,
0.002097440417855978,
-0.007313190959393978,
0.025357333943247795,
0.09753933548927307,
-0.05305160582065582,
0.09153519570827484,
-0.1336791068315506,
0.0018765488639473915,
0.02722550928592682,
0.056003257632255554,
-0.05290647968649864,
-0.05463720113039017,
-0.018745193257927895,
0.03643939644098282,
-0.053010839968919754,
0.0038826900999993086,
-0.03673343360424042,
-0.046998683363199234,
-0.006544468924403191,
0.05931244045495987,
0.025214917957782745,
-0.03764951974153519,
-0.020485030487179756,
-0.0042702858336269855,
0.03542238473892212,
-0.01516129169613123,
-0.09146993607282639,
-0.01779969595372677,
0.005013972986489534,
0.028192991390824318,
-0.10116969794034958,
0.04498516395688057,
-0.04531814157962799,
-0.00021253076556604356,
-0.024370092898607254,
0.024757426232099533,
0.027406254783272743,
-0.025664828717708588,
0.025171367451548576,
-0.0822192057967186,
-0.042798999696969986,
0.003709266195073724,
-0.07717093080282211,
-0.06124110892415047,
-0.03606916964054108,
-0.0058539146557450294,
-0.009491250850260258,
-0.01105043850839138,
-0.009947209618985653,
-0.07404907047748566,
-0.026685109362006187,
-0.04550584778189659,
0.007721499539911747,
-0.011564087122678757,
0.070167675614357,
-0.01708412356674671,
-0.08217401802539825,
-0.012096731923520565,
0.04375913366675377,
-0.0012216797331348062,
0.0014520633267238736,
-0.026480020955204964,
-0.03535729646682739,
-0.028462247923016548,
0.005399411078542471,
0.006777984090149403,
-0.016454622149467468,
0.053258661180734634,
0.054210081696510315,
0.04841608926653862,
-0.009121716022491455,
0.13224981725215912,
0.04071665555238724,
0.04797850921750069,
0.005713123362511396,
-0.009863701649010181,
-0.13190892338752747,
0.08980448544025421,
-0.013952096924185753,
-4.1141444029335617e-8,
-0.1439974457025528,
-0.0060525317676365376,
0.0062032840214669704,
-0.02650023065507412,
0.02406182885169983,
0.028412874788045883,
0.03945949673652649,
0.12299712002277374,
-0.0011766614625230432,
0.05912027135491371,
0.07109155505895615,
0.0459350161254406,
-0.05718152970075607,
-0.04752318188548088,
0.032149787992239,
0.051657188683748245,
0.05346371605992317,
0.017454111948609352,
-0.009092223830521107,
0.031734898686409,
-0.008852529339492321,
0.0036976297851651907,
-0.025091756135225296,
0.03700166568160057,
0.043663546442985535,
-0.0963323563337326,
-0.053889233618974686,
0.04904962331056595,
0.01655498519539833,
0.0020436588674783707,
-0.0587698295712471,
-0.020882951095700264,
0.002553411992266774,
-0.011286028660833836,
0.015371749177575111,
0.05209124833345413,
0.0062482538633048534,
-0.008536584675312042,
0.006257074419409037,
-0.01828969456255436,
0.039219990372657776,
0.048058610409498215,
-0.01590431109070778,
0.03665683791041374,
-0.008442430756986141,
0.005957816261798143,
0.024937139824032784,
0.03639580309391022,
0.06219645217061043,
-0.09660670161247253,
-0.08323902636766434,
-0.007114209234714508,
-0.0685153529047966,
-0.0025757441762834787,
-0.013637516647577286,
0.07255689799785614,
-0.04982716590166092,
-0.029689496383070946,
0.04300708323717117,
-0.012070532888174057,
0.021497951820492744,
-0.028134912252426147,
-0.03410235419869423,
-0.041122354567050934
] | 0
| 0
| 0
| 0
|
|
Efficient Active Imitation Learning with Random Network Distillation
|
https://openreview.net/forum?id=GFgn2LprFR
|
[
"Emilien Biré",
"Anthony Kobanda",
"Ludovic Denoyer",
"Rémy Portelas"
] |
Poster
|
Developing agents for complex and underspecified tasks, where no clear objective exists, remains challenging but offers many opportunities. This is especially true in video games, where simulated players (bots) need to play realistically, and there is no clear reward to evaluate them. While imitation learning has shown promise in such domains, these methods often fail when agents encounter out-of-distribution scenarios during deployment. Expanding the training dataset is a common solution, but it becomes impractical or costly when relying on human demonstrations. This article addresses active imitation learning, aiming to trigger expert intervention only when necessary, reducing the need for constant expert input along training. We introduce Random Network Distillation DAgger (RND-DAgger), a new active imitation learning method that limits expert querying by using a learned state-based out-of-distribution measure to trigger interventions. This approach avoids frequent expert-agent action comparisons, thus making the expert intervene only when it is useful. We evaluate RND-DAgger against traditional imitation learning and other active approaches in 3D video games (racing and third-person navigation) and in a robotic locomotion task and show that RND-DAgger surpasses previous methods by reducing expert queries.
https://sites.google.com/view/rnd-dagger
|
Active Imitation Learning, Imitation Learning, Interactive Learning, Navigation
|
Human in the loop interactive training during which the agent is deciding when expert demonstration is needed
| 11,451
|
2411.01894
|
[
-0.04094834625720978,
-0.0928860604763031,
0.059380341321229935,
0.011765371076762676,
-0.03621894493699074,
0.0075773317366838455,
0.03707238659262657,
-0.030814113095402718,
-0.010700209997594357,
0.053112924098968506,
-0.005631589330732822,
-0.01807151548564434,
0.014794346876442432,
0.04826042428612709,
0.014544776640832424,
0.016303962096571922,
0.13848638534545898,
0.10172130167484283,
-0.07963943481445312,
0.016854116693139076,
0.04191485419869423,
-0.06905015558004379,
0.07292885333299637,
0.010180745273828506,
-0.05502810329198837,
-0.04909469932317734,
0.023139476776123047,
-0.05612720921635628,
-0.0011887726141139865,
-0.005320284049957991,
0.05040288344025612,
-0.08448078483343124,
-0.01265591848641634,
-0.07367023825645447,
-0.06987971067428589,
0.019188804551959038,
-0.12105374783277512,
0.022626124322414398,
0.001762186293490231,
-0.05094137042760849,
-0.04136737063527107,
0.017911965027451515,
0.004923089873045683,
0.05573286861181259,
0.06182468309998512,
0.04628170281648636,
0.0032302988693118095,
-0.0055933245457708836,
0.11900898814201355,
0.020169328898191452,
-0.11623086780309677,
-0.02080860547721386,
0.09664355218410492,
-0.03234867751598358,
0.0352233462035656,
0.057133179157972336,
0.06160416826605797,
0.03526916354894638,
0.04603563994169235,
0.00017643201863393188,
0.04626870155334473,
-0.09890749305486679,
0.002571907825767994,
-0.024488363415002823,
-0.04838453233242035,
-0.0010534664615988731,
-0.007743991445749998,
0.044473808258771896,
0.033137332648038864,
-0.018357660621404648,
-0.01465702336281538,
0.03448573499917984,
-0.003396391635760665,
0.006908508948981762,
0.03453974798321724,
-0.0018126929644495249,
-0.039233360439538956,
-0.06978197395801544,
-0.030635764822363853,
0.014278640039265156,
0.028263622894883156,
-0.038314346224069595,
0.04468709975481033,
0.054793257266283035,
0.07498396933078766,
0.02217680960893631,
-0.03137040138244629,
0.033327918499708176,
0.06299789994955063,
0.01773926429450512,
-0.11812211573123932,
0.002117718104273081,
0.017843078821897507,
-0.042371250689029694,
-0.015775136649608612,
0.036127109080553055,
-0.03773481026291847,
-0.016428541392087936,
-0.05675471946597099,
0.09296571463346481,
0.009548352099955082,
0.008654331788420677,
-0.05021747946739197,
-0.04297187551856041,
-0.0016418116865679622,
0.00820621382445097,
0.05745639652013779,
0.0060982415452599525,
0.09160807728767395,
0.0034981288481503725,
0.002721581142395735,
0.019908051937818527,
0.01087143924087286,
0.049586180597543716,
-0.013380217365920544,
0.06871035695075989,
-0.01936422288417816,
0.04645756259560585,
-0.05714825168251991,
0.11623076349496841,
0.04483254998922348,
-0.026578443124890327,
0.029760299250483513,
0.016817225143313408,
0.10843416303396225,
0.00011313844152027741,
0.03333735838532448,
1.9937405859840754e-34,
0.0736672580242157,
0.006723043043166399,
0.037605393677949905,
0.0556870736181736,
0.012325959280133247,
-0.028293263167142868,
0.07073582708835602,
-0.041364237666130066,
-0.014812005683779716,
-0.018527960404753685,
0.009296989068388939,
0.03430875390768051,
-0.07136078178882599,
0.05285918712615967,
0.006832548417150974,
-0.0022631585597991943,
-0.06699086725711823,
-0.021468570455908775,
-0.048890430480241776,
0.006026099435985088,
0.1274460256099701,
-0.06840509921312332,
-0.021044069901108742,
-0.03052986040711403,
-0.009955385699868202,
-0.03489741310477257,
-0.0024525520857423544,
0.014214782044291496,
0.0965067595243454,
-0.007394462823867798,
-0.056003641337156296,
-0.03797291964292526,
-0.08112070709466934,
0.03394961357116699,
0.002724145073443651,
-0.011368886567652225,
-0.044694624841213226,
-0.0650700032711029,
-0.06460808217525482,
0.018196070566773415,
-0.02930689975619316,
-0.005046758335083723,
0.015045048668980598,
-0.0012984945205971599,
-0.08448351919651031,
-0.08761049062013626,
0.030403539538383484,
-0.026642385870218277,
-0.05413100868463516,
-0.045511193573474884,
0.05126471444964409,
0.04698946699500084,
0.05260371044278145,
-0.09385734051465988,
-0.02484949678182602,
-0.00430001225322485,
0.07927889376878738,
0.010908062569797039,
0.0003225996042601764,
0.004386638756841421,
-0.017756707966327667,
-0.049416713416576385,
-0.022627033293247223,
0.15020745992660522,
-0.013893332332372665,
0.07756170630455017,
0.02066347748041153,
0.08684922009706497,
0.07815510779619217,
0.0044680931605398655,
-0.02120734006166458,
0.007290669716894627,
-0.014137947000563145,
-0.04268496111035347,
0.046755608171224594,
0.0013030616100877523,
-0.0011541605927050114,
-0.09293323010206223,
0.01115159597247839,
0.023960191756486893,
-0.007871488109230995,
-0.057739268988370895,
-0.11519066989421844,
0.020034819841384888,
-0.007311864756047726,
-0.044207822531461716,
-0.050840187817811966,
-0.08695034682750702,
0.014925041235983372,
-0.030612442642450333,
-0.08412972092628479,
-0.030853435397148132,
-0.09918413311243057,
0.0760091170668602,
0.00800483301281929,
-9.492800138981387e-34,
0.017376694828271866,
0.012839586474001408,
0.02516636624932289,
0.09258074313402176,
0.008044284768402576,
-0.014562537893652916,
-0.012694944627583027,
-0.040003444999456406,
0.015383966267108917,
-0.05132146552205086,
-0.09436384588479996,
0.0011014720657840371,
0.06421331316232681,
0.011625411920249462,
-0.017574509605765343,
-0.025084948167204857,
-0.033366214483976364,
0.07034561783075333,
-0.0005163141759112477,
-0.006212745793163776,
0.017752274870872498,
0.08419403433799744,
-0.011706765741109848,
-0.016098931431770325,
-0.04019170254468918,
0.029951270669698715,
0.022462695837020874,
0.13728196918964386,
0.032969117164611816,
0.05460066348314285,
-0.01862209476530552,
0.011761362664401531,
0.055696386843919754,
-0.025184182450175285,
-0.06841734796762466,
0.14008119702339172,
0.0660935714840889,
0.06197025999426842,
-0.10141796618700027,
0.07712088525295258,
0.03935684263706207,
-0.07706481218338013,
-0.05524025484919548,
0.0038119181990623474,
-0.049330320209264755,
0.019531743600964546,
-0.05367673188447952,
-0.003506567096337676,
-0.030510911718010902,
0.034244850277900696,
0.010577134788036346,
0.051961399614810944,
-0.06427789479494095,
-0.04635442793369293,
-0.06913754343986511,
-0.037327948957681656,
-0.055405307561159134,
0.06683852523565292,
0.05262375995516777,
0.02643190510571003,
-0.01690697856247425,
0.018985653296113014,
-0.004889737814664841,
0.07820755988359451,
-0.0507209450006485,
0.024334078654646873,
-0.03036881983280182,
0.0949941873550415,
0.03741345927119255,
-0.0015992324333637953,
-0.012971348129212856,
0.039181001484394073,
0.030927713960409164,
-0.060968708246946335,
0.031055884435772896,
-0.011073295958340168,
-0.02547919563949108,
0.0016224446007981896,
0.005522170104086399,
-0.004062453750520945,
0.013642956502735615,
-0.02007465809583664,
0.052869752049446106,
0.02750316821038723,
0.03365087881684303,
0.09738387912511826,
-0.035161733627319336,
0.0723598301410675,
-0.00476897181943059,
0.03846059739589691,
0.02174489200115204,
0.05380220711231232,
0.015067724511027336,
-0.007676792796701193,
-0.03623747453093529,
-4.173346468405725e-8,
-0.056944698095321655,
0.09378255158662796,
0.11093349754810333,
0.026923784986138344,
0.04458451271057129,
0.010098195634782314,
-0.08277323842048645,
-0.000293743796646595,
-0.06580716371536255,
-0.030747544020414352,
0.11870015412569046,
-0.00006605352973565459,
0.022477246820926666,
-0.08899445831775665,
-0.006912246346473694,
0.022990036755800247,
-0.0030746308621019125,
0.010996426455676556,
-0.04078809916973114,
0.048083409667015076,
0.06875045597553253,
-0.03683605045080185,
0.011584046296775341,
-0.033444784581661224,
-0.01399676501750946,
-0.06580595672130585,
-0.0823829397559166,
0.02751505933701992,
0.045103769749403,
0.038724254816770554,
0.016373805701732635,
-0.04520805552601814,
0.06616099923849106,
-0.05020001158118248,
-0.04158009588718414,
0.02721615694463253,
0.011996020562946796,
-0.10679145902395248,
-0.024388782680034637,
-0.0583820678293705,
-0.05495135486125946,
0.009063774719834328,
0.03258935734629631,
-0.007546989247202873,
-0.07616763561964035,
-0.010104184970259666,
-0.06838962435722351,
-0.10172683000564575,
0.09911021590232849,
-0.08183524757623672,
0.012041380628943443,
-0.08493222296237946,
-0.08436280488967896,
0.08239804953336716,
0.08296220749616623,
0.05527224391698837,
-0.015663273632526398,
-0.01712813600897789,
0.00016013662389013916,
0.05474335327744484,
0.024484243243932724,
-0.026690220460295677,
-0.070391945540905,
-0.06986451894044876
] | 0
| 0
| 0
| 0
|
|
Model-based RL as a Minimalist Approach to Horizon-Free and Second-Order Bounds
|
https://openreview.net/forum?id=txD9llAYn9
|
[
"Zhiyong Wang",
"Dongruo Zhou",
"John C.S. Lui",
"Wen Sun"
] |
Poster
|
Learning a transition model via Maximum Likelihood Estimation (MLE) followed by planning inside the learned model is perhaps the most standard and simplest Model-based Reinforcement Learning (RL) framework. In this work, we show that such a simple Model-based RL scheme, when equipped with optimistic and pessimistic planning procedures, achieves strong regret and sample complexity bounds in online and offline RL settings. Particularly, we demonstrate that under the conditions where the trajectory-wise reward is normalized between zero and one and the transition is time-homogenous, it achieves nearly horizon-free and second-order bounds.
|
reinforcement learning theory, model-based reinforcement learning
| null | 11,438
|
2408.08994
|
[
-0.029758790507912636,
-0.030764847993850708,
0.03151388838887215,
0.01119093969464302,
0.0323072224855423,
0.05071152746677399,
-0.10055903345346451,
0.042064566165208817,
0.062445178627967834,
0.05968080833554268,
-0.03760131075978279,
0.032927632331848145,
0.002277367515489459,
0.07137434184551239,
0.02217979170382023,
0.031247273087501526,
-0.012423570267856121,
-0.016389157623052597,
-0.026254573836922646,
-0.019825825467705727,
0.0005042360862717032,
0.03867969289422035,
0.01889774575829506,
0.00044789776438847184,
0.0004704367311205715,
-0.0009551657713018358,
0.05216065049171448,
0.03219468146562576,
0.023455830290913582,
0.03165849298238754,
0.019802236929535866,
-0.04934750124812126,
0.02268495410680771,
-0.1025756374001503,
-0.01923990622162819,
0.03428725153207779,
-0.04852689802646637,
-0.04880690574645996,
-0.016618292778730392,
-0.007123892195522785,
0.012052679434418678,
-0.005554354749619961,
0.06561083346605301,
0.03722678869962692,
0.08451894670724869,
-0.0056699286215007305,
0.004396116826683283,
-0.02624923177063465,
0.01722831465303898,
-0.03839179128408432,
0.018414774909615517,
-0.06491783261299133,
-0.026339460164308548,
-0.006891814060509205,
-0.028581183403730392,
0.05142510309815407,
0.012153349816799164,
0.005733525846153498,
-0.00739279855042696,
-0.04206153005361557,
-0.09433867037296295,
-0.12866716086864471,
-0.029779763892292976,
-0.051816754043102264,
-0.08429349213838577,
-0.05985844507813454,
-0.03237815946340561,
0.048287972807884216,
0.0028608450666069984,
0.08567189425230026,
-0.06277764588594437,
0.00015396175149362534,
-0.00972224771976471,
-0.02373480424284935,
0.011184655129909515,
-0.032217759639024734,
0.09236928075551987,
-0.022924695163965225,
-0.006464655511081219,
-0.00679879542440176,
-0.028146132826805115,
0.009604555554687977,
-0.024229437112808228,
0.020415019243955612,
0.012243400327861309,
-0.03132231533527374,
0.01957305334508419,
0.04001946374773979,
0.20459125936031342,
-0.030185896903276443,
0.010493412613868713,
0.0508514903485775,
-0.05028657615184784,
-0.040574561804533005,
-0.05876575782895088,
0.02682988904416561,
-0.03636690601706505,
-0.07739033550024033,
-0.09850728511810303,
0.08544987440109253,
-0.028234655037522316,
0.08023940026760101,
0.010687422007322311,
0.003425284055992961,
0.03028653934597969,
-0.049081988632678986,
-0.006742300931364298,
0.05456715077161789,
0.045809581875801086,
-0.11047361046075821,
0.06251999735832214,
-0.01829717680811882,
0.10223859548568726,
0.010623961687088013,
-0.003944037016481161,
0.042551133781671524,
-0.023637577891349792,
-0.04104902595281601,
-0.0707259476184845,
0.022022739052772522,
-0.03953932598233223,
-0.0351804718375206,
0.07997699081897736,
0.04574155807495117,
-0.05089818686246872,
-0.06942465901374817,
-0.01459570787847042,
3.212962695287916e-35,
0.04624619707465172,
-0.014806842431426048,
0.08136439323425293,
-0.018379373475909233,
0.019064003601670265,
0.07490973174571991,
0.0653684139251709,
-0.0008048395393416286,
-0.021566295996308327,
0.011029528453946114,
0.031549304723739624,
-0.017725706100463867,
-0.04650624468922615,
0.023108117282390594,
0.07139468938112259,
0.013706449419260025,
-0.02868269756436348,
0.026457319036126137,
0.06537806987762451,
-0.03934865817427635,
0.04295170307159424,
-0.05863020196557045,
0.032585229724645615,
-0.080910824239254,
0.03839503973722458,
0.012152321636676788,
-0.028770046308636665,
-0.012918375432491302,
0.018150856718420982,
-0.0309365913271904,
0.01032493356615305,
0.04734610766172409,
-0.05282661318778992,
0.04513087868690491,
0.026987262070178986,
0.04785099998116493,
-0.09370328485965729,
0.020609848201274872,
-0.010333793237805367,
-0.06850598007440567,
-0.016815125942230225,
-0.01790219359099865,
0.07084479182958603,
0.01612413115799427,
-0.062109269201755524,
-0.12484285980463028,
0.03598273918032646,
-0.0271772388368845,
-0.0944586917757988,
-0.006549389101564884,
0.009674621745944023,
0.03234945237636566,
-0.03779595345258713,
-0.06634879112243652,
-0.02326924167573452,
0.02903743088245392,
-0.01097023207694292,
0.09828207641839981,
-0.04654090851545334,
0.020244194194674492,
0.027209073305130005,
-0.015351335518062115,
0.035732775926589966,
0.03599925339221954,
0.03349702060222626,
0.1481141597032547,
-0.054127585142850876,
0.008441858924925327,
0.06360723823308945,
-0.01495103444904089,
0.0743887796998024,
0.02826458215713501,
0.06524530053138733,
-0.06620720773935318,
0.10360916703939438,
-0.005313579458743334,
0.08362302929162979,
-0.06551521271467209,
0.09296351671218872,
-0.016990074887871742,
-0.0032849435228854418,
-0.038404665887355804,
-0.03936600685119629,
-0.08174073696136475,
0.049074336886405945,
-0.08564341068267822,
0.005316779017448425,
-0.04356617480516434,
0.0010447745444253087,
-0.05667288228869438,
-0.1694280356168747,
-0.010637225583195686,
-0.011347287334501743,
0.06757239252328873,
0.03230638429522514,
-8.198433921932196e-34,
0.04989646375179291,
-0.11363951861858368,
0.01118405256420374,
-0.022186946123838425,
0.004723589401692152,
-0.05165613442659378,
-0.04641696438193321,
-0.008922318927943707,
-0.0037066435907036066,
-0.018981270492076874,
-0.010809404775500298,
0.058486998081207275,
0.01839398965239525,
-0.010624941438436508,
0.02058349922299385,
-0.038345012813806534,
0.03992830961942673,
-0.051506564021110535,
0.0558704249560833,
0.03630269318819046,
-0.010359414853155613,
0.0036610320676118135,
-0.10686870664358139,
0.10013166069984436,
0.015321074053645134,
0.01982228271663189,
0.011534102261066437,
0.0979342833161354,
-0.08381485939025879,
0.006899059284478426,
-0.07953830808401108,
-0.06327826529741287,
-0.025136655196547508,
-0.01183380838483572,
-0.047753769904375076,
0.06868952512741089,
0.03444498032331467,
0.02378513664007187,
-0.04587196186184883,
0.12027569860219955,
0.0014831547159701586,
-0.05472772568464279,
-0.03440484777092934,
-0.030090313404798508,
0.00608441699296236,
0.004476964473724365,
-0.006913686636835337,
0.07709666341543198,
0.009274092502892017,
-0.029738929122686386,
-0.033454060554504395,
0.06486143916845322,
-0.05164724960923195,
0.0023653674870729446,
-0.04350325092673302,
-0.015259717591106892,
-0.0751817524433136,
-0.036730460822582245,
0.1399804800748825,
0.017026424407958984,
-0.04129193723201752,
0.028320234268903732,
-0.06347011774778366,
0.06678512692451477,
-0.02750871516764164,
0.03428187966346741,
-0.030017772689461708,
0.023357659578323364,
0.011200226843357086,
0.02765602432191372,
0.01676996238529682,
0.04843895509839058,
0.010389741510152817,
-0.043246831744909286,
0.027453724294900894,
-0.062061283737421036,
0.03060038574039936,
0.005803900770843029,
0.013778951950371265,
-0.04389369115233421,
0.07620766013860703,
0.00452578067779541,
0.06847523152828217,
0.05161081999540329,
0.060182176530361176,
0.02365082874894142,
0.01015747245401144,
0.05108251795172691,
-0.005536538083106279,
-0.042126528918743134,
-0.03491368517279625,
-0.005003632977604866,
0.07289128750562668,
0.03702741861343384,
-0.04383320361375809,
-3.9554347353032426e-8,
-0.03355943411588669,
0.020150896161794662,
0.1196901947259903,
0.006144118960946798,
0.13003113865852356,
0.046583354473114014,
0.004664587322622538,
-0.02325085550546646,
-0.018236279487609863,
0.031208232045173645,
0.0032508119475096464,
0.06720639020204544,
0.0020927588921040297,
-0.036755677312612534,
-0.07175228744745255,
0.08189509063959122,
0.029104094952344894,
-0.07774947583675385,
0.02746119163930416,
-0.027562811970710754,
0.027049921452999115,
-0.0209579449146986,
0.0037503500934690237,
-0.12865041196346283,
0.0364656075835228,
-0.10610631853342056,
-0.029533566907048225,
0.07440046966075897,
0.007663888391107321,
0.004994161892682314,
0.012416208162903786,
-0.014580096118152142,
0.06085490062832832,
0.05482226982712746,
0.07040421664714813,
0.024978704750537872,
0.06650905311107635,
-0.059503212571144104,
-0.024928484112024307,
-0.05195998027920723,
-0.0092404680326581,
0.10902650654315948,
-0.03661098703742027,
-0.03693731129169464,
-0.049801766872406006,
0.021276770159602165,
-0.012543040327727795,
-0.13371716439723969,
0.052779171615839005,
-0.04698134586215019,
0.05849192291498184,
0.016615523025393486,
-0.03263670951128006,
0.03348943963646889,
0.07945745438337326,
-0.0006904504261910915,
0.05150840803980827,
-0.057062774896621704,
-0.043350644409656525,
0.0980486273765564,
0.03105640783905983,
-0.03213347867131233,
-0.013513145036995411,
0.011930680833756924
] | 0
| 0
| 0
| 0
|
|
What Do You See in Common? Learning Hierarchical Prototypes over Tree-of-Life to Discover Evolutionary Traits
|
https://openreview.net/forum?id=4sDicVEy6M
|
[
"Harish Babu Manogaran",
"M. Maruf",
"Arka Daw",
"Kazi Sajeed Mehrab",
"Caleb Patrick Charpentier",
"Josef Uyeda",
"Wasila M Dahdul",
"Matthew J Thompson",
"Elizabeth G Campolongo",
"Kaiya L Provost",
"Wei-Lun Chao",
"Tanya Berger-Wolf",
"Paula Mabee",
"Hilmar Lapp",
"Anuj Karpatne"
] |
Poster
|
A grand challenge in biology is to discover evolutionary traits---features of organisms common to a group of species with a shared ancestor in the tree of life (also referred to as phylogenetic tree). With the growing availability of image repositories in biology, there is a tremendous opportunity to discover evolutionary traits directly from images in the form of a hierarchy of prototypes. However, current prototype-based methods are mostly designed to operate over a flat structure of classes and face several challenges in discovering hierarchical prototypes, including the issue of learning over-specific prototypes at internal nodes. To overcome these challenges, we introduce the framework of Hierarchy aligned Commonality through Prototypical Networks (HComP-Net). The key novelties in HComP-Net include a novel over-specificity loss to avoid learning over-specific prototypes, a novel discriminative loss to ensure prototypes at an internal node are absent in the contrasting set of species with different ancestry, and a novel masking module to allow for the exclusion of over-specific prototypes at higher levels of the tree without hampering classification performance. We empirically show that HComP-Net learns prototypes that are accurate, semantically consistent, and generalizable to unseen species in comparison to baselines. Our code is publicly accessible at Imageomics Institute Github site: https://github.com/Imageomics/HComPNet.
|
deep learning, interpretability, prototype-based neural network, phylogeny, computer vision
| null | 11,437
|
2409.02335
|
[
-0.037044718861579895,
-0.01844513975083828,
0.08882398903369904,
-0.0186558086425066,
0.10830161720514297,
-0.052432626485824585,
-0.07203429937362671,
-0.04913109540939331,
-0.01620868220925331,
-0.024194838479161263,
-0.016727618873119354,
-0.1581694483757019,
0.023615475744009018,
0.03981655463576317,
-0.05170274153351784,
0.015259649604558945,
-0.018899641931056976,
0.020888296887278557,
-0.04676241800189018,
-0.03200119361281395,
-0.03192996233701706,
-0.0027526048943400383,
0.08294760435819626,
0.012761790305376053,
-0.025298768654465675,
0.023516908288002014,
-0.014683404006063938,
0.00569577980786562,
0.028729412704706192,
-0.05859338119626045,
0.00886077806353569,
0.04315919056534767,
-0.013211285695433617,
0.03813885524868965,
-0.030405433848500252,
0.09785693883895874,
-0.032568421214818954,
0.022743206471204758,
0.03954698517918587,
0.003573392052203417,
0.034477848559617996,
0.07966916263103485,
-0.02884337119758129,
0.04301030933856964,
0.054207734763622284,
0.05894729122519493,
-0.03332396224141121,
-0.06937336176633835,
0.025835789740085602,
-0.0070452443324029446,
-0.04770427197217941,
-0.06866616755723953,
-0.050966471433639526,
0.09122994542121887,
0.05646439641714096,
0.06755835562944412,
0.042556215077638626,
0.0005874311318621039,
0.04887455701828003,
-0.010988562367856503,
0.018521929159760475,
-0.06823444366455078,
-0.03531409054994583,
-0.04747255519032478,
0.03205323591828346,
-0.00611213268712163,
0.04373805597424507,
0.03681604191660881,
0.06502946466207504,
-0.02726585604250431,
0.03439359366893768,
0.05361797660589218,
-0.08805779367685318,
0.12818042933940887,
0.05513438209891319,
0.09532486647367477,
0.05559319257736206,
-0.008628234267234802,
-0.01598723977804184,
-0.08041524887084961,
-0.02635088935494423,
0.09757238626480103,
0.0020555376540869474,
-0.006720430683344603,
0.12006693333387375,
0.02621932327747345,
0.009243357926607132,
0.03294984623789787,
-0.08773991465568542,
0.0325680635869503,
-0.08082173764705658,
-0.11683318018913269,
0.04470852017402649,
-0.07477497309446335,
-0.04857351630926132,
0.02030123956501484,
0.06717287749052048,
-0.09491527825593948,
-0.0025965573731809855,
0.025215625762939453,
-0.05231679230928421,
-0.02615349180996418,
0.04315226152539253,
-0.04456789046525955,
0.006904629059135914,
0.018008476123213768,
-0.049469683319330215,
-0.028929103165864944,
0.11145631968975067,
-0.06166553869843483,
-0.06810980290174484,
0.023304510861635208,
-0.03315173462033272,
0.025367209687829018,
-0.002470223931595683,
-0.003101926064118743,
0.02105550281703472,
-0.02741265296936035,
0.082709901034832,
0.02534620463848114,
0.016438279300928116,
-0.006811558734625578,
0.006523457355797291,
-0.03926948457956314,
0.03615199029445648,
-0.03374715521931648,
-0.1785595864057541,
2.2490088729099798e-33,
0.09785893559455872,
0.006799725349992514,
0.021201789379119873,
0.06810574233531952,
0.05690058320760727,
-0.07674961537122726,
-0.012432524934411049,
0.07393394410610199,
-0.09016358107328415,
0.011052009649574757,
-0.08797325938940048,
0.0121693080291152,
-0.10144153982400894,
0.036097392439842224,
0.031600724905729294,
0.0340639129281044,
-0.08814028650522232,
0.03714314475655556,
-0.06216372177004814,
-0.01308005303144455,
-0.06947796791791916,
-0.04295150190591812,
-0.03339762985706329,
-0.01860053651034832,
0.026931878179311752,
-0.03387061133980751,
0.06866104155778885,
-0.042504940181970596,
-0.0004876042657997459,
-0.0055070496164262295,
-0.06567895412445068,
0.004081300459802151,
0.07614864408969879,
0.093316450715065,
-0.02811131253838539,
-0.08333977311849594,
0.03687528148293495,
-0.08690471202135086,
0.058479152619838715,
-0.0057487571612000465,
-0.004297684412449598,
0.0032040283549576998,
0.007653599604964256,
0.018195902928709984,
0.023991283029317856,
-0.044192008674144745,
0.023740876466035843,
-0.01776125282049179,
-0.05882667005062103,
0.0011179115390405059,
-0.008922216482460499,
0.005123468115925789,
0.08210838586091995,
-0.04247463867068291,
-0.08375684171915054,
0.009977388195693493,
-0.008997729048132896,
0.010286297649145126,
0.033302608877420425,
0.01593061536550522,
0.031401071697473526,
0.07577559351921082,
-0.04984357953071594,
0.07782069593667984,
-0.010824444703757763,
0.0354941226541996,
-0.04438666254281998,
0.04483064264059067,
0.048069313168525696,
0.05037618428468704,
-0.012025553733110428,
0.062451839447021484,
-0.09018483757972717,
-0.10171107202768326,
0.02047491818666458,
0.008088191039860249,
0.05086641013622284,
-0.04742613062262535,
-0.00697599072009325,
0.012676817364990711,
-0.05366702377796173,
0.01596846804022789,
0.031572144478559494,
-0.005271551664918661,
-0.09928834438323975,
-0.058356765657663345,
0.08601095527410507,
-0.014910001307725906,
-0.03442959859967232,
0.031047943979501724,
-0.027614545077085495,
-0.021590489894151688,
0.001732722856104374,
-0.0532844103872776,
0.03632592037320137,
-2.042163560480513e-33,
-0.012732957489788532,
0.021194709464907646,
-0.035739630460739136,
-0.04340009018778801,
-0.03351608291268349,
-0.0037161505315452814,
-0.03594694286584854,
-0.0486241951584816,
-0.11773981153964996,
-0.02350378781557083,
0.04308668524026871,
0.0037569112610071898,
-0.04269258677959442,
-0.0961078554391861,
0.03287678584456444,
-0.05122324079275131,
-0.03516719117760658,
0.042599279433488846,
0.06640171259641647,
0.0280297864228487,
0.04185329005122185,
0.06870397925376892,
-0.11851254850625992,
0.05787799507379532,
-0.037461888045072556,
0.020774031057953835,
-0.005338404327630997,
0.10854369401931763,
0.010333448648452759,
0.004030783660709858,
-0.06375544518232346,
-0.008095063269138336,
-0.007468391675502062,
-0.036461543291807175,
0.012131699360907078,
0.03603638708591461,
-0.018653029575943947,
-0.06049022078514099,
0.03641344606876373,
0.051000289618968964,
-0.024439314380288124,
-0.02897636592388153,
-0.14496439695358276,
0.004348530899733305,
0.027382470667362213,
-0.06993556767702103,
-0.02125999517738819,
0.03996826708316803,
0.08915966749191284,
0.08558183908462524,
0.02619882859289646,
0.006614350713789463,
0.012932628393173218,
-0.047167424112558365,
-0.05788532271981239,
-0.011213365010917187,
0.0655960813164711,
0.0436142235994339,
0.09035398066043854,
0.021320588886737823,
-0.013571161776781082,
-0.08930616825819016,
0.002447371603921056,
0.03646517172455788,
-0.05974268540740013,
-0.03623288869857788,
-0.06093098968267441,
0.015445170924067497,
0.011903325095772743,
0.005090508610010147,
-0.02521776221692562,
0.005299868993461132,
0.029028328135609627,
-0.05947897955775261,
0.009852208197116852,
-0.01837289333343506,
-0.04313052073121071,
-0.009595881216228008,
0.034763578325510025,
-0.01836518570780754,
-0.021263808012008667,
-0.002892060438171029,
0.03326907008886337,
0.06091582030057907,
0.1082228571176529,
0.06517758965492249,
-0.08623576164245605,
0.0905773788690567,
0.048191189765930176,
0.036141231656074524,
0.006758417934179306,
-0.004939916078001261,
-0.0493345633149147,
0.04423454403877258,
-0.006674521137028933,
-4.349470117404053e-8,
-0.06177063658833504,
0.04828018695116043,
0.055576711893081665,
-0.0013653378700837493,
0.04661925882101059,
-0.017999926581978798,
-0.00566778564825654,
0.08061471581459045,
-0.012691859155893326,
0.015291474759578705,
-0.0030785780400037766,
0.0417756587266922,
0.014364930801093578,
-0.03505812957882881,
0.09797642379999161,
0.018493859097361565,
0.0022966244723647833,
0.03287898749113083,
-0.03754906356334686,
0.091904416680336,
0.022084340453147888,
-0.0006840750575065613,
-0.01608361303806305,
0.04609128087759018,
0.00868951715528965,
-0.12147145718336105,
-0.029256559908390045,
0.10135050863027573,
0.039727676659822464,
0.01637909933924675,
-0.03174107149243355,
0.008948379196226597,
0.006882831919938326,
-0.028753090649843216,
0.06642286479473114,
0.13092836737632751,
0.005079754628241062,
0.015857767313718796,
-0.0669604241847992,
-0.012365714646875858,
0.01701267622411251,
-0.05851927772164345,
-0.08820707350969315,
0.012957938946783543,
0.015079610981047153,
-0.03475797176361084,
0.028639264404773712,
-0.018069477751851082,
-0.004871412180364132,
0.005606514401733875,
0.031537529081106186,
-0.019421296194195747,
-0.02364589273929596,
0.020690366625785828,
0.0063532874919474125,
0.0007054555462673306,
0.012404215522110462,
-0.05839481204748154,
0.07847514748573303,
0.0668429285287857,
0.07327406853437424,
0.03523082286119461,
0.055259063839912415,
0.024027975276112556
] |
https://github.com/imageomics/hcompnet
| 0
| 0
| 0
| 0
|
Montessori-Instruct: Generate Influential Training Data Tailored for Student Learning
|
https://openreview.net/forum?id=9RCT0ngvZP
|
[
"Xiaochuan Li",
"Zichun Yu",
"Chenyan Xiong"
] |
Poster
|
Synthetic data has been widely used to train large language models, but their generative nature inevitably introduces noisy, non-informative, and misleading learning signals. In this paper, we propose Montessori-Instruct, a novel data synthesis framework that tailors the data synthesis ability of the teacher language model toward the student language model's learning process. Specifically, we utilize local data influence of synthetic training data points on students to characterize students' learning preferences. Then, we train the teacher model with Direct Preference Optimization (DPO) to generate synthetic data tailored toward student learning preferences. Experiments with Llama3-8B-Instruct (teacher) and Llama3-8B (student) on Alpaca Eval and MT-Bench demonstrate that Montessori-Instruct significantly outperforms standard synthesis methods by 18.35\% and 46.24\% relatively. Our method also beats data synthesized by a stronger teacher model, GPT-4o. Further analysis confirms the benefits of teacher's learning to generate more influential training data in the student's improved learning, the advantages of local data influence in accurately measuring student preferences, and the robustness of Montessori-Instruct across different student models. Our code and data are open-sourced at https://github.com/cxcscmu/Montessori-Instruct.
|
synthetic data, data influence, instruction tuning
| null | 11,434
| null |
[
-0.04745623841881752,
-0.09205710142850876,
0.035940226167440414,
0.024209655821323395,
-0.037223879247903824,
-0.03061012551188469,
-0.022950271144509315,
-0.02307109721004963,
-0.008556937798857689,
0.03493012115359306,
-0.025599736720323563,
0.0012685850961133838,
0.08015625923871994,
-0.015082246623933315,
0.03154998645186424,
0.012005376629531384,
0.11602142453193665,
0.07570240646600723,
-0.049396295100450516,
-0.023787062615156174,
0.05771505832672119,
0.000056510754802729934,
0.023889318108558655,
0.03231429308652878,
0.09002324938774109,
0.08293061703443527,
0.031109310686588287,
-0.034794677048921585,
-0.025027604773640633,
-0.011747060343623161,
-0.005004297010600567,
-0.030959656462073326,
0.042751625180244446,
-0.06372879445552826,
-0.06282857805490494,
0.006580024491995573,
-0.05775279924273491,
0.04360770061612129,
-0.03275609016418457,
-0.00478407833725214,
-0.06173333525657654,
0.026523754000663757,
0.09538502991199493,
-0.005073645152151585,
0.04021207243204117,
-0.05878683924674988,
-0.004120434634387493,
-0.06150350347161293,
0.012794295325875282,
0.06575165688991547,
-0.11162382364273071,
-0.09189249575138092,
-0.04067721217870712,
-0.03877728804945946,
0.0005223631160333753,
0.0376528836786747,
-0.007172660436481237,
0.02173052541911602,
0.02073446847498417,
-0.017651289701461792,
-0.10475184768438339,
-0.04905841127038002,
-0.038941219449043274,
-0.018855426460504532,
-0.014069057069718838,
-0.05571791157126427,
-0.004505542106926441,
0.055737078189849854,
-0.022981923073530197,
0.04870862513780594,
0.013785677962005138,
0.09697156399488449,
0.015270136296749115,
0.02929534949362278,
0.009197985753417015,
0.04065193235874176,
0.037971381098032,
-0.03382812440395355,
0.04638402536511421,
-0.047365445643663406,
-0.031126264482736588,
-0.0054375105537474155,
-0.02296431176364422,
-0.0056946417316794395,
0.10351036489009857,
0.0007450123666785657,
0.08993393182754517,
0.02762758731842041,
0.03225332126021385,
0.014805308543145657,
-0.04684637114405632,
0.07271885871887207,
-0.04727813974022865,
0.04326745867729187,
-0.005175378639250994,
0.06632320582866669,
0.021351395174860954,
-0.046358417719602585,
0.05050298199057579,
0.053122926503419876,
0.005539692007005215,
0.11998232454061508,
-0.019332459196448326,
-0.05182775855064392,
-0.0774279460310936,
-0.1193837821483612,
0.02501033805310726,
0.04438071325421333,
-0.0014291199622675776,
-0.043743059039115906,
0.058253735303878784,
0.01626361906528473,
-0.05202251300215721,
-0.02786824107170105,
-0.01108703576028347,
0.014263412915170193,
0.07638578861951828,
-0.041789278388023376,
-0.06310223042964935,
0.07484506815671921,
-0.02944052778184414,
-0.02245667390525341,
0.019564606249332428,
0.03075753152370453,
-0.07373302429914474,
-0.06363004446029663,
-0.0778471976518631,
1.4526473579068778e-33,
0.07552967220544815,
0.04307521879673004,
0.009263069368898869,
0.06352131813764572,
0.06450139731168747,
-0.04364301264286041,
0.023278018459677696,
-0.010421245358884335,
-0.08663050085306168,
0.036369819194078445,
-0.014671508222818375,
-0.01440237183123827,
-0.046384941786527634,
0.1405373364686966,
0.10238205641508102,
0.039597317576408386,
-0.029368024319410324,
0.01896863617002964,
-0.051900822669267654,
-0.025074822828173637,
0.06887178122997284,
0.06245320290327072,
0.04214499890804291,
-0.09083647280931473,
0.0032084020785987377,
0.05129203200340271,
0.050997890532016754,
0.0030453677754849195,
-0.014209131710231304,
0.016061605885624886,
-0.046736881136894226,
-0.049228962510824203,
0.0007429810939356685,
0.011140166781842709,
0.02290986478328705,
-0.04176577925682068,
-0.005737295374274254,
-0.0621284656226635,
0.00830778293311596,
0.05046169459819794,
0.09165547788143158,
-0.04015122726559639,
0.05829015374183655,
0.044947024434804916,
-0.02104608528316021,
0.01888248138129711,
0.07681688666343689,
0.04633623734116554,
0.06451219320297241,
-0.03788702189922333,
0.004172588232904673,
0.01625809073448181,
-0.05668751895427704,
0.003162862965837121,
-0.07108543813228607,
0.034622617065906525,
0.009984456934034824,
0.0446350984275341,
0.045632749795913696,
0.011214758269488811,
-0.05655224621295929,
0.11036639660596848,
0.04821772128343582,
0.044721852988004684,
-0.030546922236680984,
-0.0006706538260914385,
-0.06078560650348663,
-0.04265153035521507,
0.131952166557312,
-0.050929225981235504,
-0.05700541287660599,
-0.015096358023583889,
-0.08319105207920074,
-0.027308505028486252,
0.010679605416953564,
-0.04114741459488869,
0.004032467491924763,
-0.04207715392112732,
0.0922774150967598,
-0.06771983951330185,
0.0794413685798645,
-0.018302807584404945,
0.026023177430033684,
-0.06234487146139145,
-0.08421453833580017,
-0.03696421533823013,
0.01742810569703579,
-0.06265322864055634,
-0.018680082634091377,
-0.03331870958209038,
-0.030718078836798668,
-0.03941016644239426,
-0.021782871335744858,
-0.009133691899478436,
0.03210514783859253,
-2.3746253216457188e-33,
-0.018691841512918472,
-0.0007397194858640432,
-0.01011492870748043,
0.07195937633514404,
-0.07609494775533676,
-0.03208623081445694,
-0.00682584336027503,
-0.0225824061781168,
0.028802767395973206,
-0.10071445256471634,
-0.0724458172917366,
-0.040773384273052216,
0.07683439552783966,
-0.058163054287433624,
0.011160416528582573,
-0.0268687941133976,
0.012754268944263458,
-0.03832941874861717,
-0.02247856743633747,
-0.03992864489555359,
-0.00503661809489131,
0.08684886991977692,
-0.06463288515806198,
0.030893513932824135,
-0.045146748423576355,
-0.022699670866131783,
-0.08022650331258774,
0.04322311654686928,
0.027232637628912926,
0.06186137720942497,
-0.018240593373775482,
-0.022468961775302887,
0.00452781468629837,
0.011947249993681908,
-0.11249668151140213,
0.05595378205180168,
-0.022257089614868164,
-0.052795734256505966,
-0.016608165577054024,
0.1478368490934372,
-0.0037349581252783537,
-0.01435821782797575,
-0.04568945989012718,
0.00652902340516448,
0.021425748243927956,
0.04037940502166748,
0.011019776575267315,
-0.007375841028988361,
0.06310160458087921,
-0.013787830248475075,
0.058196745812892914,
0.006281972397118807,
-0.027368493378162384,
0.0014768807450309396,
0.04667549207806587,
-0.10819888859987259,
-0.0008830191800370812,
-0.10769390314817429,
0.0930279791355133,
-0.020583685487508774,
-0.03621429577469826,
-0.023044157773256302,
0.1297115832567215,
-0.0470147468149662,
-0.046125929802656174,
-0.021479900926351547,
-0.0734858587384224,
0.011058059521019459,
0.04536541923880577,
-0.03537432476878166,
-0.005574952811002731,
-0.036410603672266006,
-0.03534124791622162,
-0.03625031188130379,
0.012090474367141724,
0.011393968015909195,
-0.005321711767464876,
0.02709989808499813,
-0.07703614234924316,
0.03929624333977699,
0.08390317857265472,
0.01667465642094612,
0.0467788428068161,
0.008498222567141056,
0.07096035778522491,
0.11325935274362564,
0.04714633896946907,
0.09795107692480087,
-0.008642582222819328,
0.033794790506362915,
-0.05083516612648964,
0.07614131271839142,
-0.0015590355033054948,
0.07199511677026749,
-0.05835865065455437,
-4.446868828722472e-8,
-0.031705182045698166,
-0.0098688043653965,
0.01975705288350582,
0.14151237905025482,
-0.03255661204457283,
-0.026588525623083115,
-0.03467031940817833,
0.04944006726145744,
0.004527936689555645,
-0.021318932995200157,
0.04850904271006584,
0.025437351316213608,
-0.022424861788749695,
-0.08499521017074585,
0.048670265823602676,
0.05394469574093819,
0.049411240965127945,
0.0004876918683294207,
-0.0019017040031030774,
0.004060894250869751,
-0.02125517465174198,
0.026759851723909378,
0.018360307440161705,
-0.07966127246618271,
0.048251908272504807,
-0.05534405633807182,
-0.0642586201429367,
0.08795078098773956,
-0.005770362447947264,
0.08773953467607498,
0.014125893823802471,
-0.06116557866334915,
-0.01973755843937397,
-0.009276947937905788,
0.02667650580406189,
0.05205036327242851,
0.0383404865860939,
-0.013544646091759205,
0.01614188589155674,
-0.009944845922291279,
0.059178248047828674,
0.03327881172299385,
-0.05500265583395958,
0.03340619429945946,
-0.031160062178969383,
-0.025761594995856285,
-0.07210437953472137,
-0.15535886585712433,
0.008756126277148724,
0.0083609065040946,
0.007907435297966003,
-0.04435117915272713,
-0.05031795799732208,
-0.015171133913099766,
0.09360271692276001,
0.02643762156367302,
-0.09993697702884674,
-0.037513162940740585,
-0.0434904471039772,
0.036176975816488266,
0.014607016928493977,
0.06128018721938133,
-0.053803227841854095,
0.04936767369508743
] | 0
| 0
| 0
| 0
|
|
Affine Steerable Equivariant Layer for Canonicalization of Neural Networks
|
https://openreview.net/forum?id=5i6ZZUjCA9
|
[
"Yikang Li",
"Yeqing Qiu",
"Yuxuan Chen",
"Zhouchen Lin"
] |
Poster
|
In the field of equivariant networks, achieving affine equivariance, particularly for general group representations, has long been a challenge.
In this paper, we propose the steerable EquivarLayer, a generalization of InvarLayer (Li et al., 2024), by building on the concept of equivariants beyond invariants.
The steerable EquivarLayer supports affine equivariance with arbitrary input and output representations, marking the first model to incorporate steerability into networks for the affine group.
To integrate it with canonicalization, a promising approach for making pre-trained models equivariant, we introduce a novel Det-Pooling module, expanding the applicability of EquivarLayer and the range of groups suitable for canonicalization.
We conduct experiments on image classification tasks involving group transformations to validate the steerable EquivarLayer in the role of a canonicalization function, demonstrating its effectiveness over data augmentation.
|
equivariant networks, steerability, the affine group, equivariants, canonicalization
| null | 11,430
| null |
[
-0.051672644913196564,
-0.14013056457042694,
0.03607397899031639,
0.04214215278625488,
0.08800705522298813,
0.04037218540906906,
0.04844055697321892,
-0.06562680006027222,
0.010987338609993458,
-0.08423249423503876,
0.023761585354804993,
-0.014333516359329224,
0.009124566800892353,
0.07175411283969879,
-0.06707467883825302,
-0.049895789474248886,
0.0025533619336783886,
0.13958844542503357,
-0.09809739887714386,
-0.05332997813820839,
0.036970268934965134,
0.013057783246040344,
0.03547367453575134,
0.012891999445855618,
0.031534940004348755,
-0.03362696245312691,
0.0033122964669018984,
-0.010778985917568207,
0.011768168769776821,
-0.08004073053598404,
-0.024657322093844414,
0.01902429573237896,
-0.06473656743764877,
0.01732587069272995,
-0.07007896900177002,
0.07264362275600433,
0.0034658899530768394,
0.05875435099005699,
-0.016290033236145973,
-0.03579365834593773,
0.018291238695383072,
-0.0020149769261479378,
0.03321860730648041,
0.07478457689285278,
0.0675371065735817,
0.07292821258306503,
0.08011475205421448,
-0.07629433274269104,
-0.027569040656089783,
0.1045859307050705,
0.051334578543901443,
0.007157644256949425,
-0.052633944898843765,
0.06998901814222336,
0.001135164638981223,
-0.08021203428506851,
-0.018011849373579025,
-0.05027574673295021,
-0.004386425483971834,
0.007562184706330299,
-0.010458331555128098,
-0.017589973285794258,
0.00010457095777383074,
0.04998590424656868,
-0.04478686302900314,
-0.010938568040728569,
0.05058462172746658,
0.035452041774988174,
-0.0008240307797677815,
0.005534379743039608,
0.05021819472312927,
0.043495725840330124,
-0.02919502928853035,
0.04723221808671951,
0.06268835812807083,
0.060240235179662704,
0.026839490979909897,
-0.019295014441013336,
0.034393664449453354,
-0.017310114577412605,
0.013877675868570805,
0.08484932035207748,
0.07079274952411652,
-0.0018774543423205614,
0.10082712769508362,
0.04092716798186302,
-0.07759544253349304,
0.05575590208172798,
-0.017823144793510437,
0.028472093865275383,
-0.07102413475513458,
-0.0018403043504804373,
0.09801871329545975,
-0.04319912567734718,
0.04925641417503357,
-0.05385490879416466,
0.019441021606326103,
-0.09086579084396362,
-0.022067666053771973,
0.09889154136180878,
-0.04914966598153114,
-0.01572301983833313,
-0.025636224076151848,
0.046004973351955414,
0.09493698924779892,
-0.06418351083993912,
0.0901113748550415,
-0.03896133601665497,
-0.010388916358351707,
-0.15545009076595306,
-0.01561556663364172,
-0.027728920802474022,
-0.019027763977646828,
-0.05318107828497887,
-0.018090490251779556,
-0.012910060584545135,
-0.038423094898462296,
0.0210744459182024,
0.06649723649024963,
-0.014348458498716354,
-0.013844695873558521,
0.05306633561849594,
0.02579898200929165,
0.01753314770758152,
0.0843263790011406,
0.0044058579951524734,
-0.10568933188915253,
3.4592388981280736e-33,
0.01198595017194748,
0.06312060356140137,
0.059544675052165985,
0.027539173141121864,
-0.026804042980074883,
-0.029319103807210922,
-0.013990921899676323,
0.019948117434978485,
-0.07185531407594681,
0.024471919983625412,
-0.08731503039598465,
0.026198795065283775,
-0.048714760690927505,
0.11937251687049866,
0.026664694771170616,
-0.017640432342886925,
0.006793823558837175,
-0.015821216627955437,
-0.015777837485074997,
-0.09139170497655869,
-0.046879563480615616,
-0.015001963824033737,
0.00745003018528223,
-0.0029189663473516703,
-0.05623127147555351,
-0.03857932612299919,
-0.005490086041390896,
-0.04478870704770088,
-0.008450478315353394,
0.014514678157866001,
-0.06662151962518692,
0.042906731367111206,
-0.028608432039618492,
0.03976089879870415,
-0.033124253153800964,
-0.09042325615882874,
0.051729727536439896,
-0.026332441717386246,
0.08757620304822922,
0.014690011739730835,
0.03805486112833023,
0.02985985018312931,
-0.0346376858651638,
-0.03575507178902626,
-0.03953772410750389,
0.00015108751540537924,
0.006033367943018675,
0.04902325198054314,
-0.04715936630964279,
0.022949067875742912,
0.027540242299437523,
0.04430808871984482,
-0.026075240224599838,
-0.05471540614962578,
-0.056462373584508896,
0.045403365045785904,
0.038735512644052505,
0.09428657591342926,
-0.0008019023225642741,
-0.038955677300691605,
0.003108249744400382,
0.08452260494232178,
-0.030415330082178116,
-0.019797340035438538,
-0.056095223873853683,
0.04638361930847168,
0.044155631214380264,
0.03627898171544075,
0.06896167248487473,
0.009522152133286,
-0.07560715079307556,
0.059646401554346085,
-0.09408055990934372,
0.0059880749322474,
0.02272663451731205,
0.041751615703105927,
-0.044087447226047516,
-0.069668710231781,
-0.00014998568804003298,
0.04546809196472168,
-0.042791809886693954,
0.023155372589826584,
-0.04043242335319519,
-0.005986069794744253,
-0.06395266205072403,
-0.015358821488916874,
0.0474579744040966,
-0.052213817834854126,
0.003762564854696393,
-0.01919606328010559,
-0.06567373126745224,
0.012343024834990501,
-0.0332861989736557,
0.012615255080163479,
0.02318640798330307,
-3.788140013436242e-33,
0.027347678318619728,
0.05542554706335068,
-0.0218304842710495,
0.06848925352096558,
-0.0056353905238211155,
-0.0765506848692894,
0.01316576823592186,
0.05085158720612526,
-0.008872329257428646,
-0.020411815494298935,
0.005616248585283756,
0.01425702590495348,
0.004454588983207941,
-0.021064970642328262,
-0.015017295256257057,
-0.039528071880340576,
-0.16416515409946442,
0.049176786094903946,
0.020499806851148605,
0.012135738506913185,
0.027565279975533485,
0.021296393126249313,
-0.04603533446788788,
0.02078528329730034,
-0.06781309098005295,
0.07632308453321457,
-0.03443916514515877,
-0.01223908830434084,
0.05638517066836357,
-0.010703641921281815,
-0.07931223511695862,
-0.05846468731760979,
-0.025956984609365463,
-0.053817734122276306,
-0.027155719697475433,
0.05621187388896942,
-0.02190100960433483,
-0.05157116800546646,
-0.055630799382925034,
0.07969453185796738,
0.014809210784733295,
-0.024946173653006554,
-0.026662247255444527,
-0.01792534999549389,
0.03285027667880058,
-0.04367087781429291,
0.06102864816784859,
0.01892848126590252,
0.013976155780255795,
0.021764099597930908,
-0.01674407348036766,
0.08053402602672577,
-0.05056845024228096,
-0.00006545585347339511,
-0.03182289004325867,
-0.05580827221274376,
0.09363185614347458,
-0.09190672636032104,
0.12258376181125641,
-0.002984496997669339,
-0.007725093513727188,
-0.07559315860271454,
-0.010271935723721981,
-0.039490893483161926,
-0.022409610450267792,
-0.08722548931837082,
-0.007722125854343176,
-0.03748645260930061,
0.06911497563123703,
0.037435293197631836,
0.032643336802721024,
-0.046773556619882584,
-0.009608542546629906,
-0.05874219909310341,
0.007546024862676859,
-0.07010645419359207,
0.03421192616224289,
-0.012086227536201477,
0.06953888386487961,
-0.07713206112384796,
0.03269534930586815,
-0.07044191658496857,
0.04442806914448738,
0.10116580128669739,
0.11411620676517487,
0.10771235823631287,
0.0037557657342404127,
-0.00508706783875823,
0.09335672110319138,
0.013207118026912212,
0.006418914534151554,
0.019342469051480293,
0.024932431057095528,
0.09177076816558838,
-0.05241059884428978,
-4.3733408006119134e-8,
-0.09182661771774292,
0.05542130023241043,
-0.008330794982612133,
0.021658971905708313,
-0.021130086854100227,
-0.06740478426218033,
0.01865123212337494,
0.052359286695718765,
-0.00568598136305809,
-0.032268229871988297,
0.055121708661317825,
-0.0016446590889245272,
-0.031466737389564514,
-0.04605085402727127,
0.006651052739471197,
0.05072147771716118,
0.01392101775854826,
0.09770537167787552,
0.02081417106091976,
0.006847716867923737,
-0.03225475922226906,
-0.0189321581274271,
-0.006129251327365637,
-0.012506738305091858,
0.0243237167596817,
-0.1147553101181984,
-0.10212066769599915,
0.07604934275150299,
0.04957815259695053,
0.010546702891588211,
-0.03090239316225052,
0.023204542696475983,
-0.022991128265857697,
0.003767835209146142,
-0.014328633435070515,
0.07153146713972092,
-0.09256747364997864,
-0.010482512414455414,
-0.025830963626503944,
-0.02357846312224865,
0.01289061177521944,
0.07072839140892029,
-0.038847941905260086,
0.018022365868091583,
-0.0016232326161116362,
0.05553862452507019,
0.07337094098329544,
-0.05171511322259903,
-0.0890868604183197,
-0.03515009582042694,
0.05329722538590431,
-0.044047676026821136,
-0.03760763630270958,
0.11786169558763504,
-0.050097525119781494,
0.02321920171380043,
0.03638350963592529,
-0.06426001340150833,
0.030302880331873894,
-0.01374727487564087,
-0.07075872272253036,
0.03581142798066139,
-0.04890914633870125,
-0.0818360224366188
] | 0
| 0
| 0
| 0
|
|
Has the Deep Neural Network learned the Stochastic Process? An Evaluation Viewpoint
|
https://openreview.net/forum?id=2U8owdruSQ
|
[
"Harshit Kumar",
"Beomseok Kang",
"Biswadeep Chakraborty",
"Saibal Mukhopadhyay"
] |
Poster
|
This paper presents the first systematic study of evaluating Deep Neural Networks (DNNs) designed to forecast the evolution of stochastic complex systems. We show that traditional evaluation methods like threshold-based classification metrics and error-based scoring rules assess a DNN's ability to replicate the observed ground truth but fail to measure the DNN's learning of the underlying stochastic process. To address this gap, we propose a new evaluation criteria called _Fidelity to Stochastic Process (F2SP)_, representing the DNN's ability to predict the system property _Statistic-GT_—the ground truth of the stochastic process—and introduce an evaluation metric that exclusively assesses F2SP. We formalize F2SP within a stochastic framework and establish criteria for validly measuring it. We formally show that Expected Calibration Error (ECE) satisfies the necessary condition for testing F2SP, unlike traditional evaluation methods. Empirical experiments on synthetic datasets, including wildfire, host-pathogen, and stock market models, demonstrate that ECE uniquely captures F2SP. We further extend our study to real-world wildfire data, highlighting the limitations of conventional evaluation and discuss the practical utility of incorporating F2SP into model assessment. This work offers a new perspective on evaluating DNNs modeling complex systems by emphasizing the importance of capturing underlying the stochastic process.
|
evaluation, deep neural network, stochasticity, complex systems, forecasting
|
A novel evaluation criterion to assess whether DNNs modeling stochastic complex systems have learnt the underlying stochastic process
| 11,418
|
2402.15163
|
[
-0.08806861937046051,
-0.10446414351463318,
-0.03729889541864395,
0.10122725367546082,
0.0733807235956192,
-0.04531720653176308,
-0.021238362416625023,
0.01789362169802189,
0.035933639854192734,
-0.07351585477590561,
-0.09446682780981064,
-0.07006970793008804,
0.0006551364203915,
0.02292962558567524,
-0.1072370707988739,
-0.11846581846475601,
0.037063632160425186,
0.043750569224357605,
-0.09356893599033356,
0.005655130837112665,
0.039736393839120865,
0.07390531897544861,
0.04934290051460266,
-0.008945791982114315,
0.012063761241734028,
-0.03289371356368065,
-0.07278382778167725,
0.02246798388659954,
-0.08017967641353607,
-0.017279425635933876,
0.06502897292375565,
-0.030169714242219925,
-0.0034113037399947643,
-0.014156157150864601,
0.0905444473028183,
-0.0021191625855863094,
-0.02567979320883751,
-0.014377608895301819,
0.0326056107878685,
-0.016884461045265198,
0.02242022380232811,
-0.03019651025533676,
0.024011611938476562,
0.04943958297371864,
0.06347285956144333,
0.007643002551048994,
0.011695095337927341,
-0.07810740172863007,
0.00027140494785271585,
0.026037726551294327,
-0.018458539620041847,
0.015226438641548157,
-0.02414403110742569,
0.06790134310722351,
0.06603459268808365,
0.07196386903524399,
0.011758189648389816,
-0.012641883455216885,
-0.0357983373105526,
-0.005653723143041134,
0.07134941965341568,
-0.07395225018262863,
-0.04471432790160179,
-0.02186308614909649,
-0.05823918804526329,
0.0520167239010334,
-0.04670730605721474,
0.045740481466054916,
0.02701362408697605,
-0.023394955322146416,
-0.00039294545422308147,
0.14557726681232452,
-0.06671836972236633,
0.03072691336274147,
0.013943769969046116,
0.0003152253630105406,
0.04538213834166527,
0.014296287670731544,
0.037732336670160294,
-0.07077565789222717,
0.03770791366696358,
0.03353650122880936,
0.08404573053121567,
-0.004136992152780294,
0.1152423620223999,
0.02100853994488716,
0.01050947792828083,
0.02861100435256958,
0.011531912721693516,
-0.006640308070927858,
-0.056489646434783936,
-0.017653102055191994,
-0.05012162774801254,
0.01725948229432106,
0.0007222755812108517,
0.057905130088329315,
0.018659913912415504,
-0.13002271950244904,
0.0038540672976523638,
0.06438595801591873,
-0.020474813878536224,
0.000013759161447524093,
0.01716550625860691,
-0.05934319272637367,
0.09844528138637543,
-0.0010821328032761812,
0.06684345006942749,
0.010466481558978558,
0.04415372386574745,
-0.08262646943330765,
0.01643916592001915,
0.04600460082292557,
0.05887046456336975,
-0.03370169922709465,
0.018608232960104942,
0.017858514562249184,
-0.031196262687444687,
-0.014924663119018078,
0.001100495457649231,
0.07535844296216965,
-0.06545774638652802,
0.05335185304284096,
0.019586918875575066,
0.01896042563021183,
0.06914950907230377,
-0.0041964068077504635,
-0.033721935003995895,
2.619585854729606e-33,
-0.07611654698848724,
0.02867119386792183,
0.016260245814919472,
-0.004363600630313158,
0.02176305279135704,
-0.08841938525438309,
-0.07372007519006729,
0.002838799497112632,
-0.01358114555478096,
0.06178050488233566,
-0.07478684186935425,
-0.008345918729901314,
-0.11905939131975174,
0.02233489602804184,
0.08671821653842926,
0.05047166347503662,
-0.03430912643671036,
0.032245539128780365,
0.03259381651878357,
-0.055334582924842834,
0.10168928653001785,
-0.036163561046123505,
-0.010365255177021027,
-0.05857035890221596,
0.02988768368959427,
0.010424517095088959,
0.05887662619352341,
0.06841162592172623,
0.01747213676571846,
0.025838589295744896,
-0.050022102892398834,
-0.023618120700120926,
0.06158384680747986,
-0.0018366712611168623,
0.04559279605746269,
-0.04441268742084503,
-0.07460831105709076,
0.030091561377048492,
0.005086550489068031,
-0.035252563655376434,
-0.0032946113497018814,
0.006125198677182198,
-0.09975583106279373,
0.01152057759463787,
-0.0021129657980054617,
-0.0732048749923706,
-0.0029695520643144846,
-0.034345198422670364,
0.008028917945921421,
-0.06243591010570526,
0.010102890431880951,
-0.05350795015692711,
0.05785804241895676,
-0.011155386455357075,
-0.037017520517110825,
0.021343732252717018,
0.08656415343284607,
-0.0473601259291172,
-0.01564779505133629,
0.040835220366716385,
0.0010235814843326807,
0.042103417217731476,
-0.04478957876563072,
0.020394757390022278,
-0.03739451617002487,
0.06983862072229385,
0.021153567358851433,
0.1370806097984314,
0.004257388412952423,
-0.006031879689544439,
0.006874450016766787,
-0.004688098561018705,
-0.021676935255527496,
-0.0615055188536644,
0.06450393795967102,
0.011696765199303627,
0.0720273032784462,
-0.06740507483482361,
0.00011330358393024653,
0.07865609228610992,
0.01215439010411501,
0.038608796894550323,
-0.08908461779356003,
0.020478473976254463,
-0.09022119641304016,
-0.01851711980998516,
-0.006045382935553789,
-0.030496323481202126,
-0.013688502833247185,
0.02240375615656376,
-0.060348253697156906,
-0.07329300045967102,
-0.028452178463339806,
-0.026161711663007736,
-0.007196024991571903,
-2.3293493200138752e-33,
-0.08364113420248032,
0.09087633341550827,
-0.04983680695295334,
0.0675489529967308,
-0.0645180344581604,
-0.05766550451517105,
-0.12714405357837677,
-0.04275002330541611,
-0.00455948431044817,
0.022250834852457047,
-0.006796476896852255,
-0.011338060721755028,
0.010440558195114136,
-0.022610235959291458,
-0.05450136587023735,
-0.021501710638403893,
-0.06905925273895264,
-0.002876769285649061,
0.011262393556535244,
-0.017256563529372215,
0.04963403567671776,
0.0919259712100029,
-0.09732425212860107,
0.04069656506180763,
-0.06467830389738083,
0.0022296211682260036,
0.01104393694549799,
0.10805031657218933,
0.010710803791880608,
-0.053763248026371,
-0.00943460687994957,
-0.04469867795705795,
-0.01563369669020176,
0.06483370810747147,
0.02179753966629505,
0.05516801029443741,
0.10849922895431519,
-0.048227936029434204,
-0.023706190288066864,
-0.005578878801316023,
0.05386970937252045,
0.023362906649708748,
-0.061358124017715454,
-0.005879273638129234,
0.040334612131118774,
0.014103159308433533,
0.045419178903102875,
0.007056903094053268,
0.09045001864433289,
-0.0002602435415610671,
0.02848975919187069,
0.07864067703485489,
-0.12074014544487,
0.026830652728676796,
-0.09274780005216599,
-0.03150884062051773,
-0.0036954032257199287,
-0.0392361544072628,
-0.02824280597269535,
0.10132540762424469,
-0.043442633002996445,
-0.04029254987835884,
0.003165501868352294,
0.032673429697752,
-0.017058297991752625,
-0.015449083410203457,
-0.046656325459480286,
0.07479650527238846,
0.0976630225777626,
0.04350420460104942,
0.06692413985729218,
-0.004639609716832638,
-0.045643873512744904,
0.05835505947470665,
-0.03355633467435837,
-0.05162529647350311,
0.058161813765764236,
-0.02291378192603588,
0.007317979354411364,
-0.005976827349513769,
0.05166156589984894,
-0.027806155383586884,
0.04236452281475067,
0.04586553946137428,
0.054998498409986496,
0.017384160310029984,
0.010860824026167393,
-0.013013842515647411,
0.013629692606627941,
-0.00837182067334652,
-0.04502653330564499,
0.011558657512068748,
-0.07410795986652374,
0.020458534359931946,
-0.1051797941327095,
-4.7132342473332756e-8,
-0.0331641249358654,
0.08823186159133911,
0.08172847330570221,
0.027650443837046623,
0.04346483573317528,
0.01911846548318863,
-0.010366145521402359,
0.006693592760711908,
0.010092047974467278,
-0.0049740406684577465,
0.02597937360405922,
0.05044178292155266,
-0.06229411065578461,
-0.04807553440332413,
-0.041166361421346664,
0.012629344128072262,
0.018111426383256912,
0.020085636526346207,
0.023248618468642235,
0.024501293897628784,
0.08914286643266678,
0.04589107632637024,
0.010188976302742958,
-0.019401757046580315,
0.04731925204396248,
-0.13886748254299164,
-0.05531252548098564,
0.04151863977313042,
0.05108290910720825,
0.0756043866276741,
0.056073714047670364,
0.017679180949926376,
0.022833533585071564,
-0.027190949767827988,
-0.023850414901971817,
0.10151749104261398,
0.013454094529151917,
0.01725267805159092,
0.009379948489367962,
0.00978840421885252,
-0.03986065462231636,
0.10425326228141785,
-0.043114326894283295,
0.013642949983477592,
-0.08798015862703323,
-0.015241095796227455,
-0.036256905645132065,
-0.07473880052566528,
0.08788761496543884,
-0.06636346876621246,
0.06539421528577805,
-0.0632031112909317,
0.022301189601421356,
0.07174044847488403,
-0.006377992685884237,
0.0260305218398571,
-0.03418770059943199,
-0.06319840252399445,
-0.017861533910036087,
0.08952921628952026,
0.013463246636092663,
0.03338314965367317,
-0.060120902955532074,
0.03720822185277939
] |
https://github.com/harshitk11/evaluate_stochastic_process
| 0
| 0
| 0
| 0
|
LICORICE: Label-Efficient Concept-Based Interpretable Reinforcement Learning
|
https://openreview.net/forum?id=Mjn53GtMxi
|
[
"Zhuorui Ye",
"Stephanie Milani",
"Geoffrey J. Gordon",
"Fei Fang"
] |
Poster
|
Recent advances in reinforcement learning (RL) have predominantly leveraged neural network policies for decision-making, yet these models often lack interpretability, posing challenges for stakeholder comprehension and trust. Concept bottleneck models offer an interpretable alternative by integrating human-understandable concepts into policies. However, prior work assumes that concept annotations are readily available during training. For RL, this requirement poses a significant limitation: it necessitates continuous real-time concept annotation, which either places an impractical burden on human annotators or incurs substantial costs in API queries and inference time when employing automated labeling methods. To overcome this limitation, we introduce a novel training scheme that enables RL agents to efficiently learn a concept-based policy by only querying annotators to label a small set of data. Our algorithm, LICORICE, involves three main contributions: interleaving concept learning and RL training, using an ensemble to actively select informative data points for labeling, and decorrelating the concept data. We show how LICORICE reduces human labeling efforts to 500 or fewer concept labels in three environments, and 5000 or fewer in two more complex environments, all at no cost to performance. We also explore the use of VLMs as automated concept annotators, finding them effective in some cases but imperfect in others. Our work significantly reduces the annotation burden for interpretable RL, making it more practical for real-world applications that necessitate transparency. Our code is released.
|
Reinforcement Learning, Explainable Reinforcement Learning, Concept Bottleneck Models, Concept-based Explainability, Interpretability, XRL, Interpretable Reinforcement Learning, Interpretable Agent
|
We algorithmically reduce the concept annotation effort during training for interpretable reinforcement learning. We also explore the use of VLMs as automated concept annotators.
| 11,415
|
2407.15786
|
[
-0.059668730944395065,
-0.06587853282690048,
-0.04098593816161156,
0.0539814718067646,
0.06583847850561142,
-0.0003171232820022851,
0.0569692887365818,
-0.017015760764479637,
-0.007177315186709166,
-0.018110977485775948,
-0.0971565991640091,
-0.08636099100112915,
0.017843976616859436,
0.011034471914172173,
0.029436253011226654,
0.05150064453482628,
0.11623408645391464,
0.045853421092033386,
-0.11191622167825699,
-0.04583286494016647,
0.04503538832068443,
0.003244850318878889,
0.04945501685142517,
0.03373721241950989,
0.024613451212644577,
0.022668808698654175,
-0.02214011177420616,
-0.008631257340312004,
0.024580642580986023,
0.00015152104606386274,
0.022821806371212006,
-0.03736203536391258,
-0.03896496072411537,
0.0019882263150066137,
-0.0644892156124115,
0.01661856099963188,
-0.13920772075653076,
0.02265707217156887,
0.018432455137372017,
-0.01391172967851162,
0.01364125870168209,
-0.02972109243273735,
-0.03204842284321785,
0.02336260862648487,
0.10327178984880447,
-0.004028404597193003,
0.000003353579586473643,
-0.06956958025693893,
0.0011765281669795513,
0.05432989448308945,
-0.08009354770183563,
-0.0737919807434082,
0.030369386076927185,
0.0064909569919109344,
-0.04720081016421318,
0.03451130539178848,
0.026953019201755524,
0.01245178934186697,
-0.027811355888843536,
-0.025701506063342094,
-0.03416850045323372,
-0.1244814544916153,
0.006488646846264601,
0.024858303368091583,
-0.1273079663515091,
-0.08074704557657242,
0.029918843880295753,
0.07613382488489151,
0.03910199925303459,
-0.03272267058491707,
0.0734340101480484,
0.13793164491653442,
0.005336938891559839,
0.0008031359757296741,
0.0624941848218441,
-0.0023361085914075375,
0.011039139702916145,
-0.01487602386623621,
0.05546235665678978,
-0.050057023763656616,
-0.07924986630678177,
0.059345975518226624,
-0.005843211431056261,
0.04780711978673935,
0.07624958455562592,
0.01196995098143816,
0.0775463730096817,
0.07459063082933426,
0.00765385664999485,
-0.009945206344127655,
0.016376081854104996,
-0.024184200912714005,
0.04574053734540939,
-0.019729943946003914,
0.04855496063828468,
0.08030787110328674,
-0.033401936292648315,
-0.10587507486343384,
-0.08003144711256027,
0.05456635355949402,
-0.037522733211517334,
0.10463395714759827,
-0.06335368007421494,
-0.07646116614341736,
-0.038466859608888626,
-0.029058856889605522,
0.008317554369568825,
0.02250317670404911,
0.006334951613098383,
-0.06090833619236946,
-0.005864092614501715,
0.0249345600605011,
0.008784469217061996,
-0.004373753909021616,
0.03890986740589142,
0.03249571844935417,
0.04215177148580551,
-0.07017949223518372,
0.056152328848838806,
0.08393478393554688,
-0.02446463145315647,
0.015193908475339413,
0.023592039942741394,
-0.01604287698864937,
0.027861103415489197,
-0.04874025285243988,
-0.02519332617521286,
7.118581648583648e-34,
0.06442246586084366,
0.024673087522387505,
0.028300816193223,
0.005581323057413101,
0.027470840141177177,
0.00009776023216545582,
-0.016104187816381454,
0.08311955630779266,
-0.025817764922976494,
-0.01866658590734005,
0.025942794978618622,
0.04262905940413475,
-0.043682605028152466,
0.02290104143321514,
0.011400139890611172,
-0.03186444938182831,
-0.06669177860021591,
0.057644572108983994,
-0.05666957423090935,
-0.05068954452872276,
0.021243292838335037,
-0.025662638247013092,
-0.002472194144502282,
-0.003618720918893814,
0.020423363894224167,
0.08996360749006271,
0.08109521120786667,
0.01584622450172901,
0.03885343670845032,
-0.005828067660331726,
-0.03538515791296959,
-0.04255852848291397,
0.06628616154193878,
-0.016893697902560234,
0.00617351196706295,
-0.006190629210323095,
-0.016491560265421867,
0.007779605686664581,
0.0038231273647397757,
-0.005173920188099146,
-0.014283626340329647,
-0.010606483556330204,
-0.0032495311461389065,
0.0024519453290849924,
-0.06393569707870483,
-0.061359941959381104,
-0.03683092072606087,
-0.0026282775215804577,
-0.07138130813837051,
-0.036887187510728836,
0.02243397943675518,
0.04728037491440773,
0.026074059307575226,
-0.027175862342119217,
-0.03161321207880974,
-0.0003943217743653804,
-0.0389430895447731,
0.011179052293300629,
-0.058274608105421066,
-0.0424727238714695,
0.03510952740907669,
0.028665507212281227,
0.0409841388463974,
0.1021873727440834,
-0.027670012786984444,
0.10731588304042816,
-0.04114922881126404,
0.0890142023563385,
0.0767907053232193,
-0.0325649119913578,
-0.026597118005156517,
0.024165548384189606,
0.003929835744202137,
-0.09462877362966537,
0.09006741642951965,
0.017370063811540604,
0.010518945753574371,
-0.1358925849199295,
0.00316245318390429,
0.007014480885118246,
-0.006974147632718086,
-0.08136356621980667,
0.04289361461997032,
0.08634139597415924,
-0.0014352102298289537,
-0.04958074912428856,
0.05224199220538139,
-0.05084072798490524,
0.03346732258796692,
-0.021965673193335533,
-0.07183261960744858,
-0.016828011721372604,
-0.03452812880277634,
-0.018000761047005653,
0.01079599279910326,
-1.3867013901417782e-33,
0.02060762047767639,
-0.0661749541759491,
-0.0075385551899671555,
0.02409839816391468,
0.02764338254928589,
-0.04914122819900513,
-0.05248406156897545,
-0.009218729101121426,
0.06036345288157463,
-0.007066822610795498,
-0.08216620236635208,
0.007537566125392914,
0.024862002581357956,
-0.011000989936292171,
0.03113264963030815,
-0.05451374128460884,
-0.024192385375499725,
0.023622646927833557,
0.06447113305330276,
0.08482570946216583,
-0.017581842839717865,
0.0916733592748642,
-0.12918423116207123,
0.06460701674222946,
-0.014564131386578083,
0.032692357897758484,
-0.02019147388637066,
-0.0012970261741429567,
0.014777252450585365,
0.024123718962073326,
-0.08659900724887848,
-0.07587388157844543,
-0.011800681240856647,
-0.07381831109523773,
-0.07109713554382324,
0.06990709900856018,
0.08566401898860931,
-0.020577551797032356,
-0.11325259506702423,
0.06815673410892487,
0.045741964131593704,
-0.09856639057397842,
-0.08832162618637085,
-0.06477174907922745,
-0.014025391079485416,
-0.08557764440774918,
-0.059933312237262726,
0.01357848197221756,
-0.03504442423582077,
0.0977710634469986,
-0.03476806730031967,
0.0038202765863388777,
-0.0947728157043457,
0.01831497810781002,
-0.030569840222597122,
-0.06069072708487511,
-0.07205025851726532,
0.019563188776373863,
0.07667731493711472,
-0.004087979905307293,
-0.08584071695804596,
0.0573047399520874,
0.01000015065073967,
0.03996334969997406,
-0.003825538791716099,
-0.03845949470996857,
-0.01587161235511303,
0.024700572714209557,
0.054740335792303085,
-0.03266582265496254,
0.051291171461343765,
0.043729353696107864,
-0.019085107371211052,
-0.006658483762294054,
0.005510096438229084,
0.004043050576001406,
-0.007223275490105152,
-0.07818764448165894,
-0.013445567339658737,
-0.023298809304833412,
0.10752598941326141,
-0.0005996382096782327,
0.06648831069469452,
0.04126197472214699,
0.06549856066703796,
0.01262709405273199,
0.026547357439994812,
0.10994558036327362,
0.04339727759361267,
-0.013519583269953728,
-0.01844877190887928,
-0.00306120328605175,
-0.04452196881175041,
0.08493843674659729,
-0.1344722956418991,
-3.845786267220319e-8,
-0.011318367905914783,
0.06255503743886948,
0.06677218526601791,
0.10560280829668045,
0.04479954019188881,
-0.06279827654361725,
-0.056920576840639114,
0.0768560916185379,
-0.040089722722768784,
0.02044057659804821,
0.10378917306661606,
0.03948565199971199,
-0.048829954117536545,
-0.07378095388412476,
-0.0009165630326606333,
0.09182659536600113,
0.006031721364706755,
-0.037327609956264496,
0.026351772248744965,
-0.007616009097546339,
0.026272132992744446,
0.0029675252735614777,
0.007040307391434908,
-0.05206432566046715,
0.04093894362449646,
-0.07588452100753784,
-0.018328847363591194,
0.11992117017507553,
0.08204784989356995,
-0.0064699845388531685,
0.01323250588029623,
0.004180814605206251,
0.020454226061701775,
-0.011491511948406696,
0.06961003690958023,
0.03874580189585686,
0.03175074979662895,
-0.02951856330037117,
0.03960519656538963,
-0.056880295276641846,
0.008544069714844227,
0.0664142519235611,
0.0035978553351014853,
-0.0199152659624815,
0.024844666942954063,
0.011418712325394154,
-0.0780826136469841,
-0.08315874636173248,
0.08986526727676392,
-0.051800116896629333,
-0.04720374569296837,
0.010825329460203648,
-0.02328513190150261,
0.04580932483077049,
0.044118646532297134,
-0.006884775590151548,
-0.024211609736084938,
-0.04190525785088539,
0.01902739517390728,
0.04931154102087021,
0.007562259677797556,
-0.019195636734366417,
-0.014211030676960945,
0.021700626239180565
] | 0
| 0
| 0
| 0
|
|
MIND: Math Informed syNthetic Dialogues for Pretraining LLMs
|
https://openreview.net/forum?id=TuOTSAiHDn
|
[
"Syeda Nahida Akter",
"Shrimai Prabhumoye",
"John Kamalu",
"Sanjeev Satheesh",
"Eric Nyberg",
"Mostofa Patwary",
"Mohammad Shoeybi",
"Bryan Catanzaro"
] |
Poster
|
The utility of synthetic data to enhance pretraining data quality and hence to improve downstream task accuracy has been widely explored in recent large language models (LLMs). Yet, these approaches fall inadequate in complex, multi-hop and mathematical reasoning tasks as the synthetic data typically fails to add complementary knowledge to the existing raw corpus. In this work, we propose a novel large-scale and diverse Math Informed syNthetic Dialogue (MIND) generation method that improves the mathematical reasoning ability of LLMs. Specifically, using MIND, we generate synthetic conversations based on OpenWebMath (OWM), resulting in a new math corpus, MIND-OWM. Our experiments with different conversational settings reveal that incorporating knowledge gaps between dialog participants is essential for generating high-quality math data. We further identify an effective way to format and integrate synthetic and raw data during pretraining to maximize the gain in mathematical reasoning, emphasizing the need to restructure raw data rather than use it as-is. Compared to pretraining just on raw data, a model pretrained on MIND-OWM shows significant boost in mathematical reasoning (GSM8K: +13.42%, MATH: +2.30%), including superior performance in specialized knowledge (MMLU: +4.55%, MMLU-STEM: +4.28%) and general
purpose reasoning tasks (GENERAL REASONING: +2.51%).
|
pretraining, mathematical reasoning, synthetic dialogue, LLM, reasoning
|
We propose a novel large-scale and diverse Math Informed syNthetic Dialogue (MIND) generation method that improves the mathematical reasoning ability of LLMs during pretraining.
| 11,410
|
2410.12881
|
[
-0.028084492310881615,
-0.07791867852210999,
0.035893090069293976,
0.05079859122633934,
0.0015180051559582353,
-0.08810636401176453,
-0.03502485901117325,
0.018270380795001984,
0.05386605113744736,
-0.030252741649746895,
-0.08630173653364182,
-0.0461159348487854,
0.017014220356941223,
0.041011739522218704,
0.08805032819509506,
0.076088085770607,
0.0539138987660408,
0.05587749555706978,
-0.06841382384300232,
-0.051155779510736465,
0.0032144899014383554,
0.016768116503953934,
0.08049947023391724,
-0.0034260801039636135,
0.0829455628991127,
-0.0114146051928401,
0.015803584828972816,
-0.06804968416690826,
0.05226486921310425,
0.018141306936740875,
0.00741161871701479,
0.05704480782151222,
0.022117963060736656,
0.009763757698237896,
-0.11328396946191788,
0.08881440758705139,
-0.050802286714315414,
0.09271159768104553,
-0.028069430962204933,
-0.009929323568940163,
-0.0760589987039566,
-0.020310405641794205,
0.06639961153268814,
0.0018757261568680406,
0.010989264585077763,
0.006169282831251621,
-0.029653247445821762,
-0.016586432233452797,
0.0076883998699486256,
0.048858642578125,
-0.16901767253875732,
0.008770374581217766,
-0.04978834092617035,
0.08379925042390823,
0.03766753897070885,
0.016540590673685074,
0.044842805713415146,
0.06615431606769562,
-0.007953298278152943,
-0.027751393616199493,
-0.07482586801052094,
-0.11270887404680252,
-0.0024929740466177464,
-0.04025569185614586,
-0.07006145268678665,
0.0007129498990252614,
0.009164725430309772,
0.09878653287887573,
-0.033715587109327316,
0.03829526528716087,
-0.01908782310783863,
0.10041951388120651,
-0.033683378249406815,
0.034221623092889786,
-0.020191509276628494,
0.08534606546163559,
-0.002454346977174282,
-0.013454360887408257,
0.08828112483024597,
0.017702220007777214,
0.012465179897844791,
-0.003260415280237794,
-0.023256856948137283,
0.0178726427257061,
0.1189221665263176,
0.031300514936447144,
0.021679401397705078,
0.07991571724414825,
0.002203016309067607,
0.014768635854125023,
-0.030247008427977562,
-0.05220525339245796,
0.01944177970290184,
0.05702104791998863,
0.04639704152941704,
0.12222734093666077,
-0.034121073782444,
-0.022801829501986504,
0.018957069143652916,
0.07086513936519623,
0.040333908051252365,
0.08497781306505203,
0.011910444125533104,
-0.048143062740564346,
-0.030826551839709282,
-0.019116973504424095,
0.0899205133318901,
0.05709683150053024,
-0.007895644754171371,
-0.04142291843891144,
0.0769590213894844,
0.02852567471563816,
0.056500233709812164,
0.016201583668589592,
0.012968205846846104,
0.027140667662024498,
0.05262034013867378,
-0.0512312613427639,
0.008497431874275208,
0.016843976452946663,
0.0028001503087580204,
0.06454435735940933,
0.006440307479351759,
0.055418841540813446,
0.05977526679635048,
-0.01713438145816326,
-0.028813106939196587,
8.73916867015014e-34,
0.07314211875200272,
0.029597843065857887,
0.07586478441953659,
0.09149191528558731,
0.05520043894648552,
0.02033710479736328,
-0.031134674325585365,
0.017529135569930077,
-0.006576532032340765,
0.059310659766197205,
-0.031206898391246796,
-0.014384463429450989,
-0.06091063469648361,
0.06855606287717819,
0.043256163597106934,
0.0009122540941461921,
-0.10320915281772614,
-0.03156328946352005,
-0.03916323557496071,
-0.047130387276411057,
0.02676987089216709,
-0.017972424626350403,
-0.021999528631567955,
-0.0014599761925637722,
0.046187881380319595,
0.10341743379831314,
0.10149343311786652,
-0.06355325132608414,
-0.030884388834238052,
-0.012921912595629692,
-0.09755438566207886,
-0.03826466575264931,
-0.03332357108592987,
0.03345822915434837,
0.0034259853418916464,
-0.01297667995095253,
-0.012890515848994255,
-0.07717236876487732,
0.02814590372145176,
0.03273391351103783,
0.00903975497931242,
0.014350798912346363,
0.057905204594135284,
0.1019454225897789,
-0.061313796788454056,
-0.05894455313682556,
-0.02703149989247322,
0.07214677333831787,
0.004335465375334024,
0.0140526769682765,
-0.024759678170084953,
0.03189589083194733,
-0.0069627659395337105,
-0.08931948989629745,
-0.04348624870181084,
-0.07056669145822525,
0.004520968068391085,
0.031457722187042236,
0.013980235904455185,
0.03620534390211105,
-0.0349527932703495,
0.052027639001607895,
0.010446283966302872,
0.057812854647636414,
-0.047327954322099686,
0.05871778354048729,
-0.10214603692293167,
0.0758562684059143,
0.06086374446749687,
0.026201140135526657,
-0.013219278305768967,
0.006588428281247616,
-0.03611225262284279,
0.006020484492182732,
0.01306268572807312,
-0.01569323241710663,
0.04440095275640488,
-0.06659611314535141,
0.07361327856779099,
0.01729736104607582,
0.10992896556854248,
-0.004260770045220852,
-0.07763992995023727,
-0.03985610604286194,
-0.01843174360692501,
-0.06148817390203476,
0.019891662523150444,
-0.08466459065675735,
-0.030696237459778786,
0.0045923758298158646,
-0.06112964451313019,
-0.04881471022963524,
-0.04794719070196152,
-0.007269334979355335,
-0.04238696023821831,
-6.596926183870007e-34,
-0.04253483563661575,
0.023340603336691856,
-0.041669100522994995,
0.0716862678527832,
-0.06961791217327118,
0.0052933297120034695,
0.02609293721616268,
0.04229392856359482,
0.008052523247897625,
-0.04753253236413002,
-0.039380382746458054,
-0.02473444491624832,
0.038832686841487885,
-0.03745752200484276,
0.017390061169862747,
-0.06580816209316254,
-0.039795491844415665,
-0.030365241691470146,
-0.04748435690999031,
0.041366610676050186,
-0.025848055258393288,
0.08991532772779465,
-0.08079679310321808,
0.013817435130476952,
-0.021332096308469772,
0.03186316788196564,
-0.01390837226063013,
0.10444222390651703,
-0.058924123644828796,
0.05091492459177971,
-0.003348373342305422,
-0.10143908113241196,
-0.03145832568407059,
-0.05943676456809044,
-0.09027943015098572,
0.02585684508085251,
0.10746841877698898,
-0.033474575728178024,
-0.014982818625867367,
0.0026230530347675085,
0.02264280430972576,
-0.10456222295761108,
-0.04180926829576492,
0.01903894729912281,
-0.07880853116512299,
0.007297656498849392,
-0.12312863767147064,
-0.018618537113070488,
-0.012593338266015053,
0.03838507831096649,
-0.008551246486604214,
-0.02574196644127369,
-0.01046949252486229,
-0.04170242324471474,
0.019192498177289963,
-0.10780074447393417,
-0.09224944561719894,
-0.049046579748392105,
-0.04488750547170639,
-0.023852158337831497,
-0.0648449957370758,
0.0036095483228564262,
0.08216888457536697,
-0.09092112630605698,
-0.018596218898892403,
-0.06658350676298141,
-0.08120947331190109,
0.015964264050126076,
0.004547925665974617,
-0.00044297214481048286,
0.03010173887014389,
-0.005404804367572069,
-0.0020020995289087296,
-0.005264240317046642,
0.07909945398569107,
0.00397015456110239,
0.007890990003943443,
0.015032992698252201,
-0.002015892881900072,
-0.05846622958779335,
0.07993768155574799,
0.014429278671741486,
0.062197573482990265,
0.016974378377199173,
0.02715456299483776,
0.03893991559743881,
0.05347735062241554,
0.12430348247289658,
0.0010494671296328306,
0.05944237858057022,
-0.0909423828125,
0.05554952099919319,
-0.07674961537122726,
0.08075160533189774,
-0.09897646307945251,
-3.7206479674978254e-8,
-0.05324218422174454,
0.03458864241838455,
0.0360686331987381,
0.029007617384195328,
0.0251908116042614,
-0.0483679436147213,
-0.061351705342531204,
0.05056279897689819,
-0.07480181008577347,
-0.011948405764997005,
0.09216275811195374,
0.009854806587100029,
-0.05729617550969124,
-0.02857305109500885,
-0.022490672767162323,
0.05878234654664993,
0.0037012924440205097,
-0.032142527401447296,
-0.02054319903254509,
-0.06613986939191818,
0.07529693841934204,
0.06647323817014694,
-0.02276689000427723,
0.06722111254930496,
0.05698363855481148,
-0.05844632908701897,
-0.022013362497091293,
0.09763593971729279,
0.016935089603066444,
-0.01567062921822071,
-0.07068953663110733,
-0.029380636289715767,
-0.03781847655773163,
-0.017881862819194794,
-0.02168860472738743,
0.0475597083568573,
0.019196027889847755,
0.007452524732798338,
0.03944866731762886,
-0.07051695138216019,
-0.002893195953220129,
0.02131546474993229,
-0.0501818023622036,
0.019023489207029343,
0.027584388852119446,
-0.02899811789393425,
-0.076129250228405,
-0.05513690039515495,
0.003946787677705288,
0.006769836414605379,
0.018304230645298958,
-0.0014264433411881328,
-0.007180863991379738,
-0.01092209480702877,
0.06293169409036636,
-0.025440074503421783,
-0.05554194003343582,
0.01784798502922058,
-0.010620098561048508,
0.03165936470031738,
0.022167064249515533,
0.013718035072088242,
-0.023647595196962357,
-0.0008411076851189137
] | 0
| 0
| 0
| 0
|
|
SMT: Fine-Tuning Large Language Models with Sparse Matrices
|
https://openreview.net/forum?id=GbgCRJedQ7
|
[
"Haoze He",
"Juncheng B Li",
"Xuan Jiang",
"Heather Miller"
] |
Poster
|
Various parameter-efficient fine-tuning (PEFT) methods, including LoRA and its variants, have gained popularity for reducing computational costs. However, there is often an accuracy gap between PEFT approaches and full fine-tuning (FT), and this discrepancy has not yet been systematically explored. In this work, we introduce a method for selecting sparse sub-matrices that aims to minimize the performance gap between PEFT vs. full fine-tuning (FT) while also reducing both fine-tuning computational costs and memory costs. We explored both gradient-based and activation-based parameter selection methods to identify the most significant sub-matrices for downstream tasks, updating only these blocks during fine-tuning. In our experiments, we demonstrated that SMT consistently surpasses other PEFT
baselines (e.g., LoRA and DoRA) in fine-tuning popular large language models such as LLaMA across a broad spectrum of tasks, while reducing the GPU memory footprint by 67% compared to FT. We also examine how the performance of LoRA and DoRA tends to plateau and decline as the number of trainable parameters increases, in contrast, our SMT method does not suffer from such issues.
|
Parameter-efficient Finetuning, Large Language Model, Large Language Model Systm
|
We propose a novel fine-tuning method (SMT) which achieves state-of-the-art performance in parameter-efficient fine-tuning, effectively closing the gap between SMT and full fine-tuning.
| 11,409
| null |
[
-0.0172929298132658,
-0.10009724646806717,
-0.015055728144943714,
0.0550362803041935,
0.056796085089445114,
-0.0012745733838528395,
-0.01956566423177719,
0.0023568749893456697,
0.029405709356069565,
-0.08373645693063736,
-0.08573570102453232,
-0.008833084255456924,
-0.034858714789152145,
-0.0034893753472715616,
0.02713189274072647,
0.01913583278656006,
0.09711676836013794,
0.048171188682317734,
-0.07021286338567734,
-0.06281552463769913,
0.018286950886249542,
0.005899981129914522,
0.03473727032542229,
-0.025224145501852036,
0.01606658846139908,
-0.0074652330949902534,
-0.0355520136654377,
-0.041883278638124466,
0.02921328879892826,
0.003533098381012678,
-0.05766601860523224,
0.06943286210298538,
0.003927410580217838,
0.0099248131737113,
-0.05530848354101181,
0.07770991325378418,
-0.047952547669410706,
0.048244159668684006,
-0.020409129559993744,
-0.050968967378139496,
0.021782448515295982,
0.03158934414386749,
-0.016243232414126396,
0.07120189070701599,
0.10138779133558273,
-0.059434499591588974,
0.036681026220321655,
-0.047404415905475616,
-0.03227263689041138,
0.020280173048377037,
-0.08285398781299591,
-0.025088893249630928,
0.0010242459829896688,
0.04568935185670853,
-0.03976791724562645,
-0.0831741914153099,
0.029399747028946877,
-0.011847654357552528,
0.008557392284274101,
-0.04571998491883278,
-0.09537120908498764,
-0.1348406821489334,
-0.054872941225767136,
-0.03682849183678627,
-0.0009568058885633945,
-0.0249390359967947,
0.04923442378640175,
-0.0055114757269620895,
-0.048376038670539856,
0.056853145360946655,
-0.03947580233216286,
0.07702472805976868,
-0.033391956239938736,
0.08214621990919113,
0.012502975761890411,
0.03586520999670029,
0.059318337589502335,
-0.06186632066965103,
0.07483156770467758,
0.00670201238244772,
0.020159421488642693,
-0.0006090655806474388,
-0.002857749816030264,
-0.03747470676898956,
0.15027879178524017,
-0.02143251523375511,
0.0747494027018547,
0.012448551133275032,
0.05253208056092262,
-0.011623483151197433,
0.02248716540634632,
-0.02483529783785343,
0.03090035542845726,
-0.052600834518671036,
-0.0017514652572572231,
0.06683782488107681,
0.00906838197261095,
-0.053136229515075684,
-0.04121951758861542,
0.10233102738857269,
0.0439632423222065,
0.04295892268419266,
0.034864552319049835,
0.013968120329082012,
-0.08831121027469635,
-0.007654160261154175,
0.06951244920492172,
0.0909394845366478,
-0.042793530970811844,
-0.07480203360319138,
0.002795492997393012,
0.0664057806134224,
0.031173307448625565,
0.09145353734493256,
0.023788539692759514,
0.020407892763614655,
0.004959941376000643,
-0.0724807158112526,
0.050498705357313156,
0.12467016279697418,
-0.07073602080345154,
0.021466748788952827,
0.060796089470386505,
0.015800757333636284,
-0.04123164340853691,
0.030843760818243027,
-0.12043561786413193,
2.3699161810737293e-33,
0.04856139421463013,
0.044257618486881256,
0.02008853293955326,
-0.010526442900300026,
-0.020159438252449036,
-0.013223189860582352,
0.04444558173418045,
0.026078255847096443,
-0.047754526138305664,
0.03387556970119476,
-0.048337649554014206,
0.016308050602674484,
-0.022185465320944786,
0.06538514047861099,
0.012540061958134174,
-0.035925865173339844,
-0.011426943354308605,
0.07225203514099121,
-0.04296668991446495,
-0.07987219095230103,
0.06820930540561676,
0.01616717502474785,
0.01916484534740448,
-0.05487660691142082,
-0.005937639623880386,
-0.020268896594643593,
0.08352213352918625,
-0.07757321000099182,
-0.06641166657209396,
0.024418752640485764,
-0.10384435206651688,
-0.010255302302539349,
0.03458984196186066,
0.04166487231850624,
0.01892455480992794,
-0.06025644391775131,
-0.02322407253086567,
-0.02265738695859909,
0.07677683234214783,
-0.022752417251467705,
-0.00853679608553648,
0.024827662855386734,
0.09259746223688126,
-0.07079534232616425,
-0.10578028112649918,
-0.056869152933359146,
-0.0028083203360438347,
0.034911442548036575,
-0.03464953601360321,
-0.017644217237830162,
0.038895249366760254,
0.015790585428476334,
-0.0698770210146904,
0.029134884476661682,
-0.033081185072660446,
-0.025249509140849113,
0.03740369901061058,
0.03306124731898308,
0.02829149179160595,
0.06617550551891327,
0.016740305349230766,
-0.00213190377689898,
-0.03120129369199276,
0.06750572472810745,
0.07462820410728455,
0.010245892219245434,
-0.06342091411352158,
0.1449122577905655,
0.037982337176799774,
0.005491273477673531,
0.001581721124239266,
-0.03981472924351692,
0.0308701004832983,
-0.058054760098457336,
0.03954153135418892,
-0.0661250576376915,
0.08386260271072388,
-0.058941178023815155,
-0.03349614888429642,
0.027726130560040474,
0.016967134550213814,
0.05949819087982178,
0.02793090045452118,
-0.06328442692756653,
-0.08064313232898712,
-0.0990230143070221,
0.0023827929981052876,
-0.049656037241220474,
-0.03220550715923309,
-0.02694329246878624,
-0.07083233445882797,
0.020096445456147194,
-0.006286322604864836,
-0.05317973345518112,
0.010994897224009037,
-1.3732058884748729e-33,
0.002234353218227625,
-0.04270728677511215,
0.03834434598684311,
0.10046898573637009,
-0.0011726159136742353,
-0.10635902732610703,
0.04429585486650467,
-0.0019545669201761484,
-0.0005097089451737702,
-0.03138403594493866,
-0.040668390691280365,
-0.07406095415353775,
0.01851809211075306,
-0.027231259271502495,
0.09318382292985916,
-0.018966732546687126,
-0.06983313709497452,
0.015835771337151527,
0.03561701625585556,
0.029413726180791855,
-0.07046954333782196,
0.06978476047515869,
-0.03638923168182373,
0.027670525014400482,
0.002156553091481328,
0.01791451871395111,
-0.030577395111322403,
0.038140010088682175,
-0.013734251260757446,
-0.013693598099052906,
-0.12117799371480942,
0.07004901766777039,
-0.06278201192617416,
0.05350477620959282,
0.007812955416738987,
0.06610607355833054,
0.026267128065228462,
0.01288242731243372,
0.011708189733326435,
0.1270545870065689,
0.05875689163804054,
0.03134865686297417,
-0.04400848224759102,
0.02577121928334236,
0.00890393741428852,
0.020895441994071007,
-0.11857102811336517,
-0.01001470722258091,
-0.016339963302016258,
-0.037122175097465515,
-0.00627434765920043,
0.030130460858345032,
-0.028642689809203148,
0.01852450519800186,
-0.009014959447085857,
-0.06713998317718506,
-0.033955808728933334,
-0.04425088316202164,
0.030299590900540352,
-0.009740086272358894,
-0.08092474192380905,
-0.027706312015652657,
0.010407082736492157,
-0.1281684935092926,
0.04204820096492767,
-0.010069869458675385,
0.005273771937936544,
-0.016782166436314583,
0.025638863444328308,
0.03195584565401077,
-0.04755544289946556,
-0.03719327598810196,
-0.029679881408810616,
-0.01609979011118412,
-0.04494647681713104,
0.01886691339313984,
0.017470696941018105,
-0.05391637608408928,
-0.009286489337682724,
-0.022496027871966362,
0.04423614591360092,
0.04602986201643944,
-0.0529235303401947,
0.0032887523993849754,
-0.013040672056376934,
0.12889869511127472,
0.027565311640501022,
0.09356057643890381,
0.02827795408666134,
-0.03337959572672844,
-0.032964106649160385,
-0.007884275168180466,
0.11995968967676163,
0.04831916093826294,
-0.08989966660737991,
-4.485642790541533e-8,
-0.047229886054992676,
0.02088235132396221,
-0.05639638751745224,
0.08706872910261154,
0.023538140580058098,
-0.08395300060510635,
-0.018089311197400093,
0.08697373420000076,
0.021054673939943314,
0.0008480440592393279,
0.03284689784049988,
-0.029971569776535034,
-0.04652930423617363,
-0.055458150804042816,
0.04514094814658165,
0.03012634441256523,
0.007045269478112459,
0.0760861188173294,
-0.01600668393075466,
-0.025866612792015076,
0.01260643545538187,
0.06836894899606705,
-0.005149539094418287,
0.07835051417350769,
0.020258178934454918,
-0.0627087727189064,
-0.036863774061203,
0.0676581934094429,
0.1674826592206955,
0.013188845477998257,
0.009319073520600796,
0.007747844327241182,
-0.021518180146813393,
-0.0470169372856617,
0.03017699159681797,
0.07405023276805878,
0.015319748781621456,
-0.03317577391862869,
-0.03023705817759037,
0.002633538795635104,
0.09726966917514801,
0.034022822976112366,
-0.03322896361351013,
0.005935287103056908,
0.007903737016022205,
-0.011031465604901314,
-0.08839573711156845,
-0.06929459422826767,
0.02883438766002655,
-0.06620395183563232,
-0.012927055358886719,
0.005654327105730772,
-0.013343904167413712,
0.08100318163633347,
0.021028151735663414,
-0.002194430213421583,
-0.0010732432128861547,
-0.055041756480932236,
0.058377258479595184,
0.02966095693409443,
0.04007227346301079,
0.017682889476418495,
-0.056848376989364624,
-0.0003537840966600925
] | 0
| 0
| 0
| 0
|
|
KGARevion: An AI Agent for Knowledge-Intensive Biomedical QA
|
https://openreview.net/forum?id=tnB94WQGrn
|
[
"Xiaorui Su",
"Yibo Wang",
"Shanghua Gao",
"Xiaolong Liu",
"Valentina Giunchiglia",
"Djork-Arné Clevert",
"Marinka Zitnik"
] |
Poster
|
Biomedical reasoning integrates structured, codified knowledge with tacit, experience-driven insights. Depending on the context, quantity, and nature of available evidence, researchers and clinicians use diverse strategies, including rule-based, prototype-based, and case-based reasoning. Effective medical AI models must handle this complexity while ensuring reliability and adaptability. We introduce KGARevion, a knowledge graph-based agent that answers knowledge-intensive questions. Upon receiving a query, KGARevion generates relevant triplets by leveraging the latent knowledge embedded in a large language model. It then verifies these triplets against a grounded knowledge graph, filtering out errors and retaining only accurate, contextually relevant information for the final answer. This multi-step process strengthens reasoning, adapts to different models of medical inference, and outperforms retrieval-augmented generation-based approaches that lack effective verification mechanisms. Evaluations on medical QA benchmarks show that KGARevion improves accuracy by over 5.2% over 15 models in handling complex medical queries. To further assess its effectiveness, we curated three new medical QA datasets with varying levels of semantic complexity, where KGARevion improved accuracy by 10.4%. The agent integrates with different LLMs and biomedical knowledge graphs for broad applicability across knowledge-intensive tasks. We evaluated KGARevion on AfriMed-QA, a newly introduced dataset focused on African healthcare, demonstrating its strong zero-shot generalization to underrepresented medical contexts.
|
Medical Reasoning, Medical QA, Agent, Knowledge Graph, LLM
|
A knowledge graph based agent for knowledge-intensive biomedical question answering
| 11,401
|
2410.04660
|
[
-0.0019268818432465196,
0.0034273937344551086,
-0.02034565806388855,
0.036346182227134705,
-0.032350290566682816,
-0.04196270927786827,
-0.031063493341207504,
0.060446593910455704,
0.03609804809093475,
-0.010419104248285294,
-0.06425338238477707,
-0.06543448567390442,
0.05776118114590645,
0.08120840787887573,
0.0038668219931423664,
0.0447143130004406,
0.02801511250436306,
-0.019062643870711327,
-0.14320899546146393,
-0.03910578042268753,
0.02673063613474369,
0.07855382561683655,
0.07414324581623077,
-0.029114169999957085,
-0.032364748418331146,
0.000503617397043854,
0.019115449860692024,
-0.09027130901813507,
0.005616945680230856,
0.04707817733287811,
0.0968375951051712,
0.018021559342741966,
-0.024464324116706848,
0.036435991525650024,
0.015537369064986706,
0.08191333711147308,
-0.05035105347633362,
0.06365892291069031,
-0.015519093722105026,
0.05370721220970154,
0.022443024441599846,
-0.0007115962798707187,
0.004787991754710674,
0.0717376247048378,
0.13683366775512695,
0.009111159481108189,
-0.11953998357057571,
-0.0013262073043733835,
0.07039571553468704,
0.065726637840271,
-0.12750805914402008,
-0.06476429849863052,
-0.0022997637279331684,
0.04856965318322182,
0.03051500953733921,
0.011384952813386917,
-0.07492714375257492,
0.009080473333597183,
-0.04046927019953728,
-0.0640534907579422,
0.00011314264702377841,
-0.13225015997886658,
-0.0003332874330226332,
-0.03250053524971008,
-0.0038968094158917665,
-0.04688115790486336,
0.01829150877892971,
-0.03335165977478027,
-0.030951926484704018,
0.02598894201219082,
-0.038030955940485,
0.0372530072927475,
-0.05172818899154663,
0.08160863816738129,
-0.037625789642333984,
0.07612285763025284,
0.02520611323416233,
0.01239671278744936,
0.03554671257734299,
-0.046728070825338364,
0.02790602296590805,
0.008306504227221012,
0.011145180091261864,
0.05728190392255783,
0.08329415321350098,
-0.016965026035904884,
0.0590636320412159,
0.037209246307611465,
-0.037313688546419144,
0.01504542212933302,
-0.014268532395362854,
-0.052296653389930725,
0.08986062556505203,
-0.025916224345564842,
0.09993352741003036,
0.07649072259664536,
-0.005177143029868603,
-0.07366422563791275,
-0.03069201670587063,
0.04819689318537712,
0.010465743020176888,
0.06410209089517593,
0.000022528554836753756,
-0.0028951566200703382,
-0.03809697553515434,
-0.05005371570587158,
0.00786252785474062,
0.06644821912050247,
0.009963581338524818,
-0.08168640732765198,
-0.009046593680977821,
0.07896249741315842,
0.07303695380687714,
-0.033090561628341675,
0.055248454213142395,
0.03174500912427902,
0.032652597874403,
0.044744424521923065,
0.05819252133369446,
0.0033567016944289207,
-0.010143976658582687,
-0.0041014282032847404,
0.05391941964626312,
-0.05666356161236763,
0.0782947838306427,
0.03211924061179161,
-0.009257357567548752,
1.7206188357565844e-34,
0.07943183928728104,
0.059715624898672104,
0.15456727147102356,
0.021635133773088455,
0.026687046512961388,
0.02384341135621071,
-0.08754665404558182,
0.07176049798727036,
-0.024115802720189095,
-0.015604862943291664,
-0.06084006652235985,
0.017406998202204704,
-0.06616409122943878,
-0.024476580321788788,
0.007076268550008535,
0.08503054082393646,
-0.10037239640951157,
-0.004217205569148064,
-0.04657992348074913,
-0.05673769488930702,
0.09217333048582077,
-0.11340080946683884,
0.010046591982245445,
0.05936301127076149,
-0.008994139730930328,
-0.008287230506539345,
0.09506671130657196,
0.0012720123631879687,
0.0007070825668051839,
-0.039116665720939636,
-0.1753632128238678,
-0.0009180065826512873,
0.0025103786028921604,
0.014826054684817791,
-0.03667124733328819,
-0.01940588280558586,
-0.016782986000180244,
-0.09955955296754837,
0.04407874494791031,
0.023238664492964745,
-0.05804165452718735,
0.015512628480792046,
0.0647977814078331,
-0.005003044847398996,
-0.031227579340338707,
-0.02141539566218853,
-0.056223444640636444,
-0.10818110406398773,
0.0381806343793869,
-0.0455089695751667,
0.06340902298688889,
0.0099900271743536,
0.038074567914009094,
-0.011277616955339909,
0.038172248750925064,
-0.011763964779675007,
-0.01643826998770237,
0.07257851958274841,
0.01931619830429554,
0.09122268855571747,
0.07549580931663513,
0.011088426224887371,
0.013083201833069324,
0.12183767557144165,
-0.02304726094007492,
0.038163572549819946,
-0.08558987826108932,
0.09520123153924942,
0.042406681925058365,
0.0024989431258291006,
-0.004049050156027079,
0.06410978734493256,
0.021243903785943985,
-0.046566396951675415,
-0.012445926666259766,
-0.015260321088135242,
-0.0352800190448761,
-0.09792447835206985,
-0.027592280879616737,
-0.04410559684038162,
-0.017235098406672478,
-0.04055912792682648,
-0.08345149457454681,
0.0507054477930069,
-0.03021928109228611,
-0.0858110636472702,
0.02231585793197155,
-0.04537420719861984,
-0.04473237693309784,
-0.08532381057739258,
-0.09928277879953384,
0.00200158404186368,
-0.012167613953351974,
-0.002458126749843359,
0.007911574095487595,
-4.626318270601908e-34,
-0.016689280048012733,
-0.05748852714896202,
-0.02543776109814644,
0.0296833124011755,
0.06841570138931274,
-0.06387603282928467,
-0.048414140939712524,
0.01794574409723282,
0.03074592724442482,
-0.033086106181144714,
0.010813893750309944,
-0.05558706820011139,
0.05156148225069046,
-0.0020378378685563803,
-0.007663849275559187,
0.048268865793943405,
-0.07411973178386688,
-0.020301740616559982,
0.002516590990126133,
0.06117808818817139,
0.008144691586494446,
0.028081510215997696,
-0.09981013834476471,
0.028596462681889534,
-0.005277459044009447,
0.05417283624410629,
-0.021531078964471817,
0.03012082353234291,
-0.03689103201031685,
0.022533291950821877,
-0.020989693701267242,
-0.08354129642248154,
-0.034844789654016495,
-0.04576550051569939,
-0.05302098020911217,
0.040894050151109695,
0.08585825562477112,
-0.08914300054311752,
0.0003591289569158107,
0.025367943570017815,
-0.0005268697277642787,
-0.038641538470983505,
-0.07455446571111679,
0.01888001337647438,
-0.009081210941076279,
0.038918230682611465,
-0.044419240206480026,
0.009242802858352661,
0.04079362750053406,
0.035148028284311295,
0.0072176107205450535,
-0.033024292439222336,
-0.03970117121934891,
0.0033472042996436357,
-0.03376971557736397,
-0.0804922804236412,
-0.021448783576488495,
0.01927097886800766,
-0.01883920095860958,
-0.032262105494737625,
-0.09047020226716995,
-0.02365480363368988,
0.02359267883002758,
0.005502325017005205,
0.01551013719290495,
-0.0003223242820240557,
0.04530717805027962,
0.03736003488302231,
0.016416804865002632,
0.01779613271355629,
-0.04448741674423218,
0.026038162410259247,
-0.009099462069571018,
0.006197791080921888,
0.14242729544639587,
0.018330592662096024,
-0.0023576098028570414,
-0.03336368501186371,
-0.013504590839147568,
-0.07297205924987793,
0.036213211715221405,
-0.09611982107162476,
0.057553596794605255,
0.07360386103391647,
-0.011385209858417511,
0.04655580222606659,
0.055301401764154434,
0.04566176235675812,
0.004463808611035347,
0.07680422812700272,
-0.030217677354812622,
-0.011398529633879662,
-0.06279478967189789,
0.05827506631612778,
-0.09285980463027954,
-4.00447888182498e-8,
-0.045198798179626465,
0.038318511098623276,
-0.0063296351581811905,
-0.006071640644222498,
0.014093201607465744,
-0.06892821937799454,
-0.0759769156575203,
0.10902127623558044,
-0.03435030207037926,
-0.02299741469323635,
0.028935078531503677,
0.008104652166366577,
-0.019330531358718872,
-0.06080733984708786,
0.03439456224441528,
0.058403130620718,
-0.02964201755821705,
0.047936588525772095,
0.006277736742049456,
-0.05044372007250786,
0.030896306037902832,
-0.007677342277020216,
0.0025867733638733625,
0.0367872454226017,
0.08198455721139908,
-0.032078005373477936,
-0.047790005803108215,
0.09999874979257584,
-0.03019782155752182,
0.0012426439207047224,
0.007373737171292305,
-0.06066267192363739,
0.04946689307689667,
0.007553343661129475,
-0.011941385455429554,
0.0015414963709190488,
0.01995622180402279,
-0.08502300083637238,
-0.002965455874800682,
-0.09484072774648666,
0.007622257806360722,
0.03993573412299156,
-0.01994611881673336,
0.008095293305814266,
0.0027710727881640196,
-0.006466553080826998,
-0.04972529783844948,
0.012884818017482758,
0.05880147963762283,
-0.011352429166436195,
-0.04088490828871727,
-0.08126884698867798,
0.06189005449414253,
0.05872321501374245,
0.058083515614271164,
0.022068189457058907,
0.008525492623448372,
0.024789955466985703,
0.07822425663471222,
-0.040078628808259964,
0.05105447769165039,
0.027601605281233788,
0.01119192037731409,
0.0468270517885685
] | 0
| 0
| 0
| 0
|
|
Moral Alignment for LLM Agents
|
https://openreview.net/forum?id=MeGDmZjUXy
|
[
"Elizaveta Tennant",
"Stephen Hailes",
"Mirco Musolesi"
] |
Poster
|
Decision-making agents based on pre-trained Large Language Models (LLMs) are increasingly being deployed across various domains of human activity. While their applications are currently rather specialized, several research efforts are underway to develop more generalist agents. As LLM-based systems become more agentic, their influence on human activity will grow and their transparency will decrease. Consequently, developing effective methods for aligning them to human values is vital.
The prevailing practice in alignment often relies on human preference data (e.g., in RLHF or DPO), in which values are implicit, opaque and are essentially deduced from relative preferences over different model outputs. In this work, instead of relying on human feedback, we introduce the design of reward functions that explicitly and transparently encode core human values for Reinforcement Learning-based fine-tuning of foundation agent models. Specifically, we use intrinsic rewards for the moral alignment of LLM agents.
We evaluate our approach using the traditional philosophical frameworks of Deontological Ethics and Utilitarianism, quantifying moral rewards for agents in terms of actions and consequences on the Iterated Prisoner's Dilemma (IPD) environment. We also show how moral fine-tuning can be deployed to enable an agent to unlearn a previously developed selfish strategy. Finally, we find that certain moral strategies learned on the IPD game generalize to several other matrix game environments. In summary, we demonstrate that fine-tuning with intrinsic rewards is a promising general solution for aligning LLM agents to human values, and it might represent a more transparent and cost-effective alternative to currently predominant alignment techniques.
|
AI alignment, LLM fine-tuning, moral decision-making, social dilemmas
|
We propose a framework for fine-tuning LLMs with intrinsic rewards and demonstrate that it is a promising general solution for aligning LLM agents to human values.
| 11,392
|
2410.01639
|
[
-0.05939672514796257,
-0.1020951122045517,
-0.060607001185417175,
-0.08194005489349365,
-0.013159935362637043,
0.038745179772377014,
0.011239554733037949,
-0.024121863767504692,
0.07142852246761322,
0.0710766538977623,
-0.0013270805357024074,
-0.05923538655042648,
0.015525746159255505,
0.08361773937940598,
0.05669689178466797,
-0.004692445043474436,
0.11506584286689758,
0.06163202226161957,
-0.0967063456773758,
-0.027524905279278755,
0.04881248623132706,
0.01813124120235443,
0.03278389200568199,
-0.00819989200681448,
-0.0928063839673996,
-0.049370862543582916,
0.04781563952565193,
-0.03417687490582466,
0.01609031856060028,
-0.015247226692736149,
0.04600612446665764,
0.00542474677786231,
0.09410881996154785,
-0.03619903698563576,
-0.07424966990947723,
0.10868992656469345,
-0.1122320219874382,
-0.04362892359495163,
-0.042627934366464615,
-0.03978894650936127,
-0.03474575653672218,
-0.015723438933491707,
-0.007929259911179543,
0.02921336516737938,
0.027976244688034058,
-0.010934734717011452,
-0.005285064224153757,
-0.01697346940636635,
-0.022261356934905052,
-0.06252031028270721,
-0.030694713816046715,
-0.02062305249273777,
0.009658279828727245,
0.03194643557071686,
0.005478435195982456,
-0.002393921371549368,
0.039788924157619476,
0.04598293825984001,
-0.03391559049487114,
-0.1052117571234703,
0.036113154143095016,
-0.111213818192482,
-0.03232162445783615,
0.027571596205234528,
-0.004982078913599253,
0.007433674298226833,
-0.0184280127286911,
0.07042066007852554,
-0.09451384097337723,
0.01244546938687563,
-0.02006611041724682,
-0.00010730119538493454,
0.023320309817790985,
-0.043256405740976334,
-0.04372340813279152,
0.07921338081359863,
-0.0030556952115148306,
-0.012344694696366787,
0.044265180826187134,
-0.03919297829270363,
-0.041686661541461945,
0.015131905674934387,
0.041090529412031174,
0.046134103089571,
0.09080439060926437,
-0.0006574497674591839,
0.04001782834529877,
0.03912844881415367,
0.1331787258386612,
0.04789554700255394,
-0.01135300099849701,
0.0021038386039435863,
0.02716563642024994,
-0.006511069368571043,
0.01889600232243538,
0.10672622174024582,
-0.048158612102270126,
-0.034848690032958984,
-0.10430356860160828,
0.0974537655711174,
-0.004008532036095858,
0.09964635968208313,
-0.057398878037929535,
-0.03496268019080162,
0.062213409692049026,
-0.014968793839216232,
-0.05645790323615074,
0.021212201565504074,
0.052508626133203506,
-0.10728887468576431,
-0.007914245128631592,
0.08534976840019226,
0.04178669676184654,
0.13626274466514587,
0.0668480172753334,
-0.028029121458530426,
0.06624726206064224,
-0.025974396616220474,
0.03006291203200817,
-0.04033515229821205,
0.047233808785676956,
-0.017625009641051292,
0.0341593399643898,
0.018291480839252472,
0.04370862990617752,
-0.02434193715453148,
-0.013307926245033741,
1.1645410393600277e-33,
0.031022625043988228,
0.06899478286504745,
0.045038312673568726,
0.010161332786083221,
-0.003897756105288863,
-0.03065793588757515,
0.014620437286794186,
0.0385664626955986,
-0.04783935472369194,
0.0657685250043869,
0.011363403871655464,
0.06585737317800522,
-0.03378435596823692,
0.11136659234762192,
-0.07537326961755753,
-0.004634542390704155,
-0.025043442845344543,
-0.010907727293670177,
-0.030728887766599655,
-0.02487337961792946,
0.07954713702201843,
0.03351978212594986,
0.006303636822849512,
-0.07780470699071884,
0.027782699093222618,
-0.038507189601659775,
0.033920567482709885,
-0.06774910539388657,
0.0022154701873660088,
-0.02111802063882351,
-0.05767103657126427,
0.023679204285144806,
0.06354870647192001,
0.014676261693239212,
-0.0632719025015831,
0.014569957740604877,
-0.10223043709993362,
-0.028748128563165665,
0.0533861480653286,
-0.042559247463941574,
0.018160147592425346,
0.04895704612135887,
0.028511352837085724,
0.015459858812391758,
-0.046613991260528564,
-0.04038660228252411,
0.06931247562170029,
-0.03640919178724289,
-0.046523306518793106,
-0.014998327009379864,
0.07007082551717758,
0.05815168470144272,
0.023868484422564507,
-0.0645764023065567,
-0.0763547271490097,
0.005323364399373531,
0.03171129152178764,
0.03395286947488785,
-0.10004018247127533,
-0.028108561411499977,
-0.010321388021111488,
-0.03454539552330971,
0.037868671119213104,
0.012883584946393967,
0.01426483504474163,
0.02924749255180359,
-0.04063335433602333,
0.03051961585879326,
0.043737366795539856,
-0.028293009847402573,
0.005790336057543755,
0.04944075271487236,
-0.014375057071447372,
-0.03280206397175789,
-0.029451536014676094,
0.02757873199880123,
0.04122868925333023,
-0.09652625024318695,
0.08239153027534485,
-0.0535876601934433,
-0.017979290336370468,
0.042513832449913025,
-0.04806607589125633,
-0.06267789006233215,
-0.0027078301645815372,
-0.09530118852853775,
0.01779465191066265,
-0.09610007703304291,
0.04903296008706093,
0.022256946191191673,
-0.061685118824243546,
0.004334646742790937,
0.032096873968839645,
0.04360511526465416,
0.00681923096999526,
-3.129154206945421e-33,
-0.0658508762717247,
-0.0934436023235321,
-0.038581281900405884,
0.055221132934093475,
0.00234362599439919,
0.003298046998679638,
-0.025593386963009834,
0.003566386643797159,
-0.005093789659440517,
-0.02213379740715027,
-0.06910888105630875,
-0.010870617814362049,
0.05329543352127075,
0.014954457990825176,
0.04019704833626747,
-0.024639690294861794,
-0.01316016260534525,
0.022765755653381348,
0.04552797973155975,
0.00685896584764123,
0.04083263874053955,
0.1274026781320572,
-0.14627815783023834,
0.07122951745986938,
0.04750065878033638,
0.05533790960907936,
0.04981108754873276,
0.024075070396065712,
0.03855140879750252,
0.029581107199192047,
-0.06779627501964569,
-0.03492817282676697,
-0.07622245699167252,
-0.05238266661763191,
-0.007785179652273655,
0.05095551162958145,
0.032773151993751526,
-0.022360432893037796,
-0.06939929723739624,
0.09736191481351852,
-0.03958400338888168,
-0.057040851563215256,
-0.10811257362365723,
0.010155602358281612,
-0.04622602462768555,
-0.045844998210668564,
-0.05588332191109657,
-0.03469673544168472,
0.022368667647242546,
-0.04679623991250992,
0.04333510249853134,
0.022242173552513123,
-0.024295657873153687,
0.02388899400830269,
-0.053295690566301346,
-0.03597013279795647,
-0.016876693814992905,
-0.04594477266073227,
0.09399913996458054,
-0.00008851216261973605,
-0.04952677711844444,
0.047150932252407074,
0.014345826581120491,
0.018521910533308983,
-0.08214275538921356,
0.007609797641634941,
0.009418781846761703,
0.022641243413090706,
-0.06584017723798752,
-0.011394559405744076,
0.06630953401327133,
-0.03370276838541031,
-0.009947427548468113,
-0.03611219301819801,
0.06467738747596741,
-0.005752762779593468,
0.011983590200543404,
-0.052346255630254745,
-0.0031427862122654915,
-0.0558328777551651,
0.06579065322875977,
0.03724811598658562,
0.05653978884220123,
0.036322593688964844,
-0.0473252609372139,
-0.0005327308899722993,
0.08957456052303314,
0.08721558004617691,
-0.021720970049500465,
-0.007822748273611069,
-0.02555360086262226,
0.003460397245362401,
0.005036400631070137,
0.03354064002633095,
-0.059038251638412476,
-4.861073676920569e-8,
-0.08190949261188507,
-0.0000916798016987741,
0.02919948287308216,
0.06300964951515198,
-0.04778899624943733,
-0.07950375974178314,
0.004824142903089523,
-0.017356693744659424,
-0.067388154566288,
-0.02948954701423645,
0.048070937395095825,
0.1053040400147438,
-0.010524101555347443,
-0.05557120218873024,
-0.015011570416390896,
0.0888349786400795,
-0.047416139394044876,
-0.019117407500743866,
0.0062547726556658745,
0.04454550892114639,
0.089420385658741,
0.01939893700182438,
-0.0008603190071880817,
-0.10534749925136566,
0.04415394738316536,
-0.04447982460260391,
-0.08711031079292297,
0.08374164253473282,
0.10189810395240784,
0.0930829718708992,
0.07076795399188995,
-0.006793330889195204,
0.033226724714040756,
0.03322095423936844,
0.03637633100152016,
0.00845332257449627,
-0.0735102966427803,
-0.06474172323942184,
0.043746501207351685,
0.027340009808540344,
0.027713505551218987,
0.07954739034175873,
-0.03460168093442917,
0.005975162144750357,
0.005095142405480146,
-0.021841449663043022,
-0.07921670377254486,
-0.05143501237034798,
0.006157138850539923,
-0.06228286027908325,
-0.04680782929062843,
-0.010807783342897892,
-0.04105498269200325,
0.020672019571065903,
0.07747200131416321,
-0.045010313391685486,
0.000451708328910172,
-0.025813531130552292,
0.0468270517885685,
0.02024214155972004,
0.019828982651233673,
-0.037582628428936005,
-0.043227944523096085,
-0.06845885515213013
] | 0
| 0
| 0
| 0
|
|
ColPali: Efficient Document Retrieval with Vision Language Models
|
https://openreview.net/forum?id=ogjBpZ8uSi
|
[
"Manuel Faysse",
"Hugues Sibille",
"Tony Wu",
"Bilel Omrani",
"Gautier Viaud",
"CELINE HUDELOT",
"Pierre Colombo"
] |
Poster
|
Documents are visually rich structures that convey information through text, but also figures, page layouts, tables, or even fonts. Since modern retrieval systems mainly rely on the textual information they extract from document pages to index documents -often through lengthy and brittle processes-, they struggle to exploit key visual cues efficiently. This limits their capabilities in many practical document retrieval applications such as Retrieval Augmented Generation (RAG).
To benchmark current systems on visually rich document retrieval, we introduce the Visual Document Retrieval Benchmark $\textit{ViDoRe}$, composed of various page-level retrieval tasks spanning multiple domains, languages, and practical settings.
The inherent complexity and performance shortcomings of modern systems motivate a new concept; doing document retrieval by directly embedding the images of the document pages. We release $\textit{ColPali}$, a Vision Language Model trained to produce high-quality multi-vector embeddings from images of document pages. Combined with a late interaction matching mechanism, $\textit{ColPali}$ largely outperforms modern document retrieval pipelines while being drastically simpler, faster and end-to-end trainable.
We release models, data, code and benchmarks under open licenses at https://hf.co/vidore.
|
document embeddings, vision language models, late interaction, document retrieval, information retrieval
|
Document Retrieval from page images using Vision Language Models and Late Interaction
| 11,386
|
2407.01449
|
[
-0.08891657739877701,
-0.03334973007440567,
0.02438465692102909,
0.037625741213560104,
0.07844768464565277,
-0.0189353134483099,
-0.051882512867450714,
0.024218035861849785,
0.04747352749109268,
-0.06658650189638138,
-0.02103612571954727,
0.003907918464392424,
0.07872653007507324,
0.05628010630607605,
-0.016491195186972618,
0.0009153762366622686,
0.1047806441783905,
0.09643786400556564,
-0.06815049052238464,
-0.001528122927993536,
-0.014649366028606892,
-0.010626084171235561,
0.053665511310100555,
-0.10675452649593353,
0.0314759835600853,
0.0122947096824646,
-0.022936230525374413,
-0.06801013648509979,
0.015127680264413357,
-0.05677701532840729,
-0.0004172764893155545,
0.03146969899535179,
-0.033728670328855515,
0.0758613720536232,
-0.09212202578783035,
0.06856290251016617,
-0.00509255426004529,
-0.004910316783934832,
0.03874484822154045,
-0.047233812510967255,
-0.015084589831531048,
0.006495795212686062,
0.0175864789634943,
0.05677155777812004,
0.10165941715240479,
-0.013466131873428822,
-0.0875140130519867,
-0.010236397385597229,
0.025051046162843704,
0.019132720306515694,
-0.1283881813287735,
-0.03444962203502655,
-0.009216846898198128,
-0.027623288333415985,
-0.025039492174983025,
0.05307692289352417,
0.04700056463479996,
-0.07732211798429489,
-0.011658787727355957,
-0.04688270762562752,
-0.008705894462764263,
-0.009107131510972977,
0.02278892882168293,
-0.016895189881324768,
0.002058702288195491,
-0.009107558988034725,
0.07127983868122101,
-0.0455876849591732,
0.015607358887791634,
0.010842832736670971,
-0.06877393275499344,
0.05357242748141289,
-0.0069496287032961845,
0.02619800716638565,
-0.002919075544923544,
0.06925127655267715,
-0.01634260080754757,
-0.11748513579368591,
0.024294108152389526,
-0.057973891496658325,
0.012778963893651962,
0.011226349510252476,
0.08339014649391174,
0.04888404905796051,
0.12519849836826324,
-0.03750227764248848,
0.04561232402920723,
-0.07984264194965363,
0.010087213478982449,
-0.01858232170343399,
-0.02781190164387226,
-0.035488441586494446,
0.0542890727519989,
-0.05347418412566185,
0.015533151105046272,
0.01935925893485546,
0.11532513797283173,
-0.07357988506555557,
-0.024039536714553833,
0.07822439074516296,
0.02001667208969593,
0.003924414049834013,
0.05242704972624779,
-0.04051702842116356,
0.0026065350975841284,
0.017993448302149773,
0.08281306177377701,
0.006816545967012644,
0.012655864469707012,
-0.08521324396133423,
0.0661667063832283,
-0.03922249376773834,
-0.08523544669151306,
-0.05240323767066002,
-0.042235035449266434,
-0.006749771069735289,
-0.04613596200942993,
-0.06817472726106644,
0.07227671146392822,
0.0190273467451334,
0.024270715191960335,
0.06588149070739746,
-0.033810365945100784,
-0.037132613360881805,
-0.0473678894340992,
-0.012182720936834812,
-0.006300398614257574,
-7.480392036898907e-35,
0.04277830198407173,
0.006695982534438372,
0.011964788660407066,
0.04162759333848953,
0.03889385983347893,
-0.06703665852546692,
-0.06947401165962219,
0.049025360494852066,
-0.11763305962085724,
-0.04756666719913483,
-0.06036309525370598,
0.03974512591958046,
-0.023489879444241524,
0.14145170152187347,
0.037103984504938126,
-0.01661885343492031,
-0.04314780607819557,
0.011743521317839622,
0.006408133078366518,
0.017424991354346275,
0.0260767824947834,
0.0075380997732281685,
0.039567623287439346,
-0.017874285578727722,
0.010550398379564285,
0.018083661794662476,
0.05270151048898697,
-0.06677082180976868,
-0.04437198117375374,
0.006073113065212965,
-0.054700639098882675,
0.03130640834569931,
0.08975241333246231,
-0.003026696853339672,
-0.007507148198783398,
-0.028261754661798477,
-0.007647055666893721,
-0.056301530450582504,
0.03712897002696991,
0.03723231330513954,
-0.038544926792383194,
-0.012681011110544205,
-0.002550919773057103,
-0.09138353168964386,
-0.05580690875649452,
0.004599601961672306,
-0.018890388309955597,
0.017249692231416702,
-0.045951876789331436,
-0.03255845606327057,
0.05522880330681801,
0.058590639382600784,
-0.07016075402498245,
0.031936828047037125,
0.03517865017056465,
-0.050574153661727905,
0.04293069615960121,
0.05113088712096214,
0.055440645664930344,
0.04676581174135208,
-0.0016493841540068388,
0.04969927296042442,
0.012640831060707569,
0.08925631642341614,
-0.041861969977617264,
-0.009362583979964256,
0.042369719594717026,
0.07267963886260986,
0.03240872919559479,
0.029909279197454453,
-0.04340199753642082,
0.0343385711312294,
-0.009207438677549362,
-0.09960172325372696,
0.06523217260837555,
-0.01583566516637802,
0.04398580640554428,
-0.06614851951599121,
0.006420467048883438,
-0.0008147332700900733,
-0.005185161717236042,
-0.06094120442867279,
0.01570204272866249,
-0.12501248717308044,
-0.02921411395072937,
-0.03115822747349739,
0.02004144713282585,
-0.10213202238082886,
0.035266462713479996,
-0.08021186292171478,
-0.004628578666597605,
-0.021046137437224388,
-0.008090764284133911,
-0.059691786766052246,
0.05487401410937309,
-8.218814055239577e-34,
0.06589590013027191,
-0.050318337976932526,
0.030091065913438797,
0.03272358700633049,
-0.017574401572346687,
0.008229468017816544,
0.019081629812717438,
0.005265957675874233,
-0.01618814654648304,
-0.06510621309280396,
-0.028258569538593292,
0.009075724519789219,
-0.01842118427157402,
-0.01001988910138607,
0.011863668449223042,
0.04730495810508728,
0.002823379822075367,
-0.03819919377565384,
-0.0753539502620697,
0.019639326259493828,
0.006048289127647877,
-0.010571258142590523,
0.012618252076208591,
0.030343396589159966,
-0.025555027648806572,
0.04695732519030571,
0.008436042815446854,
-0.0548102892935276,
-0.10409564524888992,
-0.03859679773449898,
0.011009818874299526,
-0.03387954458594322,
-0.0013656666269525886,
0.0070942300371825695,
-0.059954751282930374,
0.06244407966732979,
0.008746368810534477,
-0.03656039386987686,
-0.06971701979637146,
0.14447829127311707,
0.06429591774940491,
-0.030058246105909348,
-0.007856453768908978,
0.04156484454870224,
-0.060185909271240234,
-0.05312945321202278,
-0.11391647160053253,
-0.004348200745880604,
0.05690442770719528,
-0.051347434520721436,
0.00048497362877242267,
-0.0263441763818264,
-0.05563319846987724,
-0.020176781341433525,
-0.008697821758687496,
-0.07965201139450073,
-0.10310754179954529,
0.046505384147167206,
0.0028957449831068516,
0.044242098927497864,
-0.01757391355931759,
0.005807333160191774,
-0.07282964140176773,
0.011130506172776222,
0.03627724573016167,
-0.03469325974583626,
-0.07788558304309845,
-0.042161718010902405,
-0.028922703117132187,
0.003670520382001996,
0.090279720723629,
-0.061861030757427216,
0.04037664830684662,
0.05278075486421585,
0.08432774245738983,
0.05097959190607071,
0.08597434312105179,
0.024171579629182816,
0.04405532032251358,
-0.05369655042886734,
-0.010246701538562775,
-0.032122571021318436,
0.0244970191270113,
0.020716190338134766,
0.008943112567067146,
0.1146547719836235,
-0.04890308901667595,
0.05374699831008911,
0.03934402018785477,
-0.01757044717669487,
-0.03388597443699837,
0.03025003895163536,
0.0007574587943963706,
0.06818704307079315,
0.03789829835295677,
-3.5230200978730863e-8,
-0.059178464114665985,
0.047531504184007645,
-0.07923687249422073,
-0.025800632312893867,
-0.04309793934226036,
0.010393471457064152,
0.032377470284700394,
0.10178697854280472,
-0.062250278890132904,
-0.041133441030979156,
0.08991274237632751,
-0.034865330904722214,
-0.06304990500211716,
-0.0503273531794548,
0.04592196270823479,
0.05768319219350815,
0.009171591140329838,
0.031743843108415604,
0.0273861363530159,
0.007655757945030928,
0.004552866797894239,
0.06075897067785263,
0.08923795819282532,
0.00859995000064373,
-0.07941748201847076,
0.07808804512023926,
-0.030201373621821404,
0.11878807097673416,
0.057336144149303436,
-0.051248300820589066,
0.029348183423280716,
0.04902156814932823,
-0.024043630808591843,
-0.022213773801922798,
0.030636826530098915,
0.0499548576772213,
-0.05330536514520645,
-0.028965024277567863,
-0.06174085661768913,
0.01891637220978737,
0.09402826428413391,
0.057157352566719055,
0.006541096605360508,
-0.00013030070113018155,
0.03840933367609978,
0.034316107630729675,
0.05244305357336998,
-0.016150446608662605,
0.05596834793686867,
-0.042057350277900696,
-0.11533113569021225,
-0.14669156074523926,
-0.006211279425770044,
0.14105944335460663,
-0.02992171049118042,
-0.035043712705373764,
0.011234519071877003,
0.007438191678375006,
0.08444997668266296,
0.04852733761072159,
0.01329723559319973,
-0.058212585747241974,
0.004776077810674906,
0.037257228046655655
] | 0
| 0
| 0
| 0
|
|
Content-Style Learning from Unaligned Domains: Identifiability under Unknown Latent Dimensions
|
https://openreview.net/forum?id=p60Y6o85Cj
|
[
"Sagar Shrestha",
"Xiao Fu"
] |
Poster
|
Understanding identifiability of latent content and style variables from unaligned multi-domain data is essential for tasks such as
domain translation and data generation. Existing works on content-style identification were often developed under somewhat stringent conditions, e.g., that all latent components are mutually independent and that the dimensions of the content and style variables are known. We introduce a new analytical framework via cross-domain *latent distribution matching* (LDM), which establishes content-style identifiability under substantially more relaxed conditions. Specifically, we show that restrictive assumptions such as component-wise independence of the latent variables can be removed. Most notably, we prove that prior knowledge of the content and style dimensions is not necessary for ensuring identifiability, if sparsity constraints are properly imposed onto the learned latent representations. Bypassing the knowledge of the exact latent dimension has been a longstanding aspiration in unsupervised representation learning---our analysis is the first to underpin its theoretical and practical viability. On the implementation side, we recast the LDM formulation into a regularized multi-domain GAN loss with coupled latent variables. We show that the reformulation is equivalent to LDM under mild conditions---yet requiring considerably less computational resource. Experiments corroborate with our theoretical claims.
|
unsupervised learning, identifiability, unknown latent dimension
| null | 11,383
|
2411.03755
|
[
-0.028135575354099274,
-0.10442738980054855,
-0.014460419304668903,
0.040909551084041595,
0.08161360770463943,
0.08558736741542816,
0.001217672019265592,
-0.12178700417280197,
0.028490550816059113,
-0.03851177543401718,
-0.0002578567073214799,
-0.031139878556132317,
0.0587739497423172,
0.0527147576212883,
0.008045341819524765,
-0.01779775507748127,
0.08111496269702911,
0.09790661185979843,
-0.08312013000249863,
-0.0008794258465059102,
-0.012272088788449764,
0.07501527667045593,
0.01808977499604225,
-0.006458108313381672,
0.054326992481946945,
0.046892520040273666,
0.09613899141550064,
-0.0018866773461923003,
0.012055574916303158,
-0.0245840884745121,
0.07689867168664932,
0.08200856298208237,
-0.035053178668022156,
-0.021153774112462997,
-0.03243197873234749,
0.12022153288125992,
-0.0883708968758583,
0.07750346511602402,
-0.015963027253746986,
0.034119848161935806,
0.008090825751423836,
0.0301457941532135,
-0.04363295063376427,
0.07049521058797836,
-0.006822782568633556,
-0.024706628173589706,
-0.0680534839630127,
-0.08943358808755875,
-0.046717606484889984,
0.030308973044157028,
0.006782257463783026,
-0.004801013041287661,
-0.04547996446490288,
0.0625876933336258,
-0.008778135292232037,
-0.027698388323187828,
0.012557467445731163,
0.009851313196122646,
0.006281664129346609,
0.004231139086186886,
-0.04955403134226799,
-0.09120172262191772,
0.009567382745444775,
-0.05027905851602554,
0.04252297058701515,
-0.050446443259716034,
-0.0362326018512249,
0.030771467834711075,
-0.0817459374666214,
-0.003479297971352935,
-0.04604463279247284,
0.008610978722572327,
-0.09107773005962372,
0.0921940803527832,
0.010991469956934452,
0.08608885109424591,
0.04944285750389099,
-0.07826267927885056,
-0.010865289717912674,
-0.061366382986307144,
-0.00017379769997205585,
0.10341714322566986,
0.04035736992955208,
0.04979947209358215,
0.08979848027229309,
0.019992923364043236,
-0.013208884745836258,
0.033153094351291656,
-0.020667651668190956,
0.02201440930366516,
-0.03669261932373047,
0.0041427877731621265,
0.05234181135892868,
0.01363436970859766,
0.08602171391248703,
-0.038518570363521576,
0.056589145213365555,
-0.06593933701515198,
0.06569095700979233,
0.06267108023166656,
-0.03659490495920181,
0.09530460834503174,
0.04118677228689194,
0.05261732265353203,
-0.02703389711678028,
-0.08931419998407364,
0.015196365304291248,
0.06428943574428558,
0.016788063570857048,
-0.1291464865207672,
0.02212550677359104,
0.0006275650230236351,
-0.001382958609610796,
0.014159678481519222,
0.043627332895994186,
-0.10328347980976105,
0.05967497080564499,
-0.036392439156770706,
0.040233537554740906,
-0.03933984041213989,
-0.02350584603846073,
-0.01807270385324955,
0.04825030639767647,
-0.00012356607476249337,
0.01532027404755354,
-0.0376238152384758,
-0.024971526116132736,
4.953039854079759e-33,
0.06984893232584,
0.08082380890846252,
0.002483709715306759,
0.046841900795698166,
0.017328917980194092,
-0.010056483559310436,
-0.032816458493471146,
-0.05485634133219719,
0.019217759370803833,
0.051022570580244064,
0.01961354725062847,
0.04155592992901802,
-0.07538408786058426,
0.060073260217905045,
-0.0331261083483696,
0.05507083609700203,
-0.022593405097723007,
0.08204923570156097,
0.013804948888719082,
0.0031120055355131626,
0.033142298460006714,
-0.01380092091858387,
0.0006848658085800707,
-0.03360971435904503,
0.011234541423618793,
-0.023894177749753,
0.07483946532011032,
-0.09422197192907333,
-0.05117466300725937,
-0.034559670835733414,
-0.09619294852018356,
-0.01450289972126484,
0.06498584896326065,
0.022154541686177254,
0.04888427257537842,
-0.045601293444633484,
0.009087630547583103,
-0.08056474477052689,
0.008840318769216537,
-0.044388044625520706,
-0.006159006152302027,
0.012206073850393295,
0.08317726105451584,
-0.06191185489296913,
-0.11060169339179993,
0.023423777893185616,
0.08045835793018341,
-0.041322920471429825,
-0.0007903019431978464,
-0.05078698322176933,
0.0896039754152298,
0.024006009101867676,
0.04611413925886154,
-0.08780402690172195,
-0.014706489630043507,
0.013493923470377922,
-0.009636148810386658,
0.005270970053970814,
0.06165110319852829,
0.02330160140991211,
-0.030710015445947647,
0.10996001213788986,
-0.006452753208577633,
0.008181867189705372,
-0.023060183972120285,
0.012520729564130306,
-0.013331753201782703,
-0.028224529698491096,
0.11556046456098557,
-0.04134094715118408,
0.005332128144800663,
0.006578341592103243,
-0.11816506832838058,
-0.06271030008792877,
0.03138357028365135,
-0.03309885412454605,
0.003590454114601016,
-0.07443366199731827,
0.08766476809978485,
0.02131744660437107,
-0.02629907988011837,
-0.001846257713623345,
0.012130440212786198,
-0.05764124542474747,
-0.06583961844444275,
-0.034641582518815994,
0.06103278324007988,
-0.11104974895715714,
-0.01841341331601143,
-0.02192446030676365,
-0.0018529616063460708,
-0.03085235133767128,
-0.0707615315914154,
-0.01329281460493803,
0.062060222029685974,
-4.4630747068738285e-33,
-0.04003734141588211,
-0.036009807139635086,
-0.02097664773464203,
0.005187080707401037,
0.034701403230428696,
-0.04970058053731918,
-0.032224588096141815,
0.0679435282945633,
0.01977100409567356,
-0.10184309631586075,
-0.0032338176388293505,
-0.050937533378601074,
-0.058718301355838776,
-0.04940979555249214,
-0.012714998796582222,
0.09702634811401367,
-0.03170791268348694,
0.05766589939594269,
-0.028023703023791313,
0.007032740395516157,
0.04450071230530739,
0.015984369441866875,
-0.09160543233156204,
0.0581725649535656,
0.03503807634115219,
0.04972340166568756,
0.007128543686121702,
-0.012199664488434792,
-0.03707744926214218,
0.029208220541477203,
0.04183081537485123,
-0.054527975618839264,
-0.04720303788781166,
-0.013006736524403095,
-0.08498260378837585,
-0.02674884721636772,
-0.0066429986618459225,
0.05889091640710831,
-0.07333629578351974,
0.0717817172408104,
0.05710069462656975,
-0.007198621053248644,
-0.059446655213832855,
0.037846945226192474,
-0.017616022378206253,
-0.05712483450770378,
-0.04296121746301651,
0.04187322407960892,
0.09367065876722336,
0.05528943985700607,
-0.08293229341506958,
0.018179338425397873,
-0.0014601371949538589,
0.0038295923732221127,
-0.025528324767947197,
-0.059918615967035294,
0.025579318404197693,
-0.013496643863618374,
0.09128406643867493,
-0.010483422316610813,
0.019192734733223915,
-0.025362316519021988,
0.06169774755835533,
-0.034955378621816635,
0.04156913980841637,
-0.13725101947784424,
0.009972660802304745,
0.0019101829966530204,
-0.011654949747025967,
-0.050514377653598785,
0.05116896703839302,
-0.013549386523663998,
-0.03988181799650192,
-0.014881025068461895,
0.05852500721812248,
-0.07061996310949326,
0.05892585217952728,
0.04994913190603256,
-0.021846342831850052,
-0.0004575818602461368,
0.040925633162260056,
0.02150951884686947,
0.052033621817827225,
0.05066429078578949,
0.018417013809084892,
0.1479790359735489,
0.063283272087574,
-0.006318097002804279,
0.005370256956666708,
-0.016807280480861664,
0.018875867128372192,
0.049627892673015594,
-0.006933567114174366,
0.06268113106489182,
-0.10950033366680145,
-4.345541526618035e-8,
-0.11981918662786484,
-0.04269912838935852,
-0.050433285534381866,
-0.05428327992558479,
-0.033509593456983566,
0.012794261798262596,
-0.0301571786403656,
0.006441470701247454,
-0.006111538037657738,
0.024439355358481407,
0.0964861586689949,
-0.02252814918756485,
-0.0020510852336883545,
-0.08657413721084595,
-0.03851160407066345,
0.0030683428049087524,
0.03901590034365654,
0.054131682962179184,
-0.018420463427901268,
0.016869138926267624,
-0.010672072879970074,
0.07367835938930511,
-0.0019480644259601831,
-0.07247254997491837,
0.054371509701013565,
-0.06069695204496384,
-0.03600335493683815,
0.04571415111422539,
0.07296815514564514,
0.02616218477487564,
-0.04223893582820892,
0.02984573319554329,
0.025762924924492836,
-0.06239013746380806,
0.03691185265779495,
0.07221929728984833,
0.08118466287851334,
-0.0663805678486824,
-0.0335797555744648,
-0.013401499949395657,
-0.039392489939928055,
0.04527335986495018,
-0.03670145943760872,
0.018234314396977425,
0.0328819677233696,
-0.013589423149824142,
-0.007944581098854542,
-0.044167306274175644,
0.0195112656801939,
-0.032213155180215836,
0.04102751240134239,
-0.09201640635728836,
0.008043558336794376,
0.06633538007736206,
0.06380606442689896,
0.01913866586983204,
-0.01391413901001215,
0.06243186816573143,
0.05239599943161011,
-0.0007697300752624869,
-0.01448643859475851,
0.017629103735089302,
0.03022358939051628,
0.0075307488441467285
] | 0
| 0
| 0
| 0
|
|
From an LLM Swarm to a PDDL-empowered Hive: Planning Self-executed Instructions in a Multi-modal Jungle
|
https://openreview.net/forum?id=QAAsnSRwgu
|
[
"Kaustubh Vyas",
"Damien Graux",
"Yijun Yang",
"Sebastien Montella",
"Chenxin Diao",
"Wendi Zhou",
"Pavlos Vougiouklis",
"Ruofei Lai",
"Yang Ren",
"Keshuang Li",
"Jeff Z. Pan"
] |
Poster
|
In response to the call for agent-based solutions that leverage the ever-increasing capabilities of the deep models' ecosystem, we introduce a comprehensive solution for selecting appropriate models and subsequently planning a set of atomic actions to satisfy the end-users' instructions.
Our system, Hive, operates over sets of models and, upon receiving natural language instructions, schedules and executes, explainable plans of atomic actions. These actions can involve one or more of the available models to achieve the overall task, while respecting end-users specific constraints. Hive is able to plan complex chains of actions while guaranteeing explainability, using an LLM-based formal logic backbone empowered by PDDL operations. We introduce the MuSE benchmark in order to offer a comprehensive evaluation of the multi-modal capabilities of agent systems. Our findings show that our framework redefines the state-of-the-art for task selection, outperforming other competing systems that plan operations across multiple models while offering transparency guarantees while fully adhering to user constraints.
|
Deep Models, Planning, PDDL, Knowledge Graphs, Benchmark, Large Language Models
|
Introducing Hive: a powerful, explainable system for selecting models & planning atomic actions based on natural language instructions. Hive leverages PDDL to deliver complex multi-modal tasks while respecting user constraints.
| 11,377
|
2412.12839
|
[
-0.0017684040358290076,
-0.06086944043636322,
0.03337903693318367,
0.012600348331034184,
0.06443209201097488,
-0.03467399254441261,
-0.04857635498046875,
0.014538329094648361,
-0.009159698151051998,
0.047576989978551865,
-0.07801283895969391,
-0.10291896015405655,
-0.018206480890512466,
0.03794986754655838,
0.08543910831212997,
-0.028084779158234596,
0.10974287986755371,
-0.0055411080829799175,
-0.10876660794019699,
-0.08202823996543884,
0.04780549928545952,
0.007349072489887476,
0.035516656935214996,
-0.026080358773469925,
-0.037099193781614304,
0.0040561966598033905,
0.0002894909994211048,
-0.07170917838811874,
0.014228749088943005,
0.0005492622149176896,
0.003879401134327054,
0.017015544697642326,
-0.015958867967128754,
0.02444203570485115,
-0.013928723521530628,
0.04496487230062485,
-0.09774550050497055,
-0.02763666957616806,
-0.03312677890062332,
-0.118047334253788,
-0.04840005934238434,
-0.04312058538198471,
-0.08007130771875381,
0.00406888360157609,
-0.00569471949711442,
-0.017445089295506477,
0.006569435354322195,
-0.010690749622881413,
-0.009339411742985249,
0.007310257758945227,
-0.06331101804971695,
-0.022342218086123466,
0.024678748100996017,
-0.03898915275931358,
-0.011329980567097664,
0.004498870577663183,
0.051685672253370285,
-0.010646278969943523,
-0.017946815118193626,
-0.08651424944400787,
-0.04210072383284569,
-0.08876343071460724,
-0.009786000475287437,
0.0005709062679670751,
0.005396510008722544,
0.019788697361946106,
0.0007243355503305793,
0.039585039019584656,
0.05180396884679794,
0.004714112263172865,
0.0005632321117445827,
-0.031019873917102814,
0.020509058609604836,
0.02597738988697529,
-0.049604110419750214,
0.04180409759283066,
0.036254383623600006,
-0.006744884420186281,
0.009529984556138515,
-0.058774664998054504,
-0.04707924276590347,
-0.014363229274749756,
-0.021469229832291603,
0.029346566647291183,
0.08979182690382004,
0.0015492521924898028,
-0.0470714271068573,
0.03620181605219841,
0.08187428116798401,
0.06288275122642517,
-0.02117353491485119,
-0.07164208590984344,
0.05614914372563362,
-0.004217108245939016,
0.0055836564861238,
0.05868037790060043,
0.014917531982064247,
-0.11653546988964081,
-0.0542549230158329,
0.09790549427270889,
-0.0009484810871072114,
0.11290092021226883,
-0.02290213294327259,
-0.0686824694275856,
0.00872416514903307,
-0.0033448287285864353,
0.02665889449417591,
0.01910073310136795,
-0.015917789191007614,
-0.03333836421370506,
0.02471073716878891,
0.010060188360512257,
0.10374297946691513,
-0.0035344979260116816,
0.0329425223171711,
-0.007915504276752472,
0.0373820923268795,
0.014685911126434803,
0.01169058121740818,
0.04178721457719803,
0.02488263137638569,
-0.031619925051927567,
0.00906907394528389,
0.05930528789758682,
0.08429564535617828,
0.009085995145142078,
-0.07156085968017578,
-2.308719348239196e-34,
0.058624930679798126,
0.030679963529109955,
0.06278117001056671,
0.07172473520040512,
0.09721605479717255,
0.0029271619860082865,
0.041522033512592316,
0.04382787644863129,
-0.08751628547906876,
0.06095348671078682,
-0.004796869587153196,
0.003909780643880367,
-0.035377807915210724,
0.046619713306427,
0.04464052617549896,
-0.016931742429733276,
-0.03546625375747681,
0.07254598289728165,
-0.01438148133456707,
-0.02969278208911419,
0.028402337804436684,
-0.0017612283118069172,
-0.05197448283433914,
-0.018226994201540947,
0.0775613784790039,
0.0683731883764267,
0.06781043112277985,
-0.03406563773751259,
0.034270137548446655,
0.021878371015191078,
-0.06225050985813141,
-0.014573448337614536,
-0.015219553373754025,
0.10223974287509918,
0.013754801824688911,
0.015457113273441792,
-0.08479270339012146,
-0.06304112821817398,
0.039346836507320404,
-0.056559380143880844,
0.04917556047439575,
0.0032073347829282284,
-0.04511076584458351,
0.01801603101193905,
-0.04793819412589073,
-0.07827524840831757,
-0.07038372755050659,
0.063809335231781,
-0.002028737450018525,
-0.04307044297456741,
0.10426697134971619,
0.023029306903481483,
0.034992512315511703,
-0.05936559662222862,
0.034359849989414215,
-0.021930614486336708,
-0.014750850386917591,
0.015957850962877274,
0.04872816428542137,
0.11162777245044708,
0.00982196070253849,
-0.008504156023263931,
-0.0808098241686821,
0.11811240762472153,
0.024608056992292404,
-0.009980316273868084,
-0.0786767527461052,
0.10837935656309128,
0.08411851525306702,
-0.02440291829407215,
-0.04539518430829048,
0.031540799885988235,
0.07268963754177094,
0.00427666213363409,
0.012166342698037624,
0.005442718043923378,
0.07313230633735657,
-0.14796708524227142,
0.03422289714217186,
-0.008944142609834671,
0.04747715964913368,
0.044633474200963974,
-0.06928325444459915,
-0.0029776731971651316,
0.01091835182160139,
-0.06529686599969864,
-0.05296831950545311,
-0.03591710701584816,
-0.05108630657196045,
-0.0012326090363785625,
-0.021453358232975006,
-0.04336687922477722,
-0.0324583500623703,
0.03489493578672409,
-0.052348654717206955,
-4.21725679759781e-34,
0.0031778132542967796,
-0.03500032797455788,
-0.09485142678022385,
0.0028015209827572107,
0.05032592639327049,
-0.09440285712480545,
-0.027672654017806053,
-0.054370392113924026,
-0.05189823359251022,
-0.08131041377782822,
-0.07311998307704926,
-0.013536198064684868,
0.04443247243762016,
-0.02802705205976963,
0.01674286462366581,
0.001967262476682663,
0.026211783289909363,
-0.08824751526117325,
0.025403404608368874,
0.08573416620492935,
-0.03746126592159271,
0.07950879633426666,
-0.07925066351890564,
0.02638556808233261,
-0.017747322097420692,
0.04077593609690666,
-0.05961884185671806,
0.0612279511988163,
0.04363897442817688,
0.0062528494745492935,
-0.009031164459884167,
-0.039342597126960754,
0.003308177925646305,
-0.06817308068275452,
0.013891010545194149,
0.039815161377191544,
0.019353481009602547,
0.07389678061008453,
-0.07873913645744324,
0.04263503476977348,
0.024489304050803185,
-0.11616279184818268,
-0.05906103551387787,
-0.0010163491824641824,
-0.01882908120751381,
0.07816357910633087,
-0.07229502499103546,
-0.0030003245919942856,
-0.053977496922016144,
0.0269165001809597,
-0.05232475697994232,
-0.07725705951452255,
-0.01823062263429165,
-0.02317057177424431,
-0.019693920388817787,
-0.052635643631219864,
0.007921869866549969,
-0.04226880148053169,
0.052294351160526276,
0.03694191575050354,
0.004063399042934179,
0.043318454176187515,
0.06816762685775757,
0.029031192883849144,
-0.011547887697815895,
-0.01392254326492548,
-0.018816543743014336,
-0.007226301357150078,
-0.025423144921660423,
-0.017481882125139236,
0.06556657701730728,
-0.062389228492975235,
-0.032973431050777435,
0.02328529767692089,
0.09038154780864716,
0.001447753282263875,
-0.010609648190438747,
-0.08304331451654434,
0.000112322268250864,
-0.026058880612254143,
-0.03080904111266136,
0.05544731765985489,
0.029236754402518272,
0.08802555501461029,
-0.016277525573968887,
0.008204180747270584,
0.027304960414767265,
0.1918758600950241,
-0.014346904121339321,
0.08163706958293915,
-0.05458240956068039,
-0.014440888538956642,
0.007968020625412464,
0.06980738043785095,
-0.05617280304431915,
-4.4391995857040456e-8,
-0.0004896515747532248,
0.038982145488262177,
0.03739098832011223,
-0.007181121502071619,
0.06043947860598564,
-0.02690725401043892,
-0.04282654821872711,
0.016856133937835693,
0.040664542466402054,
-0.01713619753718376,
0.10252237319946289,
-0.02272762544453144,
-0.031622905284166336,
-0.0314159020781517,
0.033742789179086685,
0.0702895000576973,
0.023828046396374702,
-0.0127555588260293,
-0.04279299080371857,
-0.06519005447626114,
-0.010040626861155033,
0.06558743864297867,
-0.024019302800297737,
0.011258118785917759,
0.05793343111872673,
-0.01419922336935997,
-0.09742075949907303,
0.07905781269073486,
0.04579740762710571,
0.08964461833238602,
-0.014478330500423908,
-0.030064597725868225,
0.07912153005599976,
0.04981580004096031,
-0.021407170221209526,
0.04604337364435196,
-0.009986532852053642,
-0.06652212142944336,
0.05112669989466667,
-0.09585732966661453,
0.03413581848144531,
0.08302035927772522,
-0.0646335780620575,
0.03815971687436104,
0.017505507916212082,
0.018789028748869896,
-0.05329100787639618,
-0.10893286764621735,
0.0064184884540736675,
-0.06947440654039383,
-0.055553141981363297,
-0.024415982887148857,
-0.0491604320704937,
0.1000562310218811,
0.001113218255341053,
0.10838004946708679,
0.00441159401088953,
-0.043828655034303665,
0.07601585239171982,
0.0580560639500618,
0.04494090378284454,
0.014573191292583942,
-0.05739344656467438,
-0.012461164966225624
] | 0
| 0
| 0
| 0
|
|
ZooProbe: A Data Engine for Evaluating, Exploring, and Evolving Large-scale Training Data for Multimodal LLMs
|
https://openreview.net/forum?id=T4LtGj7us1
|
[
"Yi-Kai Zhang",
"Shiyin Lu",
"Qing-Guo Chen",
"De-Chuan Zhan",
"Han-Jia Ye"
] |
Poster
|
Multimodal Large Language Models (MLLMs) are thriving through continuous fine-tuning by LLMs. Driven by the law that "scale is everything", MLLMs expand their training sets during version iterations. In this paper, we propose a large-scale training data engine built around an evaluating-exploring-evolving (E3) loop. Evaluating the data provides insights into its characteristics. Exploring quality rules helps identify which data enhances training. Together, these processes facilitate the systematic evolution of new, high-quality data. With the E3 loop, we introduce ZooProbe, an efficient data engine for MLLMs. First, the problem of data expansion is formalized as a tree of sampling and growth. ZooProbe introduces a small-scale model *zoo* to obtain comprehensive evaluations for child datasets. From multiple perspectives, visual, textual, and multimodal models cover over 50 dimensions of intrinsic and meta attributes, such as object and topic distribution, and higher-level properties, like annotation quality and scene complexity. ZooProbe constructs based on A$^\star$ search, modeling the heuristic function as a quality estimate from data evaluation results. It dynamically explores the rule of data quality based on the model state of the *probe* datasets. Additionally, it evolves new targeted data with identified high-quality rules. We also develop an extra heuristic quality ranker with the data utilized and discarded during the expansion. Our experiments show that ZooProbe significantly breaks the scaling law in multimodal instruction fine-tuning at scales of 260$k$ and below.
ZooProbe generates high-quality data that accelerates MLLM training and enhances performance, automating the evolution of large-scale training data.
|
Multimodal Large Language Model, Training Data Engine, Deep Learning
|
ZooProbe: a new training data engine for Multimodal LLMs that features the Evaluating-Exploring-Evolving (E3) loop.
| 11,376
| null |
[
0.009960784576833248,
-0.12570323050022125,
0.003696371568366885,
0.005875705275684595,
0.1676245629787445,
-0.05610217526555061,
-0.018463457003235817,
-0.05558395013213158,
0.03361276164650917,
-0.05255022272467613,
-0.0316355861723423,
-0.10641896724700928,
0.014848846010863781,
0.036772094666957855,
0.005167087540030479,
-0.021146392449736595,
0.046230655163526535,
0.051311295479536057,
-0.1197521761059761,
-0.04129556939005852,
0.06715135276317596,
0.055486954748630524,
0.07355276495218277,
-0.04117439314723015,
0.027373304590582848,
0.08566515147686005,
-0.019595010206103325,
-0.06907233595848083,
0.0376318134367466,
-0.0437483936548233,
0.02363772504031658,
0.004180831369012594,
0.02206246182322502,
0.05042977258563042,
-0.025230783969163895,
0.03885938972234726,
-0.039274267852306366,
0.04346022382378578,
-0.038858428597450256,
0.0224397461861372,
-0.01114011462777853,
0.018602201715111732,
0.027593284845352173,
-0.06515112519264221,
0.07731029391288757,
-0.00011019447993021458,
-0.08561935275793076,
-0.07224168628454208,
0.051689330488443375,
0.02326558530330658,
-0.1291908472776413,
-0.07901939004659653,
0.009082943201065063,
0.03214733302593231,
-0.009191812016069889,
-0.03076113946735859,
0.033333200961351395,
-0.03751394897699356,
0.03076842613518238,
-0.03867963328957558,
0.0239762831479311,
-0.08202113211154938,
-0.01670137234032154,
0.004268312361091375,
-0.008116248995065689,
-0.08893436938524246,
0.008757919073104858,
-0.0040476368740201,
0.04048754647374153,
0.020277580246329308,
-0.07278664410114288,
0.057499874383211136,
0.015180262736976147,
0.07966487854719162,
-0.017941275611519814,
0.10605733096599579,
0.05005156248807907,
0.004116550553590059,
0.0851968377828598,
-0.13248299062252045,
-0.026474198326468468,
-0.04091651365160942,
0.04389667510986328,
-0.018864106386899948,
0.099175363779068,
-0.0033740736544132233,
0.02780144289135933,
0.01377984881401062,
-0.04136914387345314,
0.0610458105802536,
-0.0474935844540596,
-0.01020200364291668,
0.04113645851612091,
-0.0015917763812467456,
0.0022552995942533016,
0.06637711822986603,
0.03370711952447891,
-0.13824839890003204,
-0.016668017953634262,
0.09423648566007614,
-0.021420227363705635,
0.089571513235569,
0.011515326797962189,
-0.03132512420415878,
0.009581543505191803,
-0.060924213379621506,
0.04735231027007103,
0.03673796355724335,
-0.016736118122935295,
-0.06727147102355957,
0.0036233782302588224,
0.048200443387031555,
0.021498166024684906,
-0.011688325554132462,
0.08172214031219482,
-0.017503922805190086,
0.07327788323163986,
0.004587199073284864,
0.00821194238960743,
0.0631338506937027,
0.024069946259260178,
-0.013640427961945534,
-0.018856557086110115,
-0.03951277956366539,
-0.004963599145412445,
-0.03489517420530319,
-0.08259670436382294,
8.36585921497783e-34,
0.05951642617583275,
0.0011580615537241101,
0.0360729955136776,
0.10355845838785172,
0.07196362316608429,
-0.007470122072845697,
-0.025476541370153427,
0.025059513747692108,
-0.04336503893136978,
-0.004102481994777918,
-0.0769713893532753,
0.05404242128133774,
-0.08076818287372589,
0.069083072245121,
0.02464161440730095,
-0.000048032325139502063,
-0.08842065185308456,
0.027967261150479317,
-0.007486261893063784,
0.04367506131529808,
0.03293585404753685,
-0.05863602086901665,
0.04149090498685837,
0.0038351155817508698,
0.04783262312412262,
0.03576472029089928,
0.08881020545959473,
-0.060046494007110596,
-0.05636146664619446,
0.032670244574546814,
-0.11035836488008499,
-0.05199331045150757,
0.04516807571053505,
-0.0040916502475738525,
0.03128952160477638,
-0.0579264834523201,
-0.02281343750655651,
-0.021727098152041435,
0.04627956449985504,
0.03988949581980705,
-0.029909608885645866,
-0.07246991991996765,
-0.062135130167007446,
-0.024246612563729286,
-0.08571600914001465,
0.002439697040244937,
0.05993296951055527,
0.01210353896021843,
-0.031559742987155914,
-0.01663070172071457,
0.09138772636651993,
-0.03238658607006073,
0.01048495527356863,
-0.02982516959309578,
-0.05739200860261917,
0.03983772546052933,
0.01755654253065586,
0.05162620171904564,
0.015481136739253998,
-0.03253401070833206,
-0.01016932912170887,
0.0731378123164177,
-0.02609795331954956,
0.07307154685258865,
0.042580753564834595,
0.01703277789056301,
0.029707007110118866,
0.06913350522518158,
0.05664544180035591,
-0.05494830384850502,
-0.01128050684928894,
0.061247073113918304,
0.002565272618085146,
-0.05509375408291817,
0.029283782467246056,
-0.016056440770626068,
0.021593013778328896,
-0.058611828833818436,
0.052421361207962036,
0.042845651507377625,
0.00008878337393980473,
0.02743031270802021,
0.018651900812983513,
-0.08656822890043259,
-0.0005040818941779435,
-0.08957763016223907,
0.023286769166588783,
-0.08877968788146973,
-0.02527492679655552,
-0.04933727905154228,
-0.04076749086380005,
-0.035751745104789734,
0.010617890395224094,
-0.054300952702760696,
0.033887408673763275,
-7.956137903933836e-34,
0.02564639411866665,
-0.024688787758350372,
-0.012421761639416218,
0.060425277799367905,
-0.036822110414505005,
-0.06517861783504486,
-0.039760954678058624,
0.0695122703909874,
-0.016857024282217026,
-0.06405910104513168,
-0.028126228600740433,
-0.05904584005475044,
0.04147813096642494,
-0.04316730052232742,
0.060828860849142075,
-0.02211872488260269,
-0.021768735721707344,
-0.04237259551882744,
-0.030601195991039276,
0.0504692867398262,
0.008213607594370842,
0.0736444965004921,
-0.07896856218576431,
0.03432995080947876,
-0.08225756138563156,
0.04927591234445572,
-0.09204314649105072,
0.017552178353071213,
0.01840803027153015,
-0.02620987966656685,
0.01577252894639969,
-0.01568857952952385,
0.008842382580041885,
-0.03767130896449089,
-0.04049338027834892,
0.011826456524431705,
0.09051170945167542,
-0.08206003904342651,
-0.008580696769058704,
0.09679561853408813,
0.08433114737272263,
-0.03143132105469704,
-0.055581796914339066,
-0.04276679828763008,
-0.0066132983192801476,
-0.061103735119104385,
-0.04919970780611038,
0.05416721850633621,
-0.00512529956176877,
0.04864197224378586,
0.008579070679843426,
-0.03481823205947876,
-0.06693796068429947,
0.0016681044362485409,
-0.0032741231843829155,
-0.0868445485830307,
-0.05144181475043297,
-0.03868944197893143,
0.020294418558478355,
0.02848712168633938,
-0.06336108595132828,
-0.0784904882311821,
-0.036310452967882156,
-0.005584339611232281,
-0.029047971591353416,
-0.03745073080062866,
-0.01745222695171833,
-0.028506463393568993,
0.019100934267044067,
-0.021591398864984512,
0.05517323315143585,
-0.06953675299882889,
-0.08472923189401627,
-0.01119689829647541,
-0.006161341443657875,
0.031119350343942642,
0.04622611775994301,
-0.02423311397433281,
0.06302283704280853,
-0.11032368242740631,
0.013444297015666962,
-0.02207661233842373,
0.05305927246809006,
0.05436524748802185,
0.10444401204586029,
0.1141505315899849,
0.003328792518004775,
0.1085098385810852,
0.03933774679899216,
0.05653456971049309,
-0.02602568082511425,
-0.02256052754819393,
-0.003783914726227522,
0.07463593780994415,
0.005656021647155285,
-3.812427706861854e-8,
-0.062179211527109146,
0.03950534388422966,
0.0022508082911372185,
0.05557643622159958,
0.027692917734384537,
-0.00953314732760191,
-0.028186753392219543,
0.10066543519496918,
0.04695817828178406,
0.016625547781586647,
0.08358732610940933,
0.019932661205530167,
-0.02096201293170452,
-0.016877003014087677,
0.029983002692461014,
0.06769859045743942,
-0.035512857139110565,
0.03687173128128052,
0.012564631178975105,
0.02384556643664837,
0.027128323912620544,
0.04759775847196579,
0.06916694343090057,
-0.027343211695551872,
0.07558228820562363,
-0.09385626018047333,
-0.022576954215765,
0.055803082883358,
0.0020041537936776876,
-0.06120980158448219,
-0.05791514366865158,
-0.015414792113006115,
-0.036687299609184265,
0.01831822097301483,
0.0325356125831604,
0.016348175704479218,
-0.0025532639119774103,
-0.04319315776228905,
-0.036086782813072205,
-0.03432052209973335,
0.07272646576166153,
0.1033029705286026,
-0.03127319738268852,
0.013758729211986065,
-0.006675954442471266,
0.01925060525536537,
-0.05623394623398781,
-0.07338177412748337,
-0.009413884952664375,
-0.07798245549201965,
-0.07146800309419632,
-0.05110985040664673,
0.007354648783802986,
0.05774770677089691,
0.0234999842941761,
0.0008494083303958178,
-0.0435369499027729,
0.06400369107723236,
0.09672579169273376,
0.11082761734724045,
0.008968805894255638,
0.025388583540916443,
-0.02593366988003254,
0.019128991290926933
] | 0
| 0
| 0
| 0
|
|
Wicked Oddities: Selectively Poisoning for Effective Clean-Label Backdoor Attacks
|
https://openreview.net/forum?id=1Z3C49JQVf
|
[
"Nguyen Hung-Quang",
"Ngoc-Hieu Nguyen",
"The-Anh Ta",
"Thanh Nguyen-Tang",
"Kok-Seng Wong",
"Hoang Thanh-Tung",
"Khoa D Doan"
] |
Poster
|
Deep neural networks are vulnerable to backdoor attacks, a type of adversarial attack that poisons the training data to manipulate the behavior of models trained on such data.
Clean-label attacks are a more stealthy form of backdoor attacks that can perform the attack without changing the labels of poisoned data.
Early works on clean-label attacks added triggers to a random subset of the training set, ignoring the fact that samples contribute unequally to the attack's success. This results in high poisoning rates and low attack success rates.
To alleviate the problem, several supervised learning-based sample selection strategies have been proposed.
However, these methods assume access to the entire labeled training set and require training, which is expensive and may not always be practical.
This work studies a new and more practical (but also more challenging) threat model where the attacker only provides data for the target class (e.g., in face recognition systems) and has no knowledge of the victim model or any other classes in the training set.
We study different strategies for selectively poisoning a small set of training samples in the target class to boost the attack success rate in this setting.
Our threat model poses a serious threat in training machine learning models with third-party datasets, since the attack can be performed effectively with limited information. Experiments on benchmark datasets illustrate the effectiveness of our strategies in improving clean-label backdoor attacks.
|
backdoor attack, data selection
|
We propose a strategy that select data to poison to improve the success rate of clean label backdoor attacks.
| 11,375
|
2407.10825
|
[
-0.07078180462121964,
-0.024660348892211914,
0.029682138934731483,
0.06168545037508011,
0.1326696276664734,
0.04497937485575676,
0.05532052367925644,
-0.05523398891091347,
-0.0017096524825319648,
-0.07023593038320541,
-0.020548531785607338,
-0.0775589644908905,
0.04831831529736519,
-0.03204737976193428,
-0.009020634926855564,
-0.016262508928775787,
0.0845840647816658,
0.09395817667245865,
-0.02444448508322239,
0.05881580710411072,
-0.06041008234024048,
0.03830534592270851,
0.04637564718723297,
0.01539377961307764,
-0.026988552883267403,
-0.012078150175511837,
0.04342739284038544,
-0.04839545488357544,
-0.03623584285378456,
-0.043051090091466904,
0.056446373462677,
-0.05643823370337486,
-0.024270176887512207,
-0.019598906859755516,
-0.06840215623378754,
-0.009061450138688087,
-0.055817458778619766,
0.04253634810447693,
0.053889062255620956,
0.0005431746249087155,
-0.003805275307968259,
0.029448064044117928,
-0.022424906492233276,
0.046099673956632614,
-0.03098157048225403,
-0.015004638582468033,
0.051391806453466415,
-0.002670023124665022,
0.021643204614520073,
-0.03612225130200386,
-0.060414306819438934,
-0.022251058369874954,
-0.0535104013979435,
0.03616318479180336,
-0.0874120369553566,
-0.1214125007390976,
0.02987736463546753,
-0.012010278180241585,
-0.06376014649868011,
0.08566614240407944,
-0.026040682569146156,
-0.053124357014894485,
0.005913255270570517,
-0.05274008959531784,
-0.03237249329686165,
-0.023733830079436302,
0.004888409748673439,
0.06730502098798752,
0.012470554560422897,
0.05237383395433426,
0.09489128738641739,
0.09726854413747787,
-0.0784977525472641,
-0.013726100325584412,
0.00590335251763463,
0.1001957356929779,
-0.006247315090149641,
-0.0034905504435300827,
0.07696131616830826,
-0.05644840747117996,
0.025982100516557693,
-0.012860503979027271,
0.06826400756835938,
0.024876955896615982,
0.11409914493560791,
0.01368418987840414,
-0.07860703021287918,
-0.006185085047036409,
0.07655096799135208,
0.032009899616241455,
-0.018359815701842308,
0.0260591022670269,
0.08804900199174881,
0.005905245430767536,
0.06674019992351532,
0.039600130170583725,
-0.03863215446472168,
-0.03822102025151253,
-0.020103728398680687,
0.08128076046705246,
-0.10077249258756638,
-0.017538437619805336,
-0.012464563362300396,
-0.01510106585919857,
0.06856505572795868,
0.00006677334749838337,
0.0455327033996582,
-0.015834661200642586,
0.04916481301188469,
-0.09051965177059174,
0.037180956453084946,
0.025964176282286644,
-0.04435623809695244,
-0.043011121451854706,
0.060716092586517334,
-0.019522234797477722,
0.02484104409813881,
-0.029922502115368843,
-0.0648922324180603,
0.07679243385791779,
-0.037922393530607224,
0.01609746180474758,
-0.00764629477635026,
0.007989478297531605,
-0.02963212877511978,
-0.00991292204707861,
-0.0832861065864563,
2.3452019633606677e-33,
0.005646561272442341,
0.04081423580646515,
-0.02732503041625023,
-0.025705426931381226,
0.008349369280040264,
-0.06133197620511055,
-0.037636466324329376,
0.07083988189697266,
0.05688484385609627,
0.09626194089651108,
-0.027830341830849648,
-0.08804892748594284,
-0.004433234687894583,
0.09924083203077316,
0.06432504206895828,
0.07355686277151108,
0.024322375655174255,
0.06159759685397148,
-0.04470040649175644,
-0.07489188015460968,
0.002376717049628496,
-0.03366536647081375,
0.007388255558907986,
-0.03683217987418175,
-0.019191063940525055,
0.015290438197553158,
-0.022643987089395523,
0.08152622729539871,
-0.0021691631991416216,
0.0017227445496246219,
-0.06674499064683914,
-0.017003901302814484,
0.06826508790254593,
0.02232333831489086,
0.009879685938358307,
-0.011593818664550781,
-0.07453896850347519,
0.023591259494423866,
-0.03207091987133026,
0.021385783329606056,
-0.022366423159837723,
0.06952563673257828,
0.012669926509261131,
0.01321110688149929,
0.0515456348657608,
-0.004199886228889227,
-0.09174991399049759,
-0.006353611592203379,
-0.0539555698633194,
-0.06542503833770752,
0.013887613080441952,
-0.01310204342007637,
-0.027649611234664917,
-0.009069029241800308,
-0.11404406279325485,
0.03352068364620209,
0.010132805444300175,
0.05917220935225487,
-0.012236462906002998,
0.08705835789442062,
-0.017676405608654022,
0.020873049274086952,
-0.055711839348077774,
0.06768348813056946,
-0.031590066850185394,
-0.0024275912437587976,
0.020464856177568436,
-0.061516184359788895,
0.01816299557685852,
-0.05425150692462921,
0.00749873835593462,
0.04811195656657219,
-0.05684532970190048,
-0.05084208771586418,
-0.02360393851995468,
-0.01173313520848751,
0.1148843839764595,
0.01225417573004961,
0.019266244024038315,
0.04630552604794502,
0.02568504773080349,
-0.06434506922960281,
0.0076642343774437904,
0.011268392205238342,
-0.09086010605096817,
0.04237472638487816,
0.014819249510765076,
-0.0806337222456932,
0.02154216170310974,
0.10117022693157196,
-0.06281650066375732,
-0.03211564943194389,
-0.08394750207662582,
-0.008717557415366173,
-0.04415540024638176,
-2.388504420180076e-33,
-0.036390189081430435,
0.03315838426351547,
-0.02959471009671688,
-0.008534436114132404,
-0.05009867250919342,
0.004050829913467169,
0.012114267796278,
0.09396030008792877,
-0.006558381486684084,
-0.0005588501808233559,
-0.002125333994626999,
-0.01344615500420332,
0.08099410682916641,
-0.024166587740182877,
0.0011144119780510664,
0.005786423571407795,
-0.0066450852900743484,
0.04820207133889198,
-0.05996548384428024,
0.04134101793169975,
0.016516288742423058,
0.11561224609613419,
-0.05702276900410652,
0.008627157658338547,
-0.01856270805001259,
-0.027657771483063698,
0.0016347465571016073,
0.08346084505319595,
0.044232163578271866,
-0.005233493633568287,
-0.06894322484731674,
0.008638331666588783,
0.04149659350514412,
0.015249108895659447,
-0.061542730778455734,
0.09143123775720596,
0.09383277595043182,
0.016505558043718338,
-0.0453052818775177,
0.09209053963422775,
0.05366448685526848,
0.004456717055290937,
-0.11425917595624924,
0.03996586427092552,
-0.03143051266670227,
-0.07370790839195251,
-0.0202676709741354,
0.008155489340424538,
0.005082241725176573,
-0.01588193140923977,
0.002969915047287941,
-0.034370776265859604,
-0.07500200718641281,
0.07590002566576004,
-0.06335737556219101,
-0.08685145527124405,
0.02019091695547104,
-0.026107775047421455,
0.08902014046907425,
0.06977282464504242,
-0.06430603563785553,
0.019346507266163826,
-0.03335597366094589,
0.001939108595252037,
-0.025231536477804184,
-0.0732818990945816,
-0.03510661795735359,
0.06298109889030457,
0.057795509696006775,
-0.04061242192983627,
0.05982179567217827,
0.0226183719933033,
0.030867241322994232,
0.03738054633140564,
-0.1127372682094574,
-0.043632443994283676,
-0.03020888939499855,
-0.0580500066280365,
-0.007765598129481077,
0.046091560274362564,
0.07935776561498642,
-0.07783149182796478,
-0.017192121595144272,
0.05114247649908066,
0.09109675884246826,
0.03407929092645645,
0.07427368313074112,
0.017841314896941185,
-0.002745559439063072,
-0.028898196294903755,
-0.04601544886827469,
-0.02874302864074707,
0.03663063049316406,
0.05765007808804512,
0.005542730446904898,
-4.4708702517937127e-8,
-0.0198272243142128,
0.049489278346300125,
0.06698131561279297,
0.020700059831142426,
0.035861313343048096,
0.02891257219016552,
-0.10038599371910095,
0.08945231884717941,
-0.032532043755054474,
-0.03477862477302551,
0.08789234608411789,
0.06040534749627113,
-0.02359749935567379,
-0.09077154099941254,
-0.05803621932864189,
0.051325105130672455,
-0.009691650979220867,
0.04839317873120308,
-0.01255450677126646,
0.02558383159339428,
-0.02490670420229435,
0.010669667273759842,
0.08505909144878387,
-0.059560589492321014,
0.08858458697795868,
-0.1099439263343811,
-0.018341707065701485,
0.0344589464366436,
0.04026895761489868,
0.0599660649895668,
-0.0770067647099495,
-0.032672956585884094,
0.021533053368330002,
0.024415694177150726,
0.03686337545514107,
0.1123771145939827,
-0.05411961302161217,
-0.11275254935026169,
-0.016167907044291496,
0.00924949161708355,
-0.049697503447532654,
0.008323471993207932,
-0.027117984369397163,
-0.03632725030183792,
-0.04390561953186989,
0.0029872742015868425,
0.023292550817131996,
-0.09494297951459885,
0.08189418911933899,
-0.02166435867547989,
-0.009149136953055859,
-0.05825497955083847,
0.02935801073908806,
0.06152325123548508,
0.005376010201871395,
-0.01339943241328001,
-0.025110267102718353,
-0.033906638622283936,
-0.031674228608608246,
0.08829522132873535,
0.06399001181125641,
0.02094242163002491,
-0.02362501248717308,
-0.044301506131887436
] | 0
| 0
| 0
| 0
|
|
Quantum (Inspired) D2-sampling with Applications
|
https://openreview.net/forum?id=tDIL7UXmSS
|
[
"Poojan Chetan Shah",
"Ragesh Jaiswal"
] |
Poster
|
$D^2$-sampling is a fundamental component of sampling-based clustering algorithms such as $k$-means++.
Given a dataset $V \subset \mathbb{R}^d$ with $N$ points and a center set $C \subset \mathbb{R}^d$, $D^2$-sampling refers to picking a point from $V$ where the sampling probability of a point is proportional to its squared distance from the nearest center in $C$.
The popular $k$-means++ algorithm is simply a $k$-round $D^2$-sampling process, which runs in $O(Nkd)$ time and gives $O(\log{k})$-approximation in expectation for the $k$-means problem.
In this work, we give a quantum algorithm for (approximate) $D^2$-sampling in the QRAM model that results in a quantum implementation of $k$-means++ with a running time $\tilde{O}(\zeta^2 k^2)$.
Here $\zeta$ is the aspect ratio ( i.e., largest to smallest interpoint distance) and $\tilde{O}$ hides polylogarithmic factors in $N, d, k$.
It can be shown through a robust approximation analysis of $k$-means++ that the quantum version preserves its $O(\log{k})$ approximation guarantee.
Further, we show that our quantum algorithm for $D^2$-sampling can be dequantized using the sample-query access model of Tang (PhD Thesis, Ewin Tang, University of Washington, 2023). This results in a fast quantum-inspired classical implementation of $k$-means++, which we call QI-$k$-means++, with a running time $O(Nd) + \tilde{O}(\zeta^2k^2d)$, where the $O(Nd)$ term is for setting up the sample-query access data structure.
Experimental investigations show promising results for QI-$k$-means++ on large datasets with bounded aspect ratio.
Finally, we use our quantum $D^2$-sampling with the known $ D^2$-sampling-based classical approximation scheme
to obtain the first quantum approximation scheme for the $k$-means problem with polylogarithmic running time dependence on $N$.
|
quantum, clustering, k-means++
| null | 11,362
| null |
[
-0.05472375452518463,
0.017538845539093018,
-0.0096934475004673,
-0.019993817433714867,
0.011931001208722591,
-0.01253423374146223,
-0.02276163175702095,
-0.017509132623672485,
0.09254234284162521,
-0.017545368522405624,
-0.024150004610419273,
-0.01265113614499569,
0.03231372684240341,
-0.008589144796133041,
0.003225209191441536,
-0.03452702611684799,
0.12959414720535278,
0.01266543474048376,
-0.009465247392654419,
-0.0053499434143304825,
0.018299587070941925,
-0.028067801147699356,
0.08254791796207428,
-0.03075055219233036,
0.07228732854127884,
-0.01953953318297863,
0.054328665137290955,
-0.008371647447347641,
0.0755600705742836,
0.013765557669103146,
0.037506893277168274,
0.09179595857858658,
0.017015952616930008,
-0.06747916340827942,
-0.003977846819907427,
-0.032450225204229355,
0.020188024267554283,
0.044232916086912155,
-0.02501559443771839,
0.01726529374718666,
0.035441216081380844,
0.12373649328947067,
-0.005294592585414648,
0.05753786116838455,
0.007239676080644131,
0.09028429538011551,
0.030260561034083366,
0.0289923045784235,
-0.02666037529706955,
-0.09804831445217133,
0.009565268643200397,
0.0004770950763486326,
-0.01936948671936989,
0.018110480159521103,
0.0009238097118213773,
-0.02572142332792282,
-0.021597780287265778,
0.009546460583806038,
0.0751044973731041,
-0.03564479574561119,
0.07895896583795547,
-0.031110631301999092,
0.010058282874524593,
-0.007709786761552095,
0.011498559266328812,
-0.069722481071949,
0.024027235805988312,
0.026886049658060074,
0.028957420960068703,
-0.007452992256730795,
-0.03533993288874626,
0.08291449397802353,
-0.059299029409885406,
0.03806570544838905,
0.037244439125061035,
0.010935695841908455,
-0.018244922161102295,
-0.026278769597411156,
0.01482190378010273,
-0.0503280945122242,
0.01751723699271679,
-0.044449593871831894,
0.015362301841378212,
0.004022352397441864,
0.01136130653321743,
0.03495452180504799,
-0.033977434039115906,
0.011166765354573727,
0.06047236919403076,
-0.11364340037107468,
-0.07361104339361191,
-0.019777104258537292,
-0.08905996382236481,
-0.10896573960781097,
-0.028549525886774063,
-0.06392744928598404,
0.04873422533273697,
0.06719903647899628,
0.07637625187635422,
0.0360817052423954,
0.0941104143857956,
0.07251798361539841,
-0.005563779268413782,
-0.08716146647930145,
0.07920786738395691,
-0.06501064449548721,
0.05147809907793999,
0.01946256123483181,
0.043804775923490524,
-0.03774143382906914,
0.025798795744776726,
0.00003298690353403799,
0.0009502234752289951,
0.04596841707825661,
-0.0015154449502006173,
-0.037889283150434494,
0.13041572272777557,
0.034682635217905045,
-0.01584046147763729,
-0.010711205191910267,
-0.021628515794873238,
0.00039626628858968616,
-0.013665189035236835,
-0.02010144479572773,
0.027091188356280327,
0.018465038388967514,
-0.0684957355260849,
-9.233839651850199e-34,
0.04901915788650513,
0.06176167353987694,
0.06634757667779922,
-0.010761001147329807,
0.0013103904202580452,
0.00013542156375478953,
-0.005821754224598408,
-0.05266684293746948,
0.025588618591427803,
-0.004991029389202595,
0.016334963962435722,
-0.04190772399306297,
0.048977579921483994,
-0.01843656413257122,
0.09963051229715347,
-0.02009173110127449,
-0.039949338883161545,
-0.0345725491642952,
0.005957786459475756,
0.015585036017000675,
0.07796327769756317,
-0.017756562680006027,
-0.001443448243662715,
0.06364552676677704,
-0.040036700665950775,
-0.002993078902363777,
0.03089134208858013,
0.006956214085221291,
-0.008745439350605011,
0.01618371717631817,
-0.0598461776971817,
0.046436186879873276,
-0.0894460529088974,
-0.029825136065483093,
-0.00593876838684082,
0.01721041649580002,
-0.011213076300919056,
-0.02250545471906662,
-0.010988665744662285,
-0.088372141122818,
-0.0067235855385661125,
-0.011502368375658989,
0.09089706838130951,
-0.04609417915344238,
-0.07004456967115402,
0.005185321439057589,
0.0855676606297493,
0.005033009685575962,
0.02423623576760292,
-0.06472522020339966,
0.01482713595032692,
-0.06097158044576645,
-0.020349709317088127,
0.00011518364772200584,
0.05518936738371849,
0.031742773950099945,
0.07902151346206665,
-0.06365039944648743,
0.019678790122270584,
0.09778392314910889,
-0.061942506581544876,
-0.03741771727800369,
-0.059418462216854095,
0.020815245807170868,
-0.013991782441735268,
-0.03424050658941269,
-0.005942055955529213,
-0.009287501685321331,
-0.004080235492438078,
0.11045953631401062,
-0.0031206700950860977,
0.05094706267118454,
-0.043352458626031876,
-0.011737422086298466,
0.023235462605953217,
-0.058285221457481384,
0.024676382541656494,
-0.07092387229204178,
-0.0014973437646403909,
-0.014408310875296593,
-0.030516821891069412,
-0.12084319442510605,
-0.08525515347719193,
-0.07835575193166733,
0.022529756650328636,
-0.10180162638425827,
-0.013683910481631756,
-0.09948607534170151,
-0.08327610790729523,
-0.06472358107566833,
-0.07018695026636124,
-0.0323665514588356,
0.02688177488744259,
0.06665714085102081,
-0.0669187679886818,
1.8297059742509648e-34,
-0.023472193628549576,
-0.017313212156295776,
0.04331517964601517,
0.08429031819105148,
0.05877247452735901,
-0.04272516444325447,
-0.015232635661959648,
-0.020046144723892212,
0.05060679838061333,
-0.06293099373579025,
-0.0840999186038971,
0.0017762501956894994,
0.08825033158063889,
0.09753325581550598,
0.04588628560304642,
0.05165264382958412,
0.018232811242341995,
-0.01720963604748249,
0.028540724888443947,
-0.001437784289009869,
0.013842117041349411,
-0.10707822442054749,
-0.003924139309674501,
-0.027535486966371536,
-0.046651508659124374,
0.10751865804195404,
-0.022338345646858215,
-0.0013957541668787599,
-0.0411047525703907,
-0.021754875779151917,
-0.037602901458740234,
-0.11097294837236404,
-0.09870009124279022,
-0.06466704607009888,
-0.10822347551584244,
0.03373228386044502,
0.06276395916938782,
0.015997394919395447,
-0.04650033637881279,
-0.017059028148651123,
-0.05924908444285393,
-0.02334829419851303,
-0.026355063542723656,
0.0469236858189106,
0.03562537953257561,
0.018234355375170708,
0.007242382969707251,
-0.013687881641089916,
-0.08384770154953003,
0.038716938346624374,
-0.032585520297288895,
0.053310733288526535,
0.005147022195160389,
0.21271362900733948,
-0.012752131558954716,
0.0032435592729598284,
-0.08678645640611649,
0.1080937534570694,
0.07208287715911865,
0.012177912518382072,
-0.033717576414346695,
-0.09097166359424591,
-0.0625816211104393,
0.08652099967002869,
-0.021530039608478546,
-0.05417601019144058,
0.03580782562494278,
0.06324079632759094,
-0.03864600881934166,
0.08237848430871964,
-0.032284367829561234,
-0.009916488081216812,
0.043012723326683044,
0.018214352428913116,
-0.12687522172927856,
-0.021226387470960617,
0.07568301260471344,
-0.03831740468740463,
0.0175378005951643,
0.008875950239598751,
-0.022606560960412025,
0.0160736795514822,
-0.018422508612275124,
0.008940842933952808,
0.058570653200149536,
-0.02125384286046028,
0.027904590591788292,
0.031197523698210716,
0.00611005537211895,
-0.01573229767382145,
0.04322238266468048,
0.007317833136767149,
0.04527723416686058,
0.04903305321931839,
0.008145061321556568,
-4.6391352981345335e-8,
-0.06274089962244034,
-0.07245831936597824,
-0.028600111603736877,
-0.049780748784542084,
0.09222874790430069,
-0.03679778426885605,
-0.02678791806101799,
0.09011774510145187,
-0.08764182031154633,
-0.09858718514442444,
0.02511245757341385,
0.00018194838776253164,
-0.029570013284683228,
-0.005214381963014603,
0.04505029693245888,
0.08650679141283035,
0.0400690920650959,
-0.03462929278612137,
0.0075532374903559685,
0.02801700308918953,
-0.05132425203919411,
0.012505416758358479,
0.021699169650673866,
-0.05582965537905693,
-0.09102016687393188,
0.0431676022708416,
0.02918158657848835,
-0.025255432352423668,
-0.005502121988683939,
-0.0034244763664901257,
-0.0310906320810318,
0.00846110936254263,
0.052666544914245605,
0.022821474820375443,
-0.013882603496313095,
-0.04497336968779564,
-0.09135371446609497,
-0.04603906348347664,
0.0017700721509754658,
-0.04403049126267433,
0.006365006789565086,
0.06749942898750305,
0.008466091938316822,
0.023438025265932083,
0.11365875601768494,
0.06339899450540543,
0.055975303053855896,
0.04146365076303482,
0.027792222797870636,
0.050190284848213196,
-0.018545906990766525,
-0.014594134874641895,
0.0059525188989937305,
-0.04740295559167862,
-0.017654310911893845,
0.008830584585666656,
-0.07670578360557556,
-0.07772381603717804,
0.11647849529981613,
-0.029447879642248154,
0.051203154027462006,
-0.03234095498919487,
-0.05222013220191002,
0.04596837982535362
] | 0
| 0
| 0
| 0
|
|
Non-Adversarial Inverse Reinforcement Learning via Successor Feature Matching
|
https://openreview.net/forum?id=LvRQgsvd5V
|
[
"Arnav Kumar Jain",
"Harley Wiltzer",
"Jesse Farebrother",
"Irina Rish",
"Glen Berseth",
"Sanjiban Choudhury"
] |
Poster
|
In inverse reinforcement learning (IRL), an agent seeks to replicate expert demonstrations through interactions with the environment.
Traditionally, IRL is treated as an adversarial game, where an adversary searches over reward models, and a learner optimizes the reward through repeated RL procedures.
This game-solving approach is both computationally expensive and difficult to stabilize.
In this work, we propose a novel approach to IRL by _direct policy search_:
by exploiting a linear factorization of the return as the inner product of successor features and a reward vector, we design an IRL algorithm by policy gradient descent on the gap between the learner and expert features.
Our non-adversarial method does not require learning an explicit reward function and can be solved seamlessly with existing RL algorithms.
Remarkably, our approach works in state-only settings without expert action labels, a setting which behavior cloning (BC) cannot solve.
Empirical results demonstrate that our method learns from as few as a single expert demonstration and achieves improved performance on various control tasks.
|
Inverse Reinforcement Learning, Imitation Learning, Successor Features
| null | 11,360
|
2411.07007
|
[
-0.12367302924394608,
-0.025430532172322273,
0.037895482033491135,
0.046029578894376755,
0.02657615952193737,
-0.023699788376688957,
0.014912319369614124,
-0.03400509059429169,
-0.03802718594670296,
0.037501756101846695,
-0.027698734775185585,
-0.07450824975967407,
0.037161875516176224,
0.0087313586845994,
-0.02552860975265503,
0.1083146408200264,
0.13445883989334106,
0.06670360267162323,
-0.04484398663043976,
-0.038224026560783386,
-0.07002109289169312,
-0.04601360112428665,
0.06643684953451157,
-0.013796461746096611,
-0.000034071683330694214,
-0.010990098118782043,
-0.009326891973614693,
-0.05114690586924553,
-0.03227350115776062,
-0.0337236225605011,
0.060238007456064224,
-0.036761291325092316,
0.047611430287361145,
-0.04222781956195831,
-0.11884967982769012,
0.10727370530366898,
-0.13787470757961273,
0.020925570279359818,
-0.004066638182848692,
-0.05329243838787079,
-0.013492140918970108,
0.0104609914124012,
0.019575657323002815,
0.06300964206457138,
0.056608278304338455,
0.0480211041867733,
-0.0019416498253121972,
0.004093415103852749,
0.09316451847553253,
0.012496060691773891,
-0.0007340007578022778,
-0.0189447533339262,
-0.0085104675963521,
0.0031728341709822416,
-0.009051049128174782,
0.03505325689911842,
0.07736591249704361,
-0.020059997215867043,
-0.0091658690944314,
0.00432246457785368,
0.01679150201380253,
-0.04045574367046356,
-0.0530376210808754,
-0.033214252442121506,
-0.038073230534791946,
-0.05713174119591713,
-0.02242286689579487,
0.04278382658958435,
-0.019350841641426086,
0.04384833946824074,
-0.01039445586502552,
0.02919895574450493,
-0.004044578410685062,
-0.0451233796775341,
0.06470607966184616,
-0.015600019134581089,
-0.057697925716638565,
-0.0006801061099395156,
-0.0064037214033305645,
-0.0002775209432002157,
0.004997421521693468,
-0.03870265558362007,
-0.0021991999819874763,
0.026626229286193848,
0.1179552897810936,
-0.06150868907570839,
-0.03616492822766304,
-0.029229607433080673,
0.13402287662029266,
0.030266212299466133,
-0.05319342389702797,
-0.024904247373342514,
0.06779270619153976,
-0.005433498881757259,
0.01918964460492134,
0.027939138934016228,
0.023224549368023872,
-0.08242820203304291,
-0.09687589108943939,
0.0853685736656189,
-0.030605273321270943,
0.03971046954393387,
-0.028357814997434616,
0.0015711813466623425,
-0.03803415223956108,
-0.0032467315904796124,
-0.009411448612809181,
0.027155742049217224,
0.07270777970552444,
-0.06150884926319122,
0.059481602162122726,
-0.03420846909284592,
0.026690050959587097,
0.03494296222925186,
0.04119868576526642,
0.02051379904150963,
0.0032451385632157326,
0.015325877815485,
-0.059806182980537415,
0.0504189133644104,
0.016494877636432648,
0.01285894587635994,
0.0512826107442379,
-0.010762365534901619,
0.05503147840499878,
0.027446217834949493,
-0.035634566098451614,
2.161739437620964e-33,
0.08405353873968124,
0.024785039946436882,
0.024155618622899055,
-0.0019404776394367218,
0.016175607219338417,
0.01999500021338463,
0.08328128606081009,
0.03990589454770088,
-0.026851234957575798,
-0.024427266791462898,
0.00810355693101883,
-0.027545388787984848,
-0.031100928783416748,
0.06841901689767838,
-0.000388081040000543,
0.04121719300746918,
-0.04312549903988838,
0.011951225809752941,
-0.008989546447992325,
-0.012392781674861908,
0.04268516227602959,
0.018697476014494896,
-0.030025184154510498,
-0.09002047032117844,
0.00021871496574021876,
-0.057128265500068665,
0.01417530607432127,
-0.03503042459487915,
0.025426002219319344,
-0.0024353154003620148,
-0.06177156791090965,
0.005007239989936352,
0.03309297934174538,
0.09381438791751862,
0.02761141024529934,
0.009899462573230267,
0.028440143913030624,
-0.045720577239990234,
-0.028462715446949005,
-0.007435017265379429,
-0.014597089029848576,
-0.002216739347204566,
-0.012919238768517971,
0.0018297076458111405,
-0.07331882417201996,
-0.14792321622371674,
0.0010837955633178353,
0.032861508429050446,
-0.09295935183763504,
-0.015919025987386703,
0.013855768367648125,
0.03773927316069603,
-0.010132281109690666,
-0.08954532444477081,
-0.0484248511493206,
0.00728291179984808,
-0.0038531352765858173,
0.0767211988568306,
-0.015768196433782578,
-0.013774417340755463,
0.02802828513085842,
0.052592892199754715,
0.012030809186398983,
0.09298224002122879,
-0.09994025528430939,
0.052150074392557144,
0.013752088882029057,
0.021766142919659615,
0.07468163222074509,
0.04773562401533127,
0.0032660337164998055,
0.026909539476037025,
0.010555604472756386,
-0.028352083638310432,
0.10707695782184601,
-0.017905142158269882,
0.02080646902322769,
-0.07746833562850952,
0.06365764886140823,
-0.05447341129183769,
-0.03808583691716194,
-0.012819213792681694,
-0.05051568150520325,
0.022442106157541275,
-0.0060433270409703255,
-0.10490856319665909,
0.007153488229960203,
-0.05653524398803711,
0.006456872448325157,
-0.005256599746644497,
-0.07797133177518845,
-0.029913129284977913,
-0.05814823508262634,
0.06253773719072342,
0.0806262269616127,
-8.966376116129742e-34,
0.05366942286491394,
-0.05528700351715088,
0.020669225603342056,
0.030322987586259842,
-0.05876763537526131,
-0.046657364815473557,
-0.007314479444175959,
-0.017865503206849098,
-0.013816355727612972,
-0.027677055448293686,
-0.04935881868004799,
0.0770852267742157,
0.0492561012506485,
0.03061344474554062,
-0.01860957033932209,
-0.023108866065740585,
-0.05498141050338745,
0.0858139917254448,
-0.00711951544508338,
-0.024818180128932,
0.013287438079714775,
0.08523938804864883,
-0.07056959718465805,
0.0842226967215538,
-0.0656527429819107,
-0.013333156704902649,
-0.023643096908926964,
0.08061999827623367,
0.027067072689533234,
0.046379972249269485,
-0.0646817535161972,
-0.027185868471860886,
0.015751218423247337,
-0.015813583508133888,
-0.02899019978940487,
0.16396746039390564,
0.07068827003240585,
0.010979119688272476,
-0.13525855541229248,
0.046438638120889664,
0.024809455499053,
-0.10672146081924438,
-0.051159732043743134,
-0.009233727119863033,
0.022658858448266983,
-0.03445062413811684,
-0.042028021067380905,
-0.0038524530827999115,
0.01897408999502659,
-0.05590984970331192,
-0.019741728901863098,
-0.017505347728729248,
-0.04671703279018402,
-0.016135722398757935,
-0.08827243745326996,
-0.036139123141765594,
0.01067347638309002,
0.09073297679424286,
0.13219492137432098,
0.0014043280389159918,
0.019174974411725998,
-0.015394176356494427,
0.0511980876326561,
0.03432498872280121,
0.021747898310422897,
0.08837860077619553,
0.035276662558317184,
0.16629266738891602,
0.032355520874261856,
0.0068568033166229725,
0.060598887503147125,
0.06364484876394272,
0.032248757779598236,
-0.03422049060463905,
0.07618512958288193,
0.0007125514675863087,
0.0026705481577664614,
-0.04174226149916649,
-0.03723658248782158,
-0.022581985220313072,
0.027054646983742714,
-0.022293737158179283,
0.025996878743171692,
0.048208802938461304,
0.01574854552745819,
0.06175759807229042,
0.015101650729775429,
0.06242057681083679,
-0.030233655124902725,
-0.002620604122057557,
0.013994242064654827,
0.029718812555074692,
-0.026364434510469437,
0.025995681062340736,
0.06467590481042862,
-4.143812759593857e-8,
-0.11571750789880753,
0.05565105751156807,
0.09005142003297806,
0.01408735103905201,
0.03427252173423767,
0.017942676320672035,
-0.04259544610977173,
-0.014212270267307758,
-0.05028803274035454,
-0.03456660360097885,
0.022490141913294792,
0.08206796646118164,
0.0759207233786583,
-0.0983075499534607,
-0.027961868792772293,
0.06537535041570663,
0.020593134686350822,
-0.016222551465034485,
-0.021947083994746208,
-0.02429972216486931,
0.033380646258592606,
-0.02003687061369419,
0.05467480421066284,
-0.08819667994976044,
-0.011660640127956867,
-0.10016024857759476,
-0.01748247817158699,
0.03461679443717003,
0.11783771961927414,
0.034509167075157166,
-0.006110502872616053,
-0.06274847686290741,
0.05564776808023453,
0.014130198396742344,
0.03714698180556297,
0.045065078884363174,
0.03499995544552803,
-0.07886313647031784,
-0.011040790937840939,
-0.07373759895563126,
-0.0337023101747036,
0.05030292645096779,
-0.017071567475795746,
-0.018348118290305138,
-0.01854766346514225,
0.021303392946720123,
-0.014915967360138893,
-0.1424287110567093,
0.02615046128630638,
-0.03660919517278671,
0.10153865069150925,
-0.05195586755871773,
-0.0350080244243145,
0.10609658807516098,
0.07191833853721619,
0.025104142725467682,
-0.015248981304466724,
0.023265426978468895,
0.05145483836531639,
0.07611742615699768,
0.05609329789876938,
-0.03554802015423775,
0.014804179780185223,
-0.0012582443887367845
] |
https://github.com/arnavkj1995/sfm
| 17
| 0
| 0
| 0
|
MatExpert: Decomposing Materials Discovery By Mimicking Human Experts
|
https://openreview.net/forum?id=AUBvo4sxVL
|
[
"Qianggang Ding",
"Santiago Miret",
"Bang Liu"
] |
Poster
|
Material discovery is a critical research area with profound implications for various industries. In this work, we introduce MatExpert, a novel framework that leverages Large Language Models (LLMs) and contrastive learning to accelerate the discovery and design of new solid-state materials. Inspired by the workflow of human materials design experts, our approach integrates three key stages: retrieval, transition, and generation. First, in the retrieval stage, MatExpert identifies an existing material that closely matches the desired criteria. Second, in the transition stage, MatExpert outlines the necessary modifications to transform this material formulation to meet specific requirements outlined by the initial user query. Third, in the generation state, MatExpert performs detailed computations and structural generation to create a new material based on the provided information. Our experimental results demonstrate that MatExpert outperforms state-of-the-art methods in material generation tasks, achieving superior performance across various metrics including validity, distribution, and stability. As such, MatExpert represents a meaningful advancement in computational material discovery using langauge-based generative models.
|
Material Discovery, Large Language Models (LLMs), Material Generation, Crystal Structure Generation, Contrastive Learning
|
We present MatExpert, a novel framework leveraging Large Language Models (LLMs) and contrastive learning to streamline material discovery like human expert workflows.
| 11,344
|
2410.21317
|
[
-0.12402240931987762,
-0.009302089922130108,
0.06279847770929337,
0.07663506269454956,
0.053175654262304306,
-0.015080411918461323,
0.011475702747702599,
-0.027157118543982506,
-0.0021018937695771456,
0.00506657874211669,
-0.10203998535871506,
-0.07568483054637909,
0.05177808925509453,
0.04653165116906166,
-0.014116901904344559,
0.01321270689368248,
0.03691323101520538,
0.006208631210029125,
-0.012240120209753513,
0.0025873221457004547,
0.08658698946237564,
0.05829339846968651,
0.04789896681904793,
-0.08750002831220627,
0.06339772045612335,
0.07401377707719803,
0.027617741376161575,
-0.053468119353055954,
0.09588588774204254,
-0.01637943834066391,
0.049197159707546234,
0.09728529304265976,
-0.0041692787781357765,
0.003173495875671506,
-0.025409599766135216,
0.1004551351070404,
-0.06165477633476257,
-0.0016166685381904244,
-0.019523410126566887,
-0.09013503789901733,
-0.05138176679611206,
-0.013948550447821617,
-0.006574968341737986,
0.011904382146894932,
0.08651236444711685,
-0.03559929132461548,
0.04967736080288887,
-0.06618794798851013,
0.010424125008285046,
-0.0034348152112215757,
-0.061915360391139984,
-0.05637853965163231,
0.008946080692112446,
0.021128127351403236,
0.013909652829170227,
0.07789686322212219,
0.08940249681472778,
-0.027873298153281212,
0.001368480152450502,
-0.05416957288980484,
-0.006868597120046616,
-0.08457852154970169,
-0.05435057356953621,
-0.0771913230419159,
0.05887559801340103,
-0.004253164865076542,
-0.0025590446311980486,
0.04765889421105385,
-0.005898684728890657,
-0.002136840717867017,
-0.006649177987128496,
0.0753435343503952,
0.014498699456453323,
0.10222887247800827,
0.06386015564203262,
0.08570381999015808,
0.05974810570478439,
-0.05441349744796753,
-0.0031326801981776953,
-0.08546619862318039,
-0.06389717757701874,
-0.019887059926986694,
0.034544359892606735,
0.016011131927371025,
0.02706712670624256,
0.028226029127836227,
0.004848262295126915,
0.03563883528113365,
0.0214435663074255,
0.027496017515659332,
-0.029047952964901924,
0.027306092903017998,
0.006479864474385977,
-0.07907469570636749,
-0.04193345457315445,
0.045134853571653366,
0.05074651539325714,
-0.01130234356969595,
0.006196613889187574,
0.055277686566114426,
-0.010794069617986679,
0.07648321986198425,
0.05374325439333916,
-0.07456108182668686,
-0.06866491585969925,
-0.05055434629321098,
-0.020882250741124153,
0.04117250069975853,
0.0075562987476587296,
-0.03649347648024559,
-0.012753980234265327,
0.05416933447122574,
-0.006359983701258898,
0.01716129295527935,
-0.03903365880250931,
0.018093537539243698,
-0.02591184340417385,
-0.046171143651008606,
-0.01831158436834812,
0.06122574210166931,
0.027259793132543564,
0.0375303253531456,
-0.004887098912149668,
0.021579600870609283,
-0.021796517074108124,
-0.07337377220392227,
-0.10052858293056488,
-8.469586489533319e-34,
0.05352194234728813,
-0.006455250084400177,
0.018322546035051346,
0.17134477198123932,
0.009334552101790905,
-0.006509807426482439,
0.044261615723371506,
0.024385271593928337,
-0.03961190953850746,
-0.05080021172761917,
-0.020154668018221855,
0.054799821227788925,
-0.08331910520792007,
0.11375462263822556,
0.027291661128401756,
-0.03798384964466095,
-0.04158928990364075,
-0.03180951625108719,
-0.023503078147768974,
-0.03388232737779617,
-0.018787240609526634,
-0.026464782655239105,
0.031114473938941956,
0.023438017815351486,
0.010729387402534485,
0.03375239670276642,
0.09083472192287445,
-0.03989255800843239,
-0.06633258610963821,
-0.01722870022058487,
-0.12336868792772293,
0.00477746594697237,
0.02307853475213051,
0.056419480592012405,
-0.030096188187599182,
0.010973481461405754,
-0.023913411423563957,
-0.06679149717092514,
0.02991473488509655,
-0.0053498828783631325,
0.008448156528174877,
0.018040476366877556,
0.06642582267522812,
0.024895822629332542,
-0.13648030161857605,
-0.028369780629873276,
0.021747278049588203,
0.043012022972106934,
0.004875790327787399,
0.0030460376292467117,
0.06343924254179001,
0.05193281173706055,
0.014858872629702091,
-0.031143436208367348,
0.010472145862877369,
0.01539647951722145,
0.010441665537655354,
-0.01293681375682354,
0.07306084781885147,
-0.01369557622820139,
0.04031366482377052,
0.05594087392091751,
0.0560266338288784,
0.09243518859148026,
0.0025710936170071363,
0.09306047111749649,
0.023015031591057777,
0.06991138309240341,
0.02999156527221203,
0.0034941541962325573,
-0.020000826567411423,
0.013874744065105915,
0.04695028439164162,
-0.0472598560154438,
0.11535392701625824,
0.0013708079932257533,
0.06763561069965363,
-0.07876099646091461,
0.048285823315382004,
-0.102345310151577,
-0.07746552675962448,
0.005056870169937611,
-0.06629155576229095,
-0.09119794517755508,
-0.13934755325317383,
-0.04225416108965874,
0.08707214891910553,
-0.0731755718588829,
-0.01571468450129032,
-0.026580462232232094,
-0.07817552238702774,
-0.08008428663015366,
-0.019371919333934784,
0.01555610727518797,
0.0028567626141011715,
-2.9578344460142154e-34,
-0.033127810806035995,
-0.0606488473713398,
0.014036150649189949,
0.015692438930273056,
0.003752870950847864,
-0.09608256816864014,
-0.06594645977020264,
-0.01979886367917061,
0.014884988777339458,
-0.04581993445754051,
0.08061402291059494,
-0.036172639578580856,
0.06086587905883789,
-0.041872039437294006,
-0.028506342321634293,
0.07035902142524719,
-0.01542188785970211,
-0.058373238891363144,
0.013216158375144005,
0.0210622176527977,
-0.06775256991386414,
0.04345710948109627,
-0.12158044427633286,
-0.006063621025532484,
-0.0962776467204094,
0.020938660949468613,
-0.030108235776424408,
-0.05898464843630791,
-0.01702072285115719,
-0.016436748206615448,
0.013954920694231987,
-0.07466916739940643,
-0.03240542113780975,
0.002931237919256091,
-0.02939545176923275,
0.04226331412792206,
0.03559081256389618,
-0.02158040925860405,
-0.011640697717666626,
0.003501626430079341,
0.05400794744491577,
-0.01928216964006424,
-0.06999624520540237,
0.06392685323953629,
-0.03592563048005104,
-0.046462103724479675,
-0.06552617251873016,
-0.005110302474349737,
0.0672198012471199,
0.017570432275533676,
0.08949613571166992,
0.0020463301334530115,
-0.015771007165312767,
-0.13487659394741058,
-0.030915074050426483,
-0.04092703387141228,
-0.02635980024933815,
-0.028322691097855568,
0.04933568462729454,
0.070480577647686,
-0.005414094775915146,
0.00233493116684258,
0.03774378448724747,
-0.013908070512115955,
-0.051148269325494766,
-0.04575996845960617,
-0.012110640294849873,
-0.03449704125523567,
-0.04463689774274826,
-0.014631894417107105,
0.04540320858359337,
-0.007815781980752945,
0.019855542108416557,
0.022116435691714287,
0.07424243539571762,
-0.11682836711406708,
0.012982715852558613,
-0.03262723609805107,
-0.05514127016067505,
-0.06682823598384857,
0.0479862242937088,
0.047776877880096436,
-0.02251887507736683,
0.04809030890464783,
0.018031377345323563,
0.03886482119560242,
-0.018657837063074112,
0.12872982025146484,
0.04044114798307419,
0.0042477622628211975,
-0.04133235290646553,
0.031033731997013092,
0.01192021369934082,
0.05234517529606819,
0.008908034302294254,
-4.219424099005664e-8,
-0.07576001435518265,
0.03199130669236183,
-0.00022872908448334783,
-0.014159157872200012,
-0.02631402015686035,
0.036473315209150314,
-0.033036548644304276,
0.05826403573155403,
-0.05166073143482208,
-0.028489043936133385,
0.09281068295240402,
-0.08338766545057297,
-0.01129251066595316,
-0.05954757332801819,
0.03263353183865547,
0.03159761428833008,
0.035312145948410034,
0.050573356449604034,
-0.05340712144970894,
-0.09036943316459656,
0.06852923333644867,
0.043636392802000046,
0.06949550658464432,
0.05523024871945381,
0.05667976662516594,
-0.03173064813017845,
-0.05747528001666069,
0.0009782729903236032,
0.04253856837749481,
-0.009113422594964504,
-0.007971188984811306,
0.025632858276367188,
0.02508140727877617,
0.06821023672819138,
0.023523474112153053,
0.05379023402929306,
0.017160875722765923,
-0.051862336695194244,
0.008513711392879486,
-0.0010073523735627532,
0.0414547361433506,
0.023780018091201782,
-0.09225568175315857,
0.05130445212125778,
-0.058405447751283646,
-0.009590496309101582,
-0.0019550214055925608,
-0.0324639156460762,
0.036803651601076126,
0.062361814081668854,
0.019188545644283295,
-0.09809020906686783,
0.023497212678194046,
0.002458498813211918,
0.06652005761861801,
0.08050735294818878,
-0.04354579374194145,
0.05407678708434105,
0.10381578654050827,
-0.04906830936670303,
0.021051885560154915,
-0.031681060791015625,
-0.007772024720907211,
0.004476574715226889
] | 0
| 0
| 0
| 0
|
|
SOAP: Improving and Stabilizing Shampoo using Adam for Language Modeling
|
https://openreview.net/forum?id=IDxZhXrpNf
|
[
"Nikhil Vyas",
"Depen Morwani",
"Rosie Zhao",
"Itai Shapira",
"David Brandfonbrener",
"Lucas Janson",
"Sham M. Kakade"
] |
Poster
|
There is growing evidence of the effectiveness of Shampoo, a higher-order preconditioning method, over Adam in deep learning optimization tasks. However, Shampoo's drawbacks include additional hyperparameters and computational overhead when compared to Adam, which only updates running averages of first- and second-moment quantities. This work establishes a formal connection between Shampoo (implemented with the 1/2 power) and Adafactor --- a memory-efficient approximation of Adam --- showing that Shampoo is equivalent to running Adafactor in the eigenbasis of Shampoo's preconditioner. This insight leads to the design of a simpler and computationally efficient algorithm: **S**hampo**O** with **A**dam in the **P**reconditioner's eigenbasis (SOAP).
With regards to improving Shampoo's computational efficiency, the most straightforward approach would be to simply compute Shampoo's eigendecomposition less frequently. Unfortunately, as our empirical results show, this leads to performance degradation that worsens with this frequency. SOAP mitigates this degradation by continually updating the running average of the second moment, just as Adam does, but in the current (slowly changing) coordinate basis. Furthermore, since SOAP is equivalent to running Adam in a rotated space, it introduces only one additional hyperparameter (the preconditioning frequency) compared to Adam. We empirically evaluate SOAP on language model pre-training with 360m and 660m sized models. In the large batch regime, SOAP reduces the number of iterations by over 40\% and wall clock time by over 35\% compared to AdamW, with approximately 20\% improvements in both metrics compared to Shampoo. An implementation of SOAP is available at https://github.com/nikhilvyas/SOAP.
|
Shampoo, Adam, Second Order Optimizer
|
We design and empirically study a new second order optimizer called SOAP, which runs Adam in the eigenbasis provided by Shampoo.
| 11,337
| null |
[
-0.05842620134353638,
-0.03806876018643379,
0.06821104884147644,
-0.026708081364631653,
-0.024598378688097,
-0.013903244398534298,
-0.032326214015483856,
-0.02970191463828087,
-0.04599950835108757,
-0.05720783770084381,
-0.0068288943730294704,
0.04799090325832367,
-0.019978266209363937,
0.002734011271968484,
-0.036516882479190826,
0.05499827861785889,
0.06233726069331169,
0.0696457028388977,
-0.09897176921367645,
-0.02987222746014595,
0.07304126024246216,
-0.01785983331501484,
0.013612337410449982,
-0.047988295555114746,
0.03181832283735275,
0.03700341656804085,
-0.02798098511993885,
-0.09253484755754471,
-0.003065752796828747,
0.0012403710279613733,
0.09378532320261002,
0.047766488045454025,
-0.003662481438368559,
-0.067176453769207,
-0.12454265356063843,
0.005489685572683811,
0.009108559228479862,
0.009180240333080292,
0.005638883914798498,
0.017215371131896973,
-0.020901480689644814,
-0.027431968599557877,
-0.08705390989780426,
-0.0028850389644503593,
0.08323434740304947,
-0.02166750282049179,
0.06615271419286728,
-0.0031182768288999796,
0.06162970885634422,
-0.00009610049892216921,
-0.06909116357564926,
-0.047664470970630646,
0.03697449341416359,
0.03158487752079964,
-0.03144790232181549,
0.007601280231028795,
0.004690116737037897,
-0.022103391587734222,
-0.07255090028047562,
0.06603942811489105,
-0.009548897854983807,
-0.07816372066736221,
-0.024372795596718788,
0.01823238842189312,
0.06103276088833809,
0.01442303229123354,
0.0571344755589962,
-0.0066454289481043816,
-0.008712012320756912,
0.05096858739852905,
-0.04463006183505058,
0.056431084871292114,
0.0008693236741237342,
0.1152469664812088,
0.02950928546488285,
0.1142142117023468,
0.0400310717523098,
-0.023558292537927628,
0.044448256492614746,
-0.040210362523794174,
0.034078843891620636,
0.031880516558885574,
0.014200587756931782,
0.06545714288949966,
0.10478319227695465,
0.0035820824559777975,
0.04030117392539978,
-0.06089350953698158,
0.03266797587275505,
-0.05028173327445984,
-0.028925253078341484,
-0.023613501340150833,
0.014475599862635136,
-0.06741764396429062,
0.1299736499786377,
0.022496722638607025,
-0.04042189195752144,
0.022323794662952423,
-0.043409623205661774,
0.07243789732456207,
-0.03983831778168678,
0.00645282119512558,
-0.03368482366204262,
-0.0672641471028328,
0.025251759216189384,
-0.027993272989988327,
0.004803318064659834,
0.05440343916416168,
0.06383805721998215,
-0.020801464095711708,
0.031209399923682213,
-0.012034578248858452,
-0.017104538157582283,
0.03313574567437172,
-0.0041472832672297955,
-0.05268164724111557,
0.055411022156476974,
-0.02717467211186886,
0.048127975314855576,
-0.009080201387405396,
0.0013816787395626307,
0.022632716223597527,
-0.0023013604804873466,
0.0022556944750249386,
-0.016340574249625206,
-0.015730397775769234,
-0.0695888102054596,
1.4743623201509156e-33,
0.002164569916203618,
0.05539839714765549,
0.014077939093112946,
-0.05190986394882202,
0.003128946525976062,
0.016641564667224884,
0.016825418919324875,
0.00514683173969388,
0.034799735993146896,
0.02732602134346962,
-0.029438285157084465,
-0.019611550495028496,
-0.07470736652612686,
0.03828340768814087,
0.0040933783166110516,
0.022911787033081055,
0.035583071410655975,
0.09484653174877167,
0.02352837473154068,
-0.09359216690063477,
0.05300814285874367,
-0.007930582389235497,
-0.06645555049180984,
-0.06507950276136398,
-0.0875072330236435,
-0.02187124453485012,
0.08919142931699753,
-0.010357314720749855,
-0.04131494089961052,
0.01684100553393364,
-0.0009727765573188663,
-0.052677180618047714,
-0.05499479919672012,
0.043818045407533646,
-0.0918816477060318,
-0.06926732510328293,
-0.007988420315086842,
-0.06780926883220673,
0.09304283559322357,
-0.03620404377579689,
-0.06319136917591095,
0.07403022050857544,
0.0878717303276062,
-0.00460363132879138,
-0.08022219687700272,
-0.031119147315621376,
-0.10126780718564987,
0.042669180780649185,
-0.012543815188109875,
-0.007453303784132004,
0.029824772849678993,
0.013705134391784668,
-0.00984752643853426,
0.06018952280282974,
-0.05418296530842781,
-0.04852371662855148,
0.009146207012236118,
0.014707072637975216,
0.05311809480190277,
0.12516482174396515,
-0.042650047689676285,
0.012768082320690155,
-0.10420583933591843,
-0.0341314896941185,
-0.057645175606012344,
-0.028252013027668,
-0.002026455709710717,
0.058749787509441376,
0.016910254955291748,
0.04511619731783867,
-0.017205914482474327,
0.00991128385066986,
-0.00027648545801639557,
-0.006678171921521425,
0.06422378122806549,
-0.07973571866750717,
0.10271578282117844,
-0.030019957572221756,
-0.04069051891565323,
-0.010548695921897888,
0.01614844612777233,
0.13500463962554932,
-0.01071938220411539,
0.06217958778142929,
-0.08698204904794693,
-0.04911361262202263,
-0.007992814294993877,
-0.006524119526147842,
-0.08278506994247437,
-0.06381846219301224,
0.056903280317783356,
-0.04881880059838295,
0.01092938706278801,
0.024919938296079636,
-0.02691522240638733,
-8.467158359014902e-34,
0.02673204056918621,
-0.07108315825462341,
-0.009861982427537441,
0.0607406347990036,
0.020230241119861603,
-0.05560363456606865,
0.044150251895189285,
0.028327740728855133,
-0.10002132505178452,
-0.10798081755638123,
0.03292452543973923,
-0.042683910578489304,
0.005360358394682407,
-0.00045691433479078114,
0.053700268268585205,
0.06935179233551025,
-0.03911612927913666,
-0.05608183518052101,
0.0352279469370842,
0.023663103580474854,
0.007258123252540827,
0.10708920657634735,
0.008721360005438328,
0.05069950595498085,
-0.06457116454839706,
-0.05205412954092026,
0.009028731845319271,
0.0701216459274292,
0.03502976521849632,
0.006976076867431402,
-0.04406115412712097,
-0.03882577642798424,
-0.08260194212198257,
-0.020604023709893227,
-0.020203912630677223,
0.07673473656177521,
-0.03580404072999954,
-0.08182715624570847,
0.027379188686609268,
0.047603994607925415,
0.0900481641292572,
-0.034571949392557144,
-0.0389467217028141,
-0.017579710111021996,
-0.0005667215446010232,
0.06609134376049042,
-0.08801745623350143,
-0.01774817891418934,
0.008837145753204823,
0.04798412695527077,
0.03349718824028969,
0.03387556970119476,
-0.0714167132973671,
0.010625229217112064,
0.01268349401652813,
-0.00264162034727633,
-0.0046100132167339325,
-0.029130179435014725,
-0.021328922361135483,
0.030743645504117012,
-0.08806901425123215,
-0.01704966276884079,
0.07755658030509949,
0.009903035126626492,
-0.07138489186763763,
0.008753692731261253,
0.013383516110479832,
0.061205316334962845,
-0.02409261092543602,
-0.006305769085884094,
0.015827162191271782,
-0.005606289021670818,
0.048263899981975555,
0.006996107753366232,
-0.022352052852511406,
-0.06818650662899017,
-0.0016120258951559663,
-0.08847323805093765,
-0.04653958976268768,
-0.0007553679752163589,
-0.0840424969792366,
0.004117349162697792,
-0.0038686690386384726,
0.058228958398103714,
0.08007870614528656,
0.07370832562446594,
0.036156900227069855,
0.07620850205421448,
0.05389401316642761,
0.04251423850655556,
0.007279247045516968,
0.01968626119196415,
-0.07355078309774399,
0.0727301612496376,
0.025273138657212257,
-4.139816311976574e-8,
-0.012071722187101841,
0.020987335592508316,
0.1142076700925827,
0.07199837267398834,
0.08257792145013809,
0.02523203007876873,
-0.06420377641916275,
0.10829507559537888,
-0.029150046408176422,
-0.021656308323144913,
0.07463305443525314,
0.05173797905445099,
-0.13855281472206116,
-0.005149305332452059,
-0.04568655043840408,
0.04427788406610489,
0.03408832103013992,
0.05390503630042076,
-0.025697505101561546,
-0.11847765743732452,
-0.04190916568040848,
0.08737844973802567,
0.0012120238970965147,
-0.011476252228021622,
0.09422282129526138,
-0.05815066024661064,
0.03641430661082268,
0.03973536565899849,
0.07149247825145721,
-0.030564723536372185,
-0.037276700139045715,
0.008600427769124508,
-0.07146844267845154,
-0.031191788613796234,
-0.00804184377193451,
0.09410300105810165,
-0.040360093116760254,
-0.00029602149152196944,
-0.04574916511774063,
-0.00532991997897625,
0.01636875793337822,
0.03508943319320679,
-0.0024990609381347895,
0.030287815257906914,
0.09657379984855652,
-0.044739022850990295,
-0.008245622739195824,
-0.02029508724808693,
0.004844852723181248,
-0.0071655986830592155,
0.0464092455804348,
-0.01895829103887081,
0.05110472813248634,
0.04607745632529259,
-0.001281294273212552,
-0.050987083464860916,
-0.08974102139472961,
-0.10857905447483063,
0.07069087028503418,
0.11344657093286514,
0.046597812324762344,
-0.030305970460176468,
0.007999158464372158,
0.02842608466744423
] | 0
| 0
| 0
| 0
|
|
Causal Graphical Models for Vision-Language Compositional Understanding
|
https://openreview.net/forum?id=haJHr4UsQX
|
[
"Fiorenzo Parascandolo",
"Nicholas Moratelli",
"Enver Sangineto",
"Lorenzo Baraldi",
"Rita Cucchiara"
] |
Poster
|
Recent work has empirically shown that Vision-Language Models (VLMs) struggle
to fully understand the compositional properties of the human language, usually
modeling an image caption as a “bag of words”. As a result, they perform
poorly on compositional tasks, which require a deeper understanding of the different
entities of a sentence (subject, verb, etc.) jointly with their mutual relationships
in order to be solved. In this paper, we model the dependency relations
among textual and visual tokens using a Causal Graphical Model (CGM), built using
a dependency parser, and we train a decoder conditioned by the VLM visual
encoder. Differently from standard autoregressive or parallel predictions, our decoder’s
generative process is partially-ordered following the CGM structure. This
structure encourages the decoder to learn only the main causal dependencies in
a sentence discarding spurious correlations. Using extensive experiments on five
compositional benchmarks, we show that our method significantly outperforms
all the state-of-the-art compositional approaches by a large margin, and it also improves
over methods trained using much larger datasets.
Our model weights and code are publicly available.
|
Compositionality, Vision-Language Models, Causal Learning
| null | 11,332
|
2412.09353
|
[
-0.024584151804447174,
-0.1304311454296112,
0.07078545540571213,
0.03833991661667824,
0.06440642476081848,
0.0049791159108281136,
-0.03492361307144165,
-0.009398428723216057,
0.026711545884609222,
-0.062282249331474304,
-0.043967656791210175,
-0.08652166277170181,
0.032154910266399384,
0.11507213860750198,
-0.020517194643616676,
-0.021673357114195824,
0.07637351751327515,
0.018735704943537712,
-0.05345555767416954,
-0.07401219755411148,
-0.006104981526732445,
-0.03478963300585747,
-0.02113211341202259,
-0.00677364319562912,
0.05008157715201378,
0.06271765381097794,
-0.02280295453965664,
-0.027591783553361893,
0.037720464169979095,
-0.02077217772603035,
0.0005515809752978384,
-0.021210961043834686,
-0.04840888828039169,
0.059815775603055954,
-0.06119384989142418,
0.10214825719594955,
-0.008608261123299599,
0.004555319435894489,
0.04002046585083008,
-0.08572712540626526,
-0.03995341435074806,
0.02068140357732773,
0.08202492445707321,
0.040942028164863586,
0.12753666937351227,
0.005533133167773485,
0.03851163759827614,
-0.03116139955818653,
-0.00721392035484314,
-0.01772499643266201,
-0.10043923556804657,
0.027718372642993927,
0.009841186925768852,
0.012879990041255951,
-0.03864024952054024,
0.02879137173295021,
0.10541126132011414,
-0.02120084874331951,
0.0015727576101198792,
-0.011992935091257095,
-0.07027528434991837,
-0.06062524393200874,
-0.0833234041929245,
-0.007844468578696251,
0.0030334556940943003,
-0.048841167241334915,
0.05032232031226158,
0.02771865390241146,
-0.057944156229496,
0.06833630055189133,
-0.01834011636674404,
0.055125050246715546,
0.002989786211401224,
0.02907877415418625,
0.0032145963050425053,
0.08217203617095947,
0.06004191190004349,
-0.058151330798864365,
0.04794677346944809,
-0.1427575945854187,
0.035716231912374496,
0.06351767480373383,
0.06748080253601074,
0.04545793682336807,
0.13616138696670532,
-0.006326852831989527,
0.02267915941774845,
0.03992971405386925,
-0.01916390471160412,
0.053512442857027054,
-0.04692961275577545,
-0.09137969464063644,
0.10144412517547607,
0.004208693280816078,
0.054685331881046295,
0.04811851307749748,
-0.021563952788710594,
-0.04238865524530411,
-0.013087518513202667,
0.06125323846936226,
0.010149607434868813,
0.04779110476374626,
0.051710691303014755,
-0.0631345883011818,
0.04276423156261444,
0.013291378505527973,
0.02168165147304535,
0.0673893392086029,
0.03061066009104252,
-0.08494296669960022,
-0.01860138773918152,
0.05567123368382454,
-0.051478512585163116,
0.01926574483513832,
0.010940486565232277,
-0.007560063619166613,
0.009669588878750801,
-0.09196756035089493,
0.046017564833164215,
0.029007771983742714,
-0.02954612672328949,
0.03378647193312645,
-0.049942709505558014,
0.005699119530618191,
-0.001305859419517219,
-0.09335807710886002,
0.018923476338386536,
-2.0142767937171773e-33,
0.050812385976314545,
0.006036828272044659,
0.07851125299930573,
0.0035283532924950123,
0.03737539425492287,
0.011227769777178764,
-0.06032896041870117,
0.03675539791584015,
-0.05371378734707832,
0.012321265414357185,
0.005431412253528833,
-0.051166824996471405,
-0.07323165982961655,
0.14797423779964447,
0.01581628806889057,
-0.028430676087737083,
-0.05576185882091522,
0.05205507576465607,
-0.03913760557770729,
-0.02006039395928383,
-0.057413335889577866,
-0.0003678069624584168,
-0.04819494113326073,
-0.003731949022039771,
0.00002586061782494653,
-0.005056937225162983,
0.07030821591615677,
-0.09440743923187256,
-0.05297277122735977,
0.01120257843285799,
-0.09110027551651001,
-0.026137085631489754,
0.056564923375844955,
0.041941676288843155,
0.05463579297065735,
-0.07151725143194199,
-0.046226952224969864,
0.01670459844172001,
0.012679873034358025,
-0.0169530026614666,
-0.023152677342295647,
-0.003260307479649782,
-0.04084859415888786,
-0.018739232793450356,
-0.08295373618602753,
-0.0554724857211113,
-0.028266409412026405,
0.04802635312080383,
-0.04711223021149635,
-0.04972725361585617,
0.05813303217291832,
0.06012789160013199,
-0.009560225531458855,
-0.05229805037379265,
0.045332059264183044,
0.009545575827360153,
0.028848357498645782,
0.006861678324639797,
0.005673943553119898,
-0.06434114277362823,
0.03517119586467743,
0.08140953630208969,
0.021830281242728233,
0.09472271054983139,
0.060681141912937164,
0.07191410660743713,
-0.054016269743442535,
-0.02652728743851185,
0.012023502960801125,
0.05658145621418953,
-0.06826111674308777,
0.07597298175096512,
-0.04082638770341873,
-0.07294853776693344,
0.07895807921886444,
0.060306135565042496,
-0.01217921543866396,
-0.0862545371055603,
-0.01291672233492136,
0.07706870883703232,
-0.06789757311344147,
-0.08466458320617676,
0.06156965717673302,
-0.04618275538086891,
-0.067678302526474,
-0.02968095801770687,
0.0720350369811058,
-0.05298370495438576,
-0.01748356595635414,
-0.05485547333955765,
0.013809910975396633,
-0.062103208154439926,
-0.011383315548300743,
-0.012229266576468945,
0.045000504702329636,
8.867164392920341e-34,
-0.0160074420273304,
0.07803438603878021,
0.06209129840135574,
-0.031069530174136162,
-0.03389063477516174,
-0.0836813822388649,
-0.02692083641886711,
0.03233114257454872,
0.00004720327342511155,
-0.027400223538279533,
0.002095885807648301,
-0.038809631019830704,
-0.03473572060465813,
0.0718584954738617,
0.023967759683728218,
0.02978392131626606,
0.024215437471866608,
0.027588430792093277,
-0.014732152223587036,
0.006609600502997637,
0.030546393245458603,
0.13974587619304657,
-0.012706542387604713,
0.011515078134834766,
-0.051715459674596786,
0.013289565220475197,
0.008081268519163132,
0.07940366119146347,
-0.07048249244689941,
0.02444598451256752,
-0.01584620401263237,
0.01837049424648285,
-0.007710565347224474,
-0.04982685670256615,
-0.061467789113521576,
0.06614363193511963,
-0.0005979745183140039,
-0.04565541818737984,
-0.06782057136297226,
0.035148538649082184,
0.1048186868429184,
-0.035061802715063095,
-0.060637395828962326,
0.02025960199534893,
-0.049972403794527054,
-0.051899317651987076,
-0.061151448637247086,
-0.00453958148136735,
0.027674317359924316,
0.04673462733626366,
-0.06076314300298691,
0.026340477168560028,
-0.07013218849897385,
0.027497347444295883,
-0.028141368180513382,
-0.10476167500019073,
0.004988509695976973,
-0.025485610589385033,
0.08978584408760071,
-0.05767834186553955,
-0.0726839154958725,
0.0033219102770090103,
0.051916591823101044,
-0.06903107464313507,
0.05265657231211662,
-0.08519412577152252,
-0.0319049172103405,
-0.032083507627248764,
0.046579759567976,
-0.023911569267511368,
0.07045125961303711,
-0.025701487436890602,
-0.00419125659391284,
0.032014645636081696,
0.03977448493242264,
-0.005811071954667568,
0.0647675022482872,
-0.0040306467562913895,
-0.03504309803247452,
-0.07999932020902634,
0.05697782337665558,
-0.06743892282247543,
0.0703420415520668,
0.06357160210609436,
0.01826738938689232,
0.10178175568580627,
0.021989358589053154,
0.016265148296952248,
0.07075130194425583,
0.04579479619860649,
-0.08155310899019241,
-0.02369525097310543,
-0.01273268647491932,
0.08190183341503143,
-0.005726138595491648,
-3.584733576644794e-8,
-0.06804932653903961,
0.0171135775744915,
-0.0408073328435421,
0.017449188977479935,
-0.007285328581929207,
-0.009444164112210274,
-0.03596127778291702,
0.06118689477443695,
-0.041709184646606445,
0.001020868425257504,
0.052475232630968094,
0.04372856393456459,
-0.03963540866971016,
-0.04452289268374443,
-0.04257763922214508,
0.03130229935050011,
0.031167425215244293,
-0.02547389268875122,
-0.0008540913695469499,
0.021918106824159622,
0.0557679757475853,
-0.042731963098049164,
0.025384629145264626,
0.024088053032755852,
-0.02187163010239601,
-0.0484151728451252,
-0.041832685470581055,
0.03849617391824722,
0.05036883056163788,
-0.03628084063529968,
0.07254405319690704,
0.09139854460954666,
-0.0663032978773117,
-0.06349517405033112,
-0.0043710265308618546,
0.06284046918153763,
0.036283496767282486,
-0.06491030752658844,
-0.026760052889585495,
-0.0005580809083767235,
0.07412487268447876,
0.05430322140455246,
-0.001035138382576406,
0.055125005543231964,
0.08529344201087952,
0.028523588553071022,
0.009388845413923264,
-0.0808371752500534,
0.03761933371424675,
-0.04383164271712303,
-0.06481911242008209,
-0.03537648916244507,
-0.02230054885149002,
0.07082968950271606,
0.026524001732468605,
-0.060476139187812805,
0.036245524883270264,
-0.00003569589171092957,
0.02686990611255169,
0.06959480792284012,
0.00694689154624939,
0.06843733042478561,
-0.009219259023666382,
-0.04062410816550255
] |
https://github.com/aimagelab/COGT
| 8
| 0
| 0
| 0
|
Mixture of Attentions For Speculative Decoding
|
https://openreview.net/forum?id=Rz0kozh3LE
|
[
"Matthieu Zimmer",
"Milan Gritta",
"Gerasimos Lampouras",
"Haitham Bou Ammar",
"Jun Wang"
] |
Poster
|
The growth in the number of parameters of Large Language Models (LLMs) has led to a significant surge in computational requirements, making them challenging and costly to deploy.
Speculative decoding (SD) leverages smaller models to efficiently propose future tokens, which are then verified by the LLM in parallel.
Small models that utilise activations from the LLM currently achieve the fastest decoding speeds.
However, we identify several limitations of SD models including the lack of on-policyness during training and partial observability.
To address these shortcomings, we propose a more grounded architecture for small models by introducing a Mixture of Attentions for SD.
Our novel architecture can be applied in two scenarios: a conventional single device deployment and a novel client-server deployment where the small model is hosted on a consumer device and the LLM on a server.
In a single-device scenario, we demonstrate state-of-the-art speedups improving EAGLE-2 by 9.5% and its acceptance length by 25%.
In a client-server setting, our experiments demonstrate: 1) state-of-the-art latencies with minimal calls to the server for different network conditions, and 2) in the event of a complete disconnection, our approach can maintain higher accuracy compared to other SD methods and demonstrates advantages over API calls to LLMs, which would otherwise be unable to continue the generation process.
|
large language models, speculative decoding, EAGLE
|
Using dynamical system view of LLM for better speculative decoding architecture
| 11,328
|
2410.03804
|
[
0.0140153169631958,
-0.08872541785240173,
0.04316173493862152,
-0.0104137621819973,
0.05183835327625275,
-0.08646940439939499,
0.0044533200562000275,
-0.013516864739358425,
0.0482105128467083,
-0.03319426253437996,
-0.046487923711538315,
-0.028850143775343895,
-0.00963953323662281,
-0.05534779652953148,
0.06344833970069885,
-0.010506507940590382,
0.12595508992671967,
-0.012504984624683857,
-0.05006363242864609,
-0.018058206886053085,
-0.02535483054816723,
-0.030430341139435768,
0.021385280415415764,
0.012346162460744381,
0.036235447973012924,
-0.037816502153873444,
-0.05802176147699356,
-0.024074381217360497,
-0.0076533532701432705,
0.03141280263662338,
0.09179271012544632,
0.03035297617316246,
-0.04204238951206207,
0.0728677436709404,
-0.03144966810941696,
-0.018030058592557907,
-0.03456514701247215,
-0.044260792434215546,
0.003770295064896345,
-0.09999120235443115,
-0.01416721660643816,
-0.0014454592019319534,
-0.01582183875143528,
0.09946414828300476,
0.04693688079714775,
-0.05135194957256317,
0.052628740668296814,
-0.049995310604572296,
0.005195083562284708,
-0.06627005338668823,
-0.03568509966135025,
-0.1090826541185379,
-0.046933457255363464,
0.049384962767362595,
-0.04690110310912132,
0.014587333425879478,
-0.004193875938653946,
0.027766522020101547,
-0.00020256370771676302,
0.05581537261605263,
-0.07122010737657547,
-0.09647858887910843,
-0.02337143011391163,
0.00819134060293436,
-0.04676227644085884,
0.019969791173934937,
0.10397972166538239,
-0.010325216688215733,
0.016415514051914215,
-0.016383135691285133,
-0.008632608689367771,
0.03907100856304169,
0.03469314053654671,
0.018121587112545967,
-0.023365452885627747,
0.026099398732185364,
0.06562084704637527,
0.015914252027869225,
0.08442024886608124,
-0.08729609847068787,
-0.009656660258769989,
-0.012308478355407715,
0.0033533363603055477,
-0.03292191028594971,
0.07613488286733627,
-0.03216295316815376,
0.02641182951629162,
0.0493699349462986,
0.0016885616350919008,
0.02778521552681923,
-0.04054287075996399,
-0.08160437643527985,
0.050183169543743134,
0.01195475272834301,
-0.019164541736245155,
0.08079688251018524,
-0.023431137204170227,
-0.0863170251250267,
-0.0816793218255043,
0.10288570821285248,
-0.020526835694909096,
0.08315378427505493,
0.028792448341846466,
-0.06960581988096237,
-0.05283859372138977,
-0.041695065796375275,
0.023425782099366188,
0.0881403386592865,
0.028232349082827568,
0.019035235047340393,
-0.0007812128169462085,
0.08280432969331741,
0.009994449093937874,
-0.015571783296763897,
-0.004915331024676561,
0.036260269582271576,
-0.0011221071472391486,
-0.03791988641023636,
0.01771852932870388,
0.10894899070262909,
-0.06560800969600677,
0.03300370275974274,
0.0007038316107355058,
-0.007930204272270203,
0.032743871212005615,
0.001168187242001295,
-0.0016867463709786534,
3.067722045651983e-34,
0.02479378692805767,
0.07388981431722641,
-0.06660562753677368,
0.04244257137179375,
0.010127256624400616,
-0.08630771934986115,
0.1001640260219574,
-0.0032826100941747427,
-0.04519824683666229,
-0.05566053092479706,
0.028414741158485413,
-0.025205284357070923,
-0.00095558469183743,
0.05400943011045456,
0.06893636286258698,
-0.0370212122797966,
0.00900396890938282,
0.0679270550608635,
0.00016278955445159227,
-0.04822808504104614,
-0.0022175307385623455,
-0.05971745401620865,
-0.036109112203121185,
-0.05688850209116936,
0.08553668856620789,
0.007254139520227909,
0.039752181619405746,
-0.04149024933576584,
0.00015652737056370825,
0.037665560841560364,
-0.12175460904836655,
-0.049588367342948914,
0.02972368523478508,
0.048485442996025085,
0.0573858767747879,
-0.05362886190414429,
-0.019800404086709023,
-0.09209880232810974,
0.06644947826862335,
-0.05328596383333206,
-0.018563907593488693,
0.050905391573905945,
-0.05837533250451088,
0.02891126088798046,
-0.10838507860898972,
-0.06235690787434578,
-0.04945062845945358,
-0.046389151364564896,
-0.049960192292928696,
-0.012607252225279808,
0.05380970239639282,
0.016981443390250206,
-0.025113800540566444,
0.0065261018462479115,
-0.028222166001796722,
-0.05456620454788208,
0.06493210047483444,
-0.013205580413341522,
0.02966662310063839,
0.0810803547501564,
-0.009405804798007011,
-0.004558366723358631,
-0.020719176158308983,
0.050806641578674316,
0.03923489525914192,
0.014901744201779366,
-0.03564194217324257,
0.08756682276725769,
-0.020880520343780518,
-0.0008372238371521235,
-0.04668806493282318,
-0.029596175998449326,
0.02470608614385128,
-0.0716690942645073,
0.0531516894698143,
0.038908738642930984,
0.06457660347223282,
-0.07811851054430008,
0.01090582087635994,
-0.006064081098884344,
0.030796974897384644,
0.009714273735880852,
-0.054142434149980545,
0.08975037187337875,
-0.024259211495518684,
-0.026901891455054283,
-0.007921797223389149,
-0.04954585060477257,
-0.09112826734781265,
0.01832609251141548,
-0.05779653042554855,
-0.02697492390871048,
-0.007818691432476044,
-0.04394204542040825,
-0.02425277605652809,
-1.9166445721650716e-34,
-0.08105290681123734,
-0.012171588838100433,
-0.0936620905995369,
0.0592922568321228,
-0.018586380407214165,
-0.09707812964916229,
-0.04477931186556816,
0.026545392349362373,
-0.0751987174153328,
-0.04355532303452492,
-0.11224342882633209,
-0.07099440693855286,
0.07302282005548477,
-0.014628672040998936,
-0.012940853834152222,
-0.036733340471982956,
0.013942200690507889,
0.007683794479817152,
0.07649759948253632,
0.0835413783788681,
-0.012194494716823101,
0.098549984395504,
-0.04964456334710121,
0.024379054084420204,
0.041497498750686646,
0.030835721641778946,
-0.0962802916765213,
0.10613671690225601,
0.031212080270051956,
-0.10029096156358719,
-0.04276527836918831,
0.020100006833672523,
0.0316644012928009,
-0.06189922243356705,
-0.006101618986576796,
0.11714503914117813,
0.0637604296207428,
0.059204842895269394,
-0.0044890898279845715,
0.05838184058666229,
0.11270749568939209,
-0.09036759287118912,
-0.0607515387237072,
0.029676243662834167,
-0.04543083906173706,
0.050630539655685425,
-0.06676345318555832,
-0.004185572266578674,
-0.0014046452706679702,
0.011344224214553833,
-0.005858892109245062,
0.03974195942282677,
0.013694824650883675,
0.011768839322030544,
-0.056417353451251984,
-0.053797781467437744,
-0.04261688515543938,
-0.058182697743177414,
0.013072109781205654,
-0.011746073141694069,
0.009540763683617115,
-0.012191245332360268,
0.07100170105695724,
-0.053138311952352524,
0.004292610567063093,
-0.041095998138189316,
0.04703200235962868,
-0.004726187326014042,
-0.019273534417152405,
0.004363530315458775,
0.06866263598203659,
-0.017958015203475952,
-0.0637323185801506,
0.08444304764270782,
-0.011146344244480133,
0.041938383132219315,
0.03972668573260307,
-0.031044790521264076,
0.0018820929108187556,
0.033514298498630524,
-0.01526778843253851,
0.14604708552360535,
-0.015341352671384811,
0.03581870719790459,
0.08632870763540268,
0.019521288573741913,
0.08068732917308807,
0.02202746830880642,
-0.04302972927689552,
-0.03264470398426056,
-0.05737793445587158,
0.05043758824467659,
-0.01619907282292843,
0.08003503829240799,
-0.034017499536275864,
-4.547598209114767e-8,
-0.014573115855455399,
0.06275475770235062,
0.0338108204305172,
0.057456690818071365,
0.06074182316660881,
0.007895057089626789,
-0.04473418369889259,
-0.005762363784015179,
0.06370538473129272,
0.01454357523471117,
0.051464881747961044,
-0.06689520925283432,
-0.05387650430202484,
-0.027989177033305168,
-0.009452500380575657,
0.08324475586414337,
0.003272551577538252,
-0.05464993417263031,
-0.07632147520780563,
-0.044389571994543076,
-0.005636689718812704,
0.07439867407083511,
0.03069765493273735,
-0.037512678653001785,
0.06032576784491539,
-0.00036959705175831914,
0.05381094664335251,
0.08550523221492767,
0.03877824917435646,
0.0046089510433375835,
-0.04896920174360275,
0.027210481464862823,
-0.026884036138653755,
-0.029869860038161278,
0.010577285662293434,
0.050087545067071915,
0.07937157899141312,
0.005176497157663107,
0.03862284868955612,
0.026542063802480698,
0.05941726639866829,
-0.027088314294815063,
-0.07945006340742111,
0.040035780519247055,
0.0066514830105006695,
-0.02633575163781643,
-0.05468042567372322,
-0.0632937029004097,
0.014225238002836704,
-0.021068241447210312,
0.08788493275642395,
-0.05313633754849434,
-0.0822397843003273,
0.1165163442492485,
-0.04319235309958458,
0.03223108872771263,
0.017141245305538177,
-0.025577250868082047,
0.02212158776819706,
0.10453537851572037,
0.07373683899641037,
0.046419017016887665,
-0.07036292552947998,
0.032759539783000946
] |
https://github.com/huawei-noah/hebo
| 2,596
| 0
| 0
| 0
|
VL-Cache: Sparsity and Modality-Aware KV Cache Compression for Vision-Language Model Inference Acceleration
|
https://openreview.net/forum?id=HMrcv7Q4Ub
|
[
"Dezhan Tu",
"Danylo Vashchilenko",
"Yuzhe Lu",
"Panpan Xu"
] |
Poster
|
Vision-Language Models (VLMs) have demonstrated impressive performance across a versatile set of tasks. A key challenge in accelerating VLMs is storing and accessing the large Key-Value (KV) cache that encodes long visual contexts, such as images or videos. While existing KV cache compression methods are effective for Large Language Models (LLMs), directly migrating them to VLMs yields suboptimal accuracy and speedup. To bridge the gap, we propose VL-Cache, a novel KV cache compression recipe tailored for accelerating VLM inference. In this paper, we first investigate the unique sparsity pattern of VLM attention by distinguishing visual and text tokens in prefill and decoding phases. Based on these observations, we introduce a layer-adaptive sparsity-aware cache budget allocation method that effectively distributes the limited cache budget across different layers, further reducing KV cache size without compromising accuracy. Additionally, we develop a modality-aware token scoring policy to better evaluate the token importance. Empirical results on multiple benchmark datasets demonstrate that retaining only 10% of KV cache achieves accuracy comparable to that with full cache. In a speed benchmark, our method accelerates end-to-end latency of generating 100 tokens by up to 2.33x and speeds up decoding by up to 7.08x, while reducing the memory footprint of KV cache in GPU by 90%.
|
KV Cache Compression, Vision-Language Models, Inference Acceleration, Sparsity, Modality
|
We propose VL-Cache, a novel KV cache compression recipe tailored for vision-language model inference acceleration.
| 11,322
| null |
[
0.028007708489894867,
-0.04676717147231102,
-0.011759118176996708,
0.05219687893986702,
0.09102927148342133,
-0.010772550478577614,
-0.011540191248059273,
-0.045518357306718826,
0.07389699667692184,
-0.03802994266152382,
-0.019416162744164467,
-0.07024604827165604,
-0.027679841965436935,
0.016183078289031982,
0.020507462322711945,
-0.043931517750024796,
0.09956327080726624,
0.09709666669368744,
-0.09728529304265976,
-0.06230995059013367,
0.07934358716011047,
-0.006980949081480503,
0.05214913934469223,
-0.007967191748321056,
0.0685436949133873,
0.01685483008623123,
-0.07011966407299042,
-0.027768313884735107,
0.044973623007535934,
-0.01778549514710903,
-0.0004378710291348398,
0.018570519983768463,
0.0059725139290094376,
0.05746709555387497,
0.012535098940134048,
0.0570717416703701,
-0.06875371932983398,
-0.03555794060230255,
0.010613654740154743,
-0.03218381106853485,
-0.05034153535962105,
-0.009905697777867317,
-0.024671848863363266,
0.02232135273516178,
0.07494838535785675,
0.06376058608293533,
0.08918910473585129,
-0.014493418857455254,
-0.015457158908247948,
-0.05747659504413605,
-0.04013098031282425,
-0.03456653654575348,
-0.017731040716171265,
0.0029146713204681873,
-0.10203195363283157,
0.025112496688961983,
0.007652895990759134,
0.011836354620754719,
-0.02449868991971016,
0.029707707464694977,
-0.032758597284555435,
-0.10097659379243851,
0.0065080937929451466,
0.03319316729903221,
-0.09504788368940353,
-0.02014460787177086,
0.13121725618839264,
-0.019198689609766006,
-0.0008524121949449182,
-0.019264521077275276,
-0.04022015258669853,
0.07926291227340698,
0.028366342186927795,
0.02329361066222191,
0.026509761810302734,
0.07004210352897644,
0.03925946354866028,
-0.05286848545074463,
0.10674045234918594,
-0.15614216029644012,
0.018144208937883377,
-0.03168283402919769,
0.02728619985282421,
0.008281239308416843,
0.16380265355110168,
-0.010957418009638786,
0.05902307480573654,
-0.020486949011683464,
-0.014573261141777039,
-0.022875424474477768,
-0.08876433968544006,
-0.07764624059200287,
0.08414848148822784,
-0.02790599875152111,
0.0032201504800468683,
0.0028343370649963617,
-0.02511509507894516,
-0.033558864146471024,
-0.038660045713186264,
0.03787098452448845,
-0.01589205302298069,
0.03896942734718323,
0.022838501259684563,
0.014669940806925297,
0.04014680162072182,
0.021017396822571754,
0.08435037732124329,
0.12878821790218353,
-0.015180934220552444,
-0.023718589916825294,
0.03959361091256142,
0.08809226751327515,
0.053105052560567856,
-0.03398870676755905,
0.02751004695892334,
0.0010962975211441517,
-0.039634253829717636,
-0.03481178730726242,
0.07001885026693344,
0.045179907232522964,
-0.0057394192554056644,
-0.026853742077946663,
-0.04589296877384186,
-0.05168597400188446,
-0.0039000478573143482,
-0.08438912034034729,
0.010006730444729328,
2.8614866127525013e-33,
0.02847116068005562,
0.0844118669629097,
0.0010397412115707994,
0.03346921131014824,
0.03612302243709564,
-0.03018243797123432,
0.036029864102602005,
0.03863188251852989,
-0.03818775340914726,
0.01539161242544651,
-0.03020409308373928,
-0.044701240956783295,
-0.027626583352684975,
0.12754304707050323,
0.010847669094800949,
0.03885030373930931,
-0.03251287713646889,
0.02125716768205166,
-0.043571870774030685,
-0.018149979412555695,
0.03867117315530777,
-0.06110886484384537,
0.00864954013377428,
-0.07744612544775009,
0.01271379366517067,
0.019372165203094482,
0.07413464039564133,
-0.08709973841905594,
-0.042401619255542755,
0.012872993014752865,
-0.07450902462005615,
0.030919194221496582,
0.053441163152456284,
0.0227486751973629,
0.05142715200781822,
-0.019756382331252098,
0.06431417912244797,
-0.019307071343064308,
0.0348135307431221,
0.019621841609477997,
0.013045525178313255,
0.09424954652786255,
-0.014147195033729076,
-0.08241059631109238,
-0.13614453375339508,
-0.02618425525724888,
0.03007473237812519,
0.03486992418766022,
-0.06740688532590866,
0.0490603968501091,
0.09795276820659637,
-0.020323706790804863,
-0.042894747108221054,
0.009809847921133041,
-0.008254794403910637,
-0.06609223783016205,
0.08777044713497162,
0.013587480410933495,
0.04460477828979492,
0.05909806862473488,
-0.07591498643159866,
-0.028366226702928543,
0.012907898053526878,
0.10767100006341934,
0.020797165110707283,
0.012593092396855354,
0.025862358510494232,
0.03578375652432442,
-0.07086696475744247,
-0.013632845133543015,
-0.04674365743994713,
0.025535134598612785,
-0.0002037026424659416,
-0.0872943103313446,
0.023486202582716942,
0.010163280181586742,
-0.006671181879937649,
-0.07512912154197693,
-0.020956840366125107,
0.0410267636179924,
0.001665643067099154,
-0.014847141690552235,
-0.0071947299875319,
-0.024247871711850166,
-0.05790897458791733,
-0.052673377096652985,
0.052234817296266556,
-0.07878826558589935,
0.007579931057989597,
-0.0527663379907608,
-0.013971582986414433,
-0.0701330229640007,
0.052658095955848694,
-0.08461592346429825,
-0.012948510237038136,
-1.7887745492302376e-33,
0.018818791955709457,
0.03233771771192551,
-0.021945089101791382,
0.09568023681640625,
-0.012505429796874523,
-0.05274873971939087,
0.0392153337597847,
-0.002180361421778798,
-0.052515678107738495,
-0.08671512454748154,
-0.04975669085979462,
-0.06972881406545639,
0.00796880479902029,
0.050793468952178955,
0.0036058102268725634,
-0.0016897463938221335,
-0.027366185560822487,
-0.04399517551064491,
-0.02181815728545189,
0.031177770346403122,
-0.013586023822426796,
0.03410212695598602,
-0.0573798306286335,
-0.009013372473418713,
-0.020786695182323456,
0.07281584292650223,
-0.07856262475252151,
0.032092440873384476,
-0.01890326291322708,
-0.07034754008054733,
-0.03555688634514809,
0.017903557047247887,
-0.02718372456729412,
-0.09308340400457382,
-0.009494774974882603,
0.0755336657166481,
0.025300800800323486,
-0.007317360956221819,
-0.045003969222307205,
0.11893143504858017,
0.06964190304279327,
-0.0525457002222538,
-0.06058277562260628,
0.013520085252821445,
0.010996397584676743,
-0.0404452383518219,
-0.04756012186408043,
-0.05014042556285858,
-0.002746486570686102,
0.02606479451060295,
0.010979029349982738,
0.03555462136864662,
-0.10684099793434143,
0.037730105221271515,
-0.0031357151456177235,
-0.0338534377515316,
-0.12478186935186386,
0.023987863212823868,
0.07522930204868317,
-0.05924306809902191,
0.0031345421448349953,
-0.03529438003897667,
-0.04338657110929489,
-0.0355033241212368,
0.019076557829976082,
-0.045717231929302216,
0.007946528494358063,
-0.057668041437864304,
0.019252730533480644,
-0.034270595759153366,
0.05765712633728981,
-0.08029914647340775,
-0.0043141525238752365,
0.033782098442316055,
-0.012860806658864021,
-0.03473132848739624,
0.06252199411392212,
-0.021021569147706032,
0.03085084818303585,
0.023111233487725258,
-0.001140511129051447,
-0.01222107745707035,
0.02497071959078312,
0.03865445405244827,
0.0334959402680397,
0.14849026501178741,
0.01081917155534029,
0.030323641374707222,
-0.012465234845876694,
0.006742829456925392,
-0.06132671236991882,
0.03266672044992447,
0.009353323839604855,
0.06416868418455124,
0.02314375713467598,
-3.5510772988800454e-8,
-0.04768360033631325,
0.07609689980745316,
-0.0552622489631176,
0.023232631385326385,
-0.019420888274908066,
-0.04422172158956528,
-0.03859226778149605,
0.13266432285308838,
0.020119166001677513,
-0.024211399257183075,
0.07573693990707397,
0.007132244762033224,
-0.052355796098709106,
-0.04241003468632698,
-0.005883143283426762,
0.08495991677045822,
0.0247007105499506,
0.03076864220201969,
-0.0015985966892912984,
-0.016997074708342552,
-0.03304913267493248,
0.07014832645654678,
0.028708187863230705,
-0.06576209515333176,
0.01280977763235569,
-0.053256917744874954,
0.028887467458844185,
0.09988629072904587,
0.09645653516054153,
-0.07338283210992813,
0.025059187784790993,
0.06117483600974083,
0.024166453629732132,
-0.08274458348751068,
0.032595209777355194,
0.028376199305057526,
0.011021895334124565,
0.014743833802640438,
0.054055552929639816,
0.019259478896856308,
0.03420427814126015,
0.00006607387331314385,
-0.03291323408484459,
0.032539211213588715,
-0.01052100583910942,
0.019730111584067345,
-0.004864642862230539,
-0.07799068838357925,
-0.03259117528796196,
-0.053418923169374466,
0.008691668510437012,
-0.018386641517281532,
-0.051557011902332306,
0.1245889738202095,
-0.0486651286482811,
-0.013889815658330917,
0.07526154071092606,
-0.010019583627581596,
0.04914171248674393,
0.07592981308698654,
0.047053445130586624,
-0.03246285766363144,
-0.09976600855588913,
-0.05074695870280266
] | 0
| 0
| 0
| 0
|
|
Connectome Mapping: Shape-Memory Network via Interpretation of Contextual Semantic Information
|
https://openreview.net/forum?id=PZYr22zFyE
|
[
"Kyungsu Lee",
"Haeyun Lee",
"Jae Youn Hwang"
] |
Poster
|
Contextual semantic information plays a pivotal role in the brain's visual interpretation of the surrounding environment. When processing visual information, electrical signals within synapses facilitate the dynamic activation and deactivation of synaptic connections, guided by the contextual semantic information associated with different objects. In the realm of Artificial Intelligence (AI), neural networks have emerged as powerful tools to emulate complex signaling systems, enabling tasks such as classification and segmentation by understanding visual information. However, conventional neural networks have limitations in simulating the conditional activation and deactivation of synapses, collectively known as the connectome, a comprehensive map of neural connections in the brain. Additionally, the pixel-wise inference mechanism of conventional neural networks failed to account for the explicit utilization of contextual semantic information in the prediction process. To overcome these limitations, we developed a novel neural network, dubbed the Shape Memory Network (SMN), which excels in two key areas: (1) faithfully emulating the intricate mechanism of the brain's connectome, and (2) explicitly incorporating contextual semantic information during the inference process. The SMN memorizes the structure suitable for contextual semantic information and leverages this structure at the inference phase. The structural transformation emulates the conditional activation and deactivation of synaptic connections within the connectome. Rigorous experimentation carried out across a range of semantic segmentation benchmarks demonstrated the outstanding performance of the SMN, highlighting its superiority and effectiveness. Furthermore, our pioneering network on connectome emulation reveals the immense potential of the SMN for next-generation neural networks.
|
neural representation
| null | 11,312
| null |
[
0.00246358010917902,
-0.0995352491736412,
0.039729513227939606,
0.0945492833852768,
0.01701185293495655,
-0.0008349026902578771,
0.021004147827625275,
-0.02196977287530899,
0.04065127298235893,
-0.06377661228179932,
-0.024124253541231155,
-0.026540661230683327,
0.016025127843022346,
0.05111100152134895,
-0.022647799924016,
0.050130199640989304,
0.02690587379038334,
0.05952061712741852,
-0.09334257990121841,
-0.014907052740454674,
0.057746488600969315,
-0.0024029496125876904,
0.0005830738809891045,
-0.029766103252768517,
-0.0034093125723302364,
0.07478977739810944,
0.036596257239580154,
-0.012601775117218494,
-0.010498818941414356,
-0.027191154658794403,
0.08409596979618073,
0.035220615565776825,
-0.05294210463762283,
0.031130628660321236,
-0.03752736747264862,
0.0286291241645813,
-0.05385751649737358,
0.02399466186761856,
0.0010271357605233788,
-0.08303564786911011,
0.032486479729413986,
-0.0340205617249012,
0.005929266568273306,
0.006234127562493086,
0.11417952179908752,
0.1510082483291626,
-0.01790975034236908,
-0.028870126232504845,
-0.036038096994161606,
-0.01685541681945324,
-0.06184714287519455,
0.010563448071479797,
-0.03968290612101555,
-0.013812735676765442,
0.03724590316414833,
0.08106422424316406,
0.013901502825319767,
-0.0076744044199585915,
-0.052705295383930206,
0.03707406669855118,
0.016477184370160103,
-0.06221359595656395,
0.0899941697716713,
-0.027554495260119438,
-0.028691258281469345,
-0.03776918724179268,
0.05543149635195732,
0.021256858482956886,
0.035624727606773376,
-0.005780011881142855,
0.03474310413002968,
0.07061315327882767,
-0.04204924404621124,
-0.08897159993648529,
-0.007289256900548935,
0.0940752774477005,
0.003891770029440522,
-0.03466310352087021,
0.010955684818327427,
-0.09037195891141891,
0.023961132392287254,
0.07873664796352386,
0.02936503477394581,
0.06169431656599045,
0.12792840600013733,
0.1024026870727539,
0.04550883173942566,
0.025252854451537132,
-0.09756894409656525,
0.0345967598259449,
-0.04573022574186325,
-0.09256048500537872,
-0.019719157367944717,
-0.0053858449682593346,
0.08042118698358536,
-0.019814157858490944,
-0.019043724983930588,
-0.0779159665107727,
-0.036280401051044464,
0.02233678661286831,
-0.0034223643597215414,
0.02079799398779869,
0.046772994101047516,
-0.031585779041051865,
0.044461678713560104,
-0.06011006236076355,
0.1033165380358696,
0.025278210639953613,
0.0031732902862131596,
-0.0510571151971817,
0.09475692361593246,
-0.007125035859644413,
0.010791043750941753,
0.03579220920801163,
-0.03514404967427254,
-0.05418100208044052,
-0.017428763210773468,
0.04174722731113434,
0.12017670273780823,
0.052359215915203094,
-0.051137298345565796,
-0.007192415185272694,
-0.09441307932138443,
-0.025177819654345512,
-0.07809273898601532,
-0.0898115485906601,
-0.1163562759757042,
8.317572111097958e-34,
-0.07165397703647614,
-0.06526117771863937,
0.0531722754240036,
-0.03189415857195854,
0.07381980121135712,
-0.009273583069443703,
-0.08289077132940292,
0.01827150397002697,
0.0489308163523674,
0.008471711538732052,
-0.06901608407497406,
-0.01893148571252823,
-0.04281684011220932,
0.11159998923540115,
0.04124144837260246,
-0.0700545534491539,
-0.058762211352586746,
0.02008015848696232,
-0.027958091348409653,
-0.021180417388677597,
-0.0017585604218766093,
-0.025453895330429077,
-0.025225497782230377,
-0.01954157091677189,
0.006714952178299427,
0.02403060346841812,
0.011383787728846073,
0.021324006840586662,
-0.006803658790886402,
-0.029319746419787407,
-0.06419471651315689,
0.07326357811689377,
0.0032505602575838566,
-0.06184813380241394,
0.028982164338231087,
-0.056001245975494385,
0.008700997568666935,
-0.051340680569410324,
-0.025531619787216187,
0.014958796091377735,
0.02835369110107422,
0.012968170456588268,
-0.02615787461400032,
0.02568145841360092,
-0.07873842120170593,
-0.04787139967083931,
0.029924027621746063,
0.07798244059085846,
-0.040254026651382446,
-0.11155984550714493,
0.04431001842021942,
-0.025653384625911713,
0.029067128896713257,
-0.08019892871379852,
-0.0039641233161091805,
-0.04927496239542961,
-0.04605404660105705,
0.00019499384507071227,
0.045376963913440704,
0.011160388588905334,
-0.022757094353437424,
-0.02272966131567955,
-0.024670934304594994,
0.0707627385854721,
-0.014507290907204151,
0.08195673674345016,
-0.02629660815000534,
0.07239881902933121,
-0.0070397984236478806,
0.01689162477850914,
0.02025032602250576,
0.11197364330291748,
-0.05080409720540047,
-0.038471758365631104,
-0.027168545871973038,
-0.01579836755990982,
-0.014140536077320576,
0.016349777579307556,
-0.06708674877882004,
0.09273339062929153,
0.046659257262945175,
-0.032468948513269424,
-0.08365500718355179,
0.01967182196676731,
0.0008338479092344642,
-0.003838767996057868,
0.09566713124513626,
-0.08580342680215836,
-0.08314742892980576,
-0.011896696873009205,
-0.030934881418943405,
-0.052085667848587036,
0.05684135481715202,
-0.008623583242297173,
-0.08405447751283646,
-2.613557405419811e-33,
-0.020742405205965042,
0.039953380823135376,
0.02394672855734825,
0.0041624633595347404,
-0.00047885673120617867,
-0.009197497740387917,
0.04213050752878189,
-0.03570254147052765,
-0.06337252259254456,
-0.019259363412857056,
0.0033553282264620066,
0.02796524204313755,
0.026951340958476067,
-0.034951552748680115,
0.04341946914792061,
0.03516183793544769,
-0.06543106585741043,
0.01674702949821949,
0.039704982191324234,
-0.01965818740427494,
-0.02251971699297428,
0.05172071233391762,
-0.02035897597670555,
-0.05089949443936348,
-0.03699452057480812,
0.05404645577073097,
-0.032498545944690704,
0.002372096059843898,
-0.041662171483039856,
0.07303746789693832,
-0.02953571081161499,
0.010160916484892368,
-0.04768228903412819,
0.019677095115184784,
-0.057132042944431305,
0.06338489055633545,
0.09363031387329102,
-0.09561610221862793,
-0.0687374547123909,
-0.02384103089570999,
0.024857984855771065,
-0.03626558557152748,
-0.017537754029035568,
0.05770577862858772,
-0.0310569666326046,
-0.04200237616896629,
-0.0068514044396579266,
0.04581645131111145,
-0.0699615329504013,
0.07469866424798965,
0.029359668493270874,
0.0588667206466198,
-0.08607681840658188,
-0.07818640023469925,
-0.060078565031290054,
0.011209677904844284,
-0.023899652063846588,
-0.01187707856297493,
0.09091731905937195,
-0.018662957474589348,
-0.01471058838069439,
-0.10791226476430893,
0.01857307367026806,
0.06365177780389786,
-0.02008677087724209,
0.02259085141122341,
-0.033645015209913254,
-0.02130487933754921,
-0.000935533142182976,
0.045441798865795135,
0.008872609585523605,
0.04066683351993561,
0.018364718183875084,
0.02808351069688797,
0.028686530888080597,
-0.04743260145187378,
-0.028331482782959938,
0.01569005288183689,
-0.020759964361786842,
-0.017188986763358116,
0.01794314943253994,
-0.1005696952342987,
0.02769250050187111,
0.07531357556581497,
0.06259672343730927,
0.05384272336959839,
-0.04296153411269188,
0.02922733873128891,
0.08738017827272415,
0.07936683297157288,
-0.03020474500954151,
0.01509374100714922,
-0.08799638599157333,
0.07796186953783035,
-0.08382687717676163,
-4.044863288754641e-8,
-0.0532870851457119,
-0.018064063042402267,
0.05135231465101242,
-0.03729820251464844,
0.017305660992860794,
-0.059879641979932785,
0.03392921760678291,
0.08211515843868256,
-0.013285944238305092,
-0.05019749328494072,
0.06186816096305847,
-0.0628124475479126,
-0.10364272445440292,
-0.059557631611824036,
0.03009689413011074,
0.05281743034720421,
0.08205508440732956,
-0.006245909258723259,
0.059376221150159836,
0.006165476515889168,
0.011986021883785725,
-0.03363999351859093,
-0.02139255218207836,
0.06525378674268723,
0.0804392397403717,
-0.06990855187177658,
-0.011339370161294937,
0.11342642456293106,
-0.027032485231757164,
-0.0596817247569561,
0.08346022665500641,
0.005380504764616489,
0.01697712391614914,
0.02570229209959507,
-0.08003900200128555,
0.0456407330930233,
0.022334236651659012,
-0.03765809163451195,
-0.048674073070287704,
0.0017908693989738822,
0.009461562149226665,
0.052497345954179764,
0.03710565343499184,
0.02155831642448902,
0.04625292122364044,
-0.017479168251156807,
0.09659703075885773,
-0.01990736648440361,
0.01648411713540554,
0.04320697858929634,
-0.05129909887909889,
0.0936809554696083,
0.03924596682190895,
0.11379913985729218,
0.007574457209557295,
-0.03371294587850571,
0.017264410853385925,
0.010750152170658112,
-0.02929818443953991,
0.062252528965473175,
-0.02702960930764675,
0.08013418316841125,
-0.013358798809349537,
-0.07285615056753159
] | 0
| 0
| 0
| 0
|
|
Learning Color Equivariant Representations
|
https://openreview.net/forum?id=IXyfbaGlps
|
[
"Yulong Yang",
"Felix O'Mahony",
"Christine Allen-Blanchette"
] |
Poster
|
In this paper, we introduce group convolutional neural networks (GCNNs) equivariant to color variation. GCNNs have been designed for a variety of geometric transformations from 2D and 3D rotation groups, to semi-groups such as scale. Despite the improved interpretability, accuracy and generalizability of these architectures, GCNNs have seen limited application in the context of perceptual quantities. Notably, the recent CEConv network uses a GCNN to achieve equivariance to hue transformations by convolving input images with a hue rotated RGB filter. However, this approach leads to invalid RGB values which break equivariance and degrade performance. We resolve these issues with a lifting layer that transforms the input image directly, thereby circumventing the issue of invalid RGB values and improving equivariance error by over three orders of magnitude. Moreover, we extend the notion of color equivariance to include equivariance to saturation and luminance shift. Our hue-, saturation-, luminance- and color-equivariant networks achieve strong generalization to out-of-distribution perceptual variations and improved sample efficiency over conventional architectures. We demonstrate the utility of our approach on synthetic and real world datasets where we consistently outperform competitive baselines.
|
Equivariant Neural Network, Geometric Deep Learning, Group Convolution
|
We achieve color equivariance in a convolutional neural network architecture by identifying hue and saturation transformations with the 2D rotation and 1D translation groups.
| 11,305
|
2406.09588
|
[
-0.058520250022411346,
-0.08133424818515778,
0.0721694603562355,
-0.012547054328024387,
0.059398114681243896,
0.02241750806570053,
0.03722498193383217,
-0.10208877921104431,
0.06372854858636856,
-0.09106995165348053,
-0.038269974291324615,
-0.044375400990247726,
0.001578778843395412,
0.041323624551296234,
-0.08341140300035477,
-0.012932518497109413,
0.028844544664025307,
0.1406351923942566,
-0.1535986214876175,
-0.03364818915724754,
-0.000643959385342896,
-0.05844280123710632,
-0.01744087040424347,
0.003930555656552315,
0.047290392220020294,
-0.005908282473683357,
0.01797672174870968,
-0.02311415784060955,
-0.028247898444533348,
-0.036046892404556274,
0.011509458534419537,
0.007097173947840929,
-0.004169262479990721,
-0.006654838565737009,
-0.08634698390960693,
0.04251488298177719,
0.003654481377452612,
-0.015703970566391945,
-0.09182281792163849,
-0.06176365911960602,
0.0026818844489753246,
0.02636740356683731,
0.004270292818546295,
0.008063247427344322,
0.08243295550346375,
0.06250975281000137,
0.08011915534734726,
-0.03792894259095192,
0.013744540512561798,
0.016462750732898712,
0.06383045762777328,
0.010973595082759857,
-0.07280801236629486,
0.13985860347747803,
-0.013052307069301605,
0.03362404182553291,
0.022932423278689384,
-0.05737684667110443,
0.04399813339114189,
-0.03341323137283325,
0.00986449047923088,
0.014610342681407928,
-0.003108325181528926,
0.02449670061469078,
-0.04090961441397667,
-0.021270807832479477,
0.08314479887485504,
-0.03894156217575073,
0.018002541735768318,
-0.06439999490976334,
0.004754927009344101,
0.0614410899579525,
-0.011089435778558254,
0.06311479210853577,
0.03853435814380646,
0.08383964002132416,
0.04253249987959862,
-0.010777171701192856,
-0.00495104817673564,
-0.08266622573137283,
0.0777890756726265,
0.04964384809136391,
0.0619136206805706,
0.08010455220937729,
0.09325893223285675,
0.030996156856417656,
-0.09767212718725204,
0.08220301568508148,
-0.03263768553733826,
0.016489209607243538,
0.004833484999835491,
0.05269855633378029,
-0.0009460880537517369,
-0.03454781323671341,
0.06197252497076988,
-0.022709937766194344,
0.011109155602753162,
-0.10640136897563934,
-0.01646946743130684,
0.049527857452631,
0.00509019335731864,
0.005521741230040789,
0.0002242932969238609,
-0.03642427921295166,
0.041119638830423355,
0.025337891653180122,
0.08566436916589737,
0.01125004980713129,
0.030570950359106064,
-0.10146861523389816,
0.011311346665024757,
0.015302415937185287,
0.0068949321284890175,
-0.0631602555513382,
-0.017846200615167618,
0.09837709367275238,
-0.037013351917266846,
0.025426248088479042,
0.0021893801167607307,
0.01832314394414425,
-0.030042195692658424,
-0.010568744502961636,
-0.07089708000421524,
0.017722666263580322,
0.023449750617146492,
-0.04494800418615341,
-0.06312250345945358,
2.8568698586896468e-33,
0.015539041720330715,
0.02779882214963436,
0.03523273766040802,
0.06641894578933716,
0.008408844470977783,
0.03595317155122757,
-0.017912378534674644,
0.0358135886490345,
-0.020338138565421104,
-0.08620257675647736,
-0.06000947952270508,
0.027642741799354553,
-0.07653000950813293,
0.052514754235744476,
-0.009869718924164772,
-0.0053317430429160595,
-0.010399880819022655,
-0.061542343348264694,
0.020001666620373726,
-0.01081994827836752,
-0.05482664331793785,
-0.009772171266376972,
0.029398752376437187,
0.007111279293894768,
-0.0911506712436676,
-0.01021549291908741,
0.004455473739653826,
0.06897271424531937,
0.06875171512365341,
-0.017481377348303795,
-0.006491526961326599,
0.026618996635079384,
0.0015108867082744837,
0.04723532497882843,
-0.05579257383942604,
-0.09215318411588669,
0.013088032603263855,
-0.009889882057905197,
0.07411191612482071,
0.02411773055791855,
-0.04017755389213562,
0.12787942588329315,
-0.057415157556533813,
0.028961045667529106,
-0.03192923590540886,
0.02508017048239708,
0.05582080036401749,
0.023406770080327988,
-0.061253275722265244,
0.04163036495447159,
0.04614781588315964,
0.002589380368590355,
0.007023498881608248,
-0.10269923508167267,
0.0005830817390233278,
-0.01377010066062212,
0.16100919246673584,
0.059531401842832565,
0.0491899698972702,
-0.019834045320749283,
0.02011631801724434,
0.11400312185287476,
-0.018095308914780617,
0.035310328006744385,
0.0030082722660154104,
0.004158012103289366,
0.06273876130580902,
0.09709742665290833,
-0.028661221265792847,
-0.00982392206788063,
-0.03709514066576958,
0.004165487829595804,
-0.022523634135723114,
-0.03600539267063141,
-0.013487273827195168,
-0.002949134213849902,
-0.007772084791213274,
-0.05914636701345444,
0.006774064619094133,
0.04665753245353699,
-0.029427437111735344,
0.016600528731942177,
0.010759170167148113,
-0.008831753395497799,
-0.0745539739727974,
0.026701856404542923,
-0.022063810378313065,
-0.058881521224975586,
-0.016403621062636375,
-0.007309630047529936,
-0.06564712524414062,
-0.027554720640182495,
0.014236656948924065,
-0.0396297425031662,
0.016947969794273376,
-1.801142402839819e-33,
0.011526466347277164,
0.07609067857265472,
-0.11214720457792282,
0.10824325680732727,
0.05061376467347145,
-0.06281017512083054,
-0.00044645584421232343,
0.008754058741033077,
0.04174862429499626,
0.005842836108058691,
0.08145730197429657,
-0.014780118130147457,
-0.035598501563072205,
0.04444492608308792,
0.008754263631999493,
-0.0615689717233181,
-0.052802931517362595,
0.04260127246379852,
-0.09662777930498123,
0.026987096294760704,
-0.028167862445116043,
0.037411559373140335,
-0.039158277213573456,
0.035622138530015945,
-0.10919494926929474,
0.07973212748765945,
-0.023460187017917633,
0.02587670274078846,
0.005858455318957567,
-0.007552753668278456,
-0.06330736726522446,
-0.019298987463116646,
0.0016825159545987844,
0.061467889696359634,
-0.02065412327647209,
0.01832740008831024,
0.033233702182769775,
-0.025138599798083305,
-0.07728009670972824,
0.047722939401865005,
0.0580785870552063,
-0.03217772766947746,
-0.04738348722457886,
-0.01303733792155981,
-0.05348147451877594,
-0.0019203531555831432,
0.06311634182929993,
0.011001387611031532,
0.002785619581118226,
0.09909112006425858,
-0.011279950849711895,
0.04623817652463913,
-0.06323768198490143,
0.059389546513557434,
0.08296480774879456,
-0.03580200672149658,
-0.0023854095488786697,
0.0023194425739347935,
0.10026964545249939,
0.05411577224731445,
-0.05289141461253166,
-0.057469442486763,
-0.08978594094514847,
-0.0410076379776001,
-0.054682791233062744,
0.03728007152676582,
-0.01494139526039362,
0.013204666785895824,
0.04677349701523781,
0.040796853601932526,
0.02319500967860222,
-0.1311056911945343,
0.05276074633002281,
-0.06668246537446976,
-0.06678058207035065,
-0.10568741708993912,
0.031003030017018318,
-0.02423003315925598,
-0.014210217632353306,
-0.03931388631463051,
0.028925791382789612,
-0.09898672252893448,
0.07109013944864273,
0.067405566573143,
0.06647045165300369,
0.057566575706005096,
-0.03431688994169235,
-0.007478352636098862,
0.10710344463586807,
0.02861744537949562,
-0.006322305649518967,
0.03335338458418846,
0.004387901164591312,
0.008092832751572132,
-0.00743549969047308,
-4.012825982613322e-8,
-0.025232801213860512,
0.030533839017152786,
0.04815050587058067,
-0.0018161978805437684,
0.018802590668201447,
-0.09307844936847687,
0.06913075596094131,
0.042082760483026505,
-0.003470060182735324,
0.07145354151725769,
0.0448751300573349,
-0.046038806438446045,
-0.01073960680514574,
-0.09036906063556671,
-0.06326347589492798,
0.06287255138158798,
-0.012013879604637623,
0.01802028901875019,
0.03606277331709862,
0.04818495362997055,
-0.045463062822818756,
0.0019991255830973387,
0.008900308050215244,
0.003975783009082079,
-0.002198986941948533,
-0.10016726702451706,
-0.1495247632265091,
0.040864501148462296,
0.07153642922639847,
-0.04600737243890762,
0.008654400706291199,
-0.005141458474099636,
0.04550545662641525,
-0.052219633013010025,
-0.018476812168955803,
0.029147999361157417,
-0.06222947686910629,
0.04477919638156891,
-0.015584194101393223,
-0.08138912916183472,
-0.041713811457157135,
0.0854916125535965,
-0.019711581990122795,
0.007077417802065611,
-0.022511474788188934,
0.04509346932172775,
0.07664400339126587,
0.02323189750313759,
-0.030626634135842323,
0.03196857124567032,
0.020418968051671982,
-0.04179868474602699,
-0.042891986668109894,
0.07395903021097183,
0.002889493713155389,
-0.07405827939510345,
-0.009068591520190239,
-0.006710650399327278,
0.05397466942667961,
0.03049071878194809,
-0.053407829254865646,
-0.00520473625510931,
-0.046138305217027664,
-0.05616004392504692
] |
https://github.com/cab-lab-princeton/learning-color-equivariant-representations
| 0
| 0
| 0
| 0
|
Omni-MATH: A Universal Olympiad Level Mathematic Benchmark for Large Language Models
|
https://openreview.net/forum?id=yaqPf0KAlN
|
[
"Bofei Gao",
"Feifan Song",
"Zhe Yang",
"Zefan Cai",
"Yibo Miao",
"Qingxiu Dong",
"Lei Li",
"Chenghao Ma",
"Liang Chen",
"Runxin Xu",
"Zhengyang Tang",
"Benyou Wang",
"Daoguang Zan",
"Shanghaoran Quan",
"Ge Zhang",
"Lei Sha",
"Yichang Zhang",
"Xuancheng Ren",
"Tianyu Liu",
"Baobao Chang"
] |
Poster
|
Recent advancements in large language models (LLMs) have led to significant breakthroughs in mathematical reasoning capabilities.
However, existing benchmarks like GSM8K or MATH are now being solved with high accuracy (e.g., OpenAI o1 achieves 94.8% on MATH dataset), indicating their inadequacy for truly challenging these models. To bridge this gap, we propose a comprehensive and challenging benchmark specifically designed to assess LLMs' mathematical reasoning at the Olympiad level. Unlike existing Olympiad-related benchmarks, our dataset focuses exclusively on mathematics and comprises a vast collection of 4428 competition-level problems with rigorous human annotation. These problems are meticulously categorized into over 33 sub-domains and span more than 10 distinct difficulty levels, enabling a holistic assessment of model performance in Olympiad-mathematical reasoning. Furthermore, we conducted an in-depth analysis based on this benchmark. Our experimental results show that even the most advanced models, OpenAI o1-mini and OpenAI o1-preview, struggle with highly challenging Olympiad-level problems, with 60.54% and 52.55% accuracy, highlighting significant challenges in Olympiad-level mathematical reasoning.
|
Mathematical Benchmark, LLM Evaluation, Olympic
|
We propose a comprehensive and challenging benchmark specifically designed to assess LLMs' mathematical reasoning at the Olympiad level.
| 11,296
| null |
[
-0.05661306530237198,
0.007672843988984823,
0.013061664998531342,
0.0007039877236820757,
0.048570748418569565,
-0.04021325334906578,
-0.013693670742213726,
0.043052881956100464,
0.011375559493899345,
-0.01591864787042141,
-0.1086583137512207,
-0.08523336797952652,
0.040634509176015854,
0.08688655495643616,
0.036009613424539566,
0.07003366947174072,
0.07205860316753387,
0.052568841725587845,
-0.07925919443368912,
-0.07858827710151672,
0.013848314993083477,
-0.006326746195554733,
0.03322068229317665,
-0.02591068111360073,
0.06704752147197723,
-0.06446752697229385,
-0.004076593089848757,
-0.0576225221157074,
0.005247982684522867,
0.019649064168334007,
-0.048818062990903854,
-0.020259486511349678,
0.026706524193286896,
0.05647741258144379,
-0.014423741959035397,
0.057765893638134,
-0.06719589978456497,
-0.004723259247839451,
-0.02812913991510868,
-0.034211140125989914,
-0.019692230969667435,
-0.022286735475063324,
0.0848427563905716,
-0.013720513321459293,
0.10139109939336777,
-0.03160148859024048,
-0.019955897703766823,
0.022583046928048134,
-0.004609452560544014,
0.010071277618408203,
-0.15058685839176178,
-0.042905837297439575,
-0.031006667762994766,
-0.005754042416810989,
-0.03760507330298424,
0.01466416846960783,
0.03443392738699913,
-0.02798818051815033,
0.024595875293016434,
-0.0541657879948616,
-0.07368055731058121,
-0.09287285059690475,
-0.022512609139084816,
0.02799231931567192,
-0.006369920913130045,
-0.0387289859354496,
0.05352305993437767,
0.03514140844345093,
-0.05025658756494522,
0.07740727066993713,
0.047615211457014084,
0.10255347192287445,
-0.011835875920951366,
0.06145034730434418,
0.04328981041908264,
0.08467588573694229,
0.005113835446536541,
0.020216284319758415,
0.07862566411495209,
-0.06020636111497879,
0.0034156334586441517,
-0.08233556151390076,
-0.014450686983764172,
0.0198654904961586,
0.11266151815652847,
-0.009888461790978909,
-0.02825024537742138,
0.02246808260679245,
0.03564294055104256,
0.003387471893802285,
-0.0074266488663852215,
-0.04901272803544998,
0.0684715285897255,
0.08342956751585007,
0.061074450612068176,
0.05897168442606926,
0.029454849660396576,
-0.03056427463889122,
-0.07470492273569107,
0.14470864832401276,
0.0019168724538758397,
0.10245110094547272,
0.0061571658588945866,
-0.023105919361114502,
0.019250718876719475,
0.014819533564150333,
0.08539136499166489,
-0.0017801632639020681,
0.0026986219454556704,
-0.08038262277841568,
0.07730306684970856,
-0.023365598171949387,
0.07815349102020264,
0.06162251532077789,
0.022099703550338745,
-0.007630230393260717,
0.012948616407811642,
-0.011628782376646996,
0.036968447268009186,
0.005774144548922777,
-0.013624885119497776,
0.04283219575881958,
0.022270791232585907,
-0.011792805977165699,
0.09346688538789749,
0.06332249194383621,
-0.03551333397626877,
2.3293298508886897e-33,
0.07090885192155838,
0.05977686494588852,
0.09253664314746857,
0.03664033114910126,
0.03018360026180744,
-0.03890623152256012,
-0.002722629811614752,
0.04651516675949097,
-0.0297636017203331,
0.010841796174645424,
-0.04569447785615921,
0.018908437341451645,
-0.09768633544445038,
0.016824593767523766,
0.07112973183393478,
-0.012272222898900509,
-0.017300834879279137,
-0.004280490800738335,
-0.07418643683195114,
-0.002652000403031707,
0.05039556324481964,
-0.06335568428039551,
-0.006715610157698393,
-0.03857626020908356,
0.02832740731537342,
0.08785366266965866,
0.08072169125080109,
-0.09668496996164322,
-0.05028796195983887,
0.01593305543065071,
-0.0695107951760292,
-0.037794169038534164,
-0.031298913061618805,
0.0578954741358757,
0.00587222445756197,
-0.06005145609378815,
0.0011151607614010572,
-0.05420877784490585,
0.03456403687596321,
0.0028509125113487244,
-0.07651490718126297,
0.005351774860173464,
0.022677624598145485,
0.0339200533926487,
-0.03784535080194473,
-0.03880659118294716,
-0.027757473289966583,
0.04368895664811134,
0.0051438333466649055,
0.027278374880552292,
-0.010895025916397572,
0.020115375518798828,
-0.0716329887509346,
-0.04706975445151329,
0.018051229417324066,
-0.03904716297984123,
-0.018701406195759773,
0.016191288828849792,
0.009997107088565826,
0.03857404366135597,
0.011669362895190716,
0.06349635124206543,
0.01967022754251957,
0.023762419819831848,
-0.06196289882063866,
0.08069372922182083,
-0.09267519414424896,
0.08687583357095718,
0.10446178913116455,
0.02525158040225506,
-0.030266614630818367,
-0.013650942593812943,
0.02654843218624592,
0.01609581522643566,
0.0496007576584816,
0.022262563928961754,
0.05341162905097008,
-0.0903601124882698,
0.01861480064690113,
-0.01290399581193924,
0.04909933730959892,
-0.005514319986104965,
-0.0019712287466973066,
-0.05152488872408867,
-0.08976250886917114,
-0.043513040989637375,
0.03012143075466156,
-0.06841707974672318,
-0.01826840452849865,
-0.016927603632211685,
-0.03713337704539299,
-0.02421519160270691,
-0.020784955471754074,
-0.009983344003558159,
0.000490207748953253,
-2.2613827876623075e-33,
-0.08377411961555481,
-0.04100269079208374,
-0.06650537252426147,
0.056619055569171906,
-0.007017277646809816,
-0.09933958202600479,
-0.036844633519649506,
0.026230519637465477,
0.014566163532435894,
-0.020174529403448105,
-0.018778633326292038,
-0.038026440888643265,
0.06623934209346771,
-0.004114898853003979,
0.01730247214436531,
-0.10201160609722137,
-0.04830561950802803,
0.023197971284389496,
-0.007789295632392168,
0.08577378839254379,
-0.0023894337937235832,
0.12498190999031067,
-0.08356793969869614,
-0.03467966616153717,
-0.012826616875827312,
0.021688522771000862,
-0.05098491162061691,
-0.006976888049393892,
-0.030352214351296425,
0.023783888667821884,
-0.016831614077091217,
-0.06202828511595726,
-0.08375214040279388,
-0.055585410445928574,
-0.06585846841335297,
0.03367139399051666,
0.07229716330766678,
-0.07607784122228622,
0.01760607585310936,
0.06915608048439026,
0.04550059512257576,
-0.11369947344064713,
-0.0243031308054924,
0.016365012153983116,
-0.031502000987529755,
0.07854967564344406,
-0.08083357661962509,
-0.0291584562510252,
-0.0016996421618387103,
-0.018290866166353226,
-0.03379524126648903,
-0.000492354913149029,
-0.017440905794501305,
0.023513127118349075,
0.005012144334614277,
-0.06803379952907562,
-0.0739610493183136,
-0.07371984422206879,
-0.04293142631649971,
0.009753099642693996,
-0.06026458367705345,
0.042955752462148666,
0.011333398520946503,
0.03923103213310242,
0.005091789644211531,
-0.036039553582668304,
-0.011539016850292683,
0.05602240189909935,
-0.04652094095945358,
0.006853869650512934,
-0.02416338212788105,
-0.04720976948738098,
-0.03904743492603302,
0.0683886706829071,
0.03297613188624382,
0.11030156165361404,
-0.029578203335404396,
0.06410585343837738,
0.04981237277388573,
-0.09369800984859467,
0.05167821794748306,
-0.03870399668812752,
0.029968129470944405,
0.019682561978697777,
0.032617565244436264,
0.026976609602570534,
0.11862828582525253,
0.09097812324762344,
0.01323196291923523,
0.11246241629123688,
-0.051226675510406494,
0.04919569566845894,
-0.00879812240600586,
0.060823068022727966,
-0.04458162188529968,
-4.096292016697589e-8,
-0.06233508139848709,
0.07937874644994736,
-0.00983386766165495,
-0.003121226793155074,
-0.009812776930630207,
-0.06572951376438141,
-0.08128780871629715,
0.052876733243465424,
-0.021255390718579292,
0.01723218709230423,
0.04140476882457733,
0.06576360017061234,
-0.08168023079633713,
-0.026088085025548935,
-0.07565872371196747,
0.07445463538169861,
0.023281054571270943,
0.02578841522336006,
-0.012552520260214806,
-0.058087799698114395,
0.05642854422330856,
0.0582745261490345,
0.012553590349853039,
-0.044579677283763885,
0.0007210634066723287,
-0.08518671989440918,
-0.0812210962176323,
0.11599820107221603,
0.033530402928590775,
-0.00036310844006948173,
-0.032458312809467316,
-0.025254610925912857,
0.01589886099100113,
-0.07050170749425888,
0.07285715639591217,
0.0859079360961914,
-0.005815773271024227,
-0.004974032752215862,
0.017241420224308968,
-0.05866653472185135,
-0.009665407240390778,
0.010822140611708164,
-0.040791064500808716,
-0.021450115367770195,
0.06597519665956497,
-0.025383051484823227,
-0.03795748949050903,
-0.08273939043283463,
-0.02712242305278778,
-0.027058368548750877,
-0.05342217907309532,
-0.07960841804742813,
-0.06792851537466049,
-0.0211331844329834,
0.0569591298699379,
-0.01258421316742897,
-0.04273171350359917,
-0.025111069902777672,
-0.025258038192987442,
0.01167173683643341,
0.07277251780033112,
0.0614507794380188,
-0.04255667328834534,
0.04199284687638283
] | 0
| 0
| 0
| 0
|
|
Semantic Temporal Abstraction via Vision-Language Model Guidance for Efficient Reinforcement Learning
|
https://openreview.net/forum?id=zY37C8d6bS
|
[
"Tian-Shuo Liu",
"Xu-Hui Liu",
"Ruifeng Chen",
"Lixuan Jin",
"Pengyuan Wang",
"Zhilong Zhang",
"Yang Yu"
] |
Poster
|
Extracting temporally extended skills can significantly improve the efficiency of reinforcement learning (RL) by breaking down complex decision-making problems with sparse rewards into simpler subtasks and enabling more effective credit assignment. However, existing abstraction methods either discover skills in an unsupervised manner, which often lacks semantic information and leads to erroneous or scattered skill extraction results, or require substantial human intervention. In this work, we propose to leverage the extensive knowledge in pretrained Vision-Language Models (VLMs) to progressively guide the latent space after vector quantization to be more semantically meaningful through relabeling each skill. This approach, termed **V**ision-l**an**guage model guided **T**emporal **A**bstraction (**VanTA**), facilitates the discovery of more interpretable and task-relevant temporal segmentations from offline data without the need for extensive manual intervention or heuristics. By leveraging the rich information in VLMs, our method can significantly outperform existing offline RL approaches that depend only on limited training data. From a theory perspective, we demonstrate that stronger internal sequential correlations within each sub-task, induced by VanTA, effectively reduces suboptimality in policy learning. We validate the effectiveness of our approach through extensive experiments on diverse environments, including Franka Kitchen, Minigrid, and Crafter. These experiments show that our method surpasses existing approaches in long-horizon offline reinforcement learning scenarios with both proprioceptive and visual observations.
|
Reinforcement Learning; Vision-Language Models; Temporal Abstraction
| null | 11,289
| null |
[
-0.0220831036567688,
-0.020695343613624573,
0.017441045492887497,
0.04698735103011131,
0.06777910143136978,
0.09257229417562485,
-0.013460762798786163,
-0.045234356075525284,
-0.006887189578264952,
-0.06003550440073013,
-0.007215788587927818,
-0.08824066817760468,
-0.012022728100419044,
0.1371518224477768,
0.009389365091919899,
0.003524444065988064,
0.07168224453926086,
0.05947132036089897,
-0.1099412590265274,
-0.058904729783535004,
0.09548906981945038,
0.04497264698147774,
0.03811092674732208,
0.009311480447649956,
0.07158249616622925,
0.06144046410918236,
0.00451279291883111,
-0.0982084721326828,
0.062464408576488495,
-0.024628521874547005,
0.058637771755456924,
0.005300581455230713,
0.04073885828256607,
0.011180507019162178,
-0.06451348960399628,
0.07359542697668076,
-0.06700059026479721,
0.0619521327316761,
-0.041200120002031326,
-0.0644683837890625,
-0.057593222707509995,
-0.05509967356920242,
-0.011805925518274307,
0.03253885731101036,
0.06827462464570999,
-0.011711137369275093,
0.011610250920057297,
-0.04949590563774109,
-0.004971608053892851,
-0.015329957008361816,
-0.045127127319574356,
-0.01563444547355175,
0.02252165973186493,
0.02493009716272354,
-0.05611508712172508,
0.06610815227031708,
0.054247286170721054,
0.021545924246311188,
-0.012444084510207176,
-0.06417564302682877,
-0.01969866454601288,
-0.09511318802833557,
0.009456860832870007,
-0.02328859455883503,
-0.09637545794248581,
-0.05193377658724785,
-0.03564826771616936,
0.016006985679268837,
0.03235946595668793,
0.01822548545897007,
-0.015607110224664211,
0.054261304438114166,
-0.03544684126973152,
0.005728352814912796,
0.020033055916428566,
0.05482318997383118,
0.059896986931562424,
-0.03834714740514755,
0.08771974593400955,
-0.07940161973237991,
0.00035328371450304985,
-0.005861570592969656,
-0.0027416287921369076,
-0.0021944716572761536,
0.11108139902353287,
0.026936527341604233,
-0.007159074768424034,
0.06443390995264053,
0.043267037719488144,
-0.04811982810497284,
-0.024135343730449677,
-0.06111522391438484,
0.0375894196331501,
-0.04727887362241745,
0.03729655221104622,
-0.03187237307429314,
-0.06686708331108093,
-0.09200173616409302,
-0.006188015919178724,
0.051282402127981186,
0.0363534651696682,
0.05293264985084534,
-0.05460410192608833,
-0.005437230225652456,
-0.028459087014198303,
0.0005291732959449291,
0.07505564391613007,
0.027704043313860893,
0.03365762159228325,
-0.1082557886838913,
0.015934059396386147,
0.012931859120726585,
0.06991632282733917,
0.01917138136923313,
0.006995423696935177,
0.07784830033779144,
-0.03143549710512161,
-0.031216224655508995,
0.04286633059382439,
0.03634711727499962,
0.0176057368516922,
0.0006094808923080564,
0.03068988211452961,
0.026894189417362213,
0.021363917738199234,
-0.0580841526389122,
0.014941792003810406,
2.5321672211976445e-35,
0.08158665150403976,
0.006842716597020626,
0.025156043469905853,
0.03006647527217865,
0.04703039303421974,
-0.021030576899647713,
0.04706691950559616,
0.013848048634827137,
-0.0266577135771513,
0.009953557513654232,
-0.01697775162756443,
-0.02155478298664093,
-0.04978472739458084,
0.12732373178005219,
0.034719232469797134,
0.006046748254448175,
-0.047063618898391724,
0.034556638449430466,
-0.01655183732509613,
0.006659021135419607,
0.026014257222414017,
-0.024404475465416908,
-0.0072189257480204105,
-0.0481392927467823,
0.05726305767893791,
-0.01220725104212761,
0.05483776703476906,
-0.06585422158241272,
-0.001115069491788745,
0.005100377835333347,
-0.03839206323027611,
0.031728051602840424,
-0.0060453568585217,
0.045291297137737274,
0.023414192721247673,
0.05995531380176544,
-0.009006117470562458,
-0.02085071988403797,
0.04443013668060303,
-0.010372013784945011,
0.04167494922876358,
0.0033942193258553743,
-0.04237064719200134,
-0.026135101914405823,
-0.08006428182125092,
-0.06212727352976799,
0.03504191339015961,
0.036721132695674896,
-0.08694947510957718,
0.023251961916685104,
0.04735235124826431,
0.021374326199293137,
-0.01448513101786375,
-0.13029512763023376,
-0.05393996834754944,
0.07772862911224365,
0.04481561854481697,
0.10181120038032532,
-0.026677686721086502,
0.00942591018974781,
0.0072561753913760185,
0.008984247222542763,
0.00760598573833704,
0.1136670857667923,
-0.05024218186736107,
0.02803485281765461,
-0.006254683714359999,
0.06561903655529022,
0.050212565809488297,
0.0024633719585835934,
-0.07327639311552048,
0.047063324600458145,
0.0696578174829483,
-0.07441823929548264,
0.11744268238544464,
-0.0063141584396362305,
0.05155351012945175,
-0.09092001616954803,
0.012283853255212307,
0.056984055787324905,
0.0026109637692570686,
-0.05054500699043274,
-0.023081814870238304,
-0.04610447958111763,
0.05635537579655647,
-0.041473809629678726,
0.05549246817827225,
-0.1295415461063385,
0.033131133764982224,
-0.04656510800123215,
-0.05936874449253082,
-0.04107563942670822,
-0.08223200589418411,
0.026096273213624954,
0.005634548608213663,
-1.1502620890753986e-34,
0.025341957807540894,
0.0015519083244726062,
0.024600010365247726,
0.007946755737066269,
-0.011162443086504936,
-0.05432534217834473,
0.046655796468257904,
0.007386401761323214,
0.035044703632593155,
-0.047719556838274,
-0.0510626919567585,
-0.029676510021090508,
-0.010661843232810497,
-0.011563889682292938,
0.03958127275109291,
0.017239011824131012,
-0.03188418596982956,
0.01770085096359253,
0.03181115537881851,
0.06229501962661743,
0.014188000932335854,
0.06380675733089447,
-0.14234782755374908,
-0.003583187935873866,
-0.009381011128425598,
0.05146753042936325,
0.023010017350316048,
0.059876274317502975,
-0.03128521889448166,
0.0573950931429863,
-0.03322259709239006,
-0.08104197680950165,
0.0064272102899849415,
-0.10451654344797134,
-0.09817405045032501,
0.04776836559176445,
0.029778145253658295,
-0.024820253252983093,
-0.07776181399822235,
0.0761156901717186,
0.0063224900513887405,
-0.04872419685125351,
-0.09145819395780563,
-0.02405354380607605,
-0.04919047653675079,
-0.03658014535903931,
-0.0908210501074791,
0.014162815175950527,
-0.06316521763801575,
0.04775580018758774,
-0.030217336490750313,
0.048558108508586884,
-0.10233604162931442,
0.04402811452746391,
-0.05769290775060654,
-0.03186355531215668,
-0.03582702949643135,
-0.042641352862119675,
0.05034121498465538,
-0.024247152730822563,
-0.03722238913178444,
0.023344457149505615,
-0.016361525282263756,
0.004107548855245113,
0.017471596598625183,
-0.05467444658279419,
-0.04806096851825714,
0.030228566378355026,
0.01414626743644476,
-0.056200385093688965,
0.034960951656103134,
0.029714666306972504,
0.0316212996840477,
-0.046609409153461456,
0.059348683804273605,
-0.02673405222594738,
-0.02461484633386135,
-0.027798520401120186,
0.019463051110506058,
-0.09471704065799713,
-0.0611761137843132,
-0.020154984667897224,
0.051321227103471756,
0.04825860261917114,
0.034093983471393585,
0.11634063720703125,
0.011468078009784222,
0.07931771129369736,
0.060096535831689835,
-0.03534077852964401,
-0.09234242886304855,
-0.03409286588430405,
0.03040865808725357,
0.10454872995615005,
-0.047997135668992996,
-4.616591553485705e-8,
-0.056944165378808975,
0.09462366253137589,
0.0010148370638489723,
0.0038743014447391033,
0.07615222036838531,
-0.016114864498376846,
-0.06321963667869568,
0.08748084306716919,
-0.000746306439395994,
-0.002139550633728504,
0.05118083581328392,
-0.0018015417736023664,
0.004655022174119949,
-0.14084523916244507,
-0.04758331552147865,
0.06657477468252182,
0.06248103454709053,
-0.0014106963062658906,
0.022610019892454147,
0.012631709687411785,
0.10462132841348648,
0.006993983872234821,
-0.006400849670171738,
-0.03283963352441788,
0.000058215719036525115,
-0.04438675940036774,
-0.0386291928589344,
0.14037373661994934,
0.11559443175792694,
-0.04290300980210304,
0.06037738919258118,
0.005364248063415289,
0.07018633931875229,
0.015776008367538452,
0.07321593165397644,
0.020891597494482994,
0.019654812291264534,
-0.06070530414581299,
-0.0464320071041584,
0.03452738747000694,
-0.0031425817869603634,
0.02840588055551052,
0.022439775988459587,
0.004354807082563639,
-0.00037610926665365696,
-0.005011015105992556,
-0.08534051477909088,
-0.061019908636808395,
0.038439102470874786,
-0.05497797206044197,
-0.017272144556045532,
-0.01957479864358902,
-0.04204531013965607,
0.09442569315433502,
0.08458546549081802,
-0.0008179020951502025,
0.07553538680076599,
-0.04865235090255737,
0.008798046968877316,
0.010887527838349342,
-0.00767930131405592,
0.0008418412180617452,
-0.0470765121281147,
-0.02160385623574257
] | 0
| 0
| 0
| 0
|
|
ScImage: How good are multimodal large language models at scientific text-to-image generation?
|
https://openreview.net/forum?id=ugyqNEOjoU
|
[
"Leixin Zhang",
"Steffen Eger",
"Yinjie Cheng",
"WEIHE ZHAI",
"Jonas Belouadi",
"Fahimeh Moafian",
"Zhixue Zhao"
] |
Poster
|
Multimodal large language models (LLMs) have demonstrated impressive capabilities in generating high-quality images from textual instructions. However, their performance in generating scientific images—a critical application for accelerating scientific progress—remains underexplored. In this work, we address this gap by introducing ScImage, a benchmark designed to evaluate the multimodal capabilities of LLMs in generating scientific images from textual descriptions. ScImage assesses three key dimensions of understanding: spatial, numeric, and attribute comprehension, as well as their combinations, focusing on the relationships between scientific objects (e.g., squares, circles). We evaluate seven models, GPT-4o, Llama, AutomaTikZ, Dall-E, StableDiffusion, GPT-o1 and Qwen2.5-Coder-Instruct using two modes of output generation: code-based outputs (Python, TikZ) and direct raster image generation. Additionally, we examine four different input languages: English, German, Farsi, and Chinese. Our evaluation, conducted with 11 scientists across three criteria (correctness, relevance, and scientific accuracy), reveals that while GPT4-o produces outputs of decent quality for simpler prompts involving individual dimensions such as spatial, numeric, or attribute understanding in isolation, all models face challenges in this task, especially for more complex prompts. ScImage is available: huggingface.co/datasets/casszhao/ScImage
|
LLMs, multimodality, science, image generation
|
We evaluate multimodal LLMs on scientific text-to-image generation
| 11,280
|
2412.02368
|
[
-0.020836684852838516,
-0.03701924905180931,
0.006858185864984989,
0.026850368827581406,
0.09309418499469757,
-0.08169778436422348,
-0.08508335053920746,
0.019087359309196472,
0.06956779956817627,
-0.028995798900723457,
-0.03423458710312843,
-0.10385735332965851,
0.06251800805330276,
0.06547212600708008,
-0.024663463234901428,
-0.025097932666540146,
-0.04410520941019058,
0.025040023028850555,
-0.08158326148986816,
-0.06877264380455017,
0.056762319058179855,
0.0299526434391737,
0.10527339577674866,
-0.04373301565647125,
0.10749831795692444,
0.022525792941451073,
-0.07418981939554214,
-0.031192732974886894,
0.02498415671288967,
-0.02420760691165924,
0.010796723887324333,
0.07857941836118698,
0.003468478797003627,
0.09830231219530106,
-0.01092308759689331,
0.10623876005411148,
-0.04984339699149132,
0.0022663979325443506,
-0.03218623995780945,
0.016750654205679893,
-0.033236682415008545,
0.02783484198153019,
0.0943647101521492,
-0.015196049585938454,
0.13346098363399506,
-0.005783473141491413,
-0.051894411444664,
-0.0807485580444336,
-0.028939368203282356,
0.021140003576874733,
-0.14602574706077576,
-0.03538653627038002,
-0.013429675251245499,
0.0320625938475132,
0.036486513912677765,
-0.0033940752036869526,
0.11190750449895859,
-0.08014114946126938,
-0.03244243562221527,
-0.06227193400263786,
-0.045778244733810425,
-0.07117317616939545,
-0.07693525403738022,
-0.0032038637436926365,
0.043711960315704346,
-0.02315784990787506,
-0.0029588083270937204,
-0.030205348506569862,
-0.03481470048427582,
-0.031157588586211205,
-0.0655789002776146,
0.04066669940948486,
-0.007024969905614853,
0.10029452294111252,
-0.005574787966907024,
0.05649258941411972,
0.05559615045785904,
0.008256812579929829,
0.05086490884423256,
-0.10159946978092194,
0.0823645368218422,
-0.024393219500780106,
0.027172354981303215,
-0.003981390502303839,
0.016014276072382927,
0.045379638671875,
0.02809157222509384,
0.09456335008144379,
-0.012607035227119923,
0.06722535938024521,
-0.00945174042135477,
-0.03367813304066658,
0.00994043331593275,
0.02212836779654026,
-0.03354403004050255,
0.0266491137444973,
0.09848428517580032,
-0.13587331771850586,
-0.014079751446843147,
0.03274267911911011,
0.1048719584941864,
0.009448377415537834,
0.029697222635149956,
-0.08000995218753815,
0.0035289120860397816,
-0.09427495300769806,
0.07522756606340408,
-0.012876694090664387,
0.033494383096694946,
-0.035120848566293716,
0.04852394014596939,
0.04349783435463905,
-0.06519617885351181,
0.05002879723906517,
0.012604446150362492,
0.017518265172839165,
0.036399099975824356,
0.012530156411230564,
0.037838976830244064,
-0.01222892664372921,
-0.08814427256584167,
0.025227155536413193,
-0.023744601756334305,
-0.006192451808601618,
0.0263975840061903,
-0.05430341139435768,
-0.07752450555562973,
7.914812849017702e-34,
0.03204551711678505,
0.03394853696227074,
0.009016774594783783,
0.09260936826467514,
0.05788975954055786,
-0.001662615337409079,
-0.024184374138712883,
-0.01504311803728342,
-0.028656527400016785,
-0.04713406413793564,
-0.058764610439538956,
0.051730770617723465,
-0.06700670719146729,
0.10597418248653412,
0.040443260222673416,
0.006071099080145359,
-0.03620852157473564,
0.009536651894450188,
-0.0743856355547905,
0.03442711755633354,
0.033326972275972366,
0.0012943849433213472,
0.018903857097029686,
-0.021565331146121025,
-0.0038387493696063757,
0.03398745134472847,
0.05345070734620094,
-0.09677036851644516,
-0.07249436527490616,
0.008509674109518528,
-0.173906147480011,
-0.01728752627968788,
0.05460338667035103,
0.042584378272295,
0.03788309916853905,
-0.09160403162240982,
-0.03335589915513992,
-0.05146244540810585,
0.006974979303777218,
-0.006894032936543226,
0.028422290459275246,
0.03839675337076187,
0.01886347495019436,
0.018065406009554863,
0.0022243845742195845,
0.0402679443359375,
-0.022685116156935692,
0.06673317402601242,
-0.056999046355485916,
0.035564977675676346,
-0.004453054629266262,
0.04247324913740158,
-0.018436098471283913,
-0.0013631850015372038,
0.011297612451016903,
0.026965100318193436,
0.02284204587340355,
-0.03412305936217308,
-0.008294965140521526,
0.015348454006016254,
0.030344925820827484,
0.11446155607700348,
0.005572605412453413,
0.0703708827495575,
0.032919496297836304,
0.028823886066675186,
-0.0518040731549263,
0.029552189633250237,
0.0004877661704085767,
0.07084508240222931,
-0.03947017714381218,
0.02341804839670658,
-0.03678717091679573,
-0.03344991058111191,
0.05442745238542557,
0.0012170773698017001,
0.0368792824447155,
-0.08092404901981354,
0.019425155594944954,
0.04112893342971802,
-0.04248668998479843,
0.06040840595960617,
0.00904854852706194,
-0.13473400473594666,
-0.06588363647460938,
-0.03809155523777008,
0.03773460537195206,
-0.038221947848796844,
-0.02211330458521843,
-0.05554335191845894,
-0.000340679835062474,
-0.044027794152498245,
0.04347321763634682,
-0.061663657426834106,
-0.07201676815748215,
-5.626746615098217e-34,
-0.02648192271590233,
0.009047305211424828,
-0.04293733090162277,
0.10414747148752213,
-0.03411813825368881,
-0.0447036549448967,
0.006723981350660324,
0.0211954265832901,
0.0067521375603973866,
-0.08729951083660126,
0.003730512224137783,
0.018106549978256226,
0.009733428247272968,
-0.04521806538105011,
0.061091966927051544,
-0.08005013316869736,
0.025182116776704788,
-0.005158209707587957,
-0.02998129464685917,
0.07468948513269424,
-0.03140206262469292,
0.07381753623485565,
-0.026876889169216156,
0.01743229478597641,
-0.10548730939626694,
0.0889827311038971,
0.002386084757745266,
-0.042360711842775345,
-0.0845872163772583,
0.0033220683690160513,
0.04251161962747574,
-0.017157239839434624,
-0.009759529493749142,
0.022682249546051025,
0.0013322823215276003,
0.03342381492257118,
0.09973596781492233,
-0.06885087490081787,
-0.018755070865154266,
0.0753946527838707,
0.07700669765472412,
0.03375948220491409,
-0.07273456454277039,
0.029501378536224365,
-0.03288973495364189,
-0.004013352561742067,
-0.10367172211408615,
-0.013345721177756786,
0.007499525789171457,
0.1075512170791626,
-0.0006936887511983514,
-0.0396149717271328,
-0.04411453753709793,
-0.025614721700549126,
0.0009382393909618258,
-0.12505672872066498,
-0.07083114236593246,
-0.036349039524793625,
-0.038017451763153076,
0.02714509889483452,
-0.0754089206457138,
-0.048688121140003204,
-0.012490528635680676,
-0.03511955589056015,
-0.04133402928709984,
-0.04485546797513962,
-0.015536759048700333,
0.038745518773794174,
0.007044586353003979,
0.02473701909184456,
0.040902454406023026,
-0.025786040350794792,
0.00025342986918985844,
0.04992019757628441,
0.009070537984371185,
0.07881972193717957,
0.024771295487880707,
-0.0006775652873329818,
0.031286582350730896,
-0.059601619839668274,
0.027293317019939423,
0.04968569800257683,
0.02690894901752472,
0.09134215861558914,
0.05186883360147476,
0.025292517617344856,
-0.026283299550414085,
0.057354651391506195,
0.05056982487440109,
0.04235469922423363,
0.03724782541394234,
0.0526573583483696,
0.06492111831903458,
0.10963205993175507,
0.053275544196367264,
-4.24467145876406e-8,
0.02284710668027401,
0.01363252755254507,
-0.03869868069887161,
-0.01801207661628723,
-0.04624390974640846,
-0.062162671238183975,
-0.06223273649811745,
0.021023379638791084,
-0.002045199740678072,
-0.0452020987868309,
0.03205789253115654,
-0.016133464872837067,
-0.10611115396022797,
0.0006942906766198575,
0.009742994792759418,
0.06900770962238312,
0.0034562053624540567,
0.019413474947214127,
0.020836003124713898,
0.014888901263475418,
0.02854693867266178,
0.009771368466317654,
-0.023457765579223633,
-0.007245807442814112,
-0.027504798024892807,
-0.0064964909106493,
-0.06362984329462051,
0.040575772523880005,
0.024600664153695107,
-0.07787970453500748,
0.06508579850196838,
-0.03490116074681282,
-0.06074647605419159,
0.013518285937607288,
0.0158088281750679,
-0.004881266970187426,
-0.013810190372169018,
-0.0452786460518837,
0.05228341743350029,
0.022428758442401886,
0.08475179970264435,
0.04075509309768677,
-0.046629805117845535,
0.03172432631254196,
0.014100316911935806,
-0.018069202080368996,
-0.03503844141960144,
-0.12803059816360474,
0.007071036379784346,
-0.0130871357396245,
-0.05733935534954071,
-0.02975149266421795,
-0.0012142570922151208,
0.022418677806854248,
0.01490094419568777,
-0.00889219343662262,
-0.022801542654633522,
0.01639922522008419,
0.025488754734396935,
0.09500125050544739,
0.03922227397561073,
0.04677761346101761,
-0.02729318104684353,
-0.0061033982783555984
] |
https://github.com/leixin-zhang/scimage
| 6
| 0
| 0
| 0
|
Utility-Directed Conformal Prediction: A Decision-Aware Framework for Actionable Uncertainty Quantification
|
https://openreview.net/forum?id=iOMnn1hSBO
|
[
"Santiago Cortes-Gomez",
"Carlos Miguel Patiño",
"Yewon Byun",
"Steven Wu",
"Eric Horvitz",
"Bryan Wilder"
] |
Poster
|
There is increasing interest in ``decision-focused" machine learning methods which train models to account for how their predictions are used in downstream optimization problems. Doing so can often improve performance on subsequent decision problems. However, current methods for uncertainty quantification do not incorporate any information at all about downstream decisions. We develop a framework based on conformal prediction to produce prediction sets that account for a downstream decision loss function, making them more appropriate to inform high-stakes decision-making. Our approach harnesses the strengths of conformal methods—modularity, model-agnosticism, and statistical coverage guarantees—while incorporating downstream decisions and user-specified utility functions. We prove that our methods retain standard coverage guarantees. Empirical evaluation across a range of datasets and utility metrics demonstrates that our methods achieve significantly lower decision loss compared to standard conformal methods. Additionally, we present a real-world use case in healthcare diagnosis, where our method effectively incorporates the hierarchical structure of dermatological diseases. It successfully generates sets with coherent diagnostic meaning, aiding the triage process during dermatology diagnosis and illustrating how our method can ground high-stakes decision-making on external domain knowledge.
|
Decision-focused learning, decision making, uncertainty quantification, healthcare
| null | 11,273
|
2410.01767
|
[
-0.08094391971826553,
-0.0894358679652214,
0.006679842714220285,
0.03369174525141716,
0.036078840494155884,
-0.07529303431510925,
0.0075474395416677,
0.08730379492044449,
-0.0006314243655651808,
0.022057974711060524,
-0.04353662207722664,
-0.07921897619962692,
0.08387365192174911,
0.06177983805537224,
-0.007714282721281052,
-0.013947760686278343,
0.05779968947172165,
0.04237820953130722,
-0.1365697830915451,
-0.042388319969177246,
0.037623755633831024,
0.0442105196416378,
-0.051883917301893234,
0.02802407182753086,
-0.03703130781650543,
-0.06228036433458328,
0.022284436970949173,
-0.027719441801309586,
-0.016240455210208893,
0.0334768071770668,
-0.0026096070650964975,
-0.0029437642078846693,
-0.03151145204901695,
-0.09582815319299698,
-0.019887778908014297,
0.003285409649834037,
-0.027231846004724503,
0.03350818157196045,
-0.0010480481432750821,
0.019008208066225052,
0.004435405135154724,
-0.04051872715353966,
-0.04631798714399338,
0.060046564787626266,
0.026671577244997025,
0.021702928468585014,
0.016549568623304367,
-0.030793510377407074,
0.025344310328364372,
0.027302155271172523,
0.01957402005791664,
-0.0544055812060833,
0.0186343714594841,
-0.053932394832372665,
-0.0007374896085821092,
0.024770474061369896,
0.035753823816776276,
-0.06631847470998764,
-0.10748473554849625,
-0.017356758937239647,
-0.024298889562487602,
-0.1953868567943573,
-0.0697680413722992,
-0.015307468362152576,
0.014673764817416668,
-0.03986236825585365,
0.06473971158266068,
0.006256375927478075,
-0.009688864462077618,
0.03311591595411301,
-0.06190212070941925,
0.0671822801232338,
-0.06610801815986633,
0.09545118361711502,
0.0666375607252121,
0.030897073447704315,
0.12122206389904022,
-0.009675411507487297,
0.022813187912106514,
-0.040577784180641174,
-0.0418831892311573,
0.09178131818771362,
0.06160292029380798,
0.07956778258085251,
0.12986932694911957,
-0.012807931751012802,
0.01929347775876522,
0.0465470626950264,
0.02815665490925312,
-0.0031514139845967293,
-0.041121646761894226,
-0.011369619518518448,
0.019571615383028984,
-0.029718568548560143,
0.011244118213653564,
0.08215179294347763,
-0.03817727789282799,
-0.07442895323038101,
0.00850040465593338,
0.04472823813557625,
-0.09605661779642105,
0.03205320984125137,
0.05669598653912544,
-0.07895351946353912,
0.037044525146484375,
-0.12498076260089874,
0.01716141402721405,
0.04919995740056038,
0.08706967532634735,
-0.044136859476566315,
0.059201423078775406,
0.0023335388395935297,
0.07787041366100311,
-0.061957065016031265,
0.042657751590013504,
0.030321218073368073,
0.004063883796334267,
-0.010767744854092598,
-0.00027270778082311153,
0.09062214940786362,
-0.04051806405186653,
0.06536448001861572,
0.0712851956486702,
0.017483985051512718,
0.06230675056576729,
0.035450201481580734,
-0.056686881929636,
7.286129082839172e-34,
-0.0038773168344050646,
-0.013496790081262589,
0.0838739424943924,
-0.022425321862101555,
0.02566147781908512,
-0.01709446683526039,
-0.03422427177429199,
-0.061317067593336105,
0.008074217475950718,
0.03764213249087334,
-0.02038082294166088,
0.05281950533390045,
0.005406000651419163,
0.001583190867677331,
0.09368570894002914,
0.0558091476559639,
-0.029029885306954384,
0.05167188495397568,
-0.09042750298976898,
0.0023844237439334393,
0.08888500928878784,
-0.0850699171423912,
-0.024345656856894493,
-0.0282510407269001,
0.025563012808561325,
-0.031246894970536232,
0.007056744769215584,
0.020344125106930733,
0.013440570794045925,
0.009512345306575298,
-0.11963427066802979,
0.005724991671741009,
0.06274541467428207,
0.018857400864362717,
0.020385848358273506,
-0.029777757823467255,
-0.07774795591831207,
0.007706452626734972,
-0.045696668326854706,
0.02013067714869976,
-0.020398657768964767,
-0.06813008338212967,
-0.01289489958435297,
0.017820924520492554,
-0.022126197814941406,
-0.08231477439403534,
-0.03358158841729164,
-0.022398315370082855,
-0.04510141909122467,
-0.12080416083335876,
0.061511535197496414,
0.04880392551422119,
0.02250046469271183,
-0.061924949288368225,
-0.06297798454761505,
0.04380036145448685,
-0.03346593305468559,
-0.011679582297801971,
-0.0014806205872446299,
-0.023602226749062538,
-0.0018296018242835999,
0.0016910636331886053,
-0.02170919068157673,
0.040761351585388184,
-0.026615018025040627,
0.03197380155324936,
-0.019161194562911987,
-0.047099653631448746,
0.023576030507683754,
-0.0037789198104292154,
-0.01798113062977791,
0.05986107885837555,
-0.0014405149267986417,
-0.014302903786301613,
0.09335562586784363,
0.01773260161280632,
0.04550878703594208,
0.021245073527097702,
0.025439588353037834,
-0.0329844132065773,
-0.046731822192668915,
0.05366726592183113,
-0.03305165469646454,
0.003225891850888729,
0.05608488619327545,
-0.06213053688406944,
0.02308305911719799,
-0.0067658754996955395,
-0.17525646090507507,
-0.032727666199207306,
-0.16254112124443054,
0.02209356613457203,
-0.033147573471069336,
0.028527148067951202,
-0.010146425105631351,
-3.069276912437434e-33,
-0.06178218126296997,
0.037323858588933945,
0.0384952649474144,
-0.05334581434726715,
0.00545493932440877,
-0.03733351081609726,
-0.07442572712898254,
-0.018644411116838455,
0.05353629216551781,
-0.07039304822683334,
-0.009554434567689896,
0.025025634095072746,
0.0022996566258370876,
-0.03577874228358269,
-0.03715472295880318,
0.06567567586898804,
-0.07433514297008514,
0.023431628942489624,
-0.019780131056904793,
0.003908801823854446,
-0.007711350452154875,
0.06058069318532944,
-0.11070574074983597,
0.041043490171432495,
-0.032101571559906006,
0.020260222256183624,
-0.011484949849545956,
0.031000453978776932,
-0.004089125897735357,
-0.033323317766189575,
-0.12851504981517792,
-0.04487986862659454,
-0.011423924937844276,
-0.03647224232554436,
-0.03190518915653229,
0.03274946287274361,
0.004543486051261425,
-0.014556161127984524,
-0.035887494683265686,
0.06396378576755524,
0.00885900016874075,
-0.03891700878739357,
-0.08363069593906403,
-0.029540378600358963,
0.041561514139175415,
-0.08582509309053421,
-0.0007940095383673906,
-0.05757608637213707,
0.04467732459306717,
0.0022704615257680416,
0.028420131653547287,
0.036676254123449326,
-0.04350976645946503,
0.10061869025230408,
-0.09628993272781372,
-0.029887432232499123,
-0.010283608920872211,
-0.048164863139390945,
0.04894507676362991,
0.09219050407409668,
-0.06799595803022385,
0.0345916673541069,
0.019268857315182686,
0.11271025240421295,
-0.0361725315451622,
0.03873899206519127,
0.0013800441520288587,
0.015030457638204098,
0.06032957509160042,
0.04097489267587662,
-0.0362481065094471,
0.001579268486239016,
-0.030460773035883904,
-0.007785788271576166,
0.013341580517590046,
-0.034942083060741425,
0.05181214213371277,
-0.05014359578490257,
-0.016393233090639114,
0.07218994200229645,
0.08701130002737045,
-0.061137985438108444,
0.10691150277853012,
0.07795743644237518,
0.08679322898387909,
-0.04298051819205284,
0.1016007587313652,
0.025591744109988213,
0.07784996926784515,
-0.018793985247612,
-0.06237127259373665,
-0.03922269120812416,
-0.04901470988988876,
0.03856078162789345,
-0.08117062598466873,
-4.494554417533436e-8,
0.006823445670306683,
0.050138551741838455,
0.03978549689054489,
0.02482735738158226,
0.029592515900731087,
0.02405579574406147,
-0.06910185515880585,
0.031522564589977264,
-0.030030885711312294,
0.032220568507909775,
0.01629597879946232,
0.0333225354552269,
-0.012811987660825253,
-0.041204795241355896,
0.014048771001398563,
0.02553453855216503,
0.0138776246458292,
0.06341911107301712,
-0.033715058118104935,
-0.030900251120328903,
0.0031118306796997786,
0.0024974781554192305,
0.04604233056306839,
-0.003309145336970687,
0.077996164560318,
-0.11113505065441132,
-0.04139304533600807,
0.06381650269031525,
0.008807839825749397,
0.07180558890104294,
0.014350577257573605,
0.030205097049474716,
-0.0018615570152178407,
0.07785701751708984,
-0.026763392612338066,
0.002747683087363839,
0.024830423295497894,
-0.067418172955513,
0.0020342618227005005,
0.02397291176021099,
0.0199168398976326,
0.04121380299329758,
-0.023120226338505745,
0.028258873149752617,
0.054014503955841064,
-0.04199816286563873,
0.023997163400053978,
-0.017939576879143715,
0.07474689930677414,
-0.008458921685814857,
0.0388653427362442,
0.04610958695411682,
0.05483678728342056,
0.09047810733318329,
0.04234151169657707,
0.0009350443724542856,
0.009645305573940277,
-0.06019390746951103,
-0.04725005105137825,
0.058414801955223083,
0.016903456300497055,
-0.03772817552089691,
0.023821568116545677,
0.01244760025292635
] |
https://github.com/cmpatino/utility_driven_prediction
| 0
| 0
| 0
| 0
|
On Targeted Manipulation and Deception when Optimizing LLMs for User Feedback
|
https://openreview.net/forum?id=Wf2ndb8nhf
|
[
"Marcus Williams",
"Micah Carroll",
"Adhyyan Narang",
"Constantin Weisser",
"Brendan Murphy",
"Anca Dragan"
] |
Poster
|
As LLMs become more widely deployed, there is increasing interest in directly optimizing for feedback from end users (e.g. thumbs up) in addition to feedback from paid annotators. However, training to maximize human feedback creates a perverse incentive structure for the AI to resort to manipulative or deceptive tactics to obtain positive feedback from users who are vulnerable to such strategies. We study this phenomenon by training LLMs with Reinforcement Learning with simulated user feedback in environments of practical LLM usage. In our settings, we find that: 1) Extreme forms of "feedback gaming" such as manipulation and deception are learned reliably; 2) Even if only 2% of users are vulnerable to manipulative strategies, LLMs learn to identify and target them while behaving appropriately with other users, making such behaviors harder to detect; 3) To mitigate this issue, it may seem promising to leverage continued safety training or LLM-as-judges during training to filter problematic outputs. Instead, we found that while such approaches help in some of our settings, they backfire in others, sometimes even leading to subtler manipulative behaviors. We hope our results can serve as a case study which highlights the risks of using gameable feedback sources -- such as user feedback -- as a target for RL. Our code is publicly available. Warning: some of our examples may be upsetting.
|
manipulation, deception, alignment, reward hacking, user feedback
| null | 11,268
|
2411.02306
|
[
-0.04176914691925049,
-0.020704176276922226,
0.003938887733966112,
-0.010423240251839161,
0.04872022941708565,
0.01817936822772026,
0.06398741900920868,
-0.006002084817737341,
0.0366048701107502,
-0.010519947856664658,
-0.07883729040622711,
-0.06339611113071442,
0.06457941979169846,
0.0026741644833236933,
-0.016842788085341454,
0.028193771839141846,
0.16371174156665802,
0.09680359065532684,
-0.02482665702700615,
-0.003692388767376542,
0.04260387644171715,
-0.022632881999015808,
0.09060494601726532,
-0.0011494135251268744,
-0.10119365155696869,
-0.04509666562080383,
-0.019338073208928108,
-0.05904465168714523,
-0.03133649006485939,
-0.07011339068412781,
0.03155159950256348,
-0.010165460407733917,
-0.01911160722374916,
-0.007062925957143307,
-0.06776147335767746,
-0.022815436124801636,
-0.10178102552890778,
-0.004409261979162693,
-0.016318975016474724,
-0.08628176897764206,
-0.05308929830789566,
-0.017753595486283302,
-0.0034444222692400217,
-0.013936996459960938,
-0.019099393859505653,
0.013936380855739117,
0.046659957617521286,
0.007052627857774496,
-0.06921984255313873,
0.003171002259477973,
-0.0665745660662651,
-0.027060221880674362,
0.06721257418394089,
-0.005080258008092642,
-0.03390319645404816,
-0.034916386008262634,
0.08370987325906754,
0.10900380462408066,
-0.007353239227086306,
0.015292779542505741,
0.021257245913147926,
-0.05593624338507652,
-0.02633424662053585,
-0.03735366091132164,
-0.007188559975475073,
0.03831367567181587,
-0.003004643367603421,
0.07128997147083282,
-0.027541104704141617,
0.00039215513970702887,
0.08143289387226105,
-0.008747169747948647,
-0.003524587955325842,
-0.021886738017201424,
0.017133355140686035,
0.058014705777168274,
-0.0982416570186615,
0.009089353494346142,
0.09467200934886932,
-0.014633876271545887,
0.015901096165180206,
-0.005398035980761051,
0.0194846298545599,
0.00968064647167921,
0.004401519428938627,
-0.08782398700714111,
-0.046879690140485764,
0.0038206668104976416,
0.1477435976266861,
0.03854350373148918,
-0.013065236620604992,
0.04207657277584076,
-0.004004062153398991,
0.0018171315314248204,
0.011030012741684914,
-0.028109481558203697,
-0.056870199739933014,
-0.06927264481782913,
-0.039576321840286255,
0.05922330170869827,
0.01991507038474083,
0.1303982436656952,
-0.013988219201564789,
-0.036727581173181534,
0.07784368842840195,
-0.0011310845147818327,
-0.03971840441226959,
0.015289361588656902,
0.02011737786233425,
-0.05885830521583557,
-0.018224593251943588,
0.07883239537477493,
0.012440473772585392,
-0.004230150021612644,
0.09971925616264343,
-0.03319644555449486,
-0.006659864913672209,
0.03228335082530975,
-0.026285335421562195,
0.05258902162313461,
0.028452686965465546,
-0.044482797384262085,
0.02271958626806736,
0.09853804856538773,
0.08246096968650818,
0.022580383345484734,
-0.1174883097410202,
5.7848964311963976e-33,
0.07251270860433578,
0.01254681684076786,
-0.017504872754216194,
-0.023639865219593048,
-0.03461143374443054,
0.02076791226863861,
0.03895418718457222,
0.04157540947198868,
0.01062664482742548,
-0.008444808423519135,
-0.03690630570054054,
0.011877840384840965,
0.009037287905812263,
0.11069405823945999,
0.021445736289024353,
0.010052025318145752,
0.001069581601768732,
0.008780654519796371,
-0.08131875097751617,
-0.061514586210250854,
0.0482885017991066,
-0.019080083817243576,
0.045452382415533066,
-0.0739072635769844,
-0.013465930707752705,
0.0016865123761817813,
-0.010180652141571045,
0.034002143889665604,
0.034767527133226395,
0.004912349861115217,
-0.0546194389462471,
-0.024252615869045258,
0.07569541782140732,
-0.0014836837071925402,
-0.01561221294105053,
0.06648951023817062,
-0.054621823132038116,
-0.025333337485790253,
0.004839095287024975,
0.01089652068912983,
-0.019750991836190224,
0.04297209531068802,
0.042807962745428085,
0.04357272759079933,
-0.060968395322561264,
0.004479443654417992,
-0.014858522452414036,
-0.09489303082227707,
-0.13078829646110535,
-0.0008358076447620988,
-0.0029544024728238583,
0.07855705171823502,
0.051708780229091644,
-0.04629538208246231,
-0.061811819672584534,
-0.008762244135141373,
-0.020921258255839348,
0.0001628052705200389,
-0.07570025324821472,
-0.02121243253350258,
0.011406545527279377,
-0.004234822932630777,
-0.06558649241924286,
0.01176490169018507,
-0.0731266587972641,
-0.02589881606400013,
-0.00009314851195085794,
0.037743519991636276,
0.023209886625409126,
0.008787991479039192,
-0.05287207290530205,
0.039734385907649994,
-0.010212588123977184,
-0.05995582416653633,
-0.009849063120782375,
-0.08643390238285065,
0.005364901386201382,
-0.05664927139878273,
0.02769826166331768,
-0.05910809710621834,
0.1141088530421257,
-0.01105712354183197,
-0.024065155535936356,
0.015394301153719425,
-0.041167836636304855,
-0.09174451231956482,
0.001987382071092725,
-0.07342466711997986,
-0.010209704749286175,
0.12230181694030762,
-0.08092335611581802,
-0.04153861477971077,
0.010245916433632374,
0.14761720597743988,
-0.01832835003733635,
-6.0657745615116444e-33,
-0.09250633418560028,
0.004855410661548376,
-0.0869951918721199,
0.03240801393985748,
-0.04461925849318504,
0.017791397869586945,
-0.054331790655851364,
0.008511434309184551,
-0.04779430478811264,
-0.053663209080696106,
-0.12737204134464264,
-0.037690937519073486,
0.04574069753289223,
0.020690998062491417,
-0.04595573619008064,
-0.11274927854537964,
0.0467093326151371,
0.02781418338418007,
0.026600126177072525,
-0.05672212317585945,
0.04588685557246208,
0.14886100590229034,
-0.06046442687511444,
0.026090966537594795,
-0.02555841952562332,
0.016868123784661293,
0.05157942324876785,
-0.006993281655013561,
0.05555139109492302,
0.08195256441831589,
0.006214539986103773,
0.014237895607948303,
0.01811504177749157,
-0.03251156583428383,
-0.012108311988413334,
0.03338973596692085,
0.03963179886341095,
0.04828423261642456,
-0.03793952986598015,
0.06937950849533081,
-0.005321411415934563,
-0.029809879139065742,
-0.02693042904138565,
0.018961653113365173,
0.002616331446915865,
-0.07033989578485489,
-0.010831063613295555,
-0.04964994639158249,
0.0501478910446167,
0.02425955794751644,
-0.011583089828491211,
-0.06446229666471481,
-0.0043343729339540005,
-0.026822291314601898,
-0.06843169033527374,
-0.11890480667352676,
0.025032158941030502,
-0.0012871630024164915,
0.05660780891776085,
-0.00045141385635361075,
-0.12077071517705917,
0.009693566709756851,
-0.011134718544781208,
0.0028500023763626814,
0.033767588436603546,
0.03153301402926445,
-0.00367716234177351,
0.06757968664169312,
0.08535156399011612,
-0.007578743156045675,
0.07627206295728683,
0.03695424646139145,
0.012775704264640808,
-0.026081223040819168,
0.04815032705664635,
0.044385768473148346,
-0.07893496751785278,
-0.07791416347026825,
-0.04267435148358345,
-0.037159763276576996,
0.034502509981393814,
-0.023466527462005615,
0.07090456038713455,
0.07932330667972565,
-0.02869032323360443,
0.045385561883449554,
0.028071792796254158,
0.059782855212688446,
-0.06741873919963837,
0.015869151800870895,
0.03672059625387192,
-0.018240144476294518,
-0.01071922667324543,
0.025043264031410217,
-0.016141697764396667,
-5.216050880108014e-8,
-0.013586422428488731,
-0.03820078819990158,
0.003596256487071514,
0.08862229436635971,
0.08575768023729324,
0.047597117722034454,
-0.012138579040765762,
0.05908637121319771,
-0.026878437027335167,
-0.0556267574429512,
0.10219316184520721,
0.022162675857543945,
0.02750922553241253,
-0.09633295983076096,
0.07952940464019775,
0.021414509043097496,
-0.0013150738086551428,
-0.006584190763533115,
-0.07079796493053436,
-0.007686635944992304,
0.04128551855683327,
0.01914302073419094,
-0.017500270158052444,
-0.08463510870933533,
0.029297849163413048,
-0.06536003947257996,
-0.03690708801150322,
0.1061367318034172,
0.00747340964153409,
0.07607169449329376,
0.01628485508263111,
-0.08996879309415817,
0.013332345522940159,
0.026977701112627983,
0.04197302833199501,
0.05907618626952171,
-0.03606000542640686,
-0.0626692995429039,
0.07099838554859161,
-0.02796710841357708,
-0.04250190034508705,
0.06388261914253235,
0.05094583332538605,
0.022428492084145546,
-0.038393810391426086,
-0.002740450669080019,
-0.035913627594709396,
-0.0628376305103302,
0.047113120555877686,
-0.042370229959487915,
-0.02007213979959488,
-0.05869276449084282,
-0.023573733866214752,
0.07435286045074463,
0.07382036000490189,
0.009545998647809029,
0.014708573929965496,
-0.029287371784448624,
0.0070068128407001495,
0.10374927520751953,
0.04197245463728905,
-0.003156638005748391,
-0.07628963142633438,
-0.014829631894826889
] |
https://github.com/marcus-jw/targeted-manipulation-and-deception-in-llms
| 15
| 0
| 0
| 0
|
Gaussian Differentially Private Human Faces Under a Face Radial Curve Representation
|
https://openreview.net/forum?id=K2Tqn8R9pu
|
[
"Carlos J Soto",
"Matthew Reimherr",
"Aleksandra Slavkovic",
"Mark Shriver"
] |
Poster
|
In this paper we consider the problem of releasing a Gaussian Differentially Private (GDP) 3D human face. The human face is a complex structure with many features and inherently tied to one's identity. Protecting this data, in a formally private way, is important yet challenging given the dimensionality of the problem. We extend approximate DP techniques for functional data to the GDP framework. We further propose a novel representation, face radial curves, of a 3D face as a set of functions and then utilize our proposed GDP functional data mechanism. To preserve the shape of the face while injecting noise we rely on tools from shape analysis for our novel representation of the face. We show that our method preserves the shape of the average face and injects less noise than traditional methods for the same privacy budget. Our mechanism consists of two primary components, the first is generally applicable to function value summaries (as are commonly found in nonparametric statistics or functional data analysis) while the second is general to disk-like surfaces and hence more applicable than just to human faces.
|
differential privacy, shape analysis, functional data analysis
|
We develop a representation for 3D faces, extend Gaussian differential privacy to function space, and employ the latter on the former.
| 11,266
|
2409.08301
|
[
-0.08367733657360077,
-0.00340972188860178,
0.000683590944390744,
0.008565356954932213,
0.05002526938915253,
-0.022072946652770042,
0.025427058339118958,
-0.06251909583806992,
0.02577032707631588,
-0.025834735482931137,
0.01881546527147293,
-0.03128437697887421,
0.013120237737894058,
0.022285420447587967,
0.033849406987428665,
-0.08855719119310379,
0.01820950023829937,
0.05043908953666687,
-0.017903979867696762,
0.060996245592832565,
0.004109812900424004,
0.012283127754926682,
-0.029269594699144363,
-0.009809712879359722,
0.08536017686128616,
-0.04288700595498085,
0.1615537852048874,
-0.020111912861466408,
-0.013484744355082512,
-0.05108289793133736,
0.08978801220655441,
-0.04403665289282799,
0.006860633380711079,
0.020602405071258545,
-0.05216483026742935,
-0.026750575751066208,
-0.0352838858962059,
0.12884047627449036,
-0.027358179911971092,
-0.04578862711787224,
-0.05683513358235359,
-0.012537673115730286,
-0.02721485123038292,
0.04566927254199982,
0.02528911642730236,
0.08117911964654922,
0.021427853032946587,
-0.03704962134361267,
0.021297140046954155,
-0.03278730437159538,
-0.01678486354649067,
0.000054029365855967626,
-0.04214539751410484,
0.00048178641009144485,
0.01714020036160946,
-0.017992766574025154,
0.008878749795258045,
-0.06974785774946213,
-0.01736527867615223,
0.0639057382941246,
-0.02281251922249794,
-0.06921105831861496,
0.052342936396598816,
0.01996525377035141,
-0.0168672613799572,
0.044213902205228806,
0.02195708081126213,
-0.06416323035955429,
-0.055509649217128754,
0.031002400442957878,
-0.04994894564151764,
0.06461847573518753,
-0.10131804645061493,
-0.08861760050058365,
-0.0035372176207602024,
0.029002554714679718,
-0.04210813716053963,
-0.02371205762028694,
-0.003234287491068244,
-0.036226604133844376,
0.04016700014472008,
0.02593230828642845,
0.02421668730676174,
0.036439552903175354,
0.005284431856125593,
0.03630555048584938,
-0.030331842601299286,
-0.0034707176964730024,
-0.050131671130657196,
0.04059699550271034,
-0.03768476843833923,
0.08059274405241013,
0.019315680488944054,
-0.03438369929790497,
-0.06714104115962982,
-0.038226209580898285,
-0.01406241673976183,
-0.031445976346731186,
0.04194530099630356,
0.08861071616411209,
-0.02887033484876156,
-0.024651268497109413,
0.06551384925842285,
0.02575579844415188,
0.046019211411476135,
0.021165141835808754,
0.03928953781723976,
-0.032554931938648224,
0.002312880475074053,
0.017889797687530518,
0.05847976356744766,
-0.04186876490712166,
-0.09823529422283173,
0.04642953723669052,
0.00978199765086174,
-0.08231090009212494,
0.005198885686695576,
-0.03995621204376221,
0.05497349053621292,
0.02269650809466839,
0.04651157557964325,
-0.0008532695355825126,
0.011609324254095554,
0.01597895659506321,
-0.00033389063901267946,
0.038929473608732224,
-0.06150573864579201,
1.3438705085951411e-33,
-0.036571647971868515,
0.0813884511590004,
-0.032584816217422485,
0.016257394105196,
-0.06477105617523193,
0.036216605454683304,
-0.06757599115371704,
-0.014598824083805084,
0.0014450090238824487,
0.09693647176027298,
-0.09479513019323349,
0.03333711251616478,
0.02111547254025936,
0.046248000115156174,
0.04687898978590965,
0.04264131560921669,
0.0022410701494663954,
0.06406767666339874,
0.01252767350524664,
0.01984436996281147,
0.10981747508049011,
-0.06829230487346649,
0.04004637524485588,
0.014746524393558502,
-0.026250282302498817,
0.0022610758896917105,
0.04358461871743202,
-0.09162344038486481,
0.03884977102279663,
0.0035929575096815825,
-0.043955713510513306,
0.039598189294338226,
0.05203144997358322,
-0.0022453356068581343,
0.09414040297269821,
-0.04987477511167526,
0.0008844442199915648,
-0.06637607514858246,
-0.01882818527519703,
0.021582558751106262,
0.017667854204773903,
0.0197466928511858,
0.00017860958178061992,
-0.029712138697504997,
-0.09479330480098724,
0.13793765008449554,
0.041436128318309784,
0.09819818288087845,
-0.012438115663826466,
-0.043026234954595566,
0.018439117819070816,
0.034617405384778976,
-0.08507021516561508,
-0.0561482310295105,
-0.0828021690249443,
-0.0029440089128911495,
-0.029582258313894272,
-0.05721684545278549,
0.02178371138870716,
0.02581821382045746,
-0.030087988823652267,
0.08435297757387161,
0.04901750758290291,
-0.052001938223838806,
-0.07843384146690369,
-0.023775452747941017,
0.029778584837913513,
-0.015801651403307915,
0.0015724465483799577,
-0.004376737866550684,
0.0360364206135273,
0.0486929826438427,
-0.00628101360052824,
-0.05147211626172066,
-0.05974297970533371,
-0.03250620514154434,
0.07925628870725632,
0.04289887100458145,
0.03126821666955948,
0.00032072208705358207,
-0.02400468848645687,
0.0003571515844669193,
-0.023738166317343712,
-0.05254044383764267,
0.03263659402728081,
-0.08044255524873734,
-0.009836179204285145,
0.006934658158570528,
-0.040762800723314285,
-0.005575500428676605,
-0.08044587820768356,
-0.04738418012857437,
-0.04512184113264084,
0.012783519923686981,
-0.11021941155195236,
-1.78366665893393e-33,
-0.12379701435565948,
0.029014406725764275,
-0.04500887915492058,
-0.02815243974328041,
0.06860695034265518,
-0.003338754642754793,
-0.0021572099067270756,
0.051335182040929794,
0.030255675315856934,
0.007057490758597851,
-0.06021467596292496,
-0.02756953053176403,
0.09118068218231201,
-0.006023892667144537,
0.09727482497692108,
0.0031128355767577887,
-0.016460930928587914,
-0.05474906414747238,
-0.14522434771060944,
0.004820256493985653,
-0.05381414666771889,
0.03485019877552986,
-0.07109040766954422,
0.05001584067940712,
0.01971522532403469,
0.07477163523435593,
-0.07495912164449692,
-0.009998084977269173,
-0.008350508287549019,
0.07981487363576889,
-0.11097578704357147,
0.013417276553809643,
-0.08433974534273148,
-0.09423017501831055,
-0.06873903423547745,
-0.034802358597517014,
-0.012037783861160278,
0.029757782816886902,
0.003279123455286026,
-0.015317321754992008,
-0.03349579870700836,
0.043169233947992325,
-0.004341342020779848,
0.05490022897720337,
-0.006016796920448542,
-0.005566281732171774,
-0.04360494390130043,
-0.0025000390596687794,
-0.04799872636795044,
0.05840788781642914,
-0.03931247070431709,
0.08230553567409515,
-0.03353897109627724,
0.04790276288986206,
-0.07990235090255737,
-0.010417764075100422,
-0.028517430648207664,
-0.0068236165679991245,
0.11108501255512238,
0.011325156316161156,
-0.005570927634835243,
-0.08173898607492447,
-0.07141800969839096,
0.05447753146290779,
-0.022701140493154526,
-0.03350911661982536,
0.04195941984653473,
-0.10648650676012039,
-0.04906119406223297,
0.030778739601373672,
0.07332665473222733,
-0.08753285557031631,
0.0008750546840019524,
0.03520548716187477,
-0.008537063375115395,
-0.004210831597447395,
0.03094339184463024,
0.011585523374378681,
0.006853156723082066,
0.14898864924907684,
0.09368126094341278,
-0.019432663917541504,
0.10481589287519455,
0.023511240258812904,
0.07636082172393799,
-0.04281359538435936,
0.011735068634152412,
-0.05917658656835556,
-0.015296265482902527,
0.013098376803100109,
-0.07020115852355957,
0.06000160425901413,
-0.11214927583932877,
0.09869392961263657,
0.0021746607962995768,
-3.8558138015787335e-8,
0.007034554611891508,
-0.0729755163192749,
0.039159975945949554,
-0.03382233530282974,
0.013489040546119213,
-0.013422133401036263,
0.029981108382344246,
0.020386828109622,
-0.09987789392471313,
-0.043166935443878174,
0.08937027305364609,
-0.06131109222769737,
-0.056506454944610596,
-0.006997051183134317,
0.02476995438337326,
0.12430673092603683,
-0.017687715590000153,
0.029146524146199226,
-0.041739363223314285,
-0.012112388387322426,
-0.10169681161642075,
0.0302665326744318,
-0.08200086653232574,
-0.008207728154957294,
0.05264550819993019,
-0.02158660814166069,
-0.021345285698771477,
0.03450535982847214,
0.04778849333524704,
0.07884888350963593,
0.007063455879688263,
0.0178297758102417,
0.04757848381996155,
0.03619001805782318,
-0.03239315375685692,
0.014075158163905144,
0.022657137364149094,
0.05708446726202965,
-0.011877794750034809,
0.05815193057060242,
0.026769544929265976,
0.09217850863933563,
-0.004931801930069923,
0.04018673300743103,
0.030565615743398666,
0.06782271713018417,
0.08935204893350601,
-0.09556625783443451,
0.03067845292389393,
0.07093454897403717,
0.019986532628536224,
-0.06073348969221115,
-0.041669342666864395,
0.10904320329427719,
0.03389456868171692,
0.030203398317098618,
-0.006135969422757626,
-0.012334884144365788,
0.010833305306732655,
0.019854940474033356,
-0.010099589824676514,
0.03731836378574371,
-0.020410843193531036,
-0.017598478123545647
] | 0
| 0
| 0
| 0
|
|
Diffusion Transformers for Tabular Data Time Series Generation
|
https://openreview.net/forum?id=bhOysNJvWm
|
[
"Fabrizio Garuti",
"Enver Sangineto",
"Simone Luetto",
"Lorenzo Forni",
"Rita Cucchiara"
] |
Poster
|
Tabular data generation has recently attracted a growing interest due to its different application scenarios. However,
generating time series of tabular data, where each element of the series depends on the others,
remains a largely unexplored domain.
This gap is probably due to the difficulty of jointly solving different problems, the main of which are the heterogeneity of tabular data (a problem common to non-time-dependent approaches) and the variable length of a time series.
In this paper, we propose a Diffusion Transformers (DiTs) based approach for tabular data series generation. Inspired by the recent success of DiTs in image and video generation, we extend this framework to deal with heterogeneous data and variable-length sequences.
Using extensive experiments on six datasets, we show that the proposed approach outperforms previous work by a large margin.
|
tabular data generation, time series, diffusion models
| null | 11,262
| null |
[
-0.0744175836443901,
-0.03534327819943428,
0.04586009308695793,
-0.02888806164264679,
0.02315293811261654,
-0.05240099877119064,
-0.059458937495946884,
-0.09986962378025055,
0.03136107698082924,
-0.03613455966114998,
-0.02219376340508461,
0.007294423412531614,
-0.04514121636748314,
0.04293146729469299,
-0.053081005811691284,
-0.016722004860639572,
-0.05822157487273216,
0.007014748640358448,
-0.03642605245113373,
-0.05480857938528061,
-0.03377703204751015,
-0.04174988344311714,
-0.06583581864833832,
-0.0654667317867279,
0.04069465398788452,
0.10164324194192886,
0.05650440603494644,
-0.029269356280565262,
-0.006128848996013403,
-0.028215317055583,
0.0010216679656878114,
0.021015174686908722,
-0.038291431963443756,
0.050902109593153,
-0.076809361577034,
-0.03771781921386719,
-0.05431899428367615,
0.05751635134220123,
-0.025310924276709557,
0.0023037169594317675,
-0.02053106762468815,
0.03905513882637024,
0.04244251549243927,
0.022966871038079262,
-0.011480930261313915,
0.0013798036379739642,
-0.011766266077756882,
-0.044629111886024475,
-0.0061387005262076855,
0.002365688793361187,
-0.08355265110731125,
0.08375202119350433,
-0.014547034166753292,
0.10593567788600922,
0.007819516584277153,
0.013875412754714489,
0.021221240982413292,
-0.0012847287580370903,
0.06380326300859451,
-0.03317089006304741,
-0.0837659239768982,
-0.015237020328640938,
0.006441050674766302,
-0.038915231823921204,
-0.04688723385334015,
0.00638344744220376,
0.0524270236492157,
0.07598783820867538,
-0.0019760397262871265,
-0.014137941412627697,
-0.10809558629989624,
0.08072344958782196,
-0.058842986822128296,
0.02490997686982155,
-0.06231805309653282,
0.000010196475159318652,
0.034341275691986084,
-0.05848914757370949,
-0.03466019406914711,
-0.08743289858102798,
0.042783185839653015,
-0.015043489634990692,
0.048926811665296555,
0.0606757290661335,
0.04239577427506447,
0.004322177730500698,
0.039010729640722275,
0.06854795664548874,
-0.06775203347206116,
-0.031129315495491028,
-0.06828688830137253,
0.08453678339719772,
0.029377326369285583,
-0.01069051492959261,
-0.060824960470199585,
0.03515886887907982,
0.00871041789650917,
-0.03584953770041466,
0.09265519678592682,
0.0609053410589695,
0.015994654968380928,
-0.052952904254198074,
-0.003213130636140704,
0.039761923253536224,
0.023604106158018112,
-0.0876532793045044,
0.0979563295841217,
-0.005281477700918913,
0.032187484204769135,
-0.010625295341014862,
0.04550757631659508,
-0.012149655260145664,
-0.005062354728579521,
-0.03759509325027466,
0.01858985796570778,
-0.09988400340080261,
-0.052704282104969025,
-0.03155902773141861,
0.02117365412414074,
0.035028815269470215,
0.07130132615566254,
-0.03792927414178848,
-0.11995428800582886,
0.0834796205163002,
-0.011358089745044708,
-0.00875805038958788,
0.06277186423540115,
1.36731243734443e-33,
-0.07011406868696213,
-0.04928690567612648,
0.0601649172604084,
0.05625852569937706,
0.09591671824455261,
0.03289347514510155,
-0.0031784120947122574,
-0.048724982887506485,
0.007932066917419434,
-0.01274413987994194,
-0.0847838893532753,
-0.005519653204828501,
-0.10718752443790436,
0.006509113125503063,
0.05571196228265762,
-0.041527289897203445,
-0.005113480146974325,
0.061332762241363525,
-0.013976635411381721,
0.02920113131403923,
-0.007313339971005917,
-0.031203024089336395,
0.030555324628949165,
0.007758262567222118,
0.0015377388335764408,
-0.03652864322066307,
0.05727819725871086,
-0.049007125198841095,
0.11857278645038605,
-0.018204830586910248,
-0.03822365403175354,
0.019605513662099838,
-0.0015546329086646438,
-0.02676718309521675,
0.01722000353038311,
0.007678060792386532,
-0.0352947823703289,
-0.0024355934001505375,
0.035769861191511154,
0.03941573202610016,
0.04322144761681557,
0.06812117993831635,
-0.0629599392414093,
-0.022050485014915466,
-0.04657846689224243,
0.03254279866814613,
0.12447380274534225,
0.0562942698597908,
-0.016217680647969246,
-0.009972943924367428,
0.03299148008227348,
-0.017603252083063126,
0.008527047000825405,
-0.08762270957231522,
0.032307591289281845,
0.03655347228050232,
0.04353690892457962,
-0.07488071173429489,
0.07278769463300705,
0.06695960462093353,
-0.014421237632632256,
0.026545679196715355,
-0.036364275962114334,
0.0013211960904300213,
-0.01648448035120964,
-0.050144195556640625,
0.06177140772342682,
0.07721688598394394,
0.008293076418340206,
0.008139925077557564,
0.03425796329975128,
-0.030622640624642372,
-0.010026829317212105,
-0.11650275439023972,
0.06003740802407265,
-0.019336972385644913,
0.0974392369389534,
0.07921053469181061,
-0.058124978095293045,
-0.012203178368508816,
0.015025998465716839,
-0.02741090953350067,
0.08359888941049576,
-0.13847051560878754,
-0.008684405125677586,
-0.00889354757964611,
-0.0016780435107648373,
0.0036219817120581865,
-0.09773854166269302,
-0.08958777785301208,
-0.015155977569520473,
0.05536981672048569,
-0.03088991343975067,
0.012778231874108315,
0.03882830590009689,
-1.9493108943670758e-33,
-0.00207798695191741,
0.07799652218818665,
0.03291304036974907,
0.047233954071998596,
0.04109390825033188,
0.07798314839601517,
-0.030812492594122887,
0.03845864534378052,
0.13487574458122253,
0.03617881238460541,
-0.036012135446071625,
-0.10521437972784042,
-0.04167192801833153,
-0.08721152693033218,
-0.039403121918439865,
-0.03252740576863289,
0.06010083481669426,
-0.05019856616854668,
-0.11104251444339752,
-0.015986589714884758,
-0.0020540058612823486,
0.06079520657658577,
-0.05357229337096214,
0.001953186932951212,
-0.05520985648036003,
0.00929504819214344,
0.014293896965682507,
0.08130582422018051,
-0.02459840476512909,
0.0008569282945245504,
-0.04007570445537567,
-0.04354742169380188,
0.055798519402742386,
-0.046824902296066284,
-0.10783863812685013,
0.030906569212675095,
0.06108393892645836,
0.022282183170318604,
-0.04291805252432823,
0.06886037439107895,
-0.004197007045149803,
0.02019505575299263,
0.012858354486525059,
0.027458589524030685,
-0.045542579144239426,
-0.009922569617629051,
-0.010048821568489075,
0.061262283474206924,
-0.03305695205926895,
0.06658531725406647,
0.012083400040864944,
0.06764768809080124,
-0.06901012361049652,
0.008036930114030838,
0.01763295568525791,
-0.015705034136772156,
-0.07028127461671829,
0.016728224232792854,
0.025773311033844948,
-0.036873675882816315,
-0.06876665353775024,
-0.07086067646741867,
0.044161371886730194,
-0.03791327401995659,
-0.005127502139657736,
0.012674326077103615,
-0.025502948090434074,
-0.08919953554868698,
-0.04415448382496834,
0.10241460800170898,
0.03433449938893318,
0.032541222870349884,
-0.0014390863943845034,
-0.06133754551410675,
0.0027533364482223988,
-0.03324227035045624,
-0.023082943633198738,
-0.012300443835556507,
-0.005779739934951067,
-0.02950693480670452,
-0.03850596770644188,
0.0415043905377388,
0.11349745094776154,
-0.04675975441932678,
-0.027758365496993065,
0.05682343244552612,
0.06125239282846451,
-0.013232659548521042,
0.032205019146203995,
0.06674343347549438,
-0.013008421286940575,
-0.06701138615608215,
-0.024820096790790558,
0.07881547510623932,
0.0220014825463295,
-3.8990773276736945e-8,
-0.04005005955696106,
0.020891796797513962,
-0.02351781725883484,
0.01180334109812975,
0.009259884245693684,
-0.09609785676002502,
0.11922211945056915,
0.10189233720302582,
0.07705765217542648,
-0.08249488472938538,
0.1264704018831253,
-0.040650393813848495,
-0.034240055829286575,
0.01317945308983326,
-0.014195005409419537,
0.08297671377658844,
-0.02311217039823532,
-0.058098334819078445,
0.011090588755905628,
0.06009649112820625,
-0.01105977687984705,
0.08704078197479248,
-0.0016451054252684116,
0.03879796713590622,
0.03645273670554161,
0.05286319553852081,
-0.09545254707336426,
-0.010971306823194027,
0.0942029058933258,
-0.07872574031352997,
0.015741094946861267,
0.0033742752857506275,
0.008165654726326466,
-0.08225344866514206,
-0.014484603889286518,
-0.029072310775518417,
0.018585629761219025,
0.01633176952600479,
-0.09303156286478043,
-0.06269649416208267,
0.0010750207584351301,
0.04746852070093155,
-0.013431045226752758,
-0.010782570578157902,
-0.02845323644578457,
0.021749677136540413,
0.021600069478154182,
0.0036457751411944628,
0.02013246901333332,
0.00006589490658370778,
0.0687733069062233,
-0.048170510679483414,
-0.0017837114864960313,
-0.025145437568426132,
0.1363213062286377,
0.0029690731316804886,
0.0658211037516594,
-0.004870587028563023,
0.05102551728487015,
-0.0008586418116465211,
-0.00003409127748454921,
-0.0043894886039197445,
-0.00598563440144062,
0.04510600492358208
] | 0
| 0
| 0
| 0
|
|
Non-Equilibrium Dynamics of Hybrid Continuous-Discrete Ground-State Sampling
|
https://openreview.net/forum?id=BlSIKSPhfz
|
[
"Timothee Leleu",
"Sam Reifenstein"
] |
Poster
|
We propose a general framework for a hybrid continuous-discrete algorithm that integrates continuous-time deterministic dynamics with Metropolis-Hastings (MH) steps to combine search dynamics that either preserve or break detailed balance. Our purpose is to study the non-equilibrium dynamics that leads to the ground state of rugged energy landscapes in this general setting. Our results show that MH-driven dynamics reach ``easy'' ground states more quickly, indicating a stronger bias toward these solutions in algorithms using reversible transition probabilities. To validate this, we construct a set of Ising problem instances with a controllable bias in the energy landscape that makes certain degenerate solutions more accessible than others. The constructed hybrid algorithm demonstrates significant improvements in convergence and ground-state sampling accuracy, achieving a 100x speedup on GPU compared to simulated annealing, making it well-suited for large-scale applications.
|
Combinatorial optimization, Degenerate ground-state sampling, Metropolis-Hastings algorithm, Chaotic dynamics, Wishart planted ensemble
| null | 11,258
|
2410.22625
|
[
-0.01671232283115387,
-0.0020462770480662584,
0.05404001101851463,
0.027382515370845795,
0.05399694666266441,
-0.1306636780500412,
-0.04082462564110756,
-0.03112773224711418,
0.002825751667842269,
-0.010897581465542316,
-0.10477083176374435,
-0.04801090434193611,
-0.023097269237041473,
-0.022641729563474655,
-0.001959956483915448,
0.004300132393836975,
0.06362193077802658,
-0.003020143136382103,
0.06388317793607712,
-0.01661541312932968,
-0.05110183358192444,
0.04802411422133446,
0.03678187355399132,
-0.01118149608373642,
-0.0014772278955206275,
0.03318542614579201,
0.056261077523231506,
0.003398720407858491,
0.03538064658641815,
-0.007654708344489336,
0.015076061710715294,
-0.0013213349739089608,
-0.0035771813709288836,
-0.029469823464751244,
0.03201635181903839,
0.107925184071064,
-0.07415888458490372,
0.012439301237463951,
-0.026590365916490555,
-0.003939134534448385,
0.009791492484509945,
0.07933089882135391,
0.08904831856489182,
0.012983907014131546,
0.06256823241710663,
0.008834300562739372,
0.02460210956633091,
-0.06203669309616089,
0.04527502506971359,
-0.043650902807712555,
0.029426511377096176,
0.045583516359329224,
-0.04374147951602936,
0.005491584539413452,
0.023256395012140274,
0.022640876471996307,
0.13517914712429047,
-0.0173750352114439,
0.016941484063863754,
-0.027956265956163406,
0.04861700162291527,
-0.08996568620204926,
0.018099499866366386,
-0.006090643350034952,
0.026281464844942093,
0.00861865933984518,
0.10670323669910431,
0.08446326106786728,
0.03167632967233658,
-0.02076624520123005,
-0.030940458178520203,
0.06549331545829773,
-0.006574005354195833,
-0.02572391740977764,
0.052133895456790924,
0.027087043970823288,
-0.03027532622218132,
0.025251351296901703,
0.047129999846220016,
-0.04159875586628914,
-0.03999222069978714,
-0.047830089926719666,
0.06049508601427078,
-0.027749648317694664,
-0.02506151981651783,
-0.026562483981251717,
-0.06565205007791519,
0.08357223123311996,
0.19785308837890625,
-0.06564220041036606,
-0.04093781113624573,
0.05444088578224182,
-0.07553490251302719,
-0.048920534551143646,
0.019808582961559296,
-0.001967384945601225,
0.0007539045764133334,
-0.0689433142542839,
0.02227994240820408,
0.027330275624990463,
0.014466372318565845,
0.03106447495520115,
0.08973397314548492,
0.07336723804473877,
0.009867504239082336,
-0.06429021805524826,
-0.002038492588326335,
0.02323170378804207,
-0.025888850912451744,
-0.005719760898500681,
0.050385188311338425,
-0.009629080072045326,
0.13152624666690826,
0.06431710720062256,
-0.047405824065208435,
-0.02128124237060547,
0.014436434954404831,
-0.0831177681684494,
-0.022778529673814774,
0.04289931058883667,
-0.0741456151008606,
-0.036643072962760925,
-0.02179853804409504,
0.022042084485292435,
0.05655868723988533,
0.0278744176030159,
0.012987745925784111,
-4.592508114336382e-34,
-0.006482372060418129,
-0.021227506920695305,
0.0872548446059227,
0.016895651817321777,
0.0391346737742424,
0.00014575499517377466,
0.03765963017940521,
-0.015743762254714966,
-0.07127054035663605,
0.059024278074502945,
0.0412760004401207,
0.04378427192568779,
-0.00028842463507317007,
0.01975073851644993,
0.04863424599170685,
-0.02883290871977806,
-0.022639978677034378,
0.0033691683784127235,
-0.013090427033603191,
-0.0803496390581131,
0.0444965697824955,
-0.007310761604458094,
-0.05595158413052559,
-0.033190660178661346,
-0.0067430529743433,
0.007395764347165823,
0.008798740804195404,
0.002371666720137,
0.018248673528432846,
-0.031205082312226295,
-0.0668017715215683,
0.01006622426211834,
-0.06169847771525383,
0.09829382598400116,
0.08605987578630447,
0.08207544684410095,
0.007123232819139957,
-0.03685593605041504,
-0.05984049662947655,
-0.03463314101099968,
0.019909361377358437,
0.019090035930275917,
0.10773071646690369,
-0.03252570331096649,
-0.08847245573997498,
-0.06291847676038742,
0.044052328914403915,
-0.05999171361327171,
-0.06012079864740372,
0.005720367189496756,
0.028744805604219437,
-0.015183727256953716,
-0.026229092851281166,
-0.02782857045531273,
0.019430149346590042,
0.011137604713439941,
0.043568938970565796,
0.004726817831397057,
0.015254415571689606,
0.08960747718811035,
0.011970635503530502,
-0.011524500325322151,
-0.06532593816518784,
0.00232395576313138,
0.02037816494703293,
0.12265270203351974,
-0.05982917919754982,
0.04012323543429375,
0.07274194061756134,
0.05842418596148491,
0.03248567506670952,
0.010253407061100006,
0.010609081014990807,
-0.04641118273139,
0.05887680873274803,
-0.03425494581460953,
0.03628721088171005,
0.0006218687631189823,
-0.04798083007335663,
0.0008746537496335804,
-0.021948236972093582,
0.002653163857758045,
-0.10739657282829285,
-0.026114163920283318,
-0.0067853122018277645,
-0.10942814499139786,
-0.002568383002653718,
-0.036243170499801636,
-0.0521216057240963,
-0.049481261521577835,
-0.017794601619243622,
-0.06358738243579865,
0.09041325002908707,
0.0725630670785904,
-0.014948489144444466,
-1.2452329409077198e-33,
-0.06040653586387634,
-0.03610566258430481,
0.02774118445813656,
0.05194297432899475,
-0.007263782899826765,
-0.01303081028163433,
0.06119094789028168,
-0.04861278086900711,
0.05972224101424217,
-0.13249655067920685,
-0.10502005368471146,
0.012798196636140347,
0.1126638799905777,
0.03746258467435837,
0.0029980652034282684,
0.00987281370908022,
-0.008843490853905678,
-0.003781039733439684,
0.05972250923514366,
0.028004100546240807,
0.019832413643598557,
0.05732390657067299,
-0.08438196778297424,
-0.04591156914830208,
-0.022049788385629654,
0.049163028597831726,
-0.007802185136824846,
0.03162093088030815,
0.018115170300006866,
0.0047455825842916965,
-0.04715310409665108,
0.021043969318270683,
-0.08600063621997833,
0.04604214429855347,
-0.07883621752262115,
0.05732748284935951,
-0.03211173415184021,
0.0610036738216877,
0.0061800517141819,
0.039795972406864166,
-0.010835312306880951,
-0.07652367651462555,
-0.04670368880033493,
0.02571822889149189,
0.07606740295886993,
0.008115818724036217,
-0.04824094846844673,
0.04364289343357086,
0.026568986475467682,
-0.003099168185144663,
-0.05255928635597229,
0.026416735723614693,
-0.06492477655410767,
0.10030487924814224,
-0.025948012247681618,
-0.010370760224759579,
-0.05701330676674843,
0.11788035929203033,
0.0008209606749005616,
-0.02995702438056469,
-0.06531461328268051,
-0.039967332035303116,
0.0044531743042171,
0.009365690872073174,
-0.0801425352692604,
0.0025819416623562574,
-0.05746791884303093,
-0.06324125081300735,
0.0077125816605985165,
0.01199301052838564,
-0.08483286201953888,
0.014746242202818394,
0.05105430632829666,
-0.04766939580440521,
-0.0401703342795372,
-0.051571011543273926,
0.07834728062152863,
-0.05362750217318535,
0.009614530019462109,
-0.08977663516998291,
-0.0027876500971615314,
0.07238715887069702,
-0.004570987541228533,
0.04646964371204376,
0.004725571721792221,
-0.04202309623360634,
-0.0629395842552185,
-0.048074789345264435,
-0.0011728081153705716,
-0.06969200074672699,
0.01914988085627556,
-0.042588602751493454,
0.04900423437356949,
0.01649567112326622,
-0.08601754903793335,
-4.414399512597811e-8,
-0.02736886963248253,
0.040738340467214584,
0.06702443212270737,
-0.017433511093258858,
0.07724539935588837,
0.06501106172800064,
0.04668821394443512,
-0.07608689367771149,
0.009204447269439697,
-0.0773402750492096,
0.12611745297908783,
-0.060759592801332474,
0.0214600320905447,
-0.045222654938697815,
-0.038940247148275375,
0.0199994295835495,
0.01656314544379711,
-0.09056197106838226,
-0.02921033836901188,
-0.031771134585142136,
-0.10609744489192963,
-0.01591419242322445,
-0.020625950768589973,
-0.0025877528823912144,
0.0015105052152648568,
0.014471586793661118,
0.024836571887135506,
-0.10902044177055359,
0.08439242839813232,
0.008910966105759144,
-0.025821469724178314,
0.03836315870285034,
0.03293781355023384,
0.07574187964200974,
0.011326668784022331,
0.0059623271226882935,
-0.015336830168962479,
0.1284608393907547,
0.046248286962509155,
-0.1051882654428482,
-0.001582161639817059,
0.06862593442201614,
-0.10774988681077957,
-0.030259696766734123,
-0.03243277966976166,
0.06308849155902863,
-0.060915011912584305,
-0.005773012060672045,
0.06441327184438705,
0.058224912732839584,
0.005315443500876427,
-0.013949327170848846,
0.008100000210106373,
0.0351489819586277,
0.045787740498781204,
-0.0206486564129591,
-0.02846347913146019,
-0.030568066984415054,
0.028325900435447693,
0.05713806301355362,
0.006361429579555988,
-0.06101224571466446,
-0.0827801302075386,
-0.021165242418646812
] | 0
| 0
| 0
| 0
|
|
The Belief State Transformer
|
https://openreview.net/forum?id=ThRMTCgpvo
|
[
"Edward S. Hu",
"Kwangjun Ahn",
"Qinghua Liu",
"Haoran Xu",
"Manan Tomar",
"Ada Langford",
"Dinesh Jayaraman",
"Alex Lamb",
"John Langford"
] |
Poster
|
We introduce the "Belief State Transformer", a next-token predictor that takes both a prefix and suffix as inputs, with a novel objective of predicting both the next token for the prefix and the previous token for the suffix. The Belief State Transformer effectively learns to solve challenging problems that conventional forward-only transformers struggle with, in a domain-independent fashion. Key to this success is learning a compact belief state that captures all relevant information necessary for accurate predictions.
Empirical ablations show that each component of the model is essential in difficult scenarios where standard Transformers fall short.
For the task of story writing with known prefixes and suffixes, our approach outperforms the Fill-in-the-Middle method for reaching known goals and demonstrates improved performance even when the goals are unknown. Altogether, the Belief State Transformer enables more efficient goal-conditioned decoding, better test-time inference, and high-quality text representations on small scale problems. Website: https://sites.google.com/view/belief-state-transformer/
|
representation learning, transformers, next-token prediction, reasoning, planning
|
We show that transformers trained with belief state representations solve tasks that conventional next-token predictors (e.g. GPT) cannot.
| 11,256
|
2410.23506
|
[
-0.07665221393108368,
-0.048720479011535645,
0.04245411604642868,
0.024679461494088173,
0.02387895993888378,
-0.02033967152237892,
-0.030554242432117462,
-0.00045247189700603485,
0.06899513304233551,
-0.017270438373088837,
-0.07135305553674698,
0.012451207265257835,
0.04145865514874458,
0.064084991812706,
-0.020006561651825905,
0.04211990907788277,
-0.05817610025405884,
0.05743539705872536,
-0.09087857604026794,
-0.11888639628887177,
0.058752067387104034,
0.047446515411138535,
-0.04670076444745064,
0.012503235600888729,
0.08224190771579742,
0.04198979213833809,
0.0010226621525362134,
-0.06995796412229538,
0.010806970298290253,
-0.009836636483669281,
-0.01421012170612812,
-0.03069138526916504,
0.019133897498250008,
-0.0358063280582428,
-0.006923897657543421,
0.04630301520228386,
-0.07974881678819656,
0.030492521822452545,
-0.0004409321991261095,
-0.04076419770717621,
0.01421478670090437,
-0.0863024890422821,
0.04325603321194649,
0.0409543514251709,
0.06980329751968384,
0.013102552853524685,
0.03925243392586708,
-0.03172559291124344,
-0.03567275032401085,
0.017120899632573128,
-0.047484591603279114,
-0.028259124606847763,
-0.0016448048409074545,
0.05263320729136467,
-0.06419266760349274,
0.01746089942753315,
0.028338754549622536,
0.01138217281550169,
-0.0026090939063578844,
-0.05240464210510254,
-0.07871296256780624,
-0.08246798813343048,
-0.045516517013311386,
-0.020071998238563538,
-0.050016723573207855,
-0.0113137848675251,
0.06585916876792908,
0.08025374263525009,
-0.06964144110679626,
0.01346121821552515,
-0.03762822970747948,
0.09068677574396133,
-0.01638459973037243,
0.060927968472242355,
0.06780900806188583,
0.004414848051965237,
0.04812540113925934,
-0.13173717260360718,
0.027556726709008217,
0.0010149425361305475,
0.02321658469736576,
-0.0052541703917086124,
0.0620901994407177,
0.07664987444877625,
0.0912986770272255,
-0.02097800001502037,
0.020967945456504822,
0.03394543379545212,
-0.05615909397602081,
0.03392019867897034,
-0.048973504453897476,
-0.20598585903644562,
0.07507214695215225,
0.01284047681838274,
-0.043932799249887466,
0.08760131895542145,
-0.029313795268535614,
-0.06482984870672226,
-0.04641372337937355,
0.05527786910533905,
-0.06749247014522552,
0.037202782928943634,
0.04292606562376022,
-0.08197521418333054,
0.025439485907554626,
-0.05795664340257645,
-0.010079382918775082,
-0.018289677798748016,
0.02390320599079132,
-0.040343351662158966,
0.02957199141383171,
0.017950473353266716,
-0.036459267139434814,
-0.03489099070429802,
0.048457127064466476,
0.0022289128974080086,
0.004234956577420235,
0.0004153249319642782,
-0.04055160656571388,
0.04960630461573601,
-0.0068468875251710415,
0.042677976191043854,
-0.03168468922376633,
0.0983150452375412,
0.013165579177439213,
-0.04010675847530365,
0.03479986637830734,
1.2131124659360047e-33,
-0.004739437717944384,
0.0893075168132782,
0.019959088414907455,
0.044434648007154465,
-0.009269094094634056,
0.043038129806518555,
-0.04652675241231918,
0.03067743591964245,
-0.040801674127578735,
-0.01028426457196474,
-0.018260886892676353,
0.03973517566919327,
-0.012390854768455029,
0.0065003978088498116,
0.010648110881447792,
-0.0733914002776146,
-0.06050904467701912,
0.03186797350645065,
-0.0660393163561821,
0.002264526439830661,
0.03022652678191662,
0.03425111621618271,
-0.008108649402856827,
-0.08002821356058121,
-0.011557680554687977,
0.004462259355932474,
0.07610955089330673,
-0.051651619374752045,
0.00043914272100664675,
-0.013130011036992073,
-0.07501838356256485,
0.002099005039781332,
-0.03660758212208748,
-0.004606170114129782,
0.04891600087285042,
-0.06056695058941841,
0.010351981036365032,
-0.08659067749977112,
-0.015932466834783554,
-0.02553769014775753,
-0.03855634480714798,
0.014993970282375813,
-0.030390067026019096,
0.01599769853055477,
0.004435951821506023,
-0.046021364629268646,
-0.012275266461074352,
-0.05503593385219574,
-0.009657365269958973,
-0.021470924839377403,
0.06020815670490265,
-0.010044850409030914,
-0.028025507926940918,
-0.027547329664230347,
0.05524959787726402,
-0.01349717192351818,
0.07069354504346848,
-0.029456727206707,
0.05311867594718933,
-0.0036613433621823788,
0.019950199872255325,
-0.005665548611432314,
-0.008355606347322464,
0.06566227972507477,
-0.020694704726338387,
0.04561328515410423,
-0.007355745416134596,
0.0190507173538208,
0.036159612238407135,
-0.015651080757379532,
-0.010862473398447037,
-0.028774674981832504,
0.00977762695401907,
-0.06785321980714798,
0.003881185781210661,
0.03940741345286369,
0.05101704224944115,
-0.008523236028850079,
-0.05268058925867081,
0.023956166580319405,
0.0038386175874620676,
-0.048639751970767975,
-0.027053365483880043,
0.009511088021099567,
0.06309685856103897,
-0.017749525606632233,
0.0014046241994947195,
-0.09913578629493713,
-0.06905931234359741,
-0.041938602924346924,
-0.06945574283599854,
0.03271757811307907,
-0.01014066394418478,
0.03553367406129837,
0.02544955536723137,
3.5927710615374026e-34,
-0.002635158831253648,
0.0386236347258091,
-0.000017585236491868272,
0.06861037015914917,
-0.022258147597312927,
-0.14837849140167236,
-0.011486892588436604,
0.04366525262594223,
0.053655270487070084,
-0.05255027487874031,
-0.059630539268255234,
-0.010698202066123486,
0.02462545409798622,
-0.020613301545381546,
0.026987485587596893,
-0.011904536746442318,
-0.0119852339848876,
0.02911575883626938,
-0.029138773679733276,
0.03159701079130173,
0.0015719523653388023,
0.0890229120850563,
-0.14910051226615906,
0.0164394062012434,
0.02329300530254841,
0.017974665388464928,
0.008546238765120506,
0.08034781366586685,
0.01779891550540924,
-0.05998734384775162,
-0.05844902619719505,
-0.05096030980348587,
0.05439472571015358,
0.027988990768790245,
-0.1142452135682106,
0.0509856715798378,
0.029294414445757866,
0.017481455579400063,
-0.018947992473840714,
-0.009774674661457539,
0.08634260296821594,
-0.010519962757825851,
-0.04487280920147896,
0.003494885517284274,
-0.0717259868979454,
-0.03039892390370369,
-0.011741339229047298,
0.047847166657447815,
0.030768126249313354,
0.062359556555747986,
-0.02358512580394745,
0.04640650376677513,
-0.12279053777456284,
0.000877922517247498,
-0.0018855920061469078,
-0.07034333050251007,
-0.11624374240636826,
-0.010313530452549458,
0.014501026831567287,
-0.01088661327958107,
-0.01727154105901718,
0.06995826959609985,
0.14383569359779358,
-0.007478304207324982,
0.033971283584833145,
-0.056189581751823425,
-0.022227216511964798,
0.011305635794997215,
0.03230597451329231,
0.007998323999345303,
0.05069641396403313,
0.007653344888240099,
-0.010811038315296173,
0.02561195008456707,
0.08132703602313995,
-0.010077979415655136,
-0.014388147741556168,
-0.05545993894338608,
-0.033896736800670624,
-0.0172637440264225,
-0.028019478544592857,
0.00024082386516965926,
0.03834358975291252,
0.09085696190595627,
0.07374333590269089,
0.12479233741760254,
0.07512957602739334,
0.07528609037399292,
0.09490564465522766,
0.05496387928724289,
-0.05786486342549324,
0.1359139382839203,
0.09414628148078918,
0.09356831759214401,
-0.08648781478404999,
-4.075286952343049e-8,
-0.1007174625992775,
-0.0037078834138810635,
-0.03288903459906578,
0.0056863268837332726,
0.09197455644607544,
0.016357874497771263,
0.010035219602286816,
-0.019918503239750862,
-0.04212024062871933,
-0.011969484388828278,
0.0772583931684494,
-0.03814978525042534,
-0.0992221087217331,
-0.031054344028234482,
0.029455961659550667,
0.02992927096784115,
0.11310220509767532,
-0.009912651963531971,
-0.03141877427697182,
0.021135222166776657,
0.07381024956703186,
0.09393569827079773,
-0.04580531269311905,
-0.005457178223878145,
0.0005163956084288657,
-0.04020749777555466,
-0.07011046260595322,
0.05964080989360809,
-0.0019780718721449375,
0.02667846530675888,
0.017847619950771332,
0.03698639199137688,
0.023387111723423004,
-0.025948457419872284,
-0.01652410440146923,
0.10318945348262787,
0.10083669424057007,
0.006367246620357037,
0.004104133229702711,
-0.017890287563204765,
0.14375072717666626,
0.04046926274895668,
-0.02894902415573597,
0.05251174047589302,
-0.05112392082810402,
-0.020574349910020828,
0.019926289096474648,
-0.13780416548252106,
0.03648817166686058,
-0.017304278910160065,
0.04886121675372124,
-0.0295846126973629,
-0.03832519054412842,
0.06470852345228195,
0.07782901078462601,
0.02984229288995266,
0.026281004771590233,
0.004522417671978474,
0.009766188450157642,
0.056036848574876785,
0.021421480923891068,
0.019400570541620255,
-0.015515985898673534,
0.03519507870078087
] | 0
| 0
| 0
| 0
|
|
Bridging the Data Provenance Gap Across Text, Speech, and Video
|
https://openreview.net/forum?id=G5DziesYxL
|
[
"Shayne Longpre",
"Nikhil Singh",
"Manuel Cherep",
"Kushagra Tiwary",
"Joanna Materzynska",
"William Brannon",
"Robert Mahari",
"Naana Obeng-Marnu",
"Manan Dey",
"Mohammed Hamdy",
"Nayan Saxena",
"Ahmad Mustafa Anis",
"Emad A. Alghamdi",
"Vu Minh Chien",
"Da Yin",
"Kun Qian",
"Yizhi LI",
"Minnie Liang",
"An Dinh",
"Shrestha Mohanty",
"et al. (23 additional authors not shown)"
] |
Poster
|
Progress in AI is driven largely by the scale and quality of training data. Despite this, there is a deficit of empirical analysis examining the attributes of well-established datasets beyond text. In this work we conduct the largest and first-of-its-kind longitudinal audit across modalities --- popular text, speech, and video datasets --- from their detailed sourcing trends and use restrictions to their geographical and linguistic representation. Our manual analysis covers nearly 4000 public datasets between 1990-2024, spanning 608 languages, 798 sources, 659 organizations, and 67 countries. We find that multimodal machine learning applications have overwhelmingly turned to web-crawled, synthetic, and social media platforms, such as YouTube, for their training sets, eclipsing all other sources since 2019. Secondly, tracing the chain of dataset derivations we find that while less than 33% of datasets are restrictively licensed, over 80% of the source content in widely-used text, speech, and video datasets, carry non-commercial restrictions. Finally, counter to the rising number of languages and geographies represented in public AI training datasets, our audit demonstrates measures of relative geographical and multilingual representation have failed to significantly improve their coverage since 2013. We believe the breadth of our audit enables us to empirically examine trends in data sourcing, restrictions, and Western-centricity at an ecosystem-level, and that visibility into these questions are essential to progress in responsible AI. As a contribution to ongoing improvements in dataset transparency and responsible use, we release our entire multimodal audit, allowing practitioners to trace data provenance across text, speech, and video.
|
training data, audit, speech, video, text
|
An audit of text, speech, and video training sets for their sources, use restrictions, and representation.
| 11,243
|
2412.17847
|
[
-0.03871411457657814,
-0.1399213969707489,
-0.006427028216421604,
0.022081289440393448,
0.17418010532855988,
0.04173452779650688,
-0.02252395637333393,
-0.05960540845990181,
0.022947195917367935,
-0.02173648029565811,
-0.013023417443037033,
-0.019766859710216522,
0.010857180692255497,
0.03909226506948471,
-0.012122401967644691,
-0.010766602121293545,
0.0316694900393486,
0.015257437713444233,
-0.06823974847793579,
-0.04517137631773949,
0.011274874210357666,
0.05352158099412918,
0.08251672983169556,
0.02058778516948223,
0.06906890869140625,
-0.00041928229620680213,
-0.01824631355702877,
-0.07094401121139526,
0.03365432098507881,
-0.03771768510341644,
-0.00920372549444437,
0.03410506993532181,
-0.037092700600624084,
0.04949508607387543,
-0.011360462754964828,
0.05355776101350784,
-0.01698085106909275,
0.033339302986860275,
0.010188487358391285,
0.01706184446811676,
0.029154084622859955,
-0.058008525520563126,
0.021543970331549644,
-0.01762058585882187,
0.045931149274110794,
0.029974892735481262,
-0.010516230948269367,
-0.04656941816210747,
0.005247242748737335,
0.05465029180049896,
-0.15198783576488495,
-0.01153701450675726,
0.0770219936966896,
0.012995974160730839,
-0.06458364427089691,
-0.12615856528282166,
0.06856586784124374,
0.04287632554769516,
0.024994391947984695,
-0.0015577540034428239,
0.0017892398172989488,
-0.08922206610441208,
-0.031394295394420624,
0.025418739765882492,
0.002248422708362341,
-0.004367003683000803,
0.014737212099134922,
0.053010355681180954,
-0.04294014349579811,
-0.024065932258963585,
-0.0024722483940422535,
0.07571419328451157,
-0.03157174587249756,
0.09119366109371185,
0.01853112317621708,
0.017637012526392937,
0.038716670125722885,
-0.0004973579780198634,
0.03504815697669983,
-0.17109215259552002,
0.0693192183971405,
-0.008060982450842857,
0.06602860987186432,
0.038258347660303116,
0.016510849818587303,
-0.04395049437880516,
-0.0035997626837342978,
-0.001387579832226038,
0.01651082932949066,
0.047493528574705124,
-0.04776093736290932,
0.0044939699582755566,
0.10149825364351273,
0.01794431358575821,
0.0677853375673294,
0.014575108885765076,
-0.013312955386936665,
-0.05951359122991562,
0.05394849181175232,
0.09288957715034485,
-0.05840276554226875,
0.05149989575147629,
0.01208637934178114,
-0.023540319874882698,
0.003789112903177738,
-0.04254806041717529,
0.03606194630265236,
0.002597762504592538,
0.060223959386348724,
0.04618144407868385,
0.014204343780875206,
0.07819360494613647,
-0.033679697662591934,
-0.06019052863121033,
0.10571625083684921,
-0.03275493532419205,
-0.061076559126377106,
0.0017317108577117324,
0.010799496434628963,
0.07502257823944092,
-0.04954344034194946,
0.06040771305561066,
0.06944331526756287,
-0.014950043521821499,
0.02782241441309452,
-0.02615508809685707,
-0.05220944434404373,
-7.680214941484786e-34,
0.12436007708311081,
-0.018459070473909378,
0.03940148651599884,
0.024086257442831993,
0.009142904542386532,
-0.035557180643081665,
-0.05217595770955086,
0.01844724453985691,
-0.07300637662410736,
-0.004875149577856064,
-0.014572711661458015,
0.10636408627033234,
-0.04941384494304657,
0.049493204802274704,
0.02625461108982563,
0.04403374344110489,
-0.003873012727126479,
0.03883882984519005,
0.02284820005297661,
0.0068324171006679535,
0.09098129719495773,
-0.06470153480768204,
0.07449205219745636,
0.04878465458750725,
-0.030187787488102913,
-0.05310765653848648,
-0.006806362420320511,
-0.03041536919772625,
0.02028593420982361,
0.0401112399995327,
-0.09652348607778549,
-0.01734369993209839,
0.10033395141363144,
0.029383817687630653,
0.06625139713287354,
-0.05472470819950104,
-0.06624644249677658,
-0.017550168558955193,
0.03303546831011772,
0.034734178334474564,
-0.027279779314994812,
-0.06161684915423393,
-0.025194311514496803,
-0.03095739521086216,
-0.03754723444581032,
-0.034272048622369766,
-0.017458295449614525,
-0.03236805275082588,
-0.0007808817317709327,
-0.02476426400244236,
0.00941858533769846,
0.07023956626653671,
-0.021798361092805862,
-0.04974764585494995,
0.00786236859858036,
0.047495946288108826,
-0.009949632920324802,
-0.0051226248033344746,
0.05253736674785614,
-0.0879536047577858,
-0.03247813880443573,
0.08149055391550064,
0.04333864524960518,
0.07928235828876495,
-0.04185633361339569,
0.04325712099671364,
0.007142307236790657,
0.015345534309744835,
0.07189065217971802,
0.01582701876759529,
-0.01564827933907509,
0.0034989884588867426,
-0.07314836233854294,
-0.0346522256731987,
0.027453839778900146,
-0.013875036500394344,
0.01679193787276745,
-0.05272684618830681,
0.046640194952487946,
0.06853418797254562,
0.014348741620779037,
-0.03228040039539337,
0.05390306934714317,
-0.04361974447965622,
-0.07297080010175705,
-0.04427535831928253,
-0.02259032428264618,
-0.06572272628545761,
0.05053279921412468,
0.018384402617812157,
-0.11684530973434448,
-0.043309397995471954,
-0.12109063565731049,
-0.056710634380578995,
0.0025076803285628557,
-1.0245422335709857e-33,
-0.09079064428806305,
0.006190103478729725,
-0.005163978319615126,
0.019770728424191475,
-0.06328563392162323,
-0.015500747598707676,
0.010374752804636955,
0.04223276302218437,
0.032559432089328766,
-0.011172839440405369,
0.002454275032505393,
-0.11590705066919327,
0.05716245248913765,
-0.05410228669643402,
-0.018065055832266808,
-0.009900038130581379,
0.028531385585665703,
0.006293027196079493,
-0.02584751881659031,
0.03205868974328041,
0.06531347334384918,
0.031230861321091652,
-0.07961782068014145,
0.02329557202756405,
0.0020944084972143173,
-0.027457090094685555,
-0.09479568153619766,
-0.003105638548731804,
-0.008646040223538876,
0.029689287766814232,
-0.007022778503596783,
-0.028985075652599335,
-0.06907177716493607,
-0.018953783437609673,
-0.10810815542936325,
0.009024119935929775,
0.08132337033748627,
-0.027374539524316788,
-0.02091965079307556,
0.05659281834959984,
0.0698801577091217,
0.030273696407675743,
-0.07255376875400543,
-0.06399174779653549,
-0.029067987576127052,
-0.046377237886190414,
-0.13277028501033783,
0.007829726673662663,
0.022445159032940865,
-0.057294197380542755,
0.00591302802786231,
0.03187422454357147,
-0.020917359739542007,
-0.02787746861577034,
-0.03038817271590233,
-0.11189281940460205,
-0.003151044948026538,
0.04114501550793648,
-0.004243127536028624,
0.0560666099190712,
-0.04952223226428032,
-0.01912093162536621,
-0.05011102184653282,
-0.02509274147450924,
0.004947980400174856,
-0.09050051122903824,
0.049762140959501266,
0.03706197813153267,
-0.009814238175749779,
-0.07327627390623093,
0.06845071166753769,
-0.09537465870380402,
-0.0528523251414299,
-0.0038019840139895678,
-0.07489582151174545,
0.09778077900409698,
0.048222292214632034,
0.005455870646983385,
-0.01112101599574089,
-0.047835059463977814,
0.09096143394708633,
-0.05049367621541023,
0.07677583396434784,
0.0021187327802181244,
0.13461841642856598,
0.10219065099954605,
-0.02646518684923649,
-0.013743603602051735,
-0.001909930957481265,
0.06342660635709763,
-0.052415259182453156,
-0.025889385491609573,
-0.11966332048177719,
0.07437802106142044,
0.005560982506722212,
-4.264273556486842e-8,
-0.03592044860124588,
0.054600927978754044,
0.01735966093838215,
0.02201075479388237,
0.07465038448572159,
0.010598642751574516,
-0.0059625208377838135,
0.09471302479505539,
-0.03485959395766258,
-0.01521731074899435,
0.04643243923783302,
0.006829879246652126,
-0.006270484998822212,
-0.05297161266207695,
0.005470130126923323,
0.033895112574100494,
-0.06952743232250214,
-0.01206443551927805,
0.014191717840731144,
0.022632339969277382,
0.07356783747673035,
0.0178926233202219,
0.03161149099469185,
-0.08258374035358429,
0.02281583473086357,
-0.0757964700460434,
-0.023516813293099403,
0.002934474730864167,
-0.047220379114151,
-0.025620080530643463,
-0.030314484611153603,
-0.005043338052928448,
-0.020455611869692802,
-0.09631552547216415,
0.034656740725040436,
0.03342306613922119,
-0.019313663244247437,
-0.03611687943339348,
-0.052918121218681335,
0.03687415271997452,
0.027565550059080124,
0.04499645158648491,
0.008919593878090382,
-0.003586279693990946,
0.04091700538992882,
0.03450968116521835,
-0.055197518318891525,
-0.029439613223075867,
0.0406208336353302,
-0.017652416601777077,
-0.03743185102939606,
-0.06284449994564056,
0.040559738874435425,
0.11481747776269913,
0.10469050705432892,
0.008740854449570179,
0.017303569242358208,
0.002548841992393136,
0.019968440756201744,
0.07806045562028885,
0.056155551224946976,
-0.01758601702749729,
0.020856423303484917,
-0.026073990389704704
] | 0
| 0
| 0
| 0
|
|
Free Hunch: Denoiser Covariance Estimation for Diffusion Models Without Extra Costs
|
https://openreview.net/forum?id=4JK2XMGUc8
|
[
"Severi Rissanen",
"Markus Heinonen",
"Arno Solin"
] |
Poster
|
The covariance for clean data given a noisy observation is an important quantity in many training-free guided generation methods for diffusion models. Current methods require heavy test-time computation, altering the standard diffusion training process or denoiser architecture, or making heavy approximations. We propose a new framework that sidesteps these issues by using covariance information that is available for free from training data and the curvature of the generative trajectory, which is linked to the covariance through the second-order Tweedie's formula. We integrate these sources of information using (i) a novel method to transfer covariance estimates across noise levels and (ii) low-rank updates in a given noise level. We validate the method on linear inverse problems, where it outperforms recent baselines, especially with fewer diffusion steps.
|
diffusion model, conditional generation, inverse problems, denoiser covariance estimation
|
We propose a new, efficient method for denoiser covariance estimation in diffusion models, which can be used for conditional generation and inverse problems
| 11,229
|
2410.11149
|
[
-0.028953449800610542,
-0.1184917464852333,
0.11240067332983017,
0.04207102581858635,
0.04972447082400322,
-0.04412681236863136,
-0.04492015391588211,
-0.11644601821899414,
-0.030744988471269608,
-0.00031479206518270075,
-0.017385093495249748,
-0.009010964073240757,
-0.012579112313687801,
-0.028318384662270546,
-0.07509024441242218,
0.0074291652999818325,
0.04577673226594925,
0.019588652998209,
-0.026799341663718224,
-0.036093570291996,
-0.05820846185088158,
0.07467576116323471,
-0.05935995280742645,
0.05817732587456703,
0.11804885417222977,
-0.07780612260103226,
0.06168944761157036,
-0.020101942121982574,
-0.0018676196923479438,
-0.03489115834236145,
0.1169690266251564,
0.018405457958579063,
-0.04499109461903572,
-0.027954505756497383,
-0.012693424709141254,
0.05484861508011818,
0.04500924423336983,
-0.013883453793823719,
0.09478107839822769,
0.05100811645388603,
0.03989587724208832,
0.08040116727352142,
-0.02208523638546467,
0.057881910353899,
0.0849192887544632,
-0.033572930842638016,
-0.0019012981792911887,
-0.08135487884283066,
0.008842868730425835,
0.016876183450222015,
0.05105750635266304,
0.006898799445480108,
-0.08527882397174835,
0.019427992403507233,
0.028000246733427048,
0.020351337268948555,
0.09139592200517654,
-0.00365334190428257,
0.05606747046113014,
0.038391221314668655,
-0.032537791877985,
-0.05973958596587181,
-0.07471270114183426,
-0.03412037342786789,
-0.03918176516890526,
-0.03214695677161217,
0.048379506915807724,
0.03480245918035507,
-0.052367620170116425,
0.06612052023410797,
-0.035585325211286545,
0.040545444935560226,
-0.10948026180267334,
-0.05670537054538727,
-0.030223693698644638,
-0.012486916966736317,
0.010023564100265503,
-0.02255117893218994,
0.032132893800735474,
-0.05245905742049217,
0.07415880262851715,
0.019012171775102615,
0.07045715302228928,
-0.043373920023441315,
0.08995053172111511,
-0.02160140685737133,
0.005477766506373882,
0.01640680432319641,
0.014969483017921448,
-0.016014935448765755,
-0.022238997742533684,
0.03331703320145607,
-0.10046403855085373,
0.010528958402574062,
-0.04810335487127304,
0.004041245207190514,
0.09956822544336319,
0.014712798409163952,
0.036583852022886276,
0.08796165883541107,
-0.010061006993055344,
0.005691680125892162,
-0.05598035082221031,
0.024447990581393242,
0.049483392387628555,
-0.03558119758963585,
0.13278017938137054,
0.04086365923285484,
-0.01584099419414997,
0.012949274852871895,
-0.008590753190219402,
-0.011597026139497757,
-0.024285566061735153,
0.035221636295318604,
0.01906922087073326,
-0.05851270258426666,
0.056218069046735764,
0.027579762041568756,
0.024132005870342255,
-0.02369498461484909,
0.010252006351947784,
-0.033369727432727814,
0.03769386187195778,
0.004075502976775169,
-0.015711264684796333,
0.0010111918672919273,
-0.049015454947948456,
3.366485414350487e-33,
0.028417617082595825,
0.03336351364850998,
0.03710293024778366,
0.010586249642074108,
0.042724430561065674,
0.006150232627987862,
-0.00825565867125988,
-0.09544744342565536,
0.008201271295547485,
-0.03336607664823532,
-0.039245590567588806,
-0.011578155681490898,
-0.1411932408809662,
0.007097712717950344,
0.005497477483004332,
-0.0009692812454886734,
-0.03502250090241432,
0.10284518450498581,
-0.03717006742954254,
-0.008800014853477478,
0.06311608105897903,
-0.028943033888936043,
-0.07209916412830353,
-0.023157533258199692,
-0.007470778655260801,
-0.05500480905175209,
-0.0011837589554488659,
-0.05666825547814369,
-0.026729702949523926,
0.015369539149105549,
-0.0710289403796196,
0.020104557275772095,
0.011880667880177498,
0.03801727667450905,
-0.05050371587276459,
0.03703815117478371,
-0.04617118090391159,
0.011369500309228897,
-0.029628604650497437,
-0.0375773161649704,
0.02533266879618168,
0.042552731931209564,
0.009775122627615929,
-0.08663757890462875,
-0.16753117740154266,
-0.013224673457443714,
-0.046780072152614594,
-0.009987843222916126,
0.021629739552736282,
-0.023830825462937355,
-0.043162181973457336,
-0.019365137442946434,
-0.04986285790801048,
-0.04967040196061134,
-0.0664147138595581,
-0.014334894716739655,
0.08205994963645935,
-0.059144701808691025,
0.05210471898317337,
-0.07207945734262466,
0.06797993183135986,
0.03136832267045975,
0.009903657250106335,
-0.01255054585635662,
0.09951908141374588,
0.01805100031197071,
0.03529449179768562,
0.009586438536643982,
0.06062363088130951,
0.0020698150619864464,
-0.03107871674001217,
0.026423398405313492,
-0.094527967274189,
-0.05549675598740578,
0.08340178430080414,
-0.023325514048337936,
0.010336077772080898,
-0.027239732444286346,
0.047643255442380905,
-0.03292274847626686,
-0.009184795431792736,
-0.04655253142118454,
0.01278090849518776,
-0.10762494802474976,
0.0295206680893898,
-0.02745005488395691,
-0.06576935201883316,
0.008481781929731369,
-0.08574988692998886,
-0.08141950517892838,
-0.040915347635746,
-0.03890850022435188,
-0.0615452378988266,
0.018325291574001312,
-0.027234522625803947,
-3.164464955218169e-33,
-0.02085554227232933,
0.06631940603256226,
0.10541807115077972,
0.05180090665817261,
-0.04324536770582199,
0.011248527094721794,
0.018512045964598656,
0.05803044140338898,
0.06518445909023285,
-0.056789278984069824,
-0.029802564531564713,
-0.024761442095041275,
-0.04373283311724663,
0.055897001177072525,
0.09555341303348541,
-0.06001241132616997,
-0.005214313976466656,
0.01058308593928814,
-0.01771857962012291,
-0.014393824152648449,
-0.00006351024057948962,
0.03029870055615902,
-0.013025863096117973,
0.021235322579741478,
-0.04572577029466629,
0.00004704186721937731,
0.02728842757642269,
0.08705926686525345,
0.0022761146537959576,
-0.05102201923727989,
-0.0918085128068924,
-0.00580249959602952,
0.03811102360486984,
-0.03241962194442749,
-0.06188322603702545,
0.1499418169260025,
0.03561657667160034,
0.020985398441553116,
-0.04691561684012413,
0.07077434659004211,
0.0439283661544323,
0.010121378116309643,
-0.025747602805495262,
0.00444788858294487,
0.03973114863038063,
-0.030547287315130234,
0.0031105882953852415,
0.019436022266745567,
0.03701680526137352,
0.060700539499521255,
0.061419930309057236,
0.08308614045381546,
-0.08447065204381943,
0.11074408888816833,
-0.040884412825107574,
-0.03982294723391533,
-0.0583111010491848,
-0.06435573101043701,
0.0948997512459755,
-0.06652083247900009,
-0.05405005067586899,
0.06363888829946518,
-0.08640101552009583,
0.007333411835134029,
-0.05396401137113571,
0.01521976850926876,
-0.07990091294050217,
-0.008135677315294743,
0.021731160581111908,
0.04987194016575813,
-0.058849554508924484,
-0.018447481095790863,
0.02394859492778778,
-0.014527315273880959,
-0.0201305840164423,
-0.12252242863178253,
0.0333830788731575,
-0.06315568089485168,
-0.0018992419354617596,
-0.023477904498577118,
0.04331337660551071,
0.02152954787015915,
0.07381729781627655,
0.003995313309133053,
0.0069837067276239395,
-0.01342154573649168,
0.06717003136873245,
-0.0793275386095047,
-0.025090262293815613,
-0.01628595031797886,
-0.02022075280547142,
-0.02659700997173786,
0.06446414440870285,
0.018060648813843727,
-0.05979973450303078,
-4.528385488811182e-8,
-0.06214700639247894,
0.09445637464523315,
0.08015453815460205,
0.022565849125385284,
0.1008395180106163,
0.031198685988783836,
0.02870267815887928,
0.07152123749256134,
0.0332757793366909,
-0.05072548985481262,
0.03443159908056259,
-0.03214624524116516,
-0.019456468522548676,
0.0025444314815104008,
-0.03633034601807594,
0.06443485617637634,
-0.0018884845776483417,
0.06025276333093643,
-0.007458140607923269,
-0.06002913415431976,
-0.04341817647218704,
0.08106696605682373,
0.04630960151553154,
-0.025511877611279488,
0.09035935997962952,
0.008845259435474873,
-0.0628662109375,
0.08732085675001144,
0.09187357127666473,
-0.02445230819284916,
-0.008493863046169281,
0.03556235879659653,
0.04123307019472122,
0.035752128809690475,
-0.05996601656079292,
0.07439932972192764,
0.02637084573507309,
0.01106756180524826,
-0.04484165087342262,
-0.043627046048641205,
-0.07858196645975113,
0.004530614707618952,
-0.02286597341299057,
0.007353892084211111,
0.016217494383454323,
0.0879443809390068,
0.03816783055663109,
-0.08867894113063812,
0.029081037268042564,
0.057225897908210754,
0.08214667439460754,
-0.05617137625813484,
0.057947151362895966,
0.02144671604037285,
0.027524948120117188,
-0.008102452382445335,
0.01948949508368969,
0.004204222466796637,
0.03733672946691513,
0.020390918478369713,
-0.007265614811331034,
-0.010788573883473873,
0.03680120408535004,
0.025486798956990242
] | 0
| 0
| 0
| 0
|
|
One Hundred Neural Networks and Brains Watching Videos: Lessons from Alignment
|
https://openreview.net/forum?id=LM4PYXBId5
|
[
"Christina Sartzetaki",
"Gemma Roig",
"Cees G. M. Snoek",
"Iris Groen"
] |
Poster
|
What can we learn from comparing video models to human brains, arguably the most efficient and effective video processing systems in existence? Our work takes a step towards answering this question by performing the first large-scale benchmarking of deep video models on representational alignment to the human brain, using publicly available models and a recently released video brain imaging (fMRI) dataset. We disentangle four factors of variation in the models (temporal modeling, classification task, architecture, and training dataset) that affect alignment to the brain, which we measure by conducting Representational Similarity Analysis across multiple brain regions and model layers. We show that temporal modeling is key for alignment to brain regions involved in early visual processing, while a relevant classification task is key for alignment to higher-level regions. Moreover, we identify clear differences between the brain scoring patterns across layers of CNNs and Transformers, and reveal how training dataset biases transfer to alignment with functionally selective brain areas. Additionally, we uncover a negative correlation of computational complexity to brain alignment. Measuring a total of 99 neural networks and 10 human brains watching videos, we aim to forge a path that widens our understanding of temporal and semantic video representations in brains and machines, ideally leading towards more efficient video models and more mechanistic explanations of processing in the human brain.
|
representational alignment, Representational Similarity Analysis, RSA, benchmarking, neuro-AI, video AI, neuroscience, fMRI, cognitive AI
|
We benchmark 99 image and video models on brain representational alignment to fMRI data of humans watching video.
| 11,218
| null |
[
-0.05286542326211929,
-0.14072492718696594,
0.002783597446978092,
-0.03641090914607048,
0.08122507482767105,
0.06750798225402832,
-0.03784998133778572,
-0.05176210403442383,
0.07599479705095291,
-0.09253718703985214,
-0.06398489326238632,
-0.1047041267156601,
-0.07096264511346817,
0.12545205652713776,
-0.08917917311191559,
-0.09151944518089294,
0.042851660400629044,
0.048374492675065994,
-0.12827859818935394,
-0.008857845328748226,
-0.0038924345280975103,
-0.02961501106619835,
0.011825760826468468,
0.026024924591183662,
0.01697145588696003,
0.03565961867570877,
0.058725327253341675,
-0.04137372225522995,
-0.007742636371403933,
-0.10221505165100098,
0.10198935121297836,
-0.026039235293865204,
-0.013700379058718681,
0.030975468456745148,
-0.055666711181402206,
0.04043126106262207,
-0.04531873017549515,
-0.022203445434570312,
0.03966696932911873,
-0.07197720557451248,
0.052494194358587265,
0.00913981907069683,
0.042715977877378464,
0.023621998727321625,
0.030842836946249008,
0.04499361664056778,
0.02835545688867569,
-0.1116705909371376,
0.012985695153474808,
-0.027089307084679604,
-0.08747527748346329,
0.027067823335528374,
-0.023621151223778725,
0.07987628132104874,
-0.04769919440150261,
0.0796358659863472,
0.07028917968273163,
0.05511555075645447,
-0.01923386938869953,
0.031402911990880966,
0.027061954140663147,
-0.05297257751226425,
0.01784464158117771,
0.018538527190685272,
-0.0005976973916403949,
-0.002043783199042082,
0.03029392473399639,
-0.06183505803346634,
-0.08301561325788498,
0.02193893864750862,
-0.03922265022993088,
0.0873497948050499,
-0.04759127274155617,
0.028313392773270607,
-0.009627384133636951,
0.02965226210653782,
0.09378524124622345,
-0.05055147781968117,
0.07578282803297043,
-0.13614848256111145,
0.049226414412260056,
0.04680060222744942,
0.03796861693263054,
0.05788194388151169,
0.15625080466270447,
-0.02857322059571743,
0.016527660191059113,
0.04622562974691391,
-0.003883604658767581,
0.002959114732220769,
-0.09767578542232513,
-0.035032305866479874,
-0.0004803197516594082,
-0.0443415567278862,
0.09381786733865738,
0.06424009799957275,
0.014742732979357243,
-0.03169823810458183,
-0.013929069973528385,
0.025887886062264442,
-0.03602968156337738,
0.03252425417304039,
0.022914858534932137,
0.002966836793348193,
0.07236692309379578,
-0.03202322870492935,
0.060944732278585434,
0.030475137755274773,
0.008114644326269627,
-0.09029635787010193,
0.05131061002612114,
0.002130832988768816,
-0.0031088399700820446,
0.05920450761914253,
0.02517849951982498,
-0.03925063833594322,
-0.025720544159412384,
0.012036630883812904,
-0.014866235665977001,
0.037406012415885925,
-0.027469098567962646,
0.02390790358185768,
-0.05569866672158241,
-0.00010923711670329794,
0.04330645129084587,
-0.04517608508467674,
-0.04711994528770447,
6.122236310246524e-34,
-0.012955311685800552,
-0.019859254360198975,
0.06410611420869827,
0.06173041835427284,
0.03840281441807747,
-0.031457290053367615,
-0.020686086267232895,
0.04353010281920433,
0.017461374402046204,
0.07097684592008591,
-0.05541786924004555,
0.005413437262177467,
-0.05018584057688713,
0.09676181524991989,
-0.05326645076274872,
0.00915835052728653,
-0.1108710989356041,
0.021458463743329048,
-0.04096706211566925,
0.017609581351280212,
0.05988979712128639,
0.0008338229381479323,
-0.02295655943453312,
-0.05668032914400101,
-0.05024804547429085,
-0.056455984711647034,
-0.03917689248919487,
-0.0696553885936737,
-0.04611121863126755,
-0.025516243651509285,
-0.0990741029381752,
0.03241964429616928,
0.026541663333773613,
0.0015762331895530224,
0.07496386021375656,
-0.023675674572587013,
0.006589116528630257,
-0.02405865490436554,
0.02902885526418686,
0.002134444424882531,
0.042571645230054855,
0.06520325690507889,
-0.09560618549585342,
-0.01018295157700777,
-0.028605882078409195,
-0.058325089514255524,
0.06935585290193558,
0.03993432596325874,
-0.00003539343015290797,
-0.055821359157562256,
0.04356954246759415,
0.06259774416685104,
0.036694105714559555,
-0.08046872913837433,
-0.05607077106833458,
0.028929635882377625,
0.04332352429628372,
-0.02775469422340393,
0.02505895122885704,
0.009849666617810726,
0.0013093104353174567,
-0.0017533261561766267,
-0.06863163411617279,
0.03963033854961395,
0.02579740807414055,
-0.011343463324010372,
0.02371850609779358,
0.1087193712592125,
0.028222644701600075,
0.04064478725194931,
0.00496705761179328,
0.06197356432676315,
-0.02312326245009899,
-0.1194358691573143,
0.03759998083114624,
0.05780775099992752,
0.06420044600963593,
-0.05569523945450783,
-0.03864031657576561,
0.019832873716950417,
0.006274875719100237,
-0.013332677073776722,
-0.025454018265008926,
-0.03330909460783005,
-0.0274350605905056,
-0.059008583426475525,
0.03529216721653938,
-0.061419084668159485,
-0.021007882431149483,
-0.01064244844019413,
0.0027722411323338747,
-0.04703773558139801,
0.06930331140756607,
-0.03654475882649422,
0.0009112994302995503,
-3.0631720559948353e-34,
-0.03762853890657425,
-0.029439687728881836,
0.05742238461971283,
0.001230113673955202,
-0.004307387862354517,
-0.048277147114276886,
0.042564116418361664,
0.0009078326402232051,
0.0026384585071355104,
0.016935203224420547,
0.005019401665776968,
-0.10902758687734604,
-0.008594103157520294,
-0.05920333042740822,
-0.008834866806864738,
-0.026749830693006516,
-0.04933920130133629,
0.02543654665350914,
-0.0005152333178557456,
-0.04861333221197128,
0.0388200506567955,
0.10476227104663849,
-0.07070342451334,
-0.042104899883270264,
-0.06814681738615036,
0.07247243076562881,
0.027385016903281212,
0.06111326441168785,
0.017549872398376465,
0.0024055279791355133,
-0.02128218486905098,
-0.01687331683933735,
-0.09040334075689316,
-0.04567551612854004,
-0.08397984504699707,
0.07918119430541992,
0.03686121478676796,
-0.06464691460132599,
-0.027783308178186417,
0.0596657320857048,
0.038381095975637436,
-0.0022993802558630705,
-0.04964761808514595,
0.029475798830389977,
0.014482959173619747,
-0.013618103228509426,
-0.013113480992615223,
-0.012857710942626,
-0.07474545389413834,
0.05566566810011864,
0.002164940582588315,
0.10433871299028397,
-0.0983223021030426,
-0.02291826903820038,
-0.005530244670808315,
-0.02139863185584545,
-0.07500142604112625,
0.010740330442786217,
0.10214027762413025,
0.02439863048493862,
-0.0041567496955394745,
0.022716032341122627,
-0.021009398624300957,
0.0005013311747461557,
-0.03329062834382057,
0.06459449976682663,
0.012482987716794014,
0.023073473945260048,
0.01184319518506527,
0.04146231710910797,
0.017342159524559975,
-0.05025842785835266,
0.006642797961831093,
0.011718389578163624,
-0.011699439957737923,
-0.03178517147898674,
0.06511660665273666,
0.011067945510149002,
0.05783720314502716,
-0.06799666583538055,
-0.004316625185310841,
-0.10566509515047073,
-0.05196760967373848,
-0.004823942668735981,
0.06493125110864639,
0.08656569570302963,
0.08636211603879929,
0.04414195194840431,
0.034237027168273926,
0.035802897065877914,
-0.066432885825634,
-0.01818552426993847,
-0.00951491016894579,
-0.033190906047821045,
-0.05837811529636383,
-3.516180058227292e-8,
-0.14852306246757507,
0.004051909316331148,
0.0032831549178808928,
0.0384298600256443,
-0.04620036110281944,
-0.03830711543560028,
0.08610094338655472,
0.0966305211186409,
-0.005328429862856865,
-0.0028233639895915985,
0.11918021738529205,
-0.00387384626083076,
-0.03323494642972946,
0.0163277518004179,
0.011059609241783619,
0.037866443395614624,
-0.021902887150645256,
0.01954091154038906,
0.024273458868265152,
0.04729814454913139,
0.013406837359070778,
0.06097591668367386,
0.056499846279621124,
-0.059041231870651245,
0.023738568648695946,
-0.023494882509112358,
-0.041141413152217865,
0.13444842398166656,
0.003619835479184985,
-0.007960448041558266,
0.023592863231897354,
0.05197560042142868,
0.02527010068297386,
-0.13210521638393402,
0.0011870695743709803,
-0.012016087770462036,
0.018548406660556793,
0.021747132763266563,
-0.049817197024822235,
0.02498251385986805,
-0.005492148920893669,
0.056809812784194946,
0.03632936626672745,
0.0433199368417263,
-0.0013555939076468349,
0.027097895741462708,
0.011299394071102142,
0.02268855646252632,
0.041678834706544876,
-0.0018051802180707455,
-0.01667816936969757,
0.0206851065158844,
0.032445721328258514,
0.07812456041574478,
0.04404330626130104,
-0.017960617318749428,
0.008584927767515182,
-0.07435967773199081,
0.000521401409059763,
0.03991131857037544,
0.03217430040240288,
0.07348039001226425,
-0.021145684644579887,
-0.04502248764038086
] | 0
| 0
| 0
| 0
|
|
Value-aligned Behavior Cloning for Offline Reinforcement Learning via Bi-level Optimization
|
https://openreview.net/forum?id=elTJBP7Fbv
|
[
"Xingyu Jiang",
"Ning Gao",
"Xiuhui Zhang",
"Hongkun Dou",
"Yue Deng"
] |
Poster
|
Offline reinforcement learning (RL) aims to optimize policies under pre-collected data, without requiring any further interactions with the environment. Derived from imitation learning, Behavior cloning (BC) is extensively utilized in offline RL for its simplicity and effectiveness. Although BC inherently avoids out-of-distribution deviations, it lacks the ability to discern between high and low-quality data, potentially leading to sub-optimal performance when facing with poor-quality data. Current offline RL algorithms attempt to enhance BC by incorporating value estimation, yet often struggle to effectively balance these two critical components, specifically the alignment between the behavior policy and the pre-trained value estimations under in-sample offline data. To address this challenge, we propose the Value-aligned Behavior Cloning via Bi-level Optimization (VACO), a novel bi-level framework that seamlessly integrates an inner loop for weighted supervised behavior cloning (BC) with an outer loop dedicated to value alignment. In this framework, the inner loop employs a meta-scoring network to evaluate and appropriately weight each training sample, while the outer loop maximizes value estimation for alignment with controlled noise to facilitate limited exploration. This bi-level structure allows VACO to identify the optimal weighted BC policy, ultimately maximizing the expected estimated return conditioned on the learned value function. We conduct a comprehensive evaluation of VACO across a variety of continuous control benchmarks in offline RL, where it consistently achieves superior performance compared to existing state-of-the-art methods.
|
offline reinforcement learning;bi-level optimization;value alignment
| null | 11,212
| null |
[
-0.0715307965874672,
-0.08196717500686646,
0.015743454918265343,
0.008309846743941307,
0.007527483627200127,
0.019644634798169136,
0.024926546961069107,
-0.027297070249915123,
-0.017591463401913643,
-0.010720787569880486,
-0.02818562090396881,
-0.03531820699572563,
0.0004238465044181794,
0.021809102967381477,
0.02965618297457695,
0.07482101023197174,
0.12415894865989685,
0.03976795822381973,
-0.06256623566150665,
-0.019663728773593903,
0.029536446556448936,
0.0053132507018744946,
0.037496671080589294,
0.04546230286359787,
0.007357761263847351,
0.050806619226932526,
0.011345998384058475,
-0.034105803817510605,
0.01800745725631714,
-0.038227833807468414,
0.021795805543661118,
-0.008235030807554722,
0.039570458233356476,
-0.09987138211727142,
-0.05146271735429764,
0.07878018915653229,
-0.11471839249134064,
0.00280370214022696,
0.033946841955184937,
-0.02247227542102337,
-0.011344420723617077,
-0.01616213470697403,
-0.008645165711641312,
-0.007770758122205734,
-0.010234189219772816,
0.004554165992885828,
-0.008857214823365211,
-0.04919661581516266,
0.02278544381260872,
0.05182091146707535,
-0.032096441835165024,
0.003220855724066496,
-0.01634635403752327,
0.04997784644365311,
0.03539709001779556,
0.07563290745019913,
0.10477408766746521,
0.018262987956404686,
0.10331862419843674,
0.0025500112678855658,
0.006173810921609402,
-0.06235545128583908,
0.0533180795609951,
0.0303626861423254,
-0.09535946696996689,
-0.0553702786564827,
-0.08833277225494385,
0.05334395915269852,
0.0015416416572406888,
-0.02590828202664852,
-0.04135921970009804,
0.08999665826559067,
0.012221062555909157,
0.008841270580887794,
0.03306977450847626,
0.018425369635224342,
0.027345707640051842,
-0.014343943446874619,
0.015992572531104088,
-0.041019171476364136,
-0.027888644486665726,
-0.014303830452263355,
-0.04221782833337784,
0.038482096046209335,
0.08343502134084702,
-0.03721759468317032,
0.027292385697364807,
0.05302146077156067,
0.08718595653772354,
0.03473551943898201,
-0.011457460932433605,
0.07552579045295715,
0.018203893676400185,
-0.0016488810069859028,
-0.03699411079287529,
0.026893269270658493,
-0.05141625180840492,
-0.0530083104968071,
-0.025970274582505226,
0.054268158972263336,
0.014041547663509846,
0.11034061014652252,
0.025246217846870422,
-0.001097804051823914,
-0.026281077414751053,
-0.1099095344543457,
0.03706267476081848,
-0.0030751570593565702,
0.02213112823665142,
-0.03887289762496948,
0.05586762726306915,
0.039445728063583374,
0.09130117297172546,
0.01677119918167591,
0.02299322560429573,
0.034433551132678986,
-0.0030817787628620863,
-0.04515857994556427,
-0.08754052221775055,
0.08142352849245071,
0.014003252610564232,
-0.08645886182785034,
0.007030440494418144,
0.032622478902339935,
0.021222984418272972,
-0.025866907089948654,
-0.04350927099585533,
2.3785487636303286e-34,
0.05923108011484146,
0.0034393128007650375,
0.10018827021121979,
-0.011865640990436077,
0.031174080446362495,
-0.05071694403886795,
0.048238273710012436,
-0.017039556056261063,
-0.05036891624331474,
-0.03789419308304787,
0.023946674540638924,
-0.0006699530640617013,
-0.04442765191197395,
0.0991508811712265,
-0.0054816631600260735,
-0.011628806591033936,
-0.08193423599004745,
0.02403588779270649,
-0.012710828334093094,
-0.001828246284276247,
0.05097578465938568,
-0.06301036477088928,
0.010334944352507591,
-0.0004661663551814854,
0.08750657737255096,
-0.02974858321249485,
0.009984232485294342,
-0.029640091583132744,
-0.004862408619374037,
0.002409619279205799,
-0.007669918239116669,
-0.048919301480054855,
0.03068883717060089,
0.02365770936012268,
0.017145967110991478,
-0.0006857749540358782,
0.0182722806930542,
-0.051734548062086105,
-0.05869875103235245,
0.002707489999011159,
0.03706103190779686,
0.010756513103842735,
0.038661230355501175,
0.032623540610075,
-0.04625033214688301,
-0.2079835683107376,
0.036834996193647385,
0.0017772674327716231,
-0.02154947817325592,
0.01145076472312212,
-0.01557912863790989,
-0.015122050419449806,
0.007817672565579414,
-0.08479969203472137,
-0.11916036903858185,
0.02040484920144081,
-0.009793377481400967,
0.02278694324195385,
-0.041501905769109726,
0.062354061752557755,
0.007115556392818689,
0.02460767701268196,
-0.023474015295505524,
0.07326497882604599,
0.038949303328990936,
0.03623932972550392,
0.012192579917609692,
0.008484713733196259,
0.04031229019165039,
-0.0012562907068058848,
0.010802323929965496,
0.04894082993268967,
-0.06550101935863495,
-0.02762424759566784,
0.07410190999507904,
-0.014976493082940578,
0.011225352995097637,
-0.11760331690311432,
0.03943860903382301,
0.037673138082027435,
-0.0012674363097175956,
0.05145168676972389,
-0.04693217948079109,
-0.0815550908446312,
-0.007114254403859377,
-0.1105174794793129,
0.02138059027493,
-0.06533562391996384,
-0.03911159187555313,
0.027163902297616005,
-0.06615250557661057,
-0.028422771021723747,
-0.07236873358488083,
0.014004205353558064,
0.029977994039654732,
-1.6701375522847578e-34,
0.029132720082998276,
-0.07955272495746613,
0.09863653779029846,
0.05813338980078697,
-0.07256970554590225,
0.042681820690631866,
-0.008202606812119484,
-0.03224862739443779,
-0.023888593539595604,
0.033903397619724274,
-0.0334215983748436,
-0.0029806382954120636,
0.012286189943552017,
-0.008018898777663708,
0.01943904533982277,
-0.04523169621825218,
-0.07057870924472809,
0.04910140857100487,
0.03759971633553505,
-0.0020954101346433163,
0.06856084614992142,
0.0720401182770729,
-0.05143878608942032,
0.11164558678865433,
-0.09165845066308975,
0.0182585921138525,
0.019074266776442528,
0.05429690703749657,
0.08203930407762527,
0.05410570651292801,
-0.012707202695310116,
-0.03877648711204529,
-0.032813671976327896,
0.00473755644634366,
-0.0849495679140091,
0.10842393338680267,
0.10117935389280319,
0.021817993372678757,
-0.05700424686074257,
0.05368983373045921,
-0.0035064900293946266,
-0.053983304649591446,
-0.12636083364486694,
-0.0655946433544159,
-0.007782054133713245,
-0.04960888996720314,
-0.03585168346762657,
0.03238831087946892,
-0.06029913201928139,
-0.019831396639347076,
-0.04759090021252632,
0.07487709820270538,
-0.09096859395503998,
0.03268707916140556,
-0.0358375608921051,
-0.05831032246351242,
-0.0706038549542427,
0.06203911826014519,
0.02387092635035515,
-0.0676649659872055,
0.027719272300601006,
0.04321456328034401,
0.003232188755646348,
-0.04035669565200806,
-0.011760949157178402,
0.006418113596737385,
0.054965052753686905,
0.09956901520490646,
-0.01693429984152317,
-0.043775659054517746,
-0.05254052206873894,
-0.019113533198833466,
0.03046836145222187,
-0.09605546295642853,
-0.01834796741604805,
-0.0865994244813919,
0.017209405079483986,
0.009591606445610523,
0.010103031992912292,
0.020895492285490036,
0.06913334131240845,
0.006742536555975676,
0.03875625506043434,
0.05972209945321083,
0.005423692520707846,
0.07872463762760162,
-0.0056752609089016914,
0.14964590966701508,
0.01902291364967823,
0.01385650597512722,
-0.07070916891098022,
0.028145641088485718,
-0.034849800169467926,
-0.0008205851772800088,
-0.047441355884075165,
-4.168877509869162e-8,
-0.10025779902935028,
0.037426237016916275,
0.11749444901943207,
0.09743168205022812,
-0.0000680611192365177,
-0.014069527387619019,
-0.02247542142868042,
-0.016932042315602303,
-0.020997535437345505,
0.0010530563304200768,
0.004611650947481394,
0.046482037752866745,
-0.016703449189662933,
-0.07665710151195526,
-0.04037696495652199,
0.0639604926109314,
0.023033248260617256,
-0.03390681371092796,
0.051792606711387634,
0.05382851883769035,
0.011763175018131733,
-0.01471067313104868,
-0.008820286020636559,
-0.02193240076303482,
0.0399240180850029,
-0.1200934499502182,
-0.002141071017831564,
0.07042641937732697,
0.11932104080915451,
-0.04173845425248146,
0.02495444566011429,
-0.05442476272583008,
0.03842291608452797,
-0.02720104716718197,
0.055507149547338486,
-0.0019395938143134117,
0.08092506974935532,
-0.025213968008756638,
-0.013287701644003391,
-0.01597432978451252,
-0.007284469902515411,
0.0817909762263298,
-0.024871278554201126,
-0.04348725825548172,
0.019130032509565353,
-0.019626997411251068,
-0.09654472768306732,
-0.0682171881198883,
0.0648043304681778,
-0.11016425490379333,
0.057210855185985565,
-0.04163401201367378,
0.0030019714031368494,
0.06574960798025131,
0.042253848165273666,
0.02494196966290474,
-0.016009032726287842,
-0.019628461450338364,
0.08284053206443787,
0.10373056679964066,
0.053089797496795654,
-0.017295319586992264,
-0.05930843949317932,
-0.029546411707997322
] | 0
| 0
| 0
| 0
|
|
Fugatto 1: Foundational Generative Audio Transformer Opus 1
|
https://openreview.net/forum?id=B2Fqu7Y2cd
|
[
"Rafael Valle",
"Rohan Badlani",
"Zhifeng Kong",
"Sang-gil Lee",
"Arushi Goel",
"Sungwon Kim",
"Joao Felipe Santos",
"Shuqi Dai",
"Siddharth Gururani",
"Aya Aljafari",
"Alexander H. Liu",
"Kevin J. Shih",
"Ryan Prenger",
"Wei Ping",
"Chao-Han Huck Yang",
"Bryan Catanzaro"
] |
Poster
|
Fugatto is a versatile audio synthesis and transformation model capable of following free-form text instructions with optional audio inputs. While large language models (LLMs) trained with text on a simple next-token prediction objective can learn to infer instructions directly from the data, models trained solely on audio data lack this capacity. This is because audio data does not inherently contain the instructions that were used to generate it. To overcome this challenge, we introduce a specialized dataset generation approach optimized for producing a wide range of audio generation and transformation tasks, ensuring the data reveals meaningful relationships between audio and language. Another challenge lies in achieving compositional abilities -- such as combining, interpolating between, or negating instructions -- using data alone. To address it, we propose ComposableART, an inference-time technique that extends classifier-free guidance to compositional guidance. It enables the seamless and flexible composition of instructions, leading to highly customizable audio outputs outside the training distribution. Our evaluations across a diverse set of tasks demonstrate that Fugatto performs competitively with specialized models, while ComposableART enhances its sonic palette and control over synthesis. Most notably, we highlight our framework's ability to execute emergent sounds and tasks -- sonic phenomena that transcend conventional audio generation -- unlocking new creative possibilities. \href{https://fugatto.github.io/}{Demo Website.}
|
Generative Models, Audio, Foundation Models
|
Fugatto: Foundational Generative Audio Transformer Opus 1 with ComposableART - an inference-time technique that extends classifier-free guidance into compositional guidance.
| 11,207
| null |
[
-0.02283576689660549,
-0.08374004065990448,
0.04661647975444794,
0.014088661409914494,
0.007752174511551857,
0.03218076378107071,
-0.055453456938266754,
-0.0395829938352108,
0.015280330553650856,
-0.079775370657444,
-0.07374979555606842,
-0.11177001893520355,
-0.06126006692647934,
-0.024520045146346092,
-0.02082626149058342,
-0.014571694657206535,
0.0931912437081337,
0.07186882942914963,
-0.057883959263563156,
-0.07815489917993546,
0.08420322835445404,
0.07300368696451187,
0.05119903385639191,
-0.004537209868431091,
0.06664607673883438,
0.06485406309366226,
-0.0494038388133049,
-0.03440593183040619,
0.09741242974996567,
-0.05997288227081299,
0.026944993063807487,
0.007431359961628914,
0.028644541278481483,
0.0420624203979969,
-0.09989439696073532,
0.01774047501385212,
0.008137667551636696,
-0.004405230283737183,
-0.015490257181227207,
-0.03994757682085037,
0.0009695715270936489,
0.06080447882413864,
-0.016825417056679726,
0.00944636669009924,
0.015188888646662235,
-0.0030804143752902746,
-0.012351633049547672,
-0.027048954740166664,
-0.0359795019030571,
0.10569935292005539,
-0.053279466927051544,
-0.04750152677297592,
-0.03077494353055954,
0.03079255297780037,
-0.032115623354911804,
0.0906750038266182,
0.09125715494155884,
0.04358978569507599,
0.036959193646907806,
-0.01507224328815937,
-0.0820973739027977,
-0.11042630672454834,
0.0021757641807198524,
-0.07729494571685791,
-0.04639323428273201,
-0.08251125365495682,
-0.009275116957724094,
0.06782333552837372,
-0.019209859892725945,
0.053004395216703415,
-0.05726122483611107,
0.0033190727699548006,
0.013279021717607975,
0.08553680032491684,
0.017041079699993134,
0.088426373898983,
0.061943963170051575,
-0.02827155590057373,
0.027102980762720108,
-0.06609062105417252,
0.08133509010076523,
-0.007853727787733078,
0.040149297565221786,
0.009831458330154419,
0.07438210397958755,
-0.010889583267271519,
-0.009933409281075,
0.026779381558299065,
0.0043026357889175415,
0.04313700646162033,
-0.13150158524513245,
-0.06873326748609543,
0.08569994568824768,
-0.06335136294364929,
0.0037719174288213253,
0.06567586958408356,
-0.02434198372066021,
-0.09100817888975143,
0.015093830414116383,
0.054173629730939865,
-0.011511002667248249,
0.051613591611385345,
0.04079083725810051,
-0.0681505799293518,
-0.039779383689165115,
-0.10577264428138733,
0.04922540485858917,
0.06759578734636307,
-0.017846575006842613,
-0.0856473445892334,
0.05814079940319061,
0.02006860449910164,
0.03723621740937233,
-0.08622997999191284,
0.08260934054851532,
0.02366837114095688,
-0.04379866272211075,
-0.04552014544606209,
0.012466976419091225,
0.05564359202980995,
-0.013553413562476635,
0.03804200142621994,
-0.046085648238658905,
0.09678027033805847,
-0.032994840294122696,
-0.0708530992269516,
-0.021035348996520042,
1.4428202251340034e-33,
0.0755002498626709,
0.045458920300006866,
0.07084054499864578,
0.07530421763658524,
0.0904219001531601,
-0.06184735894203186,
-0.0024131983518600464,
0.03666668012738228,
-0.04624030366539955,
0.050691455602645874,
0.00823127944022417,
-0.04963746294379234,
-0.057770535349845886,
0.06012957915663719,
0.01225840114057064,
0.04471654072403908,
-0.022448809817433357,
0.05467765033245087,
-0.027615580707788467,
-0.023067912086844444,
0.016520943492650986,
0.035582855343818665,
0.03365829586982727,
-0.0560535192489624,
0.05371453985571861,
0.05565813556313515,
0.04691844433546066,
-0.08957942575216293,
-0.04914257675409317,
-0.014645036309957504,
-0.17134085297584534,
0.007059390190988779,
0.00806653406471014,
0.00587253225967288,
0.0397903248667717,
-0.022486122325062752,
-0.01300494372844696,
-0.04939441755414009,
-0.02208278514444828,
-0.017976056784391403,
-0.0025445849169045687,
0.009658490307629108,
-0.0007338025025092065,
0.006665720138698816,
-0.07410818338394165,
-0.03656431660056114,
0.02430739440023899,
0.08669187128543854,
0.04662111774086952,
0.004398910328745842,
0.059150125831365585,
0.08083096891641617,
-0.005172863602638245,
-0.014792921021580696,
0.0483597069978714,
-0.025477595627307892,
0.0462869331240654,
0.047118671238422394,
0.05215848982334137,
0.0389639213681221,
0.008556879125535488,
0.08863827586174011,
-0.04068149998784065,
0.023852134123444557,
-0.015356984920799732,
0.06412848085165024,
-0.046214569360017776,
0.03488549217581749,
0.11030499637126923,
-0.01312874536961317,
-0.07612967491149902,
-0.010686453431844711,
-0.007470347918570042,
-0.016682583838701248,
0.0007109784055501223,
0.05338885262608528,
-0.0026362603530287743,
-0.09502405673265457,
0.01602606475353241,
0.007091294042766094,
0.007632947061210871,
0.0367128923535347,
-0.03125522658228874,
-0.015250959433615208,
-0.0084571847692132,
-0.018041150644421577,
0.05784830078482628,
-0.05857161432504654,
-0.06316270679235458,
-0.01194833219051361,
-0.11347546428442001,
-0.032851867377758026,
-0.05428434535861015,
0.01611383818089962,
0.013797556981444359,
-1.0250576143754244e-33,
0.018076684325933456,
0.05706758797168732,
-0.040681008249521255,
0.04563005641102791,
-0.027352211996912956,
0.0006670946022495627,
0.008618637919425964,
0.029869791120290756,
0.017483284696936607,
-0.03243072330951691,
-0.0033436547964811325,
-0.06611815094947815,
0.0794481560587883,
-0.06644445657730103,
-0.048379529267549515,
-0.0025340791326016188,
-0.054072439670562744,
0.0010478952899575233,
0.006565641611814499,
0.0034943395294249058,
-0.013409364968538284,
0.0775255337357521,
-0.014283939264714718,
0.0391136072576046,
-0.0601462759077549,
0.06431180983781815,
-0.030980927869677544,
0.09008757770061493,
0.033793311566114426,
0.027829693630337715,
-0.039340678602457047,
-0.004743306431919336,
0.010316463187336922,
-0.1049194484949112,
-0.059418246150016785,
0.01973242685198784,
0.031016649678349495,
0.027922114357352257,
-0.01295726839452982,
0.06047508120536804,
0.03505867347121239,
0.0010943975066766143,
0.03371252492070198,
0.027190636843442917,
-0.024655086919665337,
-0.015983503311872482,
-0.051499735563993454,
0.020699812099337578,
0.06623134762048721,
0.02243543602526188,
0.043335121124982834,
-0.005091761704534292,
-0.03037538193166256,
-0.02792593091726303,
-0.040312036871910095,
-0.11158530414104462,
-0.06552617996931076,
-0.09206149727106094,
0.06076567620038986,
0.01919715292751789,
-0.02359970286488533,
-0.0502404049038887,
0.04809904471039772,
-0.13647682964801788,
0.057451844215393066,
0.02209530770778656,
-0.03568444028496742,
-0.005957193672657013,
0.06185959652066231,
0.03749214857816696,
0.05307070165872574,
-0.032160092145204544,
0.031019942834973335,
0.011272907257080078,
0.0014127391623333097,
-0.01832962967455387,
-0.010595858097076416,
-0.06842335313558578,
-0.08116717636585236,
-0.04163670167326927,
-0.004234728869050741,
-0.020718049257993698,
0.018892375752329826,
0.025011232122778893,
0.07771294564008713,
0.1167132630944252,
0.04708736017346382,
0.050208523869514465,
0.033906884491443634,
0.12677089869976044,
-0.06892930716276169,
0.1056482121348381,
0.012790028937160969,
0.06948710978031158,
-0.029910974204540253,
-3.781077495546015e-8,
-0.014964370988309383,
0.05093598738312721,
0.004768825136125088,
0.052928656339645386,
0.004776319023221731,
-0.04425959289073944,
0.00998690165579319,
-0.016733461990952492,
0.002407425781711936,
-0.03174559772014618,
0.07598639279603958,
-0.013250083662569523,
0.0556032769382,
-0.02854534238576889,
-0.02524033561348915,
0.05224648490548134,
-0.010233122855424881,
0.026495778933167458,
-0.014516153372824192,
-0.08602249622344971,
0.033056728541851044,
0.05345946550369263,
-0.024854116141796112,
-0.05915333703160286,
0.05352141708135605,
-0.07397857308387756,
-0.002715162932872772,
0.02983103320002556,
0.05777330324053764,
0.03416326269507408,
-0.04038390517234802,
0.02282070554792881,
-0.03316795825958252,
-0.017669174820184708,
-0.03801321983337402,
0.014607089571654797,
0.0015509931836277246,
-0.04629846662282944,
-0.020392006263136864,
-0.09125111252069473,
0.0047781397588551044,
0.08784428983926773,
-0.11076459288597107,
0.0598357692360878,
-0.05279230698943138,
-0.05915987119078636,
0.05812346190214157,
-0.09485474228858948,
-0.0029573466163128614,
0.02164524979889393,
-0.01344755757600069,
0.004310665186494589,
-0.023576928302645683,
0.042492207139730453,
0.0814143717288971,
0.05519004166126251,
0.004037416540086269,
0.04261132702231407,
-0.019062895327806473,
0.04115418344736099,
-0.06167774647474289,
-0.010129893198609352,
0.014490727335214615,
-0.05102455988526344
] | 0
| 0
| 0
| 0
|
|
Balancing Bias in Two-sided Markets for Fair Stable Matchings
|
https://openreview.net/forum?id=qykpnEWf2J
|
[
"Siyuan Wu",
"Leong Hou U",
"Panagiotis Karras"
] |
Poster
|
The Balanced Stable Marriage (BSM) problem aims to find a stable matching in a two-sided market that minimizes the maximum dissatisfaction among two sides. The classical Deferred Acceptance algorithm merely produces an unfair stable marriage, providing optimal partners for one side while partially assigning pessimal partners to the other. Solving BSM is NP-hard, thwarting attempts to resolve the problem exactly. As the instance size increases in practice, recent studies have explored heuristics for finding a fair stable marriage but have not found an exact optimal solution for BSM efficiently. Nevertheless, in this paper we propose an efficient algorithm, Isorropia, that returns the exact optimal solution to practical BSM problem instances. Isorropia constructs two sets of candidate rotations from which it builds three sets of promising antichains, and performs local search on those three sets of promising antichains. Our extensive experimental study shows that Isorropia surpasses the time-efficiency of baselines that return the exact solution by up to three orders of magnitude.
|
stable marriage; fairness
|
We propose an efficient algorithm that returns the exact optimal solution to practical Balanced Stable Marriage problem instances.
| 11,199
| null |
[
-0.07844153791666031,
0.05018991976976395,
-0.0036945738829672337,
-0.03596816211938858,
-0.03772649168968201,
-0.04322991892695427,
-0.0018965419149026275,
-0.021715808659791946,
0.027398139238357544,
0.041989825665950775,
-0.04182327166199684,
0.0024815306533128023,
0.0545165091753006,
-0.002102821134030819,
0.038825005292892456,
0.03275926783680916,
0.08053512126207352,
0.05017537251114845,
0.01622064784169197,
0.0657050833106041,
-0.13148905336856842,
-0.0941832959651947,
-0.028452957049012184,
-0.0028741098940372467,
-0.026712359860539436,
-0.07288389652967453,
0.0074541219510138035,
-0.03279939666390419,
-0.04399392381310463,
-0.05400383099913597,
0.03215043619275093,
0.026734063401818275,
0.04866243153810501,
0.03270033374428749,
-0.06771270185709,
-0.05029842257499695,
-0.13051104545593262,
-0.03846759349107742,
-0.026063639670610428,
-0.04053851217031479,
0.018604781478643417,
0.037811920046806335,
-0.007763308938592672,
-0.06311609596014023,
-0.039150435477495193,
0.0383564755320549,
0.06275882571935654,
0.10167592018842697,
0.006501737050712109,
-0.02040170319378376,
-0.034773923456668854,
-0.012155897915363312,
-0.08518988639116287,
0.05765020474791527,
-0.016415515914559364,
0.05322202667593956,
0.029715780168771744,
0.004511913284659386,
-0.028491411358118057,
0.03597278147935867,
0.07203994691371918,
-0.017955835908651352,
-0.0781857967376709,
-0.027444621548056602,
0.033967532217502594,
-0.02280096337199211,
0.03885238617658615,
0.07448912411928177,
-0.11673711240291595,
0.09524504095315933,
0.08373704552650452,
0.010652110911905766,
-0.09254033863544464,
0.04571859911084175,
0.03454069048166275,
0.026867225766181946,
-0.0003877798735629767,
-0.055237554013729095,
-0.003207738744094968,
-0.01536547765135765,
-0.12356927245855331,
-0.06673385202884674,
-0.001044167671352625,
0.02076714113354683,
0.05027073249220848,
-0.0635702982544899,
-0.05336160585284233,
0.004345016088336706,
0.09556178748607635,
0.0016521181678399444,
-0.04873555526137352,
0.06382156908512115,
0.03261326998472214,
-0.0791015475988388,
0.0471731573343277,
0.05005822330713272,
0.04142152890563011,
0.09260368347167969,
0.030199313536286354,
0.08804876357316971,
-0.012413658201694489,
0.10407977551221848,
-0.048925574868917465,
-0.030135231092572212,
0.05093095824122429,
0.02879941463470459,
-0.03321000188589096,
-0.01786784455180168,
0.02718440257012844,
-0.049955397844314575,
-0.026325445622205734,
-0.042927563190460205,
0.04924803227186203,
0.029857929795980453,
-0.08799004554748535,
0.05150493606925011,
0.028972085565328598,
0.00354575552046299,
-0.03814556077122688,
0.048219695687294006,
-0.012432237155735493,
0.04309137165546417,
0.06216418370604515,
-0.06091362237930298,
-0.04177015274763107,
0.02730352245271206,
0.02184767834842205,
4.5548624485738176e-33,
-0.0794166624546051,
0.0021702651865780354,
0.03871423378586769,
-0.046570323407649994,
0.0005535665550269186,
0.007496438920497894,
-0.005419285036623478,
0.03682198002934456,
-0.08314334601163864,
-0.0008246917277574539,
0.015522098168730736,
0.07658786326646805,
0.013696696609258652,
-0.028664328157901764,
0.028789252042770386,
-0.11579976975917816,
0.007994172163307667,
0.0636046826839447,
0.01687508262693882,
-0.02496682107448578,
0.06728961318731308,
-0.01927097700536251,
-0.024138864129781723,
0.0038662690203636885,
-0.005695959087461233,
-0.02480049431324005,
-0.006577482912689447,
0.04032985493540764,
0.12966673076152802,
-0.0076024336740374565,
-0.020645437762141228,
0.04244694113731384,
0.033409588038921356,
0.01217335183173418,
0.015898821875452995,
0.005158547777682543,
-0.03679710626602173,
-0.04917532950639725,
-0.044692013412714005,
-0.020865442231297493,
-0.07927506417036057,
0.06402002274990082,
0.005452943500131369,
-0.036594029515981674,
-0.02728239819407463,
0.0018847968894988298,
-0.021639663726091385,
0.02596331015229225,
0.04265435039997101,
0.009777845814824104,
0.05774170160293579,
0.011809860356152058,
-0.06149250641465187,
-0.004699339158833027,
-0.07437723129987717,
0.00710276747122407,
-0.013665533624589443,
0.006541090086102486,
-0.04948434233665466,
0.09726671874523163,
0.06812059134244919,
-0.031247450038790703,
-0.010215654969215393,
0.011914207600057125,
-0.09962152689695358,
0.0924825668334961,
0.008087505586445332,
-0.07352960109710693,
-0.006626065354794264,
0.0000030440889986493858,
0.054781392216682434,
-0.03271123766899109,
-0.04333725944161415,
-0.010705155320465565,
0.03669166937470436,
-0.018100224435329437,
0.038485340774059296,
0.06786640733480453,
0.03450942039489746,
-0.07723241299390793,
-0.047391243278980255,
0.04214246943593025,
-0.009060697630047798,
0.006749174557626247,
-0.06846210360527039,
-0.10401608794927597,
-0.01862451434135437,
0.004315472207963467,
-0.012197576463222504,
0.026151252910494804,
-0.011483356356620789,
0.06050959229469299,
-0.035888753831386566,
0.05452960357069969,
0.09587349742650986,
-4.497868604974199e-33,
-0.04119380563497543,
-0.1276084929704666,
0.009084705263376236,
0.08083531260490417,
0.07872064411640167,
-0.0026696219574660063,
-0.015393512323498726,
-0.10660102963447571,
0.05549849197268486,
0.08182093501091003,
-0.003708973992615938,
-0.00044675107346847653,
0.11676240712404251,
0.01113252155482769,
-0.028173450380563736,
-0.10209355503320694,
0.05313129350543022,
0.011066362261772156,
0.052391406148672104,
0.050332680344581604,
0.09844580292701721,
0.08886353671550751,
-0.0335107259452343,
0.07181031256914139,
0.03058251366019249,
0.0486380010843277,
-0.034718744456768036,
-0.052156101912260056,
0.012137633748352528,
0.033466774970293045,
-0.08882804960012436,
-0.0544159822165966,
-0.10481654852628708,
-0.04103561490774155,
0.038758594542741776,
0.04321836307644844,
-0.06641095131635666,
0.09630081802606583,
-0.024061182513833046,
0.04525594040751457,
-0.024183612316846848,
-0.052086036652326584,
-0.0621657632291317,
0.00529042724519968,
0.021450934931635857,
0.023913053795695305,
-0.045960232615470886,
0.021169306710362434,
0.10362665355205536,
-0.05088472738862038,
0.029109301045536995,
0.08337210863828659,
0.0055063022300601006,
-0.00597397331148386,
0.007265892345458269,
-0.04973936080932617,
-0.08150611817836761,
0.0955861359834671,
0.07993088662624359,
0.07095040380954742,
0.012359357438981533,
0.02424895204603672,
-0.02891678176820278,
0.08099999278783798,
-0.01495972741395235,
0.020237049087882042,
-0.008353724144399166,
-0.02047666534781456,
-0.038902509957551956,
-0.010198508389294147,
0.0003099643508903682,
0.01790124550461769,
-0.019424691796302795,
-0.039619218558073044,
0.015971766784787178,
0.09275587648153305,
0.022423749789595604,
-0.04601707309484482,
0.03107287548482418,
-0.01535465195775032,
-0.07217011600732803,
0.08164608478546143,
0.06413107365369797,
0.05224403738975525,
-0.04350676015019417,
-0.08463137596845627,
0.005475928541272879,
-0.02460559271275997,
0.026629388332366943,
0.05823269858956337,
0.00607316056266427,
-0.019594360142946243,
0.08049460500478745,
-0.0688905343413353,
-0.06887607276439667,
-4.742507542232488e-8,
-0.02313225157558918,
-0.07264982908964157,
-0.01221575029194355,
-0.020116422325372696,
-0.07867283374071121,
-0.022246386855840683,
0.02402872033417225,
-0.00009645205864217132,
-0.03766504302620888,
0.0018520788289606571,
-0.016111716628074646,
0.06427060067653656,
0.0006720584933646023,
0.0049582780338823795,
-0.042441535741090775,
-0.009569980204105377,
-0.0043175495229661465,
-0.05264721065759659,
-0.05385332927107811,
-0.009142711758613586,
-0.02527775801718235,
0.025648009032011032,
-0.007930767722427845,
-0.0832456722855568,
-0.0409921258687973,
0.0512208491563797,
0.1077757477760315,
-0.002407231368124485,
0.054102953523397446,
0.02767438255250454,
-0.027286261320114136,
-0.05775383859872818,
0.004371704999357462,
-0.023088941350579262,
0.09114665538072586,
0.020335594192147255,
-0.14203263819217682,
0.05228253826498985,
0.0064317574724555016,
0.004020047839730978,
0.0004809265083167702,
0.08203224092721939,
-0.07163842767477036,
-0.010451660491526127,
0.07244685292243958,
0.025028761476278305,
-0.0042999908328056335,
0.023419659584760666,
-0.0651429295539856,
-0.0636763945221901,
0.011224598623812199,
-0.08382882922887802,
0.027632880955934525,
-0.06236819922924042,
0.015876349061727524,
0.006949778646230698,
0.006767520681023598,
-0.030151134356856346,
0.052593935281038284,
-0.017051052302122116,
0.0714106559753418,
-0.06571433693170547,
0.008716125041246414,
0.020048849284648895
] | 0
| 0
| 0
| 0
|
|
Addax: Utilizing Zeroth-Order Gradients to Improve Memory Efficiency and Performance of SGD for Fine-Tuning Language Models
|
https://openreview.net/forum?id=QhxjQOMdDF
|
[
"Zeman Li",
"Xinwei Zhang",
"Peilin Zhong",
"Yuan Deng",
"Meisam Razaviyayn",
"Vahab Mirrokni"
] |
Poster
|
Fine-tuning language models (LMs) with the standard Adam optimizer often demands excessive memory, limiting accessibility. The ``in-place'' version of Stochastic Gradient Descent (IP-SGD) and Memory-Efficient Zeroth-order Optimizer (MeZO) have been proposed as solutions to improve memory efficiency. However, IP-SGD still requires a decent amount of memory, and MeZO suffers from slow convergence and degraded final performance due to its zeroth-order nature. This paper introduces Addax, a novel method that improves both memory efficiency and algorithm performance of IP-SGD by integrating it with MeZO. Specifically, Addax computes the zeroth-order or first-order gradient of the data points in the minibatch based on their memory consumption and combines zeroth- and first-order gradient estimates to obtain the updated direction in each step.
By computing the zeroth-order order gradient of data points that require more memory and the first-order gradient of the ones that require less memory, Addax overcomes the slow convergence of MeZO and excessive memory requirement of IP-SGD. Additionally, the zeroth-order gradient acts as a regularizer for the first-order gradient, further enhancing the model's final performance.
Theoretically, we establish the convergence of Addax under mild assumptions, demonstrating faster convergence and less restrictive hyper-parameter choices than MeZO. Our extensive experiments with diverse LMs and tasks show that Addax consistently outperforms MeZO in terms of accuracy and convergence speed, while having a comparable memory footprint.
In particular, our experiments using one A100 GPU on OPT-13B model reveal that, on average, Addax outperforms MeZO in terms of accuracy/F1 score by 14%, and runs $15\times$ faster, while having a comparable memory footprint to MeZO. In our experiments on the larger OPT-30B model, on average, Addax outperforms MeZO in terms of accuracy/F1 score by >16% and runs $30\times$ faster on a single H100 GPU. Moreover, Addax surpasses the performance of standard fine-tuning approaches, such as IP-SGD and Adam, in most tasks in terms of Accuracy/F1 score with significantly less memory requirement.
|
Large Language Models, memory efficient Fine-tuning, Zeroth order optimization
| null | 11,183
|
2410.06441
|
[
0.006370393093675375,
-0.05011407658457756,
0.000927680463064462,
0.0008569128694944084,
0.0166670773178339,
0.016517190262675285,
-0.03996936231851578,
-0.009532533586025238,
0.02882635034620762,
-0.010026977397501469,
0.06173190101981163,
0.0858178660273552,
0.015467955730855465,
0.002626944100484252,
0.022797200828790665,
0.018712185323238373,
0.04732329398393631,
0.10020337253808975,
-0.05441606044769287,
-0.1269310861825943,
0.03282942995429039,
0.01454065553843975,
-0.009612871333956718,
-0.07304659485816956,
0.022821538150310516,
0.017627285793423653,
-0.05611454322934151,
0.016650378704071045,
0.06339200586080551,
-0.030566390603780746,
0.11043235659599304,
0.011149627156555653,
0.014794519171118736,
0.019211070612072945,
-0.08354559540748596,
0.018481362611055374,
-0.05344046279788017,
0.02047552540898323,
-0.06568104773759842,
-0.033815864473581314,
-0.01364300400018692,
0.041982803493738174,
0.04344915226101875,
0.005747821647673845,
0.08431505411863327,
-0.02189926989376545,
0.002860124222934246,
-0.03377799317240715,
0.07959092408418655,
-0.014147501438856125,
-0.04437601566314697,
-0.01684049889445305,
-0.07761150598526001,
0.04073995351791382,
-0.01722712628543377,
0.03097536973655224,
-0.014069241471588612,
0.024545200169086456,
-0.051687028259038925,
-0.010448753833770752,
-0.032612141221761703,
-0.15521402657032013,
0.044208742678165436,
0.003043566830456257,
-0.036629028618335724,
-0.09849222749471664,
0.0724690780043602,
0.021187253296375275,
-0.006660263519734144,
0.07765302062034607,
0.01348284725099802,
0.05819927901029587,
0.01709352433681488,
0.05476567521691322,
-0.04751461371779442,
0.07376718521118164,
0.05722403898835182,
-0.04611428454518318,
0.028837041929364204,
-0.07335717231035233,
-0.005275299772620201,
0.0050900643691420555,
0.07464161515235901,
-0.018926464021205902,
0.054218094795942307,
0.027296682819724083,
0.044151630252599716,
-0.01997857168316841,
0.058452896773815155,
0.017919255420565605,
-0.012324071489274502,
0.011730805039405823,
-0.11220838874578476,
-0.023877054452896118,
0.016391629353165627,
0.02250429429113865,
0.03382465988397598,
-0.006028269417583942,
-0.12971313297748566,
0.0571349561214447,
0.055775225162506104,
0.024012381210923195,
-0.056451618671417236,
0.003127892268821597,
0.022986724972724915,
-0.03562771901488304,
0.01689794473350048,
0.035672225058078766,
0.0013134818291291595,
-0.021205328404903412,
0.01464487798511982,
0.057206541299819946,
0.02349270135164261,
0.014668271876871586,
-0.0013238649116829038,
-0.029151789844036102,
0.006007345858961344,
-0.08287503570318222,
0.02311437577009201,
0.04612131416797638,
-0.0819622278213501,
-0.002123427577316761,
0.0018418926047161222,
0.05843913555145264,
-0.05289355665445328,
0.05006231740117073,
0.007427821401506662,
2.2267442752214364e-33,
-0.04801896959543228,
0.05150189623236656,
-0.008385968394577503,
-0.08781585842370987,
0.03962281718850136,
-0.0012891801306977868,
0.031879689544439316,
0.03467335179448128,
-0.10226273536682129,
0.013251523487269878,
-0.045380283147096634,
-0.09486541152000427,
-0.03633582592010498,
0.03577607870101929,
0.03760068118572235,
-0.04447994381189346,
0.021618759259581566,
0.032447800040245056,
0.04627285897731781,
-0.029435276985168457,
0.07142328470945358,
-0.03705475106835365,
0.02192828617990017,
-0.03438131883740425,
0.04045107960700989,
0.022408502176404,
0.10209745913743973,
-0.017096292227506638,
-0.0195341557264328,
0.004160265438258648,
-0.0710204690694809,
-0.036363329738378525,
-0.0485709123313427,
-0.036243464797735214,
0.045555878430604935,
-0.03946144878864288,
0.0028505942318588495,
-0.03551690652966499,
0.01607782579958439,
-0.07126551866531372,
-0.04655684158205986,
0.042062677443027496,
-0.009934724308550358,
-0.11656475812196732,
-0.14758048951625824,
-0.03112224116921425,
-0.0035123201087117195,
0.08834931999444962,
0.06109800562262535,
-0.04694719240069389,
-0.035314809530973434,
0.06520844250917435,
-0.060774967074394226,
-0.01666933484375477,
0.01043684221804142,
-0.0012037763372063637,
0.02665417082607746,
0.061910249292850494,
0.030676180496811867,
0.11730657517910004,
-0.03936026990413666,
0.04250994324684143,
-0.027120305225253105,
0.0917954221367836,
0.06697700917720795,
-0.039014264941215515,
0.005719335284084082,
0.0395641028881073,
0.07476870715618134,
-0.02765580639243126,
0.009446719661355019,
-0.05081944912672043,
0.0457145981490612,
0.028109019622206688,
0.04076763242483139,
-0.008166876621544361,
0.06445108354091644,
-0.12052669376134872,
0.037964411079883575,
0.002591256983578205,
0.016781600192189217,
0.050951242446899414,
0.007008716464042664,
-0.027547141537070274,
-0.05448263883590698,
-0.08201740682125092,
0.047609634697437286,
0.012744663283228874,
0.03694700077176094,
-0.041395965963602066,
-0.06476138532161713,
-0.023631565272808075,
0.023427540436387062,
0.007884166203439236,
-0.057781122624874115,
-2.606957188310936e-33,
-0.06381132453680038,
-0.040139906108379364,
-0.013425786048173904,
0.07204723358154297,
0.02555461972951889,
-0.07648611813783646,
0.004595608450472355,
-0.009787007234990597,
-0.0104605033993721,
-0.04795071855187416,
-0.06135166808962822,
0.011755534447729588,
0.02782939374446869,
0.0038103326223790646,
0.08594076335430145,
0.010466045700013638,
-0.06981110572814941,
0.02626548893749714,
0.03667273744940758,
0.02688208408653736,
-0.04672372341156006,
0.12486082315444946,
-0.02135736122727394,
0.01897410862147808,
-0.006934017408639193,
0.007681039161980152,
-0.013592049479484558,
0.06423117220401764,
-0.07235129922628403,
-0.03392895311117172,
-0.0051124487072229385,
0.07717512547969818,
-0.0371893011033535,
0.03090212680399418,
-0.01660553552210331,
0.05838986858725548,
-0.013301018625497818,
-0.02630804292857647,
0.027459222823381424,
0.11314541846513748,
0.1339976191520691,
0.05730031430721283,
-0.024894917383790016,
-0.050008922815322876,
0.027429644018411636,
-0.02366762049496174,
-0.16596171259880066,
0.05660269036889076,
-0.043651215732097626,
-0.013073933310806751,
-0.08448422700166702,
0.02776275761425495,
-0.04284646362066269,
-0.0018870059866458178,
0.02878830023109913,
-0.05534123256802559,
-0.0486183725297451,
0.007531691808253527,
0.09546098858118057,
-0.019777754321694374,
-0.03423108160495758,
-0.08160310983657837,
-0.00851816963404417,
-0.08261916041374207,
-0.009907877072691917,
-0.00405842624604702,
0.025080285966396332,
-0.06185450032353401,
-0.002204212360084057,
0.0034031476825475693,
0.0536910779774189,
0.07020563632249832,
-0.015729021281003952,
0.04551032930612564,
-0.04135740175843239,
-0.036987245082855225,
0.0402471087872982,
0.00737712811678648,
-0.03930081054568291,
-0.09882855415344238,
0.011343267746269703,
0.08581817150115967,
-0.008417939767241478,
0.03898689150810242,
0.014798406511545181,
0.05077725648880005,
0.021086907014250755,
0.07066608220338821,
0.06465361267328262,
-0.02169310487806797,
-0.022676803171634674,
0.05286148935556412,
0.017842987552285194,
0.005909428466111422,
-0.08876916021108627,
-4.376046547349688e-8,
-0.07901035249233246,
-0.08154979348182678,
0.07556623220443726,
0.05616419017314911,
0.08364799618721008,
-0.035545021295547485,
-0.03141401708126068,
0.06546023488044739,
0.1113077774643898,
-0.05899190902709961,
0.06317069381475449,
-0.0064944978803396225,
-0.06249057501554489,
-0.04999714717268944,
-0.06592250615358353,
0.09934284538030624,
0.007580914068967104,
-0.031185531988739967,
0.017662404105067253,
-0.030895818024873734,
-0.0490085706114769,
0.09504761546850204,
-0.012740137986838818,
0.0016634090570732951,
0.04898190125823021,
-0.07166582345962524,
0.04930030554533005,
0.03573049604892731,
0.11085746437311172,
-0.030622778460383415,
-0.01719166338443756,
0.016222840175032616,
-0.006700605154037476,
0.007076931651681662,
0.04090837016701698,
0.07630420476198196,
0.056110773235559464,
0.08210161328315735,
-0.055490706115961075,
-0.05114240199327469,
-0.004544361960142851,
0.04273594170808792,
0.026117097586393356,
0.007725004106760025,
-0.005935331806540489,
-0.04285114258527756,
-0.08756697922945023,
-0.12930645048618317,
-0.0198719073086977,
-0.09089282900094986,
0.0070268502458930016,
0.010040797293186188,
-0.0033664321526885033,
0.02082301676273346,
0.008161241188645363,
-0.012172512710094452,
-0.054079100489616394,
-0.0398155115544796,
0.018348049372434616,
0.06435061246156693,
0.016641516238451004,
0.002443246077746153,
-0.06082667410373688,
-0.02752782218158245
] | 0
| 0
| 0
| 0
|
|
Seeing Eye to AI: Human Alignment via Gaze-Based Response Rewards for Large Language Models
|
https://openreview.net/forum?id=uZgK0tcPqd
|
[
"Ángela López-Cardona",
"Carlos Segura",
"Alexandros Karatzoglou",
"Sergi Abadal",
"Ioannis Arapakis"
] |
Poster
|
Advancements in Natural Language Processing (NLP), have led to the emergence of Large Language Models (LLMs) such as GPT, Llama, Claude, and Gemini, which excel across a range of tasks but require extensive fine-tuning to align their outputs with human expectations. A widely used method for achieving this alignment is Reinforcement Learning from Human Feedback (RLHF), which, despite its success, faces challenges in accurately modelling human preferences. In this paper, we introduce GazeReward, a novel framework that integrates implicit feedback -- and specifically eye-tracking (ET) data -- into the Reward Model (RM). In addition, we explore how ET-based features can provide insights into user preferences. Through ablation studies we test our framework with different integration methods, LLMs, and ET generator models, demonstrating that our approach significantly improves the accuracy of the RM on established human preference datasets. This work advances the ongoing discussion on optimizing AI alignment with human values, exploring the potential of cognitive data for shaping future NLP research.
|
reward model, RLHF, visual attention, LLMs, eye tracking, implicit feedback
|
GazeReward: A novel framework that integrates implicit feedback into Reward Model (RM)
| 11,182
|
2410.01532
|
[
-0.08809125423431396,
-0.1130041629076004,
-0.0018646344542503357,
0.006162967998534441,
0.05521254241466522,
0.04502131789922714,
0.07490932941436768,
-0.003226320259273052,
0.023950116708874702,
-0.011393899098038673,
-0.021959491074085236,
-0.09162760525941849,
0.04421762004494667,
-0.007945547811686993,
0.04106529802083969,
-0.010536706075072289,
0.02919636107981205,
-0.00015863495355006307,
-0.05642647668719292,
-0.05375995859503746,
0.03253750875592232,
0.0033916395623236895,
0.0669446736574173,
-0.022655755281448364,
-0.029978839680552483,
0.05487898364663124,
0.01962824910879135,
-0.0685562863945961,
-0.021359071135520935,
-0.019689254462718964,
0.051431477069854736,
-0.04356508329510689,
0.012858029454946518,
0.003656513523310423,
-0.1062714084982872,
0.03813306987285614,
-0.11580691486597061,
-0.03166526183485985,
-0.06195942312479019,
-0.09646487981081009,
-0.06445368379354477,
-0.0021389955654740334,
0.007346229627728462,
0.05499159172177315,
0.05663071945309639,
-0.00314681651070714,
-0.026207195594906807,
-0.02381003089249134,
-0.032203786075115204,
0.01759113371372223,
-0.0720648244023323,
-0.034583937376737595,
0.06341284513473511,
0.04960804805159569,
-0.004789751023054123,
0.08587241172790527,
-0.0072180768474936485,
0.03233333304524422,
0.06809408217668533,
-0.07003750652074814,
-0.012375347316265106,
-0.10745750367641449,
-0.025894775986671448,
0.003049812512472272,
-0.05107150599360466,
-0.042182281613349915,
-0.0038714767433702946,
-0.03153097629547119,
-0.06673811376094818,
0.015863871201872826,
-0.08596504479646683,
0.0453995056450367,
0.035314206033945084,
-0.010797590017318726,
-0.0009850282222032547,
0.10827324539422989,
0.009399313479661942,
-0.09259378165006638,
0.07070574909448624,
-0.028580637648701668,
-0.02198607847094536,
0.016713393852114677,
0.10345160961151123,
0.01550382561981678,
0.09322888404130936,
-0.022297298535704613,
0.052930884063243866,
0.03812868893146515,
0.0009206451359204948,
0.0011728768004104495,
-0.05904589593410492,
-0.07565203309059143,
0.010323707014322281,
-0.032706134021282196,
-0.021657289937138557,
0.1344340592622757,
-0.049981120973825455,
-0.050645504146814346,
-0.08793997019529343,
0.039513275027275085,
0.055805549025535583,
0.08580294251441956,
-0.07096545398235321,
-0.035022616386413574,
-0.0262420903891325,
-0.04247329384088516,
0.012672067619860172,
0.042162343859672546,
0.06763274222612381,
-0.10073893517255783,
-0.013589667156338692,
0.00521314749494195,
0.014319250360131264,
0.0654635801911354,
0.038025639951229095,
-0.033994320780038834,
-0.007208039052784443,
-0.04146403446793556,
0.05599066987633705,
0.021727290004491806,
0.07493078708648682,
0.03049069084227085,
0.05246046558022499,
0.014205507934093475,
0.0806373879313469,
0.02095317654311657,
0.03729638829827309,
-8.581300697189669e-34,
0.046803079545497894,
0.033955417573451996,
0.04036051034927368,
0.008993498049676418,
0.06790634989738464,
-0.042071424424648285,
-0.0037093935534358025,
0.053748100996017456,
-0.09885170310735703,
-0.0037342964205890894,
-0.04282026365399361,
0.005112866871058941,
-0.0643821731209755,
0.1506347507238388,
0.015514110215008259,
-0.010571223683655262,
-0.0614013746380806,
0.06405087560415268,
-0.06983321905136108,
-0.05883066728711128,
-0.013845687732100487,
0.017879413440823555,
0.07036793231964111,
-0.014925491996109486,
0.05901816859841347,
-0.00910251121968031,
-0.013671817258000374,
-0.060252513736486435,
-0.0347960963845253,
0.006249264348298311,
-0.033860210329294205,
0.03742517903447151,
0.0017489456804469228,
0.00009059612057171762,
-0.02738519199192524,
-0.027042830362915993,
0.019679438322782516,
-0.047527577728033066,
0.018363913521170616,
0.047548528760671616,
0.0008899719105102122,
0.0878070741891861,
0.028072377666831017,
0.051632482558488846,
-0.10676410794258118,
0.0027092164382338524,
0.06745283305644989,
0.06452376395463943,
-0.0104376757517457,
-0.03877165913581848,
0.0529119074344635,
0.03995104134082794,
0.013822576031088829,
-0.06058291718363762,
-0.028476819396018982,
0.014388441108167171,
-0.00670945318415761,
0.05434652045369148,
-0.045042701065540314,
-0.06974443048238754,
0.06799829751253128,
0.025556370615959167,
0.03990891948342323,
0.01578470505774021,
0.0008394171018153429,
0.04656794294714928,
0.007834130898118019,
0.061392102390527725,
0.026329154148697853,
0.009716368280351162,
0.02999907173216343,
0.05122337490320206,
-0.0019945891108363867,
-0.03235677257180214,
0.003357905661687255,
0.05476462468504906,
0.07561453431844711,
-0.12655192613601685,
0.07982874661684036,
-0.0447314977645874,
-0.01224928442388773,
-0.04168173298239708,
-0.03655727580189705,
-0.08023051172494888,
-0.014882218092679977,
-0.12461066991090775,
0.04878082871437073,
-0.06658042222261429,
-0.01984638348221779,
-0.026333047077059746,
0.01265689730644226,
0.046967387199401855,
-0.00934602227061987,
-0.010791542939841747,
-0.05101490765810013,
3.23578284033103e-34,
-0.05038118734955788,
-0.03273972123861313,
0.03366312384605408,
-0.01565566286444664,
-0.04446034133434296,
0.04469268396496773,
0.039019059389829636,
-0.015146927908062935,
0.041738711297512054,
-0.039590898901224136,
-0.015792321413755417,
-0.05288223549723625,
0.030772823840379715,
0.0001529397559352219,
0.0068608676083385944,
-0.03736118599772453,
-0.020801033824682236,
0.05162700265645981,
0.006893211975693703,
0.02595183439552784,
0.06096004322171211,
0.06428433209657669,
-0.12365085631608963,
-0.03168412670493126,
-0.0027187534142285585,
0.04063661769032478,
0.035300035029649734,
0.0025242262054234743,
-0.055195700377225876,
0.006962694227695465,
-0.04625696316361427,
0.017070038244128227,
-0.11800059676170349,
-0.046528562903404236,
-0.04740992933511734,
0.09844411909580231,
0.026241352781653404,
-0.06665868312120438,
-0.07005108892917633,
0.11405113339424133,
-0.004668375942856073,
-0.011707817204296589,
-0.057855457067489624,
0.017546307295560837,
-0.019157815724611282,
-0.0009938398143276572,
-0.08035283535718918,
0.013515998609364033,
-0.053175825625658035,
0.029032956808805466,
-0.025957640260457993,
0.014314640313386917,
-0.08915981650352478,
0.0308450348675251,
-0.12856870889663696,
-0.04907629266381264,
-0.02675642818212509,
-0.04801931977272034,
0.02578422985970974,
-0.041486479341983795,
-0.05563601851463318,
0.020457856357097626,
0.05375423654913902,
0.02081359177827835,
-0.013883222825825214,
0.006399865262210369,
0.05047088861465454,
0.0053365714848041534,
0.024535629898309708,
-0.05282190442085266,
0.10166437923908234,
-0.05022389441728592,
0.01668642833828926,
0.022007737308740616,
0.053959500044584274,
0.05499187111854553,
0.07495588064193726,
-0.028438081964850426,
0.004649435169994831,
-0.11546134948730469,
0.0794493779540062,
0.02954716980457306,
0.03978735953569412,
0.02689368650317192,
0.03984601050615311,
0.08730176836252213,
0.09628076106309891,
0.09379038959741592,
0.005464744754135609,
0.017672529444098473,
-0.09269462525844574,
-0.019997209310531616,
0.01660829409956932,
0.09022698551416397,
-0.056012943387031555,
-3.9062303613945915e-8,
-0.12178463488817215,
-0.019927093759179115,
-0.00642762565985322,
0.13673464953899384,
-0.0000551458761037793,
0.007446603383868933,
-0.0185161791741848,
0.018982185050845146,
-0.03690306842327118,
-0.07199613749980927,
0.03755246475338936,
0.0801633670926094,
0.04579436033964157,
-0.030639663338661194,
0.03604654595255852,
0.027668224647641182,
-0.02158360928297043,
0.062330085784196854,
-0.02138727717101574,
-0.01070308592170477,
0.05268673226237297,
0.06952468305826187,
-0.005319149699062109,
-0.006904270965605974,
0.030474446713924408,
-0.03269919380545616,
-0.05197707936167717,
0.12234077602624893,
-0.01486992184072733,
0.006006925832480192,
0.06133195757865906,
0.030793797224760056,
0.008611506782472134,
-0.05052994191646576,
0.07247316837310791,
0.01877214014530182,
0.00803239643573761,
-0.013856490142643452,
0.022664958611130714,
0.00824346300214529,
-0.01435865554958582,
0.08349092304706573,
-0.029239414259791374,
0.011279369704425335,
0.006968461908400059,
0.028805440291762352,
-0.04842035472393036,
-0.1187676265835762,
0.042466405779123306,
-0.04885258898139,
0.004448436200618744,
-0.06347035616636276,
0.011915143579244614,
0.031552135944366455,
0.05365361273288727,
-0.0659618005156517,
-0.0047778175212442875,
0.009142624214291573,
0.026533953845500946,
0.004909839015454054,
0.05668604373931885,
0.00737660052254796,
-0.04132360592484474,
0.006907361559569836
] |
https://github.com/telefonica-scientific-research/gaze_reward
| 2
| 0
| 0
| 0
|
Revisiting Large-Scale Non-convex Distributionally Robust Optimization
|
https://openreview.net/forum?id=JYwVijuNA7
|
[
"Qi Zhang",
"Yi Zhou",
"Simon Khan",
"Ashley Prater-Bennette",
"Lixin Shen",
"Shaofeng Zou"
] |
Poster
|
Distributionally robust optimization (DRO) is a powerful technique to train robust machine learning models that perform well under distribution shifts. Compared with empirical risk minimization (ERM), DRO optimizes the expected loss under the worst-case distribution in
an uncertainty set of distributions. This paper revisits the important problem of DRO with non-convex smooth loss functions. For this problem, Jin et al. (2021) showed that its dual problem is generalized $(L_0, L_1)$-smooth condition and gradient noise satisfies the affine variance condition, designed an algorithm of mini-batch normalized gradient descent with momentum, and proved its convergence and complexity. In this paper, we show that the dual problem and the gradient noise satisfy simpler yet more precise partially generalized smoothness condition and partially affine variance condition by studying the optimization variable and dual variable separately, which further yields much simpler algorithm design and convergence analysis. We develop a double stochastic gradient descent with clipping (D-SGD-C) algorithm that converges to an $\epsilon$-stationary point with $\mathcal O(\epsilon^{-4})$ gradient complexity, which matches with results in Jin et al. (2021). Our algorithm does not need to use momentum, and the proof is much simpler, thanks to the more precise characterization of partially generalized smoothness and partially affine variance noise. We further design a variance-reduced method that achieves a lower gradient complexity of $\mathcal O(\epsilon^{-3})$. Our theoretical results and insights are further verified numerically on a number of tasks, and our algorithms outperform the existing DRO method (Jin et al., 2021).
|
distributionally robust optimization, generalized smoothness, non-convex optimization, variance-reduced method
| null | 11,176
| null |
[
-0.08615785092115402,
-0.05737082660198212,
0.03374025225639343,
0.034878525882959366,
-0.021107738837599754,
0.0018725177505984902,
-0.007859477773308754,
-0.03129778802394867,
-0.04612456262111664,
-0.008367004804313183,
0.017027154564857483,
0.03443772718310356,
0.016320202499628067,
0.006747254636138678,
-0.013943317346274853,
-0.012969130650162697,
0.07578372955322266,
0.11380252242088318,
-0.07157595455646515,
0.02634441666305065,
-0.024329237639904022,
0.08602502197027206,
-0.04118124395608902,
0.007704383693635464,
0.11691755801439285,
-0.05632038787007332,
-0.020014403387904167,
-0.014042272232472897,
-0.00838619377464056,
-0.01232471875846386,
0.11497765779495239,
-0.08316344767808914,
0.009759943932294846,
-0.008966182358562946,
0.05552220717072487,
0.014415879733860493,
-0.04416739568114281,
0.07107583433389664,
0.11661478132009506,
-0.001450532698072493,
-0.031572822481393814,
-0.021258343011140823,
-0.017395423725247383,
-0.08538538217544556,
0.017058158293366432,
-0.05109785497188568,
0.007936172187328339,
-0.1066381186246872,
0.08125581592321396,
0.030927900224924088,
-0.005024780984967947,
-0.050536662340164185,
-0.0666743591427803,
0.00910467840731144,
-0.026223622262477875,
-0.00098217255435884,
0.01204452570527792,
-0.0046499501913785934,
-0.004528685472905636,
0.0012745184358209372,
-0.0032759818714112043,
-0.1464463323354721,
0.023314891383051872,
0.030522650107741356,
-0.07275127619504929,
-0.07081837207078934,
0.0012067988282069564,
-0.01283519621938467,
0.0544896200299263,
0.032818105071783066,
-0.02788928709924221,
0.09644155949354172,
-0.048819247633218765,
0.02636769786477089,
0.016403954476118088,
0.023196306079626083,
0.13480062782764435,
-0.00883958674967289,
-0.026048416271805763,
-0.04042651131749153,
0.08102082461118698,
0.09272465109825134,
0.07793255895376205,
-0.04075999930500984,
0.05779070407152176,
0.019641244783997536,
-0.054338738322257996,
0.03384772315621376,
0.016061903908848763,
-0.02008926309645176,
-0.047138575464487076,
0.0370018295943737,
-0.07606919854879379,
0.06158246845006943,
0.00024351835600100458,
-0.05026089772582054,
0.0322427898645401,
-0.04555370658636093,
0.003992632497102022,
0.13047389686107635,
0.007563586346805096,
0.09640824794769287,
0.029512599110603333,
-0.05149373039603233,
0.026093121618032455,
-0.08325240015983582,
0.009658310562372208,
-0.055882904678583145,
-0.026979869231581688,
-0.08319615572690964,
0.061509259045124054,
0.07748290151357651,
0.0038680280558764935,
-0.026426928117871284,
-0.06065785139799118,
-0.013877109624445438,
0.012400870211422443,
-0.0015012193471193314,
-0.02141628973186016,
0.0529622919857502,
-0.1060992181301117,
-0.038572315126657486,
0.007206592708826065,
0.06714664399623871,
-0.00830045249313116,
0.02792862430214882,
-0.008118906058371067,
2.6102509602161385e-33,
-0.010983902029693127,
-0.0031097284518182278,
0.01404543686658144,
-0.09431227296590805,
0.04994985833764076,
0.0053160591050982475,
0.011426075361669064,
-0.050014566630125046,
-0.03679090365767479,
-0.012281094677746296,
0.062270306050777435,
-0.1102328822016716,
-0.06999542564153671,
0.008612846955657005,
0.0027340534143149853,
-0.052475444972515106,
-0.03070182166993618,
0.026583056896924973,
0.07890377193689346,
-0.05487179756164551,
0.03908758983016014,
-0.08545402437448502,
-0.02731364034116268,
0.022977903485298157,
0.012491305358707905,
-0.008166390471160412,
0.09688931703567505,
-0.018438657745718956,
0.005560221616178751,
0.053439490497112274,
-0.11889979243278503,
0.024621743708848953,
0.02739914506673813,
-0.01978374272584915,
0.06500028073787689,
-0.005084456410259008,
-0.03572212904691696,
-0.05649910122156143,
-0.016603944823145866,
-0.0021724903490394354,
-0.07381846010684967,
0.113939568400383,
0.042275071144104004,
-0.01347747165709734,
-0.09610547870397568,
-0.05908827856183052,
0.0033096640836447477,
-0.02373381331562996,
0.02052300050854683,
-0.11122268438339233,
-0.056513577699661255,
-0.007810315582901239,
-0.00650969659909606,
-0.01900659129023552,
0.02527860924601555,
-0.00870042946189642,
-0.010326048359274864,
0.0005546919419430196,
-0.0002680766920093447,
0.017417941242456436,
-0.020189154893159866,
0.034642938524484634,
-0.031192444264888763,
-0.0028899069875478745,
0.021260716021060944,
0.02832510694861412,
0.04572003334760666,
0.0647747665643692,
-0.03172015771269798,
-0.015725867822766304,
0.030050070956349373,
0.010133882984519005,
0.035070329904556274,
-0.0976790264248848,
0.0405086949467659,
0.00801975466310978,
0.1081123948097229,
0.019822608679533005,
-0.0035705850459635258,
-0.07981008291244507,
-0.015867222100496292,
0.07085183262825012,
-0.013311090879142284,
0.0016804727492854,
-0.02872881479561329,
-0.011460963636636734,
-0.004952353425323963,
-0.020704327151179314,
-0.04726328328251839,
-0.0036516704130917788,
-0.11097817867994308,
0.04596410691738129,
-0.041228409856557846,
0.021462012082338333,
-0.02172526903450489,
-1.587422468838957e-33,
-0.07923151552677155,
0.06308247894048691,
-0.08487575501203537,
0.0400690920650959,
0.013873755931854248,
-0.008731845766305923,
-0.04857780784368515,
0.01715204305946827,
-0.07555535435676575,
-0.06259894371032715,
-0.09261855483055115,
0.012602207250893116,
-0.0028488319367170334,
-0.021042298525571823,
-0.004872473422437906,
0.003807769389823079,
0.03742087259888649,
0.13625557720661163,
-0.017012841999530792,
-0.004275643732398748,
0.046864040195941925,
0.02941557951271534,
-0.04891359806060791,
0.06107970327138901,
0.035791222006082535,
0.02181883715093136,
-0.14753390848636627,
0.11499442160129547,
-0.00493379682302475,
0.012623914517462254,
-0.06369702517986298,
0.07755874842405319,
0.013196771964430809,
-0.002534172497689724,
-0.03430238366127014,
0.08823417872190475,
-0.10134398937225342,
0.011083872057497501,
0.013553127646446228,
0.07828249037265778,
-0.0094178207218647,
0.07162369787693024,
-0.009487603791058064,
-0.054537974298000336,
0.090306855738163,
0.03710399940609932,
0.04044681042432785,
-0.06670541316270828,
0.06588172167539597,
-0.03593029826879501,
-0.04928232729434967,
-0.0004383110790513456,
-0.06199268251657486,
0.10923954099416733,
-0.03678646311163902,
0.022960856556892395,
-0.02395104244351387,
-0.06275837123394012,
0.022114738821983337,
0.028733177110552788,
-0.03639496862888336,
-0.024390995502471924,
-0.05795004963874817,
0.047982752323150635,
-0.025721393525600433,
0.04087698832154274,
-0.01203179731965065,
0.0028415629640221596,
-0.02053573541343212,
0.052428729832172394,
0.02274456061422825,
0.013387210667133331,
0.06634141504764557,
0.002771932166069746,
-0.041062403470277786,
0.019394434988498688,
0.03245040774345398,
0.007590306922793388,
-0.04208255559206009,
0.06714822351932526,
0.15474380552768707,
0.030645636841654778,
0.030394306406378746,
0.061971601098775864,
0.09752834588289261,
0.06204013153910637,
0.022683115676045418,
0.00602461863309145,
0.032839301973581314,
-0.04365367069840431,
-0.02278711646795273,
0.02144572325050831,
0.046283092349767685,
-0.002664909465238452,
0.011253999546170235,
-4.4356958994740125e-8,
-0.02650712989270687,
0.011162742041051388,
0.038590189069509506,
-0.05497784912586212,
0.07267262041568756,
-0.00372729217633605,
0.0011020522797480226,
-0.04526906460523605,
-0.004268775228410959,
-0.08374948799610138,
-0.029877424240112305,
-0.038401469588279724,
0.006732161156833172,
-0.04462682828307152,
-0.07985921204090118,
0.07426561415195465,
-0.04762010648846626,
0.06392696499824524,
0.03512765094637871,
0.033084724098443985,
-0.032466158270835876,
0.031190868467092514,
0.0439787395298481,
-0.01692756451666355,
0.04552995786070824,
-0.046126365661621094,
0.060846514999866486,
-0.0021074977703392506,
0.14477446675300598,
-0.02075495384633541,
-0.09186108410358429,
-0.017763782292604446,
0.039881832897663116,
0.030024034902453423,
-0.03553902730345726,
0.06815444678068161,
0.0025707923341542482,
-0.04387586936354637,
-0.028354747220873833,
-0.022901657968759537,
0.019728893414139748,
0.05911211296916008,
0.03086136467754841,
0.018444472923874855,
-0.022264281287789345,
0.04448069632053375,
0.07701960951089859,
-0.023420898243784904,
-0.012462734244763851,
-0.01849593222141266,
0.0785553976893425,
0.03445049375295639,
0.06953653693199158,
0.07765804976224899,
0.032539572566747665,
-0.022465331479907036,
-0.037998735904693604,
-0.03519944101572037,
-0.02161840908229351,
0.018970469012856483,
-0.007866554893553257,
-0.0007083886885084212,
-0.08061483502388,
0.03159938007593155
] | 0
| 0
| 0
| 0
|
|
Beyond single neurons: population response geometry in digital twins of mouse visual cortex
|
https://openreview.net/forum?id=kSISSDUYFh
|
[
"Dario Liscai",
"Emanuele Luconi",
"Alessandro Marin Vargas",
"Alessandro Sanzeni"
] |
Poster
|
Hierarchical visual processing is essential for cognitive functions like object recognition and spatial localization. Traditional studies of the neural basis of these computations have focused on single-neuron activity, but recent advances in large-scale neural recordings emphasize the growing need to understand computations at the population level. Digital twins-computational models trained on neural data-have successfully replicated single-neuron behavior, but their effectiveness in capturing the joint activity of neurons remains unclear. In this study, we investigate how well digital twins describe population responses in mouse visual cortex. We show that these models fail to accurately represent the geometry of population activity, particularly its differentiability and how this geometry evolves across the visual hierarchy. To address this, we explore how dataset, network architecture, loss function, and training method affect the ability of digital twins to recapitulate population properties. We demonstrate that improving model alignment with experiments requires training strategies that enhance robustness and generalization, reflecting principles observed in biological systems. These findings underscore the need to evaluate digital twins from multiple perspectives, identify key areas for refinement, and establish a foundation for using these models to explore neural computations at the population level.
|
neuroscience, representation learning, population responses, cortical hierarchy, computational biology, visual perception
| null | 11,170
| null |
[
-0.07402868568897247,
-0.15268611907958984,
0.05837438255548477,
0.02661052905023098,
-0.011529163457453251,
0.03297920897603035,
-0.052613843232393265,
-0.010414990596473217,
0.030467180535197258,
-0.027988649904727936,
-0.06429214030504227,
-0.10600782930850983,
-0.020351959392428398,
0.022149890661239624,
-0.030469968914985657,
-0.0535229854285717,
0.008561457507312298,
0.11162494122982025,
-0.08501389622688293,
-0.016180269420146942,
0.07638003677129745,
-0.046887245029211044,
0.07042005658149719,
0.039788082242012024,
0.02181958220899105,
0.03027794510126114,
-0.010575252585113049,
0.0015839807456359267,
-0.036868929862976074,
-0.03991296514868736,
0.09650609642267227,
-0.03369089215993881,
-0.049467869102954865,
0.061157915741205215,
-0.040274735540151596,
0.03858625888824463,
0.024453451856970787,
0.03201596438884735,
0.010022718459367752,
-0.07489418238401413,
0.03486637771129608,
0.05942065268754959,
0.04729073867201805,
0.017041852697730064,
0.05219631642103195,
-0.011129843071103096,
0.044281259179115295,
-0.01649312674999237,
0.001894232234917581,
-0.011806235648691654,
-0.006332531571388245,
-0.0259622223675251,
-0.023247648030519485,
0.06151651218533516,
0.06934836506843567,
0.1579105705022812,
0.026175402104854584,
0.0030267355032265186,
-0.015887096524238586,
0.03761762008070946,
-0.0017631486989557743,
-0.026042195037007332,
-0.01969192922115326,
-0.0009179964545182884,
0.03825368359684944,
-0.020053070038557053,
-0.05511001497507095,
-0.0545814111828804,
-0.07214366644620895,
0.013563510961830616,
0.002271991688758135,
0.039536669850349426,
-0.010059240274131298,
0.02465876005589962,
0.0464593879878521,
-0.04425865039229393,
0.017073635011911392,
0.03994951769709587,
0.045115720480680466,
-0.09552029520273209,
-0.0030807103030383587,
-0.020024169236421585,
-0.02160324528813362,
0.01762622781097889,
0.07136701792478561,
0.04053393006324768,
-0.06780291348695755,
0.06602358818054199,
0.02495516464114189,
-0.04706807807087898,
-0.05405309051275253,
0.00792049989104271,
0.01171998493373394,
0.0030058552511036396,
0.01721033826470375,
0.01799994707107544,
0.052406392991542816,
-0.029895547777414322,
0.0035777941811829805,
0.02344648540019989,
-0.03100619651377201,
0.022398030385375023,
0.143020361661911,
0.01911373808979988,
0.04842117428779602,
-0.021853048354387283,
0.05812910571694374,
0.027241937816143036,
0.022810138761997223,
-0.07883317768573761,
0.030068043619394302,
-0.019298678264021873,
0.024850480258464813,
0.15859927237033844,
0.0071308910846710205,
-0.021020717918872833,
0.02515086531639099,
-0.015639817342162132,
0.09530941396951675,
0.03163423761725426,
-0.020923476666212082,
0.023049162700772285,
-0.021159103140234947,
0.014778879471123219,
-0.006797904148697853,
-0.0716983750462532,
-0.07868026942014694,
1.3532966875917897e-33,
-0.03589148074388504,
0.019078049808740616,
0.07973940670490265,
0.07285343110561371,
0.053500138223171234,
-0.05045754462480545,
-0.013163546100258827,
-0.04298795014619827,
0.019028257578611374,
0.06865932792425156,
-0.07867976278066635,
-0.015215267427265644,
-0.02396140620112419,
0.13081903755664825,
-0.020368902012705803,
0.034825243055820465,
-0.035494279116392136,
0.0599554143846035,
0.009620984084904194,
0.0058533139526844025,
-0.001138158142566681,
0.017858276143670082,
-0.02185039035975933,
-0.02691921964287758,
0.006532539613544941,
-0.02612595073878765,
-0.0155451325699687,
-0.006483608856797218,
-0.014592474326491356,
0.002925568027421832,
-0.04864519089460373,
0.0555560439825058,
-0.017702624201774597,
0.01062772423028946,
-0.013571515679359436,
-0.021329646930098534,
0.057014040648937225,
-0.040914878249168396,
-0.04013878107070923,
0.04042927548289299,
-0.031115587800741196,
0.02595951221883297,
0.006128388922661543,
-0.07525704801082611,
-0.04770747944712639,
-0.048484161496162415,
0.05659700185060501,
0.037699177861213684,
-0.07360998541116714,
-0.03745824471116066,
0.04211485758423805,
0.034080684185028076,
-0.026586292311549187,
-0.04627887159585953,
-0.05422036722302437,
0.03657566383481026,
-0.038615092635154724,
-0.022039633244276047,
-0.01159018836915493,
0.0035295269917696714,
0.030991651117801666,
-0.022240860387682915,
-0.005431764759123325,
0.021789325401186943,
0.036224763840436935,
0.024917960166931152,
0.044794630259275436,
0.06596981734037399,
0.08339016139507294,
0.06057148426771164,
-0.010170405730605125,
0.032734934240579605,
-0.07239469885826111,
-0.1137673631310463,
0.04993109777569771,
-0.023856254294514656,
0.08671873062849045,
-0.04484570771455765,
-0.05163862183690071,
0.02254456654191017,
-0.0006469369400292635,
-0.038515083491802216,
-0.0842902809381485,
0.007846279069781303,
-0.03632054850459099,
-0.05238179489970207,
0.06726078689098358,
-0.03288584202528,
-0.09538210183382034,
0.04290610924363136,
-0.013389994390308857,
-0.05461473390460014,
-0.015936505049467087,
-0.1291794329881668,
0.010414554737508297,
-2.2001117976646266e-33,
-0.10751383751630783,
-0.018944144248962402,
-0.024092117324471474,
-0.025843894109129906,
-0.08261232823133469,
-0.014928258955478668,
-0.00592750683426857,
-0.024526162073016167,
-0.02544652484357357,
-0.07498019933700562,
-0.01891809143126011,
0.0001755587145453319,
0.005460989195853472,
-0.013352302834391594,
-0.007190515752881765,
-0.03970539569854736,
-0.016366958618164062,
0.06615938991308212,
0.041879355907440186,
-0.05530577525496483,
0.03571092337369919,
0.09139036387205124,
-0.08493392169475555,
-0.0011997739784419537,
-0.009653962217271328,
0.06594228744506836,
-0.07475276291370392,
0.04355187714099884,
0.03074723295867443,
-0.008797486312687397,
-0.03637711703777313,
0.019122911617159843,
-0.018887959420681,
-0.049901414662599564,
0.0474124401807785,
0.06652447581291199,
0.030494175851345062,
-0.056779880076646805,
-0.03973422199487686,
-0.00034780980786308646,
-0.000044807700760429725,
-0.03820003569126129,
-0.06628978252410889,
0.055359914898872375,
0.026182040572166443,
-0.013896127231419086,
0.010037411004304886,
0.029293714091181755,
-0.07066995650529861,
0.05326896905899048,
-0.021233733743429184,
0.08833032101392746,
-0.146114781498909,
-0.014365156181156635,
-0.011838830076158047,
-0.022049661725759506,
0.04044371098279953,
-0.03914544731378555,
0.1287279725074768,
0.0015362657140940428,
-0.055019889026880264,
-0.08998602628707886,
-0.12005457282066345,
0.06808695197105408,
-0.04047418758273125,
0.017289943993091583,
-0.025773556903004646,
0.01187962293624878,
0.049859076738357544,
0.07094776630401611,
0.0472508929669857,
-0.0031268056482076645,
0.061045654118061066,
-0.03324002027511597,
-0.0018450457137078047,
-0.06949236243963242,
0.0018146849470213056,
0.03788989782333374,
-0.0010901519563049078,
0.027489593252539635,
0.03920194134116173,
-0.03188067674636841,
-0.013062541373074055,
0.055511754006147385,
0.09809306263923645,
0.0038037430495023727,
0.05270087346434593,
0.0778757855296135,
-0.0022431870456784964,
-0.060999978333711624,
-0.053220346570014954,
0.018364181742072105,
-0.13599316775798798,
0.010146998800337315,
-0.10840784758329391,
-3.899600287127214e-8,
-0.045895785093307495,
0.07682693749666214,
0.053169894963502884,
0.038843948394060135,
-0.020021609961986542,
0.014316430315375328,
-0.009108268655836582,
0.08592846989631653,
-0.11138156056404114,
-0.04643062874674797,
0.004076278768479824,
-0.012296652421355247,
0.06357107311487198,
-0.009428465738892555,
0.086168572306633,
0.06229972839355469,
-0.03214612230658531,
-0.010670743882656097,
-0.041722506284713745,
0.1085602194070816,
-0.025156009942293167,
0.046741072088479996,
0.026592912152409554,
-0.012776168994605541,
0.048063721507787704,
-0.042070988565683365,
-0.026368174701929092,
0.09763900190591812,
0.04237916320562363,
-0.029449813067913055,
0.03950251266360283,
0.0051381574012339115,
0.0443943552672863,
0.015126812271773815,
0.023624736815690994,
-0.0066565596498548985,
0.008058838546276093,
0.05071991682052612,
-0.025851871818304062,
-0.07243935018777847,
-0.06396409869194031,
0.04785146936774254,
0.040051963180303574,
0.07293682545423508,
0.08114270865917206,
0.020833292976021767,
0.08416545391082764,
-0.0670829564332962,
-0.010727099142968655,
-0.02200235426425934,
-0.10595511645078659,
0.05663582682609558,
0.030905375257134438,
0.06805426627397537,
-0.045992884784936905,
-0.017479833215475082,
-0.0660252571105957,
-0.046427447348833084,
-0.007335820700973272,
0.12964153289794922,
0.024031883105635643,
0.07990056276321411,
-0.07389616966247559,
-0.038805846124887466
] | 0
| 0
| 0
| 0
|
|
State Space Models are Provably Comparable to Transformers in Dynamic Token Selection
|
https://openreview.net/forum?id=QFgbJOYJSE
|
[
"Naoki Nishikawa",
"Taiji Suzuki"
] |
Poster
|
Deep neural networks based on state space models (SSMs) are attracting significant attention in sequence modeling since their computational cost is much smaller than that of Transformers. While the capabilities of SSMs have been demonstrated through experiments in various tasks, theoretical understanding of SSMs is still limited. In particular, most theoretical studies discuss the capabilities of SSM layers without nonlinear layers, and there is a lack of discussion on their combination with nonlinear layers. In this paper, we explore the capabilities of SSMs combined with fully connected neural networks, and show that they are comparable to Transformers in extracting the essential tokens depending on the input. As concrete examples, we consider two synthetic tasks, which are challenging for a single SSM layer, and demonstrate that SSMs combined with nonlinear layers can efficiently solve these tasks. Furthermore, we study the nonparametric regression task, and prove that the ability of SSMs is equivalent to that of Transformers in estimating functions belonging to a certain class.
|
State Space Model, Transformer, Nonparametric regression
| null | 11,158
|
2405.19036
|
[
-0.15315477550029755,
-0.10616875439882278,
0.07385160773992538,
0.06089956685900688,
-0.0061090923845767975,
0.07947439700365067,
-0.06531909853219986,
-0.009956465102732182,
0.05970165133476257,
-0.057013437151908875,
-0.06015457957983017,
-0.023579804226756096,
-0.025532633066177368,
0.05550511181354523,
-0.009683393873274326,
0.02358110062777996,
0.05851178988814354,
0.05661993473768234,
-0.0627782791852951,
-0.028172286227345467,
0.07399231940507889,
0.007949160411953926,
-0.0055278027430176735,
-0.003900042036548257,
0.09457912296056747,
0.07554126530885696,
0.023624666035175323,
-0.02730594202876091,
-0.040339451283216476,
0.007918119430541992,
0.013576250523328781,
-0.04727410525083542,
-0.03425726667046547,
0.020885517820715904,
-0.058502499014139175,
0.0252873245626688,
-0.06994755566120148,
-0.04893200099468231,
0.019165022298693657,
-0.10526230186223984,
0.031765080988407135,
-0.034188613295555115,
0.025271635502576828,
0.0649036094546318,
0.07543172687292099,
0.005445521790534258,
0.04759078100323677,
-0.09278319776058197,
-0.033250030130147934,
0.01607390120625496,
-0.06591406464576721,
0.06854328513145447,
0.0033988668583333492,
0.1700744330883026,
-0.001365985837765038,
0.029895270243287086,
0.06917218118906021,
0.03874362260103226,
-0.01998469978570938,
-0.02825205586850643,
-0.0320303849875927,
-0.07870304584503174,
0.07694020867347717,
-0.04094236344099045,
-0.023379536345601082,
0.012733140029013157,
-0.016000624746084213,
0.1300935447216034,
-0.024981394410133362,
0.03374361991882324,
-0.01479762140661478,
0.1070333868265152,
-0.013162310235202312,
-0.0007111677550710738,
0.06681771576404572,
0.0016815861454233527,
0.10037364810705185,
0.007717193569988012,
0.08126261830329895,
-0.07518911361694336,
0.0286521278321743,
-0.006594525184482336,
0.013553408905863762,
0.02860807068645954,
0.08640364557504654,
0.022279808297753334,
-0.05925121530890465,
-0.0016896093729883432,
-0.0019753093365579844,
0.03152855485677719,
-0.021625204011797905,
-0.07753400504589081,
0.06270453333854675,
-0.0845075249671936,
0.0038069477304816246,
-0.00965094193816185,
-0.04047999158501625,
-0.05224945768713951,
-0.010398760437965393,
0.0654260516166687,
-0.0425446517765522,
0.03816297650337219,
0.06707421690225601,
-0.064718097448349,
0.03625287115573883,
0.016792280599474907,
0.006356451194733381,
0.02359229512512684,
0.06258880347013474,
-0.0946098044514656,
0.026099318638443947,
0.010807648301124573,
0.008072386495769024,
0.03637625277042389,
-0.003787448164075613,
-0.04993923380970955,
0.04138287156820297,
-0.08486535400152206,
-0.05554720759391785,
0.09750422835350037,
-0.017923057079315186,
-0.021789956837892532,
-0.07563307136297226,
0.08541777729988098,
-0.06639142334461212,
0.03902224078774452,
-0.04815085977315903,
2.5330810417816016e-33,
-0.04600280895829201,
-0.0017998337280005217,
0.06865585595369339,
-0.021347109228372574,
-0.02383442595601082,
-0.03365681692957878,
0.04027680307626724,
0.03589530661702156,
-0.01894608698785305,
0.05288190767168999,
-0.08318576961755753,
0.02418864332139492,
-0.04252764955163002,
0.08407795429229736,
0.07289083302021027,
-0.05256332829594612,
-0.014204513281583786,
0.023744914680719376,
0.010136404074728489,
-0.07131905108690262,
-0.0383530855178833,
0.0261029452085495,
-0.012326279655098915,
-0.08770745992660522,
0.013532075099647045,
0.01907367818057537,
0.07960929721593857,
-0.017906131222844124,
-0.0475446954369545,
-0.018018776550889015,
-0.07906010001897812,
0.040121085941791534,
-0.033800870180130005,
0.024550776928663254,
0.12022902071475983,
0.009793641977012157,
-0.013885182328522205,
-0.00415253359824419,
-0.020616408437490463,
-0.03729932755231857,
0.002315016696229577,
0.025647643953561783,
-0.04702598601579666,
0.06477192044258118,
-0.07086820900440216,
-0.08215418457984924,
-0.0032527840230613947,
-0.07001440972089767,
0.0010265980381518602,
-0.018671156838536263,
0.014270626939833164,
0.01606062427163124,
-0.07261735945940018,
-0.06313695013523102,
0.08060190826654434,
-0.02720240131020546,
0.03583667799830437,
0.06564591825008392,
0.038137540221214294,
0.05861775577068329,
-0.04702719673514366,
-0.02599780447781086,
-0.02820497378706932,
0.026258165016770363,
0.004828593693673611,
0.04212527349591255,
0.016342146322131157,
0.07389538735151291,
0.030408861115574837,
-0.07571908831596375,
0.03559451177716255,
0.028537198901176453,
-0.015281043015420437,
-0.03924622759222984,
0.06644944101572037,
-0.02311934158205986,
0.088894322514534,
-0.08289756625890732,
-0.1009608805179596,
0.038901373744010925,
0.019904671236872673,
-0.013343813829123974,
-0.03319161385297775,
0.003797387471422553,
0.004713583737611771,
0.019146716222167015,
0.03993064537644386,
-0.05803674831986427,
-0.01612676866352558,
-0.030943460762500763,
-0.04608702287077904,
-0.03060644492506981,
-0.03465203940868378,
0.03155366703867912,
-0.04634333774447441,
-3.0090596429226698e-33,
-0.023614637553691864,
0.07247845828533173,
-0.053352709859609604,
0.049479879438877106,
-0.0880359411239624,
-0.05262698978185654,
-0.025893107056617737,
0.023973366245627403,
-0.0624702014029026,
0.04708033427596092,
-0.018201207742094994,
-0.04817371442914009,
0.0533372201025486,
-0.03787430748343468,
0.06286962330341339,
0.04118313640356064,
-0.08332651853561401,
-0.04717116802930832,
0.041084203869104385,
-0.0035736458376049995,
0.041154101490974426,
0.07827834784984589,
-0.08010908216238022,
0.04775840789079666,
-0.09422224014997482,
0.04356798157095909,
-0.09938853979110718,
0.11745889484882355,
0.03215437009930611,
-0.0027264931704849005,
-0.035479046404361725,
-0.06540392339229584,
-0.05895332992076874,
0.002548235934227705,
-0.037041932344436646,
0.02944920025765896,
0.03377257660031319,
0.07643131166696548,
-0.02060048282146454,
-0.022554805502295494,
0.01490226574242115,
-0.07544051110744476,
-0.013578593730926514,
0.07968993484973907,
-0.028481094166636467,
-0.03189561888575554,
0.01283401157706976,
0.018943483009934425,
-0.028922205790877342,
0.02496705763041973,
-0.009287528693675995,
0.0540749728679657,
-0.163652703166008,
-0.021228907629847527,
-0.06048523634672165,
-0.03494330495595932,
-0.056924134492874146,
-0.007565150503069162,
0.07738130539655685,
-0.038496922701597214,
-0.0002099845587508753,
-0.03392474725842476,
0.06275749206542969,
-0.07464738190174103,
-0.04586423933506012,
-0.04579601436853409,
-0.019759580492973328,
0.02606598474085331,
0.02602272666990757,
0.0016742425505071878,
0.017762981355190277,
-0.00569163728505373,
0.06781404465436935,
0.006729517597705126,
-0.020178986713290215,
-0.03559516742825508,
0.020748361945152283,
-0.08416277170181274,
0.01354170497506857,
-0.026615209877490997,
-0.018894650042057037,
0.036342065781354904,
0.016112985089421272,
-0.0160675048828125,
0.033303286880254745,
0.13815326988697052,
0.07134859263896942,
0.049154091626405716,
0.028774835169315338,
-0.027232788503170013,
-0.022455718368291855,
0.06636762619018555,
-0.006111006252467632,
0.0427522137761116,
-0.014321268536150455,
-4.5813887794565744e-8,
-0.0277114138007164,
0.03969527408480644,
0.05319489166140556,
-0.023313648998737335,
0.03804487735033035,
-0.054777842015028,
0.03384682536125183,
0.04898469150066376,
-0.009910640306770802,
0.0019241111585870385,
0.02538098767399788,
0.0027320212684571743,
-0.047671180218458176,
-0.08587933331727982,
-0.04411698877811432,
0.03583400323987007,
0.03645358979701996,
0.03368142619729042,
0.00698723504319787,
0.003905899589881301,
0.0342361219227314,
0.07867944985628128,
-0.026628781110048294,
0.020912718027830124,
0.056643351912498474,
-0.06659755110740662,
-0.03813963383436203,
0.022010615095496178,
0.0170509722083807,
0.04316651076078415,
0.03273306041955948,
0.03887903690338135,
-0.005331652704626322,
-0.05252043157815933,
0.0012601276393979788,
0.04180217161774635,
0.060368962585926056,
-0.042199842631816864,
-0.00552370585501194,
-0.024310987442731857,
-0.03351791948080063,
0.05721859261393547,
-0.07447738200426102,
0.039316240698099136,
-0.0619531013071537,
-0.0017189756035804749,
0.05328894779086113,
-0.13104002177715302,
0.011682292446494102,
-0.023845065385103226,
0.058550845831632614,
-0.06292323768138885,
-0.06415130943059921,
0.05023684725165367,
0.041875842958688736,
0.042481694370508194,
0.06074536591768265,
-0.0370369628071785,
-0.0120506277307868,
0.11191978305578232,
-0.021708248183131218,
0.015337047167122364,
-0.06433267146348953,
0.012456431984901428
] | 0
| 0
| 0
| 0
|
|
Consistency Models Made Easy
|
https://openreview.net/forum?id=xQVxo9dSID
|
[
"Zhengyang Geng",
"Ashwini Pokle",
"Weijian Luo",
"Justin Lin",
"J Zico Kolter"
] |
Poster
|
Consistency models (CMs) offer faster sampling than traditional diffusion models, but their training is resource-intensive. For example, as of 2024, training a state-of-the-art CM on CIFAR-10 takes one week on 8 GPUs. In this work, we propose an effective scheme for training CMs that largely improves the efficiency of building such models. Specifically, by expressing CM trajectories via a particular differential equation, we argue that diffusion models can be viewed as a special case of CMs. We can thus fine-tune a consistency model starting from a pretrained diffusion model and progressively approximate the full consistency condition to stronger degrees over the training process. Our resulting method, which we term Easy Consistency Tuning (ECT), achieves vastly reduced training times while improving upon the quality of previous methods: for example, ECT achieves a 2-step FID of 2.73 on CIFAR10 within 1 hour on a single A100 GPU, matching Consistency Distillation trained for hundreds of GPU hours. Owing to this computational efficiency, we investigate the scaling laws of CMs under ECT, showing that they obey the classic power law scaling, hinting at their ability to improve efficiency and performance at larger scales. Our [code](https://github.com/locuslab/ect) is available.
|
Consistency Models, Efficient Generative Models, Diffusion Models
|
Consistency Models Made Easy
| 11,157
|
2406.14548
|
[
-0.09460057318210602,
-0.06650859862565994,
0.00859422143548727,
0.10524976998567581,
0.023386070504784584,
-0.09543876349925995,
-0.07905957847833633,
-0.015499000437557697,
-0.02301289513707161,
-0.1015646755695343,
-0.04779822379350662,
-0.002351396484300494,
-0.005635905545204878,
0.026715833693742752,
-0.019510669633746147,
-0.09977971017360687,
0.05015439912676811,
0.08819115161895752,
-0.1251552700996399,
-0.02748280204832554,
-0.07829324901103973,
0.019756576046347618,
-0.04253487288951874,
0.023445971310138702,
0.00564557034522295,
0.0308931116014719,
0.06270638853311539,
-0.07513059675693512,
0.04603363573551178,
-0.06134791672229767,
0.028992868959903717,
-0.07381134480237961,
-0.07041606307029724,
0.02582336775958538,
-0.024389300495386124,
0.012375989928841591,
-0.036383118480443954,
-0.03792017698287964,
0.040347158908843994,
0.020748596638441086,
0.035274166613817215,
0.009722170419991016,
0.008387030102312565,
0.05911184847354889,
0.06676121056079865,
-0.05074646696448326,
0.029223380610346794,
0.030508248135447502,
-0.025987310335040092,
-0.05147784575819969,
0.01618211157619953,
-0.06675054877996445,
-0.045067258179187775,
0.023741349577903748,
-0.016154976561665535,
0.010266914032399654,
0.08250196278095245,
0.030833717435598373,
0.011080154217779636,
-0.03327861800789833,
-0.03695480525493622,
-0.07210106402635574,
-0.10461629182100296,
0.009808458387851715,
-0.05117904394865036,
-0.0252547450363636,
0.10216305404901505,
0.041406504809856415,
0.06486594676971436,
0.02846822515130043,
-0.027277490124106407,
0.08325719088315964,
-0.09198549389839172,
0.006886949762701988,
-0.002010084455832839,
0.034533366560935974,
0.04184691980481148,
0.008393113501369953,
0.003755380166694522,
0.013373753987252712,
-0.016527650877833366,
-0.03183531016111374,
0.049466345459222794,
-0.07948838174343109,
0.08726231008768082,
-0.06027663126587868,
0.03433108329772949,
0.038505733013153076,
0.0700235366821289,
-0.05238829553127289,
0.00913742184638977,
0.1566762775182724,
-0.0272745992988348,
-0.03439946100115776,
0.04090271517634392,
0.06271364539861679,
0.0501846969127655,
-0.017916623502969742,
0.011964392848312855,
0.11240893602371216,
0.009453119710087776,
0.09324493259191513,
0.0049210116267204285,
0.03533431142568588,
0.0351424515247345,
-0.011926545761525631,
0.038827598094940186,
0.1015685498714447,
0.023524444550275803,
-0.0251897145062685,
0.062290456146001816,
0.02667211927473545,
-0.042813751846551895,
0.03220021352171898,
0.051581282168626785,
-0.026916295289993286,
-0.023622330278158188,
-0.047031719237565994,
0.024358689785003662,
0.0685303658246994,
-0.052826423197984695,
0.013991733081638813,
-0.031312279403209686,
-0.04298712685704231,
0.001552209840156138,
-0.00471770903095603,
-0.031352315098047256,
4.529797602936425e-33,
-0.03406062722206116,
0.005956354085355997,
0.01055553462356329,
-0.049991242587566376,
0.009196404367685318,
0.00901192706078291,
-0.014389947056770325,
-0.012140176258981228,
-0.001974984537810087,
-0.011896111071109772,
-0.020861245691776276,
0.029960155487060547,
-0.0935375913977623,
0.06370794773101807,
0.021904565393924713,
-0.04151273891329765,
-0.005016912706196308,
0.09365154057741165,
0.006772154476493597,
-0.01543718297034502,
0.1434360146522522,
-0.03446138650178909,
-0.019915521144866943,
-0.02315390855073929,
-0.0207064189016819,
0.02862587757408619,
0.05425770580768585,
0.04800119623541832,
0.02680099941790104,
0.03150135278701782,
-0.023375028744339943,
-0.019709505140781403,
-0.01006869412958622,
0.008949851617217064,
0.024756524711847305,
-0.007653042208403349,
0.03674965351819992,
-0.0517086423933506,
0.019687624648213387,
-0.06890328228473663,
-0.02813904359936714,
0.07369764894247055,
0.02447495236992836,
-0.10238714516162872,
-0.12260734289884567,
0.025875894352793694,
0.03071175143122673,
-0.061845164746046066,
-0.04187827184796333,
-0.049889713525772095,
0.00990708451718092,
0.05393815413117409,
0.00018751778407022357,
-0.08802998065948486,
0.0004834699793718755,
-0.03273230791091919,
0.04917050153017044,
0.012485179118812084,
0.03776777535676956,
0.030094709247350693,
0.07052914053201675,
0.006355859339237213,
-0.07747915387153625,
0.08925744146108627,
0.027548179030418396,
0.03225089982151985,
-0.05938894674181938,
0.08217521756887436,
0.0006468530045822263,
-0.0040921238251030445,
-0.05155947804450989,
-0.03026101179420948,
-0.04873334988951683,
-0.13379240036010742,
0.1265670359134674,
-0.11242230236530304,
0.06036848947405815,
-0.05959917604923248,
-0.04904855042695999,
-0.015796612948179245,
-0.12955833971500397,
0.011747329495847225,
-0.06748980283737183,
-0.07349894195795059,
-0.03978031128644943,
-0.04907689988613129,
-0.06462844461202621,
0.027964729815721512,
-0.05051449313759804,
-0.06699961423873901,
-0.031507305800914764,
0.026790276169776917,
0.07341837882995605,
0.006384086795151234,
0.013976039364933968,
-4.70407823209926e-33,
0.013271572068333626,
0.011862783692777157,
0.058035336434841156,
0.16790175437927246,
0.008600865490734577,
0.0018327628495171666,
-0.01672968454658985,
0.04488438740372658,
0.052467674016952515,
-0.0888960137963295,
-0.02700560726225376,
-0.006759390234947205,
-0.010813434608280659,
-0.03323064371943474,
0.01015639677643776,
0.0013892113929614425,
-0.04269814118742943,
-0.010264824144542217,
-0.018877509981393814,
0.04011128842830658,
0.06333182752132416,
-0.005823811050504446,
-0.05721670016646385,
-0.011631110683083534,
-0.008448426611721516,
-0.009487752802670002,
-0.0996711403131485,
0.04450186714529991,
0.03374350816011429,
0.008269807323813438,
-0.017726872116327286,
0.016183974221348763,
0.001340443966910243,
-0.07379081100225449,
-0.06786920130252838,
0.047118689864873886,
0.038814712315797806,
0.033917129039764404,
0.022072643041610718,
0.10440439730882645,
0.0534980446100235,
-0.04722261428833008,
-0.05265036225318909,
0.041027531027793884,
0.027483414858579636,
0.060775186866521835,
-0.0405203141272068,
-0.035066936165094376,
-0.049667712301015854,
0.03159438818693161,
0.04520479962229729,
0.041842784732580185,
-0.05608527362346649,
0.04782567173242569,
-0.017139988020062447,
-0.015910794958472252,
-0.04853486269712448,
0.01710355281829834,
-0.06810121983289719,
-0.04196937754750252,
-0.03455677255988121,
-0.037361256778240204,
-0.025154337286949158,
-0.018819153308868408,
-0.049269575625658035,
-0.024330811575055122,
-0.02866552770137787,
0.04491224139928818,
0.03987780958414078,
0.09494653344154358,
-0.0771825984120369,
0.011201772838830948,
-0.015408581122756004,
0.011881907470524311,
-0.03561098128557205,
-0.055703435093164444,
0.07490929961204529,
-0.1487821340560913,
0.014502348378300667,
0.00805720780044794,
-0.000536348030436784,
0.0763433575630188,
0.010000301524996758,
-0.019709311425685883,
0.044160112738609314,
0.011126840487122536,
0.06341113895177841,
-0.008908122777938843,
0.0436866469681263,
0.07034676522016525,
-0.05408577248454094,
-0.014040345326066017,
0.09129813313484192,
0.05425010994076729,
-0.12560108304023743,
-5.230682731394154e-8,
-0.0958142951130867,
0.028017811477184296,
0.03148568049073219,
0.11371053755283356,
0.008240943774580956,
-0.011358936317265034,
0.017415491864085197,
0.025170257315039635,
0.0034049900714308023,
-0.022881168872117996,
0.12050309032201767,
-0.04129716381430626,
0.008062049746513367,
-0.03385603800415993,
0.022080477327108383,
0.06887652724981308,
-0.0006142266793176532,
0.006172800902277231,
-0.006667987909168005,
-0.01371628325432539,
-0.06583603471517563,
0.054447613656520844,
0.04340263828635216,
-0.011434294283390045,
0.08112736791372299,
0.010395062156021595,
-0.014249620959162712,
0.09592273831367493,
0.0654703676700592,
0.042594630271196365,
-0.007289393339306116,
-0.04091896861791611,
0.043504174798727036,
0.001775938319042325,
-0.04196961969137192,
0.04616842046380043,
0.026780014857649803,
-0.021033961325883865,
0.012524641118943691,
-0.077541284263134,
0.04829567298293114,
0.03276243433356285,
-0.013474905863404274,
0.01814570650458336,
-0.009935677982866764,
0.023469844833016396,
-0.04443987458944321,
-0.029756277799606323,
-0.012387673370540142,
0.006082651671022177,
-0.025809690356254578,
0.012733113020658493,
-0.06546162068843842,
0.036987774074077606,
0.09488488733768463,
0.000053304243920138106,
-0.03325004130601883,
-0.047764141112565994,
-0.022189011797308922,
-0.005620159208774567,
0.031517039984464645,
-0.01907357946038246,
-0.01565811224281788,
0.03429035842418671
] |
https://github.com/locuslab/ect
| 277
| 0
| 0
| 0
|
OpenMathInstruct-2: Accelerating AI for Math with Massive Open-Source Instruction Data
|
https://openreview.net/forum?id=mTCbq2QssD
|
[
"Shubham Toshniwal",
"Wei Du",
"Ivan Moshkov",
"Branislav Kisacanin",
"Alexan Ayrapetyan",
"Igor Gitman"
] |
Poster
|
Mathematical reasoning continues to be a critical challenge in large language model (LLM) development with significant interest. However, most of the cutting-edge progress in mathematical reasoning with LLMs has become closed-source due to lack of access to training data. This lack of data access limits researchers from understanding the impact of different choices for synthesizing and utilizing the data. With the goal of creating a high-quality finetuning (SFT) dataset for math reasoning, we conduct careful ablation experiments on data synthesis using the recently released Llama3.1 family of models. Our experiments show that: (a) solution format matters, with excessively verbose solutions proving detrimental to SFT performance, (b) data generated by a strong teacher outperforms on-policy data generated by a weak student model, (c) SFT is robust to low-quality solutions, allowing for imprecise data filtering, and (d) question diversity is crucial for achieving data scaling gains. Based on these insights, we create the OpenMathInstruct-2 dataset which consists of 14M question-solution pairs (≈ 600K unique questions), making it nearly eight times larger than the previous largest open-source math reasoning dataset. Finetuning the Llama-3.1-8B-Base using OpenMathInstruct-2 outperforms Llama3.1-8B-Instruct on MATH by an absolute 15.9% (51.9% → 67.8%). Finally, to accelerate the open-source efforts, we release the code, the finetuned models, and the OpenMathInstruct-2 dataset under a commercially permissive license.
|
Math Reasoning, Synthetic Data
|
We create a massive, high-quality math instruction data to support open-source efforts on math reasoning.
| 11,155
| null |
[
-0.01912691257894039,
-0.07356803119182587,
0.014801381155848503,
0.03603972867131233,
0.0457826666533947,
-0.04344674199819565,
-0.040108080953359604,
0.015413054265081882,
0.021480882540345192,
0.01651669852435589,
-0.024056943133473396,
-0.07915744185447693,
0.02722679078578949,
0.013843270018696785,
0.059084489941596985,
0.11441372334957123,
0.013387882150709629,
0.028670623898506165,
-0.11728453636169434,
-0.07066874951124191,
0.040038589388132095,
0.012401944026350975,
0.048767805099487305,
-0.01749098114669323,
0.06609943509101868,
0.047716300934553146,
0.0799439400434494,
-0.09549716114997864,
0.0397401787340641,
-0.008092006668448448,
-0.041451964527368546,
0.008593980222940445,
-0.022790107876062393,
0.0034822700545191765,
0.001283019781112671,
0.04413093253970146,
-0.023640550673007965,
0.06672725826501846,
-0.03655839338898659,
-0.0028193278703838587,
-0.03131658956408501,
0.028973842039704323,
0.09239941835403442,
-0.004134461283683777,
0.0328264981508255,
-0.040327176451683044,
-0.01662163808941841,
-0.07064566761255264,
0.005681910086423159,
0.03959919884800911,
-0.1277809888124466,
-0.045196715742349625,
-0.07291433215141296,
0.03636777400970459,
0.046835482120513916,
-0.0672919899225235,
-0.019799817353487015,
0.02055567316710949,
0.0034997127950191498,
-0.040809717029333115,
-0.10137960314750671,
-0.08575083315372467,
-0.025633912533521652,
-0.037159841507673264,
-0.011765473522245884,
-0.021136410534381866,
-0.011623872444033623,
0.05943355709314346,
-0.02988327108323574,
0.03376910835504532,
-0.050978656858205795,
0.1109616607427597,
-0.0014205977786332369,
0.06482041627168655,
0.027469871565699577,
0.06698516756296158,
0.02318851463496685,
-0.007739913649857044,
0.07754257321357727,
-0.025685634464025497,
0.008780327625572681,
-0.06283886730670929,
0.0001400475448463112,
0.041814520955085754,
0.07822827994823456,
-0.0002047245216090232,
0.03710193559527397,
0.05072135850787163,
0.017639394849538803,
-0.0015972391702234745,
-0.0057225702330470085,
-0.012236419133841991,
0.030128609389066696,
0.05865965783596039,
0.060820095241069794,
0.09103399515151978,
0.01592731662094593,
-0.029386717826128006,
0.016266297549009323,
0.08865942060947418,
-0.001413044985383749,
0.13138267397880554,
0.009065378457307816,
0.0030993130058050156,
-0.06122453883290291,
-0.03346555680036545,
0.08516841381788254,
0.03349214419722557,
0.010384796187281609,
-0.08814144134521484,
0.07569263130426407,
-0.007908298633992672,
0.027558084577322006,
0.013635645620524883,
-0.0071492427960038185,
0.0019612032920122147,
0.014797894284129143,
-0.06585665792226791,
-0.013940979726612568,
0.015615594573318958,
-0.01080563198775053,
0.05299532040953636,
0.035312265157699585,
0.014679068699479103,
0.044249046593904495,
-0.0011316481977701187,
-0.0999937579035759,
3.75793862482774e-33,
0.11782795935869217,
0.08470321446657181,
0.04799475893378258,
0.04742640256881714,
0.07081069052219391,
-0.014375437051057816,
-0.009913967922329903,
0.09812464565038681,
-0.05522006005048752,
0.039340898394584656,
-0.018851066008210182,
-0.0029274020344018936,
-0.08609183132648468,
0.03914141654968262,
0.0850415751338005,
-0.030101826414465904,
-0.03199908882379532,
0.007707112934440374,
-0.08650734275579453,
-0.05445311218500137,
0.056677840650081635,
-0.06131451204419136,
0.013620151206851006,
-0.008147673681378365,
0.0019496482564136386,
0.039973095059394836,
0.061454061418771744,
-0.03671194985508919,
0.0045712958090007305,
-0.01802186109125614,
-0.09213876724243164,
-0.01306453812867403,
-0.07920774072408676,
0.10524896532297134,
0.025200940668582916,
-0.027311179786920547,
-0.011613592505455017,
-0.0962471067905426,
0.00988085102289915,
-0.005267631262540817,
0.0405057817697525,
0.04674806818366051,
0.05944375321269035,
0.05827765539288521,
0.010318703018128872,
-0.06251394748687744,
0.014194304123520851,
0.01630561612546444,
0.06209123507142067,
-0.025418885052204132,
0.01679723709821701,
0.040728628635406494,
-0.0698888897895813,
-0.05504326894879341,
-0.024602506309747696,
-0.015402759425342083,
-0.015092657878994942,
0.009504481218755245,
0.03189246729016304,
0.08816254138946533,
-0.03704648092389107,
0.04188493639230728,
0.03324528783559799,
0.0658419206738472,
-0.03373071923851967,
0.10681353509426117,
-0.11788693070411682,
0.05802535265684128,
0.13778017461299896,
0.0017386075342074037,
-0.007462318055331707,
-0.029449211433529854,
0.023811129853129387,
-0.035318341106176376,
0.018240438774228096,
0.002648205030709505,
0.025810077786445618,
-0.07293781638145447,
0.05435171350836754,
-0.0726570263504982,
0.07712497562170029,
0.008443121798336506,
-0.0019508698023855686,
-0.05035960301756859,
-0.08115863054990768,
-0.07971455901861191,
0.030801426619291306,
-0.0544947050511837,
-0.008248894475400448,
-0.0443347692489624,
-0.028768423944711685,
-0.03758102282881737,
-0.06113926321268082,
-0.03662664443254471,
-0.019697772338986397,
-2.8372385520179376e-33,
-0.07246079295873642,
-0.03287911042571068,
-0.0921952873468399,
0.04402195289731026,
-0.043304841965436935,
-0.0728222206234932,
0.008086027577519417,
0.02525174617767334,
0.0671091079711914,
-0.08414547145366669,
-0.011581849306821823,
-0.032507337629795074,
0.05856982618570328,
-0.03910573571920395,
-0.013020757585763931,
-0.07894604653120041,
-0.021145863458514214,
-0.011710369028151035,
-0.04194864258170128,
0.046881671994924545,
-0.0727354884147644,
0.11206142604351044,
-0.07703358680009842,
-0.043756961822509766,
-0.03817611560225487,
0.046906378120183945,
-0.031209755688905716,
0.008235066197812557,
-0.022930976003408432,
0.0820065438747406,
-0.04082946106791496,
-0.08739077299833298,
-0.04730663821101189,
-0.047254305332899094,
-0.11066138744354248,
0.03801239654421806,
0.07716599106788635,
-0.012881631031632423,
-0.044098757207393646,
0.02880670130252838,
-0.015593133866786957,
-0.04241255670785904,
-0.06789694726467133,
0.02406456507742405,
-0.011039103381335735,
0.058571506291627884,
-0.07802997529506683,
0.0018823501886799932,
0.04478578269481659,
0.019707033410668373,
0.01867254078388214,
-0.05930094048380852,
0.03584880381822586,
-0.03400542214512825,
0.03563646227121353,
-0.04596070945262909,
-0.02957252226769924,
-0.02824930101633072,
-0.046986375004053116,
0.030633332207798958,
-0.0610806941986084,
-0.015055611729621887,
0.08165622502565384,
-0.04835481941699982,
0.025191739201545715,
-0.033220674842596054,
-0.08502040058374405,
0.031886760145425797,
0.004431867506355047,
0.015372025780379772,
-0.0640314593911171,
-0.06917595118284225,
0.029051652178168297,
0.037123069167137146,
0.018984084948897362,
0.010342096909880638,
0.017863180488348007,
-0.024558374658226967,
-0.011151023209095001,
0.0001258695119759068,
0.023238398134708405,
-0.035698600113391876,
0.01842287741601467,
0.06512948125600815,
0.0023160683922469616,
0.005061246454715729,
0.09840346872806549,
0.10834003239870071,
0.032171495258808136,
0.08488922566175461,
-0.09580640494823456,
0.04022004455327988,
-0.009155385196208954,
0.06093355268239975,
-0.0405738539993763,
-4.669910325105775e-8,
-0.03984798118472099,
0.0840563178062439,
-0.004327255766838789,
0.06537497043609619,
0.041140373796224594,
-0.021770931780338287,
-0.08514706045389175,
0.05915466323494911,
0.002075520344078541,
0.042567506432533264,
0.046127382665872574,
0.025529811158776283,
-0.05320214852690697,
-0.05307600647211075,
-0.006928787566721439,
0.1292552649974823,
0.04384029656648636,
-0.050745755434036255,
-0.00802256166934967,
-0.05663701891899109,
-0.0027103391475975513,
0.03580404818058014,
-0.024899499490857124,
0.04782896488904953,
0.05252254754304886,
-0.02896955795586109,
-0.05856238305568695,
0.147057443857193,
0.005626159254461527,
0.04747937619686127,
-0.01918851025402546,
-0.07571923732757568,
0.03700637444853783,
-0.07537278532981873,
0.028237123042345047,
0.026711339130997658,
0.016799403354525566,
0.00788975041359663,
0.007091914769262075,
-0.034017983824014664,
0.050393253564834595,
0.04366857558488846,
-0.0536305196583271,
0.021676789969205856,
0.02545025758445263,
-0.0029629082418978214,
-0.09233221411705017,
-0.05875925347208977,
-0.015542244538664818,
0.026914801448583603,
-0.009718157351016998,
-0.02630954049527645,
-0.04152347892522812,
0.016533391550183296,
0.07938635349273682,
0.0076362369582057,
-0.09298164397478104,
0.009899621829390526,
-0.016583889722824097,
0.006667933892458677,
0.053210530430078506,
0.02334686741232872,
-0.044178079813718796,
0.04465452954173088
] | 0
| 0
| 0
| 0
|
|
MAPS: Advancing Multi-Modal Reasoning in Expert-Level Physical Science
|
https://openreview.net/forum?id=GR0y0F3Ipd
|
[
"Erle Zhu",
"Yadi Liu",
"Zhe Zhang",
"Xujun Li",
"JinZhou",
"Xinjie Yu",
"Minlie Huang",
"Hongning Wang"
] |
Poster
|
Pre-trained on extensive text and image corpora, current Multi-Modal Large Language Models (MLLM) have shown strong capabilities in general visual reasoning tasks.
However, their performance is still lacking in physical domains that require understanding diagrams with complex physical structures and quantitative analysis based on multi-modal information.
To address this, we develop a new framework, named **M**ulti-Modal Scientific Re**A**soning with **P**hysics Perception and **S**imulation (**MAPS**) based on an MLLM.
MAPS decomposes expert-level multi-modal reasoning task into physical diagram understanding via a Physical Perception Model (PPM) and reasoning with physical knowledge via a simulator.
The PPM module is obtained by fine-tuning a visual language model using carefully designed synthetic data with paired physical diagrams and corresponding simulation language descriptions.
At the inference stage, MAPS integrates the simulation language description of the input diagram provided by PPM and results obtained through a Chain-of-Simulation process with MLLM to derive the underlying rationale and the final answer.
Validated using our collected college-level circuit analysis problems, MAPS significantly improves reasoning accuracy of MLLM and outperforms all existing models.
The results confirm MAPS offers a promising direction for enhancing multi-modal scientific reasoning ability of MLLMs.
We will release our code, model and dataset used for our experiments upon publishing of this paper.
|
multi-modal reasoning, scientific reasoning, physical simulation
|
improving multi-modal scientific reasoning capability with physics perception model and simulation assistance
| 11,154
|
2501.10768
|
[
0.07016774266958237,
-0.10268090665340424,
0.05814119428396225,
0.019180001690983772,
0.02025211602449417,
-0.08408497273921967,
0.015404412522912025,
0.04016266018152237,
-0.009808110073208809,
0.019715050235390663,
0.057148661464452744,
-0.10171081870794296,
0.06405080854892731,
0.10267655551433563,
0.019499260932207108,
0.056725770235061646,
0.013187980279326439,
0.056691933423280716,
-0.10102326422929764,
-0.04691638424992561,
0.039640553295612335,
0.0014318068278953433,
0.03981313109397888,
-0.01747829280793667,
-0.017680268734693527,
0.051755137741565704,
-0.0019908712711185217,
-0.0014670583186671138,
-0.01429009810090065,
-0.03336950019001961,
0.03585439920425415,
0.00700994161888957,
0.006337256170809269,
0.006941633298993111,
-0.012432231567800045,
0.11566152423620224,
-0.04138442501425743,
-0.030849752947688103,
-0.041361868381500244,
-0.06430323421955109,
-0.09319658577442169,
-0.024683449417352676,
0.04592152684926987,
0.016348516568541527,
0.09787070006132126,
0.04356250539422035,
0.004241836257278919,
-0.020960448309779167,
0.032165367156267166,
0.04792435094714165,
-0.11962699145078659,
-0.04644736647605896,
0.033716149628162384,
-0.03488767147064209,
0.027801811695098877,
0.039084065705537796,
0.08697953075170517,
-0.05852746590971947,
0.033521950244903564,
-0.027886712923645973,
-0.046858515590429306,
-0.011380928568542004,
-0.03579030558466911,
-0.02747231163084507,
0.010230393148958683,
0.014475385658442974,
0.029068581759929657,
0.003356657223775983,
0.04531574621796608,
-0.0361027829349041,
0.013455568812787533,
0.0371912457048893,
0.006452030036598444,
-0.0554211363196373,
0.05391181260347366,
0.04639892280101776,
-0.06306899338960648,
-0.026768770068883896,
0.036451518535614014,
-0.04560410603880882,
0.009975808672606945,
0.018856879323720932,
-0.004594620782881975,
-0.01584995910525322,
0.0416511595249176,
0.036089759320020676,
-0.038347575813531876,
0.06215090677142143,
0.011757166124880314,
-0.009566769935190678,
-0.03336272016167641,
-0.09625693410634995,
0.04046660289168358,
0.06496820598840714,
0.03258143737912178,
0.046300120651721954,
-0.020057525485754013,
-0.11184898018836975,
-0.02399713732302189,
0.05243159458041191,
0.00610612565651536,
0.034887686371803284,
-0.006215672008693218,
0.02937704510986805,
0.019539112225174904,
-0.02846650592982769,
0.06556430459022522,
0.04517717286944389,
0.00841036532074213,
-0.04863094165921211,
-0.03034355118870735,
-0.040009189397096634,
0.012991008348762989,
0.0830608680844307,
-0.06880906969308853,
0.014986269176006317,
0.04138029366731644,
0.04989127814769745,
0.05512719601392746,
-0.01804252713918686,
-0.024544423446059227,
-0.027101565152406693,
0.007109982892870903,
0.011620060540735722,
0.10349830985069275,
-0.10650516301393509,
-0.08479253202676773,
-6.161719145109729e-34,
0.042540233582258224,
0.055984869599342346,
0.12419620156288147,
0.12561237812042236,
0.07269743084907532,
-0.02645270526409149,
-0.04107319936156273,
-0.0327455997467041,
0.02783043123781681,
0.014110070653259754,
-0.03563324734568596,
0.039360933005809784,
-0.11177387833595276,
0.1412024199962616,
0.006231406237930059,
0.015587598085403442,
-0.02406006120145321,
0.008594192564487457,
-0.06984958052635193,
0.017660733312368393,
-0.009678992442786694,
-0.018238145858049393,
-0.018439358100295067,
-0.017839783802628517,
0.04221872240304947,
0.11138509958982468,
0.05274345725774765,
0.002011464908719063,
-0.03277503699064255,
-0.018285488709807396,
-0.0841318815946579,
0.014939537271857262,
-0.04879729077219963,
-0.03201080113649368,
0.05082279443740845,
-0.04232653230428696,
0.04219978675246239,
-0.05285126715898514,
-0.07003508508205414,
0.020316066220402718,
-0.0016138592036440969,
-0.03313041850924492,
-0.03133384510874748,
0.0016728139016777277,
-0.07382504642009735,
-0.05566544085741043,
-0.03596217930316925,
0.09372269362211227,
-0.08087417483329773,
-0.029306354001164436,
0.01703319326043129,
0.015915459021925926,
0.025124209001660347,
-0.10351771861314774,
-0.03328978270292282,
-0.03498920425772667,
-0.034111667424440384,
0.03131439536809921,
0.07798761874437332,
-0.0047453660517930984,
-0.025271734222769737,
0.05379270389676094,
0.02400433085858822,
0.11422328650951385,
-0.014144157990813255,
-0.002636000979691744,
-0.08891508728265762,
0.10482781380414963,
0.08790625631809235,
0.038785774260759354,
-0.07121771574020386,
0.031053796410560608,
-0.006005701143294573,
-0.02314493991434574,
0.004706159699708223,
-0.011377088725566864,
-0.02464551292359829,
-0.08694203943014145,
0.047185394912958145,
0.06763853132724762,
-0.013975465670228004,
0.01905863918364048,
-0.07462561130523682,
-0.08137794584035873,
0.027871539816260338,
-0.056463755667209625,
0.027249710634350777,
-0.06933614611625671,
0.0028830876108258963,
-0.10578269511461258,
-0.07343598455190659,
-0.023282313719391823,
-0.027212774381041527,
-0.0008051703334785998,
-0.03101555071771145,
-7.144760734295926e-34,
-0.04014911130070686,
0.0006711942260153592,
-0.043556198477745056,
0.08005698025226593,
0.003303813049569726,
-0.11032663285732269,
0.07525819540023804,
0.01388847641646862,
-0.0024788486771285534,
-0.02804459258913994,
-0.005488961935043335,
-0.04648317024111748,
-0.03125052899122238,
0.021748337894678116,
0.05289285257458687,
0.03582073375582695,
-0.038063302636146545,
0.03226574882864952,
0.027460267767310143,
0.04966909438371658,
-0.008174355141818523,
0.06242981180548668,
-0.09821193665266037,
-0.010149029083549976,
-0.11631020903587341,
0.06130751594901085,
-0.06108655408024788,
0.016160940751433372,
-0.03034045919775963,
0.08696520328521729,
-0.0022802602034062147,
-0.052162833511829376,
0.08656983077526093,
-0.003104994772002101,
-0.027312971651554108,
0.03581548109650612,
0.08171813935041428,
-0.03434780240058899,
0.005526289343833923,
0.013530759140849113,
0.013472015969455242,
-0.06302588433027267,
-0.07190550118684769,
-0.06133493408560753,
0.002429074142128229,
0.027647709473967552,
0.0322243869304657,
-0.04703764617443085,
0.014522607438266277,
-0.020393075421452522,
-0.01204686053097248,
-0.03253237530589104,
-0.007367403246462345,
-0.08802814036607742,
0.041889797896146774,
-0.059683624655008316,
-0.03923484683036804,
0.03003162331879139,
0.045302536338567734,
-0.023404354229569435,
-0.04975183308124542,
-0.07311315089464188,
0.06566987931728363,
0.027310047298669815,
-0.024160634726285934,
0.017221182584762573,
-0.04895207658410072,
-0.062180619686841965,
0.033629290759563446,
-0.012153474614024162,
-0.011696815490722656,
0.00513796741142869,
0.0038482528179883957,
0.00021714117610827088,
0.07842933386564255,
-0.03238800913095474,
-0.028185339644551277,
0.06459635496139526,
-0.0067121428437530994,
-0.0569692999124527,
0.017803365364670753,
-0.09036000818014145,
0.03468555957078934,
0.12248320132493973,
0.026216870173811913,
0.026590976864099503,
-0.009797558188438416,
0.03781012073159218,
0.059784773737192154,
0.007884158752858639,
0.018685709685087204,
0.06832829117774963,
0.024644356220960617,
0.09232643246650696,
-0.040795084089040756,
-3.989434915752099e-8,
-0.0358123779296875,
0.0729961097240448,
0.041329652070999146,
-0.021611720323562622,
0.053038813173770905,
-0.005559395533055067,
-0.04098021984100342,
0.019094394519925117,
-0.043200161308050156,
-0.03917849063873291,
0.0631125196814537,
-0.048115748912096024,
-0.02490389719605446,
-0.06473962217569351,
0.0614950992166996,
0.12284431606531143,
0.02598758228123188,
0.022391535341739655,
-0.001279170042835176,
0.0019598868675529957,
0.051578979939222336,
-0.03745894506573677,
0.06084943190217018,
0.07333086431026459,
0.01889314502477646,
-0.10512980073690414,
-0.06232414394617081,
0.07377473264932632,
-0.013846216723322868,
-0.023502100259065628,
-0.0008229303639382124,
-0.06917738914489746,
0.013697859831154346,
0.009178129956126213,
0.008023886941373348,
-0.04972432926297188,
0.050815124064683914,
-0.02448956109583378,
-0.0007816753932274878,
0.010323954746127129,
-0.009430242702364922,
0.053735602647066116,
-0.07315068691968918,
0.043833568692207336,
0.12319157272577286,
0.061994388699531555,
-0.043801773339509964,
-0.09933990985155106,
-0.03574877232313156,
0.05617618188261986,
-0.056311022490262985,
0.046940818428993225,
-0.04228933900594711,
0.09616869688034058,
-0.024099834263324738,
0.06903979927301407,
0.04463188722729683,
0.030987737700343132,
0.01725349947810173,
-0.006861050613224506,
0.025096265599131584,
0.03657156974077225,
-0.06450553983449936,
-0.07937256991863251
] | 0
| 0
| 0
| 0
|
|
MAESTRO: Masked Encoding Set Transformer with Self-Distillation
|
https://openreview.net/forum?id=FEZOLWexPb
|
[
"Matthew Eric Lee",
"Jaesik Kim",
"Matei Ionita",
"Jonghyun Lee",
"Michelle L. McKeague",
"YONGHYUN NAM",
"Irene Khavin",
"Yidi Huang",
"Victoria Fang",
"Sokratis Apostolidis",
"Divij Mathew",
"Shwetank",
"Ajinkya Pattekar",
"Zahabia Rangwala",
"Amit Bar-Or",
"Benjamin A Fensterheim",
"Benjamin A. Abramoff",
"Rennie L. Rhee",
"Damian Maseda",
"Allison R Greenplate",
"et al. (2 additional authors not shown)"
] |
Poster
|
The interrogation of cellular states and interactions in immunology research is an ever-evolving task, requiring adaptation to the current levels of high dimensionality. Cytometry enables high-dimensional profiling of immune cells, but its analysis is hindered by the complexity and variability of the data. We present MAESTRO, a self-supervised set representation learning model that generates vector representations of set-structured data, which we apply to learn immune profiles from cytometry data. Unlike previous studies only learn cell-level representations, whereas MAESTRO uses all of a sample's cells to learn a set representation. MAESTRO leverages specialized attention mechanisms to handle sets of variable number of cells and ensure permutation invariance, coupled with an online tokenizer by self-distillation framework. We benchmarked our model against existing cytometry approaches and other existing machine learning methods that have never been applied in cytometry. Our model outperforms existing approaches in retrieving cell-type proportions and capturing clinically relevant features for downstream tasks such as disease diagnosis and immune cell profiling.
|
self-supervision, representation learning, immunology, biology, single-cell, cytometry, set, set representations
|
We developed a self-supervised set representation learning model for vector-sized representations of single-cell data
| 11,153
| null |
[
-0.01827620156109333,
-0.002393121365457773,
0.04596405103802681,
-0.002500524977222085,
-0.05293096974492073,
0.01857776753604412,
0.011411445215344429,
0.013281106017529964,
0.03400804102420807,
-0.03639298677444458,
-0.09573864191770554,
-0.06829866021871567,
0.02789447084069252,
0.03277023136615753,
-0.08042678236961365,
0.03487200289964676,
0.06318163126707077,
0.09378471970558167,
-0.06458109617233276,
-0.07816363126039505,
0.015864375978708267,
0.013975149020552635,
0.04001566022634506,
0.06107214093208313,
0.027697019279003143,
-0.018672270700335503,
0.035040196031332016,
0.0008537992835044861,
-0.03235223516821861,
-0.004068730399012566,
-0.029157783836126328,
-0.0020013710018247366,
-0.05717942863702774,
0.03482089936733246,
-0.08606924116611481,
-0.000570305623114109,
-0.06115792691707611,
0.09338730573654175,
-0.019733265042304993,
-0.06449747085571289,
0.06265456974506378,
0.0037823591846972704,
0.04737841710448265,
0.08006224036216736,
0.06318386644124985,
0.062738798558712,
0.006727386265993118,
0.00410604989156127,
0.0049909502267837524,
-0.039081819355487823,
-0.030502108857035637,
0.03202466294169426,
-0.029695086181163788,
0.08657067269086838,
-0.027172498404979706,
0.021463384851813316,
-0.020228220149874687,
0.0004716356052085757,
0.013285218738019466,
-0.016602501273155212,
-0.04366152733564377,
-0.08648017793893814,
0.02595117874443531,
0.03598567098379135,
-0.03396271541714668,
-0.0010425165528431535,
-0.00011315621668472886,
0.1247115209698677,
-0.10610125958919525,
-0.016973786056041718,
-0.024253973737359047,
0.05308825895190239,
-0.06445609778165817,
0.0743444561958313,
0.0744662657380104,
0.10171520709991455,
0.04847540706396103,
-0.0160237904638052,
0.09236502647399902,
0.028850281611084938,
0.011182456277310848,
0.040821999311447144,
0.1316441148519516,
0.03235951066017151,
0.10981887578964233,
-0.04215528815984726,
0.010753022506833076,
0.05976633355021477,
0.0007384007913060486,
0.01722545735538006,
0.008467943407595158,
-0.028046313673257828,
0.09022517502307892,
-0.006003960035741329,
0.02777251787483692,
0.008966298773884773,
0.06523501873016357,
-0.001043457305058837,
0.07230192422866821,
0.056955788284540176,
-0.0750475823879242,
0.013488777913153172,
0.047127898782491684,
0.013292865827679634,
0.005219475831836462,
-0.03974078595638275,
0.028967002406716347,
-0.008380512706935406,
0.050433095544576645,
-0.10297233611345291,
0.04942576214671135,
-0.009484034031629562,
-0.040424928069114685,
-0.025360524654388428,
0.08919763565063477,
0.008493737317621708,
0.018179072067141533,
0.0068788365460932255,
0.11814523488283157,
0.03157234191894531,
0.013297822326421738,
-0.10442902147769928,
-0.05362170189619064,
0.06256099045276642,
0.045128218829631805,
0.0706099271774292,
-0.07757275551557541,
1.0250682672929787e-33,
0.0008439467055723071,
0.03854263946413994,
0.04705167934298515,
0.04221140593290329,
-0.05958809703588486,
-0.04584578424692154,
0.04921150952577591,
-0.033167362213134766,
-0.008326505310833454,
0.08555800467729568,
-0.01419251598417759,
0.04356566071510315,
-0.02237207628786564,
0.1208968460559845,
-0.02638419345021248,
0.030004756525158882,
-0.03751427307724953,
-0.016097771003842354,
-0.06834698468446732,
-0.1280427724123001,
-0.00841763336211443,
0.025098804384469986,
-0.021611256524920464,
-0.08335623890161514,
0.021511055529117584,
0.02184239961206913,
0.0514150932431221,
-0.011747181415557861,
-0.05292190611362457,
-0.012006456963717937,
-0.07843520492315292,
0.05446843057870865,
0.03982366621494293,
0.029242735356092453,
0.05256453901529312,
-0.01759311929345131,
0.03295915573835373,
-0.040179383009672165,
0.077984519302845,
-0.05811263248324394,
-0.0099715581163764,
-0.03206172585487366,
0.009250737726688385,
-0.06697062402963638,
-0.01626051403582096,
-0.0913621038198471,
-0.07180629670619965,
-0.05156894400715828,
-0.04818024858832359,
-0.0064699468202888966,
0.08504994213581085,
0.020596394315361977,
-0.0371311791241169,
-0.04695860669016838,
-0.017674710601568222,
0.04191937670111656,
-0.008085846900939941,
0.07681664079427719,
0.03375709056854248,
0.025060143321752548,
-0.022592389956116676,
-0.011499742977321148,
-0.022179991006851196,
0.07579770684242249,
0.03469565138220787,
0.013661772012710571,
-0.05232822149991989,
0.014943022280931473,
-0.015869775786995888,
0.038652487099170685,
0.014493493363261223,
-0.007100361865013838,
-0.05990183725953102,
-0.09678975492715836,
0.019370950758457184,
0.031711772084236145,
0.11124126613140106,
-0.004568188916891813,
-0.1250961720943451,
0.06105778366327286,
0.03844570368528366,
0.012074131518602371,
-0.12367600947618484,
0.001537692965939641,
-0.025411689653992653,
-0.08129674196243286,
0.02516624890267849,
-0.056999944150447845,
-0.050573043525218964,
-0.028064832091331482,
0.011516774073243141,
-0.006600722670555115,
-0.006787972990423441,
0.03668137267231941,
-0.09557370096445084,
-2.1909679209831678e-33,
-0.05354386568069458,
0.05728155001997948,
-0.021644441410899162,
-0.011329053901135921,
0.04471920058131218,
-0.04470194876194,
0.028484681621193886,
0.06029779836535454,
0.03983587026596069,
-0.042076218873262405,
-0.029224185273051262,
0.003068306017667055,
-0.02437668852508068,
0.01602674089372158,
-0.0349150113761425,
0.01480326522141695,
-0.10059283673763275,
0.05846794322133064,
-0.05357655510306358,
0.06913349777460098,
-0.11541011929512024,
0.06751711666584015,
-0.045801762491464615,
0.08059956133365631,
-0.03602829948067665,
0.03167705610394478,
0.01028510183095932,
0.047310683876276016,
-0.002488447818905115,
-0.07307959347963333,
-0.12398259341716766,
-0.011383641511201859,
0.00980285368859768,
-0.03032800555229187,
-0.04577864706516266,
0.011548972688615322,
-0.003398793749511242,
-0.06419533491134644,
0.04231595993041992,
0.0252692811191082,
-0.034309014678001404,
-0.047386396676301956,
-0.09423687309026718,
0.04400916025042534,
0.0658380314707756,
-0.005108128767460585,
0.029678136110305786,
0.023790499195456505,
0.10765775293111801,
0.07986073195934296,
-0.024885304272174835,
0.04166802018880844,
-0.05615147575736046,
0.014351780526340008,
-0.03974389657378197,
-0.08248936384916306,
0.037111274898052216,
-0.06919824331998825,
0.0868450254201889,
-0.01432510931044817,
-0.06908704340457916,
-0.078461654484272,
0.05971889942884445,
0.006306175608187914,
-0.004429191816598177,
0.005737503990530968,
0.0011779103660956025,
0.0008664472843520343,
-0.04215064272284508,
0.01704963482916355,
0.01696387305855751,
-0.016877057030797005,
-0.007509380578994751,
-0.0996987372636795,
-0.026734769344329834,
-0.08660591393709183,
-0.03130244463682175,
-0.03662792220711708,
-0.062189843505620956,
0.0696655884385109,
0.044521450996398926,
-0.08802789449691772,
0.03444285690784454,
0.034512922167778015,
0.0846564844250679,
0.03890754282474518,
0.05285421758890152,
0.04325631633400917,
0.0174825768917799,
0.025968467816710472,
-0.00854581780731678,
0.01769624650478363,
-0.021141834557056427,
0.032824136316776276,
-0.05885167047381401,
-3.927725344965438e-8,
0.030599871650338173,
-0.02234717272222042,
-0.01651480235159397,
-0.08714226633310318,
-0.00885383877903223,
0.016650551930069923,
-0.08415979892015457,
0.07650675624608994,
-0.048423413187265396,
0.022269314154982567,
0.0934034138917923,
-0.0009295640047639608,
-0.04642724618315697,
-0.052234161645174026,
0.07258903235197067,
0.03540880233049393,
0.008316382765769958,
0.036266881972551346,
-0.003040080890059471,
-0.005238352343440056,
-0.03316211700439453,
0.031053297221660614,
0.012115060351788998,
-0.044237419962882996,
0.08963928371667862,
-0.07394468784332275,
-0.058703672140836716,
0.04737268015742302,
0.02853567525744438,
-0.026407061144709587,
-0.02451985701918602,
0.05180777236819267,
0.06911416351795197,
-0.0015890784561634064,
-0.06001615896821022,
0.06501919031143188,
-0.015093286521732807,
-0.06033872812986374,
0.003890182124450803,
-0.08500628918409348,
-0.04048677906394005,
0.06124415248632431,
-0.08075463026762009,
-0.0034185140393674374,
-0.05522874742746353,
0.01484051626175642,
0.021029217168688774,
-0.02617018297314644,
0.005156717728823423,
-0.06450504809617996,
-0.03465011715888977,
-0.06593425571918488,
-0.0016721300780773163,
0.024923892691731453,
0.05409885570406914,
0.08115667849779129,
-0.0019456680165603757,
-0.02365841343998909,
0.08489574491977692,
-0.013962315395474434,
-0.02201710268855095,
0.011077557690441608,
-0.04114791378378868,
-0.033558882772922516
] | 0
| 0
| 0
| 0
|
|
AttriBoT: A Bag of Tricks for Efficiently Approximating Leave-One-Out Context Attribution
|
https://openreview.net/forum?id=9kJperA2a4
|
[
"Fengyuan Liu",
"Nikhil Kandpal",
"Colin Raffel"
] |
Poster
|
The influence of contextual input on the behavior of large language models (LLMs) has prompted the development of context attribution methods that aim to quantify each context span's effect on an LLM's generations. The leave-one-out (LOO) error, which measures the change in the likelihood of the LLM's response when a given span of the context is removed, provides a principled way to perform context attribution, but can be prohibitively expensive to compute for large models. In this work, we introduce AttriBoT, a series of novel techniques for efficiently computing an approximation of the LOO error for context attribution. Specifically, AttriBoT uses cached activations to avoid redundant operations, performs hierarchical attribution to reduce computation, and emulates the behavior of large target models with smaller proxy models. Taken together, AttriBoT can provide a 300x speedup while remaining more faithful to a target model's LOO error than prior context attribution methods. This stark increase in performance makes computing context attributions for a given response $30\times$ faster than generating the response itself, empowering real-world applications that require computing attributions at scale. We release a user-friendly and efficient implementation of AttriBoT to enable efficient LLM interpretability as well as encourage future development of efficient context attribution methods.
|
Large Language Model, Context Attribution, Interpretability
|
A suite of methods for efficiently attributing an LLM's response back to parts of its context
| 11,149
|
2411.15102
|
[
-0.07222282886505127,
-0.03873450309038162,
0.014512727968394756,
0.059435971081256866,
0.15053923428058624,
-0.04614368453621864,
-0.001165090361610055,
-0.00668746093288064,
0.06789914518594742,
-0.05841811001300812,
-0.026719458401203156,
-0.08942873030900955,
0.0981755182147026,
-0.07142085582017899,
0.011634823866188526,
0.038380734622478485,
0.10448474436998367,
0.10648445039987564,
-0.08529103547334671,
-0.07177060097455978,
0.030076520517468452,
0.041275255382061005,
0.054356902837753296,
-0.006764204706996679,
0.07732358574867249,
-0.03986763954162598,
-0.04920986667275429,
-0.02666573040187359,
0.024357985705137253,
0.0030894442461431026,
0.020376360043883324,
0.05192641541361809,
-0.039177875965833664,
0.03661760315299034,
-0.03713419288396835,
0.0727752074599266,
-0.08029311895370483,
0.03203411027789116,
0.0552169568836689,
-0.05612321197986603,
0.02896720916032791,
0.029668645933270454,
0.006867464166134596,
-0.0032197905238717794,
0.042926520109176636,
-0.029216181486845016,
0.036243993788957596,
0.017618976533412933,
-0.023540055379271507,
0.048658885061740875,
-0.05493538826704025,
-0.03577801585197449,
-0.006203947588801384,
-0.004699521232396364,
-0.08609235286712646,
-0.08676277846097946,
0.030454762279987335,
0.048099298030138016,
-0.028176315128803253,
-0.014816895127296448,
-0.07523704320192337,
-0.08681754022836685,
-0.08969350904226303,
0.020063916221261024,
-0.017926456406712532,
-0.025370435789227486,
0.022991754114627838,
0.016037149354815483,
-0.0695161446928978,
0.04836893081665039,
0.019815947860479355,
0.026561344042420387,
-0.04370254650712013,
0.10684249550104141,
0.022085972130298615,
0.017609506845474243,
0.0379139669239521,
0.04151518642902374,
0.06410365551710129,
-0.07511202991008759,
-0.01708681881427765,
-0.015212303958833218,
0.07156728953123093,
-0.02044343389570713,
0.08406130969524384,
-0.008541530929505825,
0.055050767958164215,
0.07872141152620316,
0.051230426877737045,
0.07047952711582184,
-0.03622401878237724,
-0.07830554246902466,
0.13717997074127197,
0.025978712365031242,
0.019664978608489037,
0.030422812327742577,
0.012500481680035591,
-0.0771225318312645,
0.013757011853158474,
0.10292301326990128,
0.0017843259265646338,
0.04812018573284149,
-0.020542580634355545,
-0.07798910140991211,
0.04235956072807312,
-0.021311217918992043,
-0.004596828483045101,
0.025162549689412117,
-0.008604789152741432,
-0.04240024462342262,
-0.025466185063123703,
0.05304800719022751,
0.03606759384274483,
-0.0036258073523640633,
0.048976801335811615,
-0.059874553233385086,
0.034205976873636246,
-0.021330323070287704,
0.006808900739997625,
0.046907033771276474,
0.024621907621622086,
0.024555431678891182,
0.032545678317546844,
0.0172545425593853,
-0.049487605690956116,
-0.04942680895328522,
-0.06863134354352951,
2.9384102283863425e-33,
0.07132264226675034,
0.011239703744649887,
0.0099147018045187,
0.0959753766655922,
0.0348774790763855,
0.04638025537133217,
0.04786078631877899,
0.06642576307058334,
-0.013409551233053207,
-0.026223009452223778,
-0.010150961577892303,
-0.00030536350095644593,
-0.0254898089915514,
0.0032178594265133142,
0.054782938212156296,
0.04304586723446846,
0.010159574449062347,
0.04792972654104233,
-0.026375267654657364,
0.017465468496084213,
0.01783834956586361,
-0.055453017354011536,
-0.005299821496009827,
-0.02063486911356449,
0.0758824273943901,
0.0596412755548954,
0.013287745416164398,
-0.05964295193552971,
-0.11527804285287857,
0.048880111426115036,
-0.06325408071279526,
-0.01277901791036129,
-0.012416073121130466,
0.018973974511027336,
-0.006557621993124485,
-0.08960393071174622,
-0.013168258592486382,
-0.03770702704787254,
-0.027966110035777092,
-0.03340822085738182,
-0.0802336111664772,
0.05138339847326279,
-0.0030908980406820774,
-0.002755959052592516,
-0.1220371350646019,
-0.10113772749900818,
-0.04045097157359123,
0.028011992573738098,
0.029185116291046143,
0.029147397726774216,
0.10706131905317307,
0.07239564508199692,
0.03802039474248886,
-0.0171654149889946,
-0.07713320106267929,
-0.04891036078333855,
-0.024782337248325348,
0.03184859827160835,
0.02729889005422592,
0.020956413820385933,
0.06776228547096252,
0.0016152510652318597,
-0.00875047966837883,
0.10739200562238693,
0.10367696732282639,
-0.002772887470200658,
-0.03755068778991699,
0.018243705853819847,
0.03974807262420654,
0.015603871084749699,
0.0075110625475645065,
0.0008407014538533986,
-0.07150790840387344,
-0.05348746478557587,
-0.03470001742243767,
-0.04216759279370308,
-0.005695395637303591,
-0.037282247096300125,
0.06307563185691833,
0.027499644085764885,
0.05934940278530121,
0.037164535373449326,
0.0033361290115863085,
-0.05364418774843216,
-0.061497919261455536,
-0.019059201702475548,
0.008696947246789932,
-0.060867808759212494,
-0.05812075734138489,
-0.033395472913980484,
-0.002312947530299425,
-0.0734114870429039,
-0.04027220234274864,
-0.0941908210515976,
-0.017423409968614578,
-2.5094418503865654e-33,
-0.07420618832111359,
-0.034836072474718094,
0.023620983585715294,
0.03954419121146202,
-0.013212650083005428,
-0.08606600761413574,
0.006115851458162069,
0.06093380227684975,
-0.03576914593577385,
0.010099779814481735,
-0.10464977473020554,
-0.031035376712679863,
0.04493316635489464,
0.00004399910176289268,
0.08005652576684952,
0.0050332327373325825,
0.05411709472537041,
-0.03946153447031975,
-0.023677479475736618,
0.04295564070343971,
0.0280145350843668,
0.0522947795689106,
-0.04762604832649231,
-0.012621806934475899,
-0.07214200496673584,
-0.01096361968666315,
-0.08956670016050339,
0.004666765686124563,
0.0014392239972949028,
-0.006183573044836521,
-0.08112185448408127,
0.07187901437282562,
-0.008867103606462479,
-0.08174775540828705,
-0.05004976689815521,
0.03747176378965378,
0.02473391778767109,
0.10737381130456924,
-0.07976420968770981,
0.05950009450316429,
0.10560440272092819,
-0.031241565942764282,
-0.01443488523364067,
-0.011588423512876034,
0.011414856649935246,
-0.010414142161607742,
-0.13480985164642334,
-0.037748150527477264,
0.04862917959690094,
0.011377664282917976,
-0.027030406519770622,
-0.02571599744260311,
-0.0021771932952106,
0.01582382060587406,
-0.052783019840717316,
-0.09991089999675751,
-0.026808498427271843,
-0.03376758471131325,
-0.00010296736581949517,
-0.07624351978302002,
-0.08084741979837418,
-0.025958016514778137,
-0.010674964636564255,
-0.041927918791770935,
0.05175788328051567,
0.00846180785447359,
-0.022467976436018944,
-0.03234090656042099,
-0.014085747301578522,
-0.06342175602912903,
0.06367624551057816,
-0.050150416791439056,
-0.055177606642246246,
0.005153898615390062,
0.01017434149980545,
0.007790910545736551,
0.009592934511601925,
-0.07332689315080643,
0.03098917379975319,
-0.006687604356557131,
-0.008814695291221142,
0.025871286168694496,
-0.0007493090815842152,
0.0029078172519803047,
0.01030388567596674,
-0.017627915367484093,
0.0789538025856018,
0.09024889767169952,
0.06121362745761871,
0.04918990284204483,
0.0034901462495326996,
0.047858692705631256,
-0.003551861736923456,
0.041273750364780426,
-0.04149235785007477,
-4.3084465772835756e-8,
-0.008103854022920132,
0.033673159778118134,
0.008396171033382416,
0.11853446066379547,
0.04961429163813591,
0.06323646754026413,
-0.03238552063703537,
0.08078858256340027,
-0.01679639145731926,
-0.0540652796626091,
0.039892226457595825,
0.03523632138967514,
0.01051214337348938,
-0.04190092533826828,
0.039826422929763794,
0.06469163298606873,
-0.02047032304108143,
0.011356442235410213,
-0.034486498683691025,
0.0013896467862650752,
0.04010292887687683,
0.05865413695573807,
0.02909238450229168,
-0.0007634361973032355,
0.03966955095529556,
-0.06445758789777756,
-0.03087611123919487,
0.09459718316793442,
0.0980684757232666,
-0.04764177277684212,
-0.03480693697929382,
-0.04722018167376518,
-0.10597285628318787,
0.004442373290657997,
-0.02888192981481552,
0.0632501095533371,
-0.033417269587516785,
-0.06971912831068039,
0.049949273467063904,
0.006124759092926979,
0.05984264612197876,
0.020981108769774437,
-0.013419394381344318,
0.08820129930973053,
0.03477073833346367,
0.02060331590473652,
-0.03785596415400505,
-0.1195065975189209,
0.03340516611933708,
-0.09585349261760712,
-0.02351015992462635,
-0.06834752857685089,
-0.09192657470703125,
0.04026016220450401,
-0.03125031664967537,
-0.05137529596686363,
0.012225854210555553,
-0.05556309595704079,
0.06719907373189926,
0.02065415307879448,
0.06706023216247559,
0.0450332909822464,
-0.0333627313375473,
-0.008976745419204235
] |
https://github.com/r-three/AttriBoT
| 6
| 0
| 0
| 0
|
Provably Safeguarding a Classifier from OOD and Adversarial Samples
|
https://openreview.net/forum?id=kwCHcaeHrf
|
[
"Nicolas Atienza",
"Johanne Cohen",
"Christophe Labreuche",
"Michele Sebag"
] |
Poster
|
This paper aims to transform a trained classifier into an abstaining classifier, such
that the latter is provably protected from out-of-distribution and adversarial samples. The proposed Sample-efficient Probabilistic Detection using Extreme Value
Theory (SPADE) approach relies on a Generalized Extreme Value (GEV) model
of the training distribution in the latent space of the classifier. Under mild assumptions, this GEV model allows for formally characterizing out-of-distribution
and adversarial samples and rejecting them. Empirical validation of the approach
is conducted on various neural architectures (ResNet, VGG, and Vision Transformer) and considers medium and large-sized datasets (CIFAR-10, CIFAR-100,
and ImageNet). The results show the stability and frugality of the GEV model and
demonstrate SPADE’s efficiency compared to the state-of-the-art methods.
|
OOD detection, Adversarial detection, Extreme Value Theory
| null | 11,146
|
2501.10202
|
[
-0.07710856944322586,
-0.019681531935930252,
-0.021165641024708748,
-0.02997209131717682,
0.14902955293655396,
0.03451482951641083,
-0.00441526435315609,
0.02066478133201599,
-0.014620193280279636,
-0.06194232404232025,
0.011286601424217224,
0.017396198585629463,
0.07133103162050247,
-0.0005662827170453966,
-0.10073915123939514,
-0.04972568899393082,
0.0956772193312645,
0.06940881907939911,
-0.09589385986328125,
-0.004873740021139383,
0.03216947242617607,
0.031425345689058304,
-0.03164346516132355,
-0.003975787665694952,
0.011392035521566868,
-0.013627719134092331,
0.009445899166166782,
0.004405553452670574,
-0.005110712721943855,
-0.03726597875356674,
0.03384183347225189,
-0.011009473353624344,
-0.0610639862716198,
0.047002796083688736,
-0.08681073039770126,
0.050711389631032944,
-0.026353009045124054,
0.037114933133125305,
0.06067954748868942,
0.0004673390940297395,
-0.0022054368164390326,
-0.057022858411073685,
-0.016666047275066376,
-0.057579196989536285,
0.053050801157951355,
0.0015072876121848822,
0.04607759416103363,
-0.09048029035329819,
-0.001935314037837088,
-0.0298302061855793,
0.01856047660112381,
0.011630043387413025,
-0.04596122354269028,
0.0973859503865242,
-0.04913211613893509,
-0.040145762264728546,
0.022871727123856544,
-0.04716856777667999,
-0.023198025301098824,
0.08191598951816559,
-0.004869229160249233,
-0.01826186291873455,
0.0013945541577413678,
0.018476491793990135,
-0.04913186654448509,
-0.041412632912397385,
0.017221473157405853,
0.028288578614592552,
-0.024117259308695793,
-0.026627425104379654,
0.026094241067767143,
0.08618444204330444,
-0.05661516264081001,
0.026639513671398163,
0.05711738392710686,
0.04941408708691597,
0.02944011799991131,
0.009811611846089363,
0.10613491386175156,
-0.036398839205503464,
0.07300154864788055,
-0.011315402574837208,
0.05654667317867279,
-0.044907283037900925,
0.16208219528198242,
0.06856673955917358,
-0.08471028506755829,
0.06965159624814987,
0.07353847473859787,
0.0016423651250079274,
-0.036986492574214935,
-0.04169969633221626,
0.02283918857574463,
0.06022612005472183,
0.02928214520215988,
-0.016601433977484703,
-0.03515156731009483,
-0.0803334191441536,
-0.04371440410614014,
0.09593375027179718,
-0.08175656944513321,
-0.03235131502151489,
0.05874617397785187,
-0.05290813744068146,
0.1307932734489441,
-0.11580464243888855,
0.09371575713157654,
-0.01280947495251894,
0.09058136492967606,
-0.0591849610209465,
-0.008686572313308716,
0.00571623258292675,
0.02224125526845455,
-0.066789411008358,
0.05065067484974861,
-0.02534632943570614,
-0.1064940094947815,
0.06730201840400696,
-0.10613775998353958,
0.002372479997575283,
-0.050431109964847565,
0.07626565545797348,
0.004373101983219385,
0.02935067005455494,
0.06816739588975906,
0.04066474735736847,
-0.023583969101309776,
3.3239982730659694e-33,
-0.016494978219270706,
-0.04754805192351341,
0.04084070771932602,
-0.07033419609069824,
0.044589798897504807,
-0.014348351396620274,
0.01866188645362854,
0.050945647060871124,
0.030796565115451813,
0.07486766576766968,
-0.05053302273154259,
0.01017377432435751,
-0.04640159383416176,
0.07614968717098236,
0.05285276472568512,
0.040770597755908966,
-0.020897429436445236,
-0.029695795848965645,
-0.008387526497244835,
-0.03500582277774811,
0.05266997590661049,
-0.09634758532047272,
0.056992098689079285,
-0.0683082640171051,
-0.04457065835595131,
0.011334523558616638,
-0.007286465726792812,
0.0019103396916761994,
-0.0030963660683482885,
0.04431450739502907,
-0.1451088935136795,
0.03405199572443962,
0.020817717537283897,
-0.008339382708072662,
0.0154977822676301,
-0.013585114851593971,
-0.07087548822164536,
-0.01946466788649559,
-0.00795791856944561,
-0.007579764351248741,
-0.0557335689663887,
0.07329262048006058,
0.007038341835141182,
-0.014012266881763935,
0.0295098964124918,
0.06006426736712456,
0.10715141892433167,
0.022719256579875946,
-0.07440266758203506,
-0.06880649924278259,
0.008841495960950851,
-0.01710224710404873,
-0.06980963796377182,
-0.03380731865763664,
-0.09236587584018707,
0.05200137570500374,
0.010454444214701653,
0.07802891731262207,
-0.023387519642710686,
0.02931729517877102,
-0.08656900376081467,
0.014351293444633484,
0.02234778180718422,
0.05579924210906029,
-0.0013952868757769465,
-0.03936294466257095,
-0.01851634308695793,
-0.02191804349422455,
0.03782914578914642,
0.06623855978250504,
0.07891835272312164,
0.028667816892266273,
-0.06474535912275314,
-0.07884470373392105,
-0.024361753836274147,
0.011518876999616623,
0.08161649107933044,
0.027716178447008133,
-0.0011049640597775578,
0.044559694826602936,
-0.048444658517837524,
0.02031373418867588,
0.041884250938892365,
-0.015300559811294079,
-0.04101962968707085,
0.0766969546675682,
0.03781638666987419,
-0.03641703724861145,
-0.013452055864036083,
0.03998921811580658,
-0.10359282046556473,
0.0009571705013513565,
-0.026880457997322083,
-0.05956827849149704,
-0.010921388864517212,
-1.836925369233811e-33,
-0.029644466936588287,
0.09158710390329361,
-0.06404901295900345,
0.05900278687477112,
-0.029285358265042305,
-0.009844735264778137,
-0.0026793305296450853,
0.08306767791509628,
-0.08863843977451324,
-0.01782252825796604,
0.007700695656239986,
-0.012093178927898407,
0.07129458338022232,
-0.061287280172109604,
0.021638544276356697,
-0.03514125570654869,
-0.044797129929065704,
0.05955532565712929,
0.033666230738162994,
-0.04843466728925705,
0.06185213103890419,
0.10843175649642944,
-0.07733021676540375,
0.037002310156822205,
-0.028680646792054176,
-0.015765005722641945,
-0.049139536917209625,
0.08999577909708023,
0.025579659268260002,
-0.021706098690629005,
-0.010064893402159214,
-0.021992892026901245,
-0.012590308673679829,
0.012181865982711315,
-0.006787675898522139,
0.0027752809692174196,
0.05296426638960838,
-0.005806711968034506,
-0.011633436195552349,
0.11786343157291412,
0.05529172718524933,
0.03691120818257332,
-0.1036631241440773,
0.022580940276384354,
0.002034040866419673,
-0.05172271654009819,
0.042859308421611786,
-0.04744701087474823,
0.08552782982587814,
0.005428205244243145,
-0.01170552521944046,
-0.06666437536478043,
-0.08411478996276855,
0.04695431515574455,
-0.0017126547172665596,
-0.07850553095340729,
-0.0007264751475304365,
-0.033114101737737656,
0.01711701974272728,
0.027393581345677376,
-0.00515891844406724,
-0.004021539352834225,
-0.060441598296165466,
0.011854036711156368,
-0.0014677185099571943,
0.0036812464240938425,
-0.019947074353694916,
0.04435725882649422,
0.0029706342611461878,
0.032645631581544876,
0.058387432247400284,
0.023428576067090034,
0.030281679704785347,
0.021697085350751877,
-0.05910416692495346,
-0.03689735382795334,
0.015306944027543068,
-0.012015183456242085,
0.04002867639064789,
-0.04551996290683746,
0.017294587567448616,
-0.06621487438678741,
0.03559629246592522,
0.08363400399684906,
0.05105535313487053,
0.0825023427605629,
0.06913989037275314,
-0.09322620183229446,
0.058054324239492416,
-0.002207785611972213,
-0.06350471824407578,
0.0787142813205719,
-0.06674657762050629,
0.005184209905564785,
0.015055647119879723,
-4.5658183012164955e-8,
-0.04206293821334839,
0.05109245330095291,
0.04294300079345703,
-0.0014033688930794597,
0.09046092629432678,
-0.002678914461284876,
0.007367409765720367,
0.039584144949913025,
-0.027072234079241753,
0.009688305668532848,
0.05514134094119072,
-0.06256032735109329,
0.008633368648588657,
-0.056351277977228165,
-0.031979478895664215,
0.0014751586131751537,
0.03791697695851326,
0.025686688721179962,
0.00820962619036436,
-0.01288581732660532,
0.002946541178971529,
-0.02540350705385208,
0.036091357469558716,
-0.038349758833646774,
-0.008243873715400696,
-0.06704647094011307,
-0.00909875612705946,
-0.03756290674209595,
0.05482247844338417,
0.035017434507608414,
-0.0799785628914833,
0.005964537151157856,
0.04385392367839813,
0.01685185357928276,
0.04922405257821083,
0.12441488355398178,
-0.045194320380687714,
-0.014715800993144512,
-0.017159661278128624,
0.07727628946304321,
-0.03871089592576027,
0.0256117545068264,
0.002083694329485297,
-0.037930164486169815,
-0.02051543816924095,
0.03542347997426987,
0.08710379153490067,
-0.06409094482660294,
0.02472700923681259,
0.04275689274072647,
0.028075022622942924,
-0.02582797221839428,
0.06275921314954758,
0.11017869412899017,
0.01339380256831646,
-0.09679452329874039,
-0.01364277582615614,
-0.06604459881782532,
-0.0420454666018486,
0.06962133944034576,
0.02621537633240223,
-0.023062288761138916,
-0.06182081252336502,
-0.001429729163646698
] | 0
| 0
| 0
| 0
|
|
Learning Chaos In A Linear Way
|
https://openreview.net/forum?id=Llh6CinTiy
|
[
"Xiaoyuan Cheng",
"Yi He",
"Yiming Yang",
"Xiao Xue",
"Sibo Cheng",
"Daniel Giles",
"Xiaohang Tang",
"Yukun Hu"
] |
Poster
|
Learning long-term behaviors in chaotic dynamical systems, such as turbulent flows and climate modelling, is challenging due to their inherent instability and unpredictability. These systems exhibit positive Lyapunov exponents, which significantly hinder accurate long-term forecasting. As a result, understanding long-term statistical behavior is far more valuable than focusing on short-term accuracy. While autoregressive deep sequence models have been applied to capture long-term behavior, they often lead to exponentially increasing errors in learned dynamics. To address this, we shift the focus from simple prediction errors to preserving an invariant measure in dissipative chaotic systems. These systems have attractors, where trajectories settle, and the invariant measure is the probability distribution on attractors that remains unchanged under dynamics. Existing methods generate long trajectories of dissipative chaotic systems by aligning invariant measures, but it is not always possible to obtain invariant measures for arbitrary datasets. We propose the Poincaré Flow Neural Network (PFNN), a novel operator learning framework designed to capture behaviors of chaotic systems without any explicit knowledge of the invariant measure. PFNN employs an auto-encoder to map the chaotic system to a finite-dimensional feature space, effectively linearizing the chaotic evolution. It then learns the linear evolution operators to match the physical dynamics by addressing two critical properties in dissipative chaotic systems: (1) contraction, the system’s convergence toward its attractors, and (2) measure invariance, trajectories on the attractors following a probability distribution invariant to the dynamics.
Our experiments on a variety of chaotic systems, including Lorenz systems, Kuramoto-Sivashinsky equation and Navier–Stokes equation, demonstrate that PFNN has more accurate predictions and physical statistics compared to competitive baselines including the Fourier Neural Operator and the Markov Neural Operator.
|
Dynamical systems, operator learning, chaos, physics-informed learning
| null | 11,142
| null |
[
-0.09400177747011185,
-0.0719592273235321,
0.07757939398288727,
0.021343406289815903,
0.002247635740786791,
0.03733472153544426,
-0.02836664579808712,
-0.13641704618930817,
0.06120259314775467,
-0.05348457768559456,
-0.038894012570381165,
-0.06665265560150146,
-0.03854185342788696,
0.035641126334667206,
-0.11145723611116409,
0.02670409344136715,
-0.05093781277537346,
0.07453992962837219,
-0.001528496271930635,
-0.04530598223209381,
-0.013868066482245922,
0.007788211572915316,
-0.025731369853019714,
0.10642234236001968,
0.03986673429608345,
0.008059322834014893,
-0.030425116419792175,
0.015493903309106827,
-0.05391408130526543,
0.001972859026864171,
0.09326578676700592,
-0.027396442368626595,
-0.057086411863565445,
-0.0328189991414547,
-0.021309543401002884,
0.041156597435474396,
-0.0986800566315651,
0.020768769085407257,
-0.020324748009443283,
0.00008213164255721495,
0.07746101170778275,
0.023128168657422066,
0.029814014211297035,
0.0606868714094162,
0.04838263615965843,
0.01232520304620266,
0.06577921658754349,
-0.11855816096067429,
0.021803969517350197,
0.038189250975847244,
-0.04085417464375496,
0.07552646845579147,
0.00252152094617486,
0.02829081378877163,
0.01892600953578949,
0.009318957105278969,
-0.01593301072716713,
0.01750962622463703,
-0.028760654851794243,
-0.04136808589100838,
0.01584482565522194,
-0.04158475250005722,
-0.02488703466951847,
0.003962268587201834,
-0.1248168796300888,
0.007893597707152367,
-0.03679153323173523,
0.12494292855262756,
-0.014666521921753883,
0.03751278668642044,
-0.07819192111492157,
0.0801151767373085,
0.0030725328251719475,
-0.007104072254151106,
0.035624176263809204,
0.08842576295137405,
0.016369247809052467,
-0.053020261228084564,
-0.01038403157144785,
-0.018212798982858658,
0.06523755192756653,
0.05164989084005356,
0.05904968082904816,
0.016535956412553787,
0.10385142266750336,
0.019483532756567,
0.03381280601024628,
0.03359479084610939,
0.03736850619316101,
-0.0732230544090271,
0.03112957440316677,
-0.05720535293221474,
0.05594702064990997,
-0.00605815090239048,
0.08808289468288422,
0.07191668450832367,
0.040996309369802475,
-0.053139496594667435,
0.001397909363731742,
0.061478354036808014,
-0.026211319491267204,
0.017885521054267883,
-0.021567853167653084,
0.047155946493148804,
0.044993992894887924,
-0.02167722024023533,
0.030105413869023323,
-0.017654675990343094,
0.06215113401412964,
-0.05398024991154671,
0.000556772924028337,
0.009372884407639503,
0.06874457746744156,
0.0888364389538765,
0.07781311869621277,
-0.049429669976234436,
-0.06382594257593155,
0.026647912338376045,
-0.03856175020337105,
0.1029621884226799,
-0.06844381988048553,
-0.01959240809082985,
-0.06634382158517838,
0.03361218795180321,
0.025035617873072624,
0.014111639931797981,
-0.022833537310361862,
1.7852660659350177e-33,
0.010045069269835949,
-0.008054873906075954,
0.05577056109905243,
0.009218764491379261,
0.06734917312860489,
-0.012857822701334953,
-0.03517448157072067,
-0.023266833275556564,
0.018126191571354866,
0.05916759744286537,
-0.032714858651161194,
0.013045689091086388,
-0.08020389080047607,
0.07565361261367798,
-0.05679614841938019,
-0.01151266135275364,
-0.08983627706766129,
-0.049469783902168274,
-0.002903282642364502,
-0.07307548820972443,
0.01675417274236679,
-0.07769817858934402,
-0.05322239547967911,
-0.08349286019802094,
0.041003044694662094,
0.034936707466840744,
0.07543966174125671,
0.12548011541366577,
-0.016830192878842354,
0.007946365512907505,
-0.04454518482089043,
-0.005555362906306982,
0.014769579283893108,
0.004550958052277565,
0.009068788029253483,
-0.0354064479470253,
-0.018630748614668846,
0.003678728826344013,
0.015693115070462227,
-0.050277937203645706,
0.015221789479255676,
0.013371848501265049,
-0.09570878744125366,
0.042018745094537735,
-0.03193492069840431,
-0.06835512816905975,
0.0832131952047348,
0.04994126409292221,
-0.03925755247473717,
-0.03996225818991661,
0.03344490006566048,
-0.06767518818378448,
-0.08725288510322571,
-0.05284661054611206,
-0.03584592044353485,
0.01524566300213337,
0.09239523112773895,
0.02668396756052971,
-0.0008195327827706933,
0.04031632840633392,
-0.040790628641843796,
-0.0040466743521392345,
0.04827162250876427,
-0.04446316882967949,
0.05842401459813118,
0.06572634726762772,
0.017241885885596275,
0.09234520047903061,
0.049355778843164444,
-0.08292527496814728,
0.05938819423317909,
-0.0033236807212233543,
-0.024228807538747787,
-0.01917245425283909,
0.0406704917550087,
0.010069106705486774,
0.0009268177673220634,
-0.06317432969808578,
-0.060078930109739304,
0.08783045411109924,
-0.03589032590389252,
-0.038058508187532425,
0.020391836762428284,
0.044419776648283005,
-0.06236949563026428,
-0.059911783784627914,
0.01392525713890791,
-0.030686868354678154,
-0.006745317485183477,
0.013888760469853878,
-0.0591520331799984,
-0.07562734931707382,
0.08991096168756485,
0.015572827309370041,
-0.053526367992162704,
-3.788602129652909e-33,
-0.0487651526927948,
0.016609251499176025,
-0.025104692205786705,
0.05881341174244881,
0.009487980045378208,
-0.033044952899217606,
-0.018108056858181953,
0.06736250221729279,
-0.0064244638197124004,
-0.04610253497958183,
-0.036928944289684296,
0.010875738225877285,
0.03831465169787407,
0.006802903488278389,
0.1155349463224411,
-0.0367031954228878,
0.010031662881374359,
-0.005009224638342857,
0.014864261262118816,
-0.025320038199424744,
-0.0513809509575367,
-0.006605737376958132,
-0.102317214012146,
0.03140363097190857,
-0.014437252655625343,
0.01480026450008154,
0.0407237745821476,
0.047009944915771484,
-0.05607682466506958,
0.009441640228033066,
-0.09571672230958939,
-0.0746140107512474,
-0.017683574929833412,
-0.007201062049716711,
-0.01517392136156559,
0.030121464282274246,
0.03781992197036743,
-0.04644525423645973,
-0.03800807148218155,
0.06803526729345322,
0.009066727943718433,
-0.06138286739587784,
0.030509741976857185,
-0.015356650575995445,
0.01892864890396595,
-0.015397671610116959,
0.009142854250967503,
0.031187064945697784,
0.13137708604335785,
0.06923586130142212,
0.0272921621799469,
0.07495363801717758,
-0.12820909917354584,
0.02076704241335392,
-0.014010596089065075,
0.06345890462398529,
-0.06151501089334488,
-0.017344117164611816,
0.03176943585276604,
0.022558454424142838,
-0.08395812660455704,
-0.05404377728700638,
0.004596106242388487,
0.07368981093168259,
-0.07597123086452484,
0.01853320188820362,
-0.05451788753271103,
0.021583693102002144,
0.05571512505412102,
-0.002134411595761776,
0.08127698302268982,
0.007548973895609379,
-0.013882780447602272,
0.020176930353045464,
-0.01568182185292244,
-0.10193204134702682,
-0.020014310255646706,
-0.02516247145831585,
0.008559091947972775,
0.02070300094783306,
0.01437497790902853,
-0.018154576420783997,
0.049206849187612534,
0.023393729701638222,
-0.018441539257764816,
-0.02114068530499935,
0.05880464240908623,
-0.058905329555273056,
0.03743765130639076,
0.03371468931436539,
0.06750219315290451,
0.003506798529997468,
-0.06272289901971817,
0.0036464931908994913,
-0.053931061178445816,
-4.396352082380872e-8,
-0.03626173734664917,
0.011409267783164978,
0.10312496870756149,
0.017449159175157547,
-0.003953130915760994,
-0.01654759608209133,
0.034653421491384506,
0.027538271620869637,
0.027008431032299995,
-0.010324877686798573,
0.1388772428035736,
-0.03529336303472519,
-0.07501543313264847,
-0.04947557672858238,
-0.04940207675099373,
0.016108091920614243,
0.05445578321814537,
0.03995654359459877,
-0.005413508974015713,
-0.00018295212066732347,
0.03545846417546272,
0.050294291228055954,
-0.04559914022684097,
0.0045847538858652115,
0.06450607627630234,
-0.013234744779765606,
-0.0962088331580162,
0.047374315559864044,
0.035412274301052094,
-0.01792137138545513,
0.0452888123691082,
-0.0020125964656472206,
0.01875443011522293,
-0.09980319440364838,
-0.12763705849647522,
0.0347115583717823,
0.019491907209157944,
-0.05185239762067795,
0.030466917902231216,
-0.06508582830429077,
-0.024570079520344734,
0.08954298496246338,
0.0043671573512256145,
-0.041782770305871964,
-0.0878605991601944,
0.008829215541481972,
0.03301037847995758,
-0.12092457711696625,
0.04713091999292374,
-0.00013304321328178048,
0.10104604065418243,
-0.01104535162448883,
0.042634349316358566,
0.03371448442339897,
0.04599865525960922,
-0.07167574763298035,
-0.031107209622859955,
-0.06377852708101273,
-0.03375813364982605,
0.07213441282510757,
-0.08123181760311127,
0.023000337183475494,
-0.008148623630404472,
-0.027779346331954002
] | 0
| 0
| 0
| 0
|
|
Data Taggants: Dataset Ownership Verification Via Harmless Targeted Data Poisoning
|
https://openreview.net/forum?id=6ldD8Y4gBQ
|
[
"Wassim Bouaziz",
"Nicolas Usunier",
"El-Mahdi El-Mhamdi"
] |
Poster
|
Dataset ownership verification, the process of determining if a dataset is used in a model's training data, is necessary for detecting unauthorized data usage and data contamination.
Existing approaches, such as backdoor watermarking, rely on inducing a detectable behavior into the trained model on a part of the data distribution.
However, these approaches have limitations, as they can be harmful to the model's performances or require unpractical access to the model's internals.
Most importantly, previous approaches lack guarantee against false positives.\
This paper introduces *data taggants*, a novel non-backdoor dataset ownership verification technique.
Our method uses pairs of out-of-distribution samples and random labels as secret *keys*, and leverages clean-label targeted data poisoning to subtly alter a dataset, so that models trained on it respond to the key samples with the corresponding key labels.
The keys are built as to allow for statistical certificates with black-box access only to the model.\
We validate our approach through comprehensive and realistic experiments on ImageNet1k using ViT and ResNet models with state-of-the-art training recipes.
Our findings demonstrate that data taggants can reliably detect models trained on the protected dataset with high confidence, without compromising validation accuracy, and show their superiority over backdoor watermarking.
We demonstrate the stealthiness and robustness of our method
% shows to be stealthy and robust
against various defense mechanisms.
|
dataset watermarking, dataset ownership verification, data poisoning, backdoor attack
|
We elaborate a harmless and stealthy targeted data poisoning approach to mark a model trained on a protected dataset and detect it.
| 11,140
|
2410.09101
|
[
-0.12496203184127808,
-0.02752074785530567,
-0.005941196344792843,
0.013678377494215965,
0.10081500560045242,
-0.06807167828083038,
0.03600439056754112,
-0.036226049065589905,
-0.003019590862095356,
-0.031387172639369965,
0.028168784454464912,
-0.03513456881046295,
0.16688139736652374,
-0.027231086045503616,
-0.07318685948848724,
-0.0005413683829829097,
0.05647340416908264,
0.09533219784498215,
-0.07738517224788666,
-0.039799004793167114,
-0.05493776127696037,
-0.048112429678440094,
0.02858675830066204,
0.0065232692286372185,
0.0010608281008899212,
-0.025314703583717346,
0.05279206857085228,
-0.001960411202162504,
-0.01853834092617035,
-0.05546524003148079,
0.08109138906002045,
-0.037770580500364304,
-0.009399864822626114,
0.06587804108858109,
0.016747646033763885,
0.01934073120355606,
-0.004392487462610006,
-0.013171804137527943,
0.014981327578425407,
-0.042715054005384445,
0.022738661617040634,
-0.04553094878792763,
-0.028438888490200043,
0.048140957951545715,
0.0037238067016005516,
0.039460498839616776,
0.05883636698126793,
0.02457326278090477,
-0.04799259826540947,
-0.00568955997005105,
-0.01260136254131794,
-0.006077818106859922,
-0.04053196683526039,
0.07857175171375275,
-0.04023302346467972,
-0.027036797255277634,
0.007352874614298344,
-0.03167131915688515,
-0.08205919712781906,
0.059490419924259186,
-0.0021203095093369484,
-0.02527870237827301,
-0.05867777392268181,
-0.007283725775778294,
0.011020999401807785,
0.05227024108171463,
-0.046648554503917694,
0.03649211674928665,
0.059086039662361145,
-0.02173120342195034,
0.06784959137439728,
0.07085687667131424,
-0.03613472357392311,
0.0425490066409111,
-0.03268164023756981,
0.09468889981508255,
-0.015986450016498566,
0.03117256611585617,
0.061093974858522415,
-0.15797768533229828,
0.02387451007962227,
-0.005147907882928848,
0.05950712412595749,
0.005610920488834381,
0.047151897102594376,
0.015788855031132698,
-0.008014731109142303,
0.03939081355929375,
0.0560263991355896,
0.05241614952683449,
-0.0495128408074379,
0.019071316346526146,
0.14810970425605774,
0.007443415001034737,
0.05071777477860451,
0.05544334277510643,
0.004627597518265247,
-0.051505062729120255,
0.004694035276770592,
0.0695158988237381,
-0.1477576494216919,
0.014253370463848114,
-0.005292466375976801,
-0.08733457326889038,
0.16305437684059143,
0.023969311267137527,
0.04649225249886513,
-0.03246133774518967,
0.07497642189264297,
-0.06844506412744522,
-0.0012058551656082273,
0.046789489686489105,
-0.07515755295753479,
-0.041374679654836655,
0.06860239803791046,
0.012666645459830761,
-0.07339712977409363,
0.029923832044005394,
-0.04886806011199951,
0.027772651985287666,
-0.02239328809082508,
0.002309005009010434,
-0.007212179247289896,
-0.05367047339677811,
-0.032913994044065475,
-0.02224005199968815,
-0.09950073063373566,
3.389305433119794e-33,
-0.01698269136250019,
0.04656745493412018,
0.0000712098553776741,
-0.02969483658671379,
0.07179278135299683,
-0.006515045650303364,
-0.029019393026828766,
0.02377321571111679,
-0.06430936604738235,
0.020640388131141663,
-0.013970030471682549,
-0.023522034287452698,
-0.06754454970359802,
0.0568455308675766,
0.08028914779424667,
0.08215799182653427,
-0.02782905101776123,
0.03626666218042374,
-0.0021082989405840635,
0.02366296760737896,
0.07092241942882538,
-0.07888329029083252,
0.011901731602847576,
-0.024052483960986137,
0.037802644073963165,
-0.03910740464925766,
-0.018429920077323914,
0.0295348409563303,
0.07291682809591293,
0.036591824144124985,
-0.06772210448980331,
0.027748407796025276,
0.08984973281621933,
0.039841506630182266,
0.050929855555295944,
0.003458387916907668,
-0.010704354383051395,
-0.04490245133638382,
0.033732496201992035,
0.030141744762659073,
0.032027460634708405,
-0.043378088623285294,
-0.0332392118871212,
-0.007944842800498009,
-0.07644903659820557,
-0.028188420459628105,
-0.04412245750427246,
0.03082716278731823,
-0.025858866050839424,
0.051672495901584625,
0.06673955172300339,
0.020986177027225494,
-0.02053987607359886,
-0.09590102732181549,
-0.14803652465343475,
-0.029773017391562462,
0.014542926102876663,
-0.028992848470807076,
0.06195487454533577,
0.035625506192445755,
-0.06496831029653549,
-0.0027330012526363134,
-0.0035197827965021133,
0.07028524577617645,
-0.044972289353609085,
-0.028724197298288345,
0.06597989797592163,
-0.052647631615400314,
0.005318053066730499,
-0.031307000666856766,
-0.04297588765621185,
0.013095060363411903,
-0.09447558224201202,
-0.019333675503730774,
-0.006418618373572826,
-0.05917426571249962,
0.05766812711954117,
0.023955512791872025,
-0.045247793197631836,
0.0430876761674881,
0.03640628233551979,
-0.014485455118119717,
0.058023083955049515,
0.008438876830041409,
-0.11480066180229187,
-0.010022414848208427,
-0.027799038216471672,
-0.05347549915313721,
-0.03157670795917511,
0.04435001313686371,
-0.012582107447087765,
-0.03822164237499237,
-0.03365813195705414,
-0.02497568540275097,
-0.01360701397061348,
-2.3835017733623656e-33,
-0.015532775782048702,
0.05498833209276199,
-0.028625363484025,
0.038036324083805084,
-0.022890159860253334,
-0.05967976525425911,
0.009818937629461288,
0.07195190340280533,
0.008929603733122349,
-0.03916052356362343,
0.010276620276272297,
-0.04834509268403053,
0.010023687034845352,
-0.0839630514383316,
-0.020278407260775566,
-0.012667618691921234,
-0.05368442460894585,
0.019081758335232735,
-0.03422076255083084,
0.016646387055516243,
0.0005543521838262677,
0.09330060333013535,
-0.11888681352138519,
0.09606475383043289,
0.01084249746054411,
0.0321492925286293,
-0.04754142835736275,
0.10052844882011414,
0.04108061641454697,
0.017307966947555542,
-0.04909399151802063,
0.032308273017406464,
0.007992865517735481,
-0.11348175257444382,
-0.06929898262023926,
-0.04271192103624344,
0.1215357780456543,
0.0359768345952034,
-0.01941186562180519,
0.07770651578903198,
0.06823250651359558,
0.025862865149974823,
-0.10328880697488785,
0.028922121971845627,
-0.07343999296426773,
-0.027929604053497314,
-0.007004591636359692,
0.005550774745643139,
0.040735531598329544,
-0.007612696848809719,
0.011298338882625103,
-0.028902238234877586,
-0.004485620651394129,
0.03846783563494682,
0.000722382974345237,
-0.0294436514377594,
0.06084354221820831,
0.02967423014342785,
0.03841903433203697,
0.04451727494597435,
-0.03578416258096695,
-0.051471054553985596,
-0.004096906166523695,
0.053236138075590134,
-0.0049537792801856995,
-0.07512456923723221,
-0.02344541996717453,
0.030932577326893806,
-0.04039556533098221,
0.013438494876027107,
0.09482541680335999,
-0.05282249301671982,
-0.006450552958995104,
0.04822740703821182,
0.008786017075181007,
-0.06461386382579803,
-0.02433578111231327,
-0.0372612401843071,
-0.026773659512400627,
0.011373844929039478,
0.05195054039359093,
-0.06207390874624252,
0.049149297177791595,
0.07324052602052689,
0.14033330976963043,
0.020483197644352913,
0.06138040870428085,
-0.062231361865997314,
-0.040588829666376114,
0.05427682027220726,
-0.07279647141695023,
-0.001377327018417418,
-0.06638149172067642,
0.1080692708492279,
0.046158935874700546,
-4.2038912795305805e-8,
-0.0018545440398156643,
0.024985281750559807,
0.009323658421635628,
0.029729802161455154,
0.011512055061757565,
0.022091727703809738,
-0.030888685956597328,
0.08744246512651443,
0.024600636214017868,
-0.05140110105276108,
0.06676139682531357,
-0.014926271513104439,
-0.0869465321302414,
-0.11524378508329391,
-0.053354911506175995,
0.060965340584516525,
0.00338348257355392,
-0.01136997900903225,
-0.0362403430044651,
0.010630382224917412,
-0.025415724143385887,
0.02426225319504738,
0.0226019024848938,
-0.02737272158265114,
0.06291701644659042,
-0.02569139376282692,
-0.00778889050707221,
0.09514494240283966,
0.03311707451939583,
0.07428675144910812,
0.010605958290398121,
-0.0706244558095932,
0.07432065159082413,
0.014565248042345047,
0.004036956932395697,
0.11918966472148895,
-0.025668915361166,
-0.02718086913228035,
-0.02673979103565216,
0.01813298650085926,
0.0002327628171769902,
0.05019956827163696,
-0.07010539621114731,
-0.03525683283805847,
0.002879603998735547,
-0.022074757143855095,
0.05493977665901184,
-0.03938484564423561,
0.019573858007788658,
0.007502967957407236,
0.006503446958959103,
-0.03323360159993172,
-0.05240240320563316,
0.08155962824821472,
-0.026856299489736557,
0.031084585934877396,
0.0033871394116431475,
0.02436077781021595,
0.023731954395771027,
0.061517491936683655,
0.059121523052453995,
-0.0180902648717165,
0.0028536072932183743,
-0.027891693636775017
] | 0
| 0
| 0
| 0
|
|
Do Contemporary Causal Inference Models Capture Real-World Heterogeneity? Findings from a Large-Scale Benchmark
|
https://openreview.net/forum?id=Q2bJ2qgcP1
|
[
"Haining Yu",
"Yizhou Sun"
] |
Poster
|
We present unexpected findings from a large-scale benchmark study evaluating Conditional Average Treatment Effect (CATE) estimation algorithms. By running 16 modern CATE models across 43,200 datasets, we find that: (a) 62\% of CATE estimates have a higher Mean Squared Error (MSE) than a trivial zero-effect predictor, rendering them ineffective; (b) in datasets with at least one useful CATE estimate,
80\% still have higher MSE than a constant-effect model; and (c) Orthogonality-based models outperform other models only 30\% of the time, despite widespread optimism about their performance. These findings expose significant limitations in current CATE models and suggest ample opportunities for further research.
Our findings stem from a novel application of \textit{observational sampling}, originally developed to evaluate Average Treatment Effect (ATE) estimates from observational methods with experiment data. To adapt observational sampling for CATE evaluation, we introduce a statistical parameter, $Q$, equal to MSE minus a constant and preserves the ranking of models by their MSE. We then derive a family of sample statistics, collectively called $\hat{Q}$, that can be computed from real-world data. We prove that $\hat{Q}$ is a consistent estimator of $Q$ under mild technical conditions. When used in observational sampling, $\hat{Q}$ is unbiased and asymptotically selects the model with the smallest MSE. To ensure the benchmark reflects real-world heterogeneity, we handpick datasets where outcomes come from field rather than simulation. By combining the new observational sampling method, new statistics, and real-world datasets, the benchmark provides a unique perspective on CATE estimator performance and uncover gaps in capturing real-world heterogeneity.
|
causal inference
|
We find contemporary CATE model fail to recover real-world heterogeneity through a large-scale real-world benchmark
| 11,132
|
2410.07021
|
[
-0.05231182277202606,
-0.0473858118057251,
0.09960228204727173,
0.11375508457422256,
0.06226827949285507,
0.03745103254914284,
-0.07038950175046921,
0.08671849220991135,
0.038211144506931305,
0.015872782096266747,
-0.06701593846082687,
0.0016313265077769756,
0.03224660083651543,
0.05196443572640419,
-0.043027181178331375,
-0.009316988289356232,
0.11312636733055115,
0.006632499862462282,
-0.06059838831424713,
-0.03338487073779106,
-0.048540808260440826,
-0.032151930034160614,
0.051569871604442596,
-0.035208217799663544,
0.015786610543727875,
-0.07003316283226013,
-0.06412500143051147,
-0.04697662591934204,
-0.020620565861463547,
0.003927033860236406,
-0.01815270446240902,
0.0683191791176796,
-0.09920093417167664,
-0.051254983991384506,
-0.06619685143232346,
-0.04928765073418617,
0.05514821410179138,
-0.029352443292737007,
-0.05046107620000839,
0.005968368146568537,
0.04331284761428833,
-0.0008720808546058834,
-0.0013464702060446143,
-0.010223007760941982,
0.05525539070367813,
-0.05949223041534424,
0.08816076070070267,
-0.023444484919309616,
-0.06664232164621353,
-0.00958381500095129,
-0.03810495138168335,
-0.04676598682999611,
0.05006366968154907,
0.014929189346730709,
0.015988238155841827,
-0.08039914816617966,
0.07983103394508362,
-0.023195631802082062,
0.0166698656976223,
0.07461075484752655,
-0.07289011776447296,
-0.05488130450248718,
-0.07245807349681854,
0.007985956035554409,
0.006795108783990145,
-0.014332078397274017,
-0.024403497576713562,
0.0601065419614315,
-0.07018358260393143,
0.02665618620812893,
-0.04680284857749939,
0.03422468528151512,
-0.07749659568071365,
-0.006765528116375208,
0.02125377207994461,
0.01782083325088024,
0.011894820258021355,
0.10064516216516495,
0.11036837100982666,
-0.07710949331521988,
0.022208426147699356,
-0.03322002664208412,
-0.024696100503206253,
-0.03576324135065079,
0.08971630781888962,
0.006192442029714584,
0.046807754784822464,
0.04707958176732063,
0.0312657430768013,
0.07714680582284927,
0.052526336163282394,
0.051210079342126846,
-0.01286994107067585,
-0.04719095677137375,
0.048567552119493484,
0.008785486221313477,
0.012186040170490742,
-0.028064630925655365,
-0.013609219342470169,
0.08375132083892822,
0.04908044636249542,
0.05647094547748566,
0.005634103436022997,
0.004835388623178005,
0.007924137637019157,
-0.05424577370285988,
0.02369721233844757,
0.021010030061006546,
0.05052034184336662,
-0.0697428360581398,
-0.06492658704519272,
-0.009316637180745602,
-0.008657503873109818,
0.04626286029815674,
-0.04031141847372055,
-0.0948154404759407,
0.040176086127758026,
-0.0020003574900329113,
0.0789564922451973,
0.010106207802891731,
0.024497486650943756,
-0.022854939103126526,
0.005932921543717384,
0.02438613586127758,
0.03324538469314575,
0.06608084589242935,
-0.11778250336647034,
3.511522682800741e-33,
0.0941326841711998,
0.005374636501073837,
0.0865984559059143,
-0.03549756854772568,
-0.00502442754805088,
-0.03228437528014183,
-0.05688254162669182,
0.002337757032364607,
0.052681826055049896,
-0.018896939232945442,
-0.01692725159227848,
-0.021525045856833458,
-0.01731807552278042,
-0.022059787064790726,
0.020140470936894417,
0.09822888672351837,
-0.0908166840672493,
0.10280311852693558,
-0.08043017983436584,
-0.03571896255016327,
0.06588849425315857,
-0.07104486972093582,
-0.024751368910074234,
0.041145991533994675,
-0.021195819601416588,
-0.0050505660474300385,
-0.009365295991301537,
-0.026022830978035927,
-0.06266847997903824,
-0.011709025129675865,
-0.032941803336143494,
-0.007062158081680536,
0.0974118784070015,
0.02007996290922165,
-0.06360644102096558,
-0.06278405338525772,
-0.04651074483990669,
-0.05930277332663536,
0.0026173957157880068,
0.010189524851739407,
-0.0371125265955925,
-0.08448995649814606,
0.08405344933271408,
-0.027052205055952072,
-0.00852199550718069,
-0.09370553493499756,
-0.056263670325279236,
0.012645473703742027,
-0.036268990486860275,
-0.006290796212852001,
0.015429669059813023,
0.07670864462852478,
0.02242271415889263,
-0.07190483808517456,
-0.07372786849737167,
0.05614404007792473,
-0.07908481359481812,
0.026250820606946945,
-0.0443231500685215,
-0.009260168299078941,
0.0005615082918666303,
0.03791182488203049,
-0.0001997327053686604,
-0.03393752500414848,
0.01966203562915325,
0.07773126661777496,
-0.022841332480311394,
-0.049034152179956436,
-0.007685365155339241,
0.09461253136396408,
-0.02781868912279606,
0.0368640199303627,
-0.03208518773317337,
0.01582878641784191,
0.06168650835752487,
-0.11727539449930191,
-0.005969015881419182,
-0.07196909934282303,
0.017893213778734207,
0.03719564527273178,
0.025095447897911072,
-0.00576408114284277,
-0.051811590790748596,
-0.05049928277730942,
-0.008902054280042648,
-0.0025540466886013746,
-0.016000809147953987,
-0.08219869434833527,
-0.01760951243340969,
-0.07088521122932434,
0.05179612711071968,
-0.020104069262742996,
-0.045974310487508774,
-0.0022721937857568264,
0.04185957461595535,
-5.4192963822555575e-33,
-0.005645868368446827,
-0.0021624930668622255,
0.08682150393724442,
-0.05029689148068428,
0.0065220072865486145,
-0.018506111577153206,
0.043041668832302094,
-0.030347120016813278,
0.0923672616481781,
-0.06638321280479431,
0.06411672383546829,
0.029592953622341156,
0.09266737848520279,
0.043976157903671265,
0.044489286839962006,
0.03765988349914551,
-0.01924605667591095,
-0.047192174941301346,
0.04487919807434082,
0.0002323917142348364,
0.09976211190223694,
0.0046274312771856785,
0.058870647102594376,
0.10032349079847336,
0.0018490551738068461,
-0.05695480853319168,
-0.01324809156358242,
0.056996047496795654,
0.0007436450687237084,
-0.08180345594882965,
-0.021874569356441498,
0.004199856426566839,
-0.047247160226106644,
-0.05086670070886612,
-0.035209301859140396,
0.07109559327363968,
0.07946998625993729,
0.001717478851787746,
-0.053928885608911514,
-0.02126391790807247,
-0.005110849626362324,
0.01644553802907467,
-0.07328161597251892,
0.03408040106296539,
0.02005143277347088,
0.015709398314356804,
-0.10719803720712662,
0.044607579708099365,
0.015463756397366524,
-0.0582682229578495,
-0.06572066992521286,
0.05690181255340576,
-0.03419782966375351,
0.056482262909412384,
-0.024473197758197784,
-0.13855820894241333,
-0.014730648137629032,
-0.0750739723443985,
-0.056391917169094086,
0.02567717432975769,
-0.05343038961291313,
0.006927500478923321,
-0.0650671124458313,
0.018565084785223007,
-0.04118713364005089,
0.013053844682872295,
0.06698481738567352,
-0.01725415326654911,
-0.009156489744782448,
-0.028610099107027054,
0.04100564122200012,
-0.015018132515251637,
-0.03900296241044998,
-0.04235334321856499,
-0.009633323177695274,
0.07420665770769119,
0.07409963756799698,
-0.04485076293349266,
-0.12790198624134064,
0.015677057206630707,
0.03949906304478645,
-0.04660821706056595,
0.04350699484348297,
-0.12022151798009872,
0.020547298714518547,
0.026749340817332268,
0.060722630470991135,
0.06147615611553192,
0.02787407487630844,
0.04014958068728447,
-0.10843983292579651,
-0.039777133613824844,
0.030675576999783516,
0.056517086923122406,
-0.04783928394317627,
-5.385354739928516e-8,
0.08735476434230804,
0.054344117641448975,
0.011792604811489582,
0.09138568490743637,
-0.040586937218904495,
-0.05900491774082184,
-0.12988214194774628,
0.04920407384634018,
-0.05834418162703514,
-0.013006186112761497,
0.009011653251945972,
0.013322402723133564,
-0.005298346281051636,
-0.017179882153868675,
0.053524672985076904,
0.042788323014974594,
-0.026739707216620445,
0.0006277902866713703,
-0.014000659808516502,
-0.013524461537599564,
-0.024207552894949913,
0.04106561094522476,
0.05296315625309944,
-0.05083533376455307,
0.09227075427770615,
-0.0118148447945714,
0.050150129944086075,
0.05250225588679314,
0.03816023841500282,
-0.028561411425471306,
0.03366087004542351,
0.01929680071771145,
-0.006115636322647333,
-0.07430458068847656,
0.02605786733329296,
-0.0017781120259314775,
0.07744037359952927,
0.010465167462825775,
0.015201478265225887,
-0.010647107847034931,
0.012698694132268429,
0.05685700476169586,
-0.025559956207871437,
0.027493709698319435,
0.07950467616319656,
0.013563734479248524,
-0.029902372509241104,
-0.032529182732105255,
0.06002530828118324,
-0.0524313785135746,
0.06678393483161926,
0.0005304189980961382,
-0.017463266849517822,
-0.024536432698369026,
0.06086844950914383,
0.03591504320502281,
0.06249960884451866,
-0.027023937553167343,
0.02508370205760002,
-0.035895224660634995,
0.1317177414894104,
-0.04460688680410385,
0.003035597503185272,
0.019185105338692665
] | 0
| 0
| 0
| 0
|
|
Harnessing Webpage UIs for Text-Rich Visual Understanding
|
https://openreview.net/forum?id=IIsTO4P3Ag
|
[
"Junpeng Liu",
"Tianyue Ou",
"Yifan Song",
"Yuxiao Qu",
"Wai Lam",
"Chenyan Xiong",
"Wenhu Chen",
"Graham Neubig",
"Xiang Yue"
] |
Poster
|
Text-rich visual understanding—the ability to interpret both textual content and visual elements within a scene—is crucial for multimodal large language models (MLLMs) to effectively interact with structured environments. We propose leveraging webpage UIs as a naturally structured and diverse data source to enhance MLLMs’ capabilities in this area. Existing approaches, such as rule-based extraction, multimodal model captioning, and rigid HTML parsing, are hindered by issues like noise, hallucinations, and limited generalization. To overcome these challenges, we introduce MultiUI, a dataset of 7.3 million samples spanning various UI types and tasks, structured using enhanced accessibility trees and task taxonomies. By scaling multimodal instructions from web UIs through LLMs, our dataset enhances generalization beyond web domains, significantly improving performance in document understanding, GUI comprehension, grounding, and advanced agent tasks. This demonstrates the potential of structured web data to elevate MLLMs’ proficiency in processing text-rich visual environments and generalizing across domains.
|
Multimodal, Instruction-tuning, Large Language Model
|
Leveraging structured webpage UIs, the MultiUI dataset enhances multimodal large language models’ ability to interpret text-rich visual environments, overcoming challenges in existing methods and boosting performance across diverse tasks.
| 11,128
|
2410.13824
|
[
-0.005468417424708605,
-0.1483776569366455,
0.010870314203202724,
0.02518511191010475,
0.16049082577228546,
-0.06693218648433685,
-0.056503910571336746,
-0.04274355620145798,
-0.041382454335689545,
-0.05717822536826134,
-0.026727696880698204,
-0.09672042727470398,
0.056528687477111816,
0.050892096012830734,
0.046727266162633896,
-0.012966889888048172,
0.0650349035859108,
0.08180312812328339,
-0.12521375715732574,
-0.006048986688256264,
0.03468455374240875,
0.028597988188266754,
0.04429034888744354,
-0.03804826736450195,
-0.0010963245294988155,
0.04777456074953079,
-0.0007254125666804612,
-0.016073230654001236,
0.061999812722206116,
-0.07535497099161148,
0.012343879789113998,
-0.026210829615592957,
0.030264871194958687,
0.09822830557823181,
-0.010065081529319286,
0.1177903264760971,
-0.03411330655217171,
0.05454026162624359,
0.007180580869317055,
-0.04584317281842232,
-0.05727476254105568,
-0.02842945232987404,
0.04889373108744621,
-0.0011106213787570596,
0.04681088402867317,
-0.06357060372829437,
-0.10477551817893982,
-0.10969111323356628,
0.035045325756073,
-0.0105818472802639,
-0.0693655014038086,
-0.040650684386491776,
0.06534060835838318,
0.0437728688120842,
-0.05881543830037117,
0.032194849103689194,
0.05943792313337326,
-0.04581727087497711,
0.021653372794389725,
-0.05625580623745918,
-0.05447481572628021,
-0.033422380685806274,
0.028942324221134186,
0.007219242863357067,
0.003907056991010904,
-0.005124945659190416,
0.004616487305611372,
-0.015190702863037586,
-0.010693036019802094,
-0.018282560631632805,
-0.0723533183336258,
-0.00435268459841609,
0.045023079961538315,
0.040520329028367996,
0.0431441105902195,
0.007233086042106152,
-0.032692838460206985,
-0.08676107972860336,
0.04531766474246979,
-0.07413410395383835,
0.02990926243364811,
0.0075661358423531055,
0.016473935917019844,
0.06543935835361481,
0.08401278406381607,
0.015760328620672226,
-0.045274145901203156,
0.06636729091405869,
-0.0034040219616144896,
0.0555911585688591,
-0.019426189363002777,
-0.09747976064682007,
0.10490337759256363,
0.010449448600411415,
0.006845897529274225,
0.05272276699542999,
0.05439944565296173,
-0.11586855351924896,
-0.06078166887164116,
0.053348664194345474,
0.0028648043517023325,
0.05926782265305519,
-0.014928075484931469,
-0.0599302276968956,
0.0030995612032711506,
-0.012413542717695236,
0.0834784284234047,
0.05470670387148857,
-0.001252835732884705,
-0.05669098347425461,
0.07090062648057938,
0.034751661121845245,
-0.019606666639447212,
-0.015468544326722622,
0.014224237762391567,
-0.10628578811883926,
0.02653018943965435,
-0.06004524603486061,
0.07301011681556702,
0.030145827680826187,
0.01101788878440857,
0.10091143101453781,
-0.01531188189983368,
-0.04905484616756439,
0.07754091918468475,
-0.009776226244866848,
-0.02269567921757698,
-1.639485664649853e-33,
0.12397442758083344,
-0.0007656968664377928,
0.015475478023290634,
0.07687098532915115,
0.0789797231554985,
-0.028972841799259186,
-0.09294087439775467,
0.02739131450653076,
-0.07408886402845383,
0.012062475085258484,
0.011213249526917934,
0.02445029653608799,
-0.06416648626327515,
0.06777609884738922,
-0.001242499565705657,
0.040393322706222534,
-0.06685129553079605,
0.01260808389633894,
-0.015782492235302925,
0.020285863429307938,
-0.02590007893741131,
-0.006553571671247482,
0.02401290088891983,
0.024210548028349876,
0.057217393070459366,
0.041108302772045135,
0.05845693126320839,
-0.03852580115199089,
-0.00906357727944851,
0.03659853711724281,
-0.13088935613632202,
-0.010844315402209759,
0.04354311525821686,
0.006151142530143261,
0.019448067992925644,
-0.03566271439194679,
-0.006575825624167919,
-0.07626736164093018,
0.0424073226749897,
-0.0013054326409474015,
-0.09011955559253693,
-0.07781343162059784,
-0.006496984977275133,
-0.042012158781290054,
-0.1250976175069809,
0.004310131072998047,
-0.003962601535022259,
0.0684879720211029,
-0.022664571180939674,
0.01615089178085327,
0.037608981132507324,
0.020214878022670746,
0.019411202520132065,
-0.0015052908565849066,
-0.01733318343758583,
-0.028682289645075798,
0.004401722457259893,
0.030256981030106544,
0.09692402929067612,
-0.017400335520505905,
0.019577134400606155,
0.08692439645528793,
0.0015501017915084958,
0.08697400242090225,
0.018938498571515083,
0.028345737606287003,
-0.054094716906547546,
0.09955867379903793,
0.0340140163898468,
-0.03287353739142418,
-0.0495198629796505,
0.029895596206188202,
-0.009936590678989887,
0.005525849759578705,
-0.001391479279845953,
0.030063768848776817,
0.06190072372555733,
-0.07311951369047165,
0.06935268640518188,
0.02881881594657898,
0.006993516348302364,
-0.02015296369791031,
0.01418889220803976,
-0.04914363473653793,
-0.03347837179899216,
-0.07993794232606888,
0.0282266978174448,
-0.12337035685777664,
0.00546127138659358,
-0.08452063798904419,
-0.018565015867352486,
-0.041938040405511856,
-0.054027676582336426,
-0.079630047082901,
-0.015633832663297653,
1.0562910502897495e-33,
0.018947632983326912,
-0.01500027161091566,
-0.03652315214276314,
-0.0021621035411953926,
0.01044065784662962,
-0.028750140219926834,
-0.014859705232083797,
0.021141983568668365,
0.027897171676158905,
-0.06476173549890518,
-0.014673835597932339,
0.03820396959781647,
-0.04905208200216293,
-0.03039662353694439,
0.015583721920847893,
0.004673072602599859,
0.06706199795007706,
0.0299778301268816,
-0.06201784685254097,
0.048559948801994324,
0.006441548466682434,
0.03461650386452675,
-0.05168062448501587,
0.004325603600591421,
-0.06276388466358185,
0.06224917247891426,
-0.07421868294477463,
0.03682907298207283,
-0.04641539230942726,
0.010917231440544128,
0.011010854505002499,
0.02858732081949711,
0.029728543013334274,
-0.0753113329410553,
-0.01539572887122631,
0.05887197330594063,
0.016368262469768524,
-0.05736566707491875,
-0.026066232472658157,
0.06100476533174515,
0.06256914883852005,
-0.03705712407827377,
-0.030836788937449455,
-0.023077966645359993,
-0.013760029338300228,
0.036189593374729156,
-0.10113722085952759,
-0.025337109342217445,
0.04929082468152046,
0.06588821858167648,
-0.033389799296855927,
-0.015561545267701149,
-0.05804690346121788,
-0.03624711558222771,
-0.05015011131763458,
-0.07659195363521576,
-0.027882445603609085,
-0.017358331009745598,
0.014725127257406712,
-0.008785871788859367,
-0.06148088723421097,
0.014569607563316822,
-0.08449015766382217,
0.05167604982852936,
0.00925371889024973,
-0.08306968212127686,
-0.05794994533061981,
-0.07074083387851715,
-0.03163037449121475,
-0.040040988475084305,
0.04602810740470886,
-0.11377494782209396,
0.0036879517138004303,
0.008607585914433002,
0.013651890680193901,
-0.00480486499145627,
0.051090896129608154,
0.00559049379080534,
0.023472096771001816,
-0.08743772655725479,
0.09990474581718445,
-0.03459109738469124,
0.05953781306743622,
0.027537986636161804,
0.11150205880403519,
0.05760825052857399,
-0.012729644775390625,
0.08180734515190125,
0.04280085489153862,
-0.0248201135545969,
-0.02949158474802971,
0.014171303249895573,
-0.0069186813198029995,
0.09792754799127579,
-0.05771711468696594,
-3.507406276526126e-8,
-0.08712292462587357,
0.03437042608857155,
-0.053720518946647644,
0.014002839103341103,
-0.020475512370467186,
-0.00908505730330944,
-0.03994845598936081,
0.04705321788787842,
0.04249652102589607,
-0.030362194404006004,
0.07333865016698837,
-0.02866075560450554,
-0.031659770756959915,
-0.049083661288022995,
0.029200255870819092,
0.09891947358846664,
-0.06532543152570724,
0.07106388360261917,
0.005286898463964462,
-0.03817038610577583,
0.043920014053583145,
-0.0067627448588609695,
-0.0033193391282111406,
0.0009843757143244147,
0.03996425122022629,
0.018780961632728577,
-0.06733057647943497,
0.08878562599420547,
0.0022160732187330723,
-0.018706312403082848,
0.03729749843478203,
0.008645792491734028,
-0.04542385786771774,
0.02702517993748188,
0.01880849525332451,
0.03433993458747864,
-0.015094575472176075,
-0.034745775163173676,
0.016835862770676613,
-0.04290211200714111,
0.027553630992770195,
0.05124390125274658,
-0.022884173318743706,
0.04245127737522125,
0.043887581676244736,
0.06431601941585541,
-0.03580528497695923,
-0.040271155536174774,
0.07422427088022232,
-0.025693386793136597,
-0.05586927384138107,
-0.05521523207426071,
0.08836544305086136,
0.07286175340414047,
0.05875442177057266,
-0.024377230554819107,
0.05255242809653282,
0.05403619259595871,
0.09481734782457352,
0.04027934372425079,
0.03318711742758751,
0.044305749237537384,
-0.0045211357064545155,
0.01633402146399021
] | 0
| 2
| 3
| 1
|
|
Fast unsupervised ground metric learning with tree-Wasserstein distance
|
https://openreview.net/forum?id=FBhKUXK7od
|
[
"Kira Michaela Düsterwald",
"Samo Hromadka",
"Makoto Yamada"
] |
Poster
|
The performance of unsupervised methods such as clustering depends on the choice of distance metric between features, or ground metric. Commonly, ground metrics are decided with heuristics or learned via supervised algorithms. However, since many interesting datasets are unlabelled, unsupervised ground metric learning approaches have been introduced. One promising option employs Wasserstein singular vectors (WSVs), which emerge when computing optimal transport distances between features and samples simultaneously. WSVs are effective, but can be prohibitively computationally expensive in some applications: $\mathcal{O}(n^2m^2(n \log(n) + m \log(m))$ for $n$ samples and $m$ features. In this work, we propose to augment the WSV method by embedding samples and features on trees, on which we compute the tree-Wasserstein distance (TWD). We demonstrate theoretically and empirically that the algorithm converges to a better approximation of the standard WSV approach than the best known alternatives, and does so with $\mathcal{O}(n^3+m^3+mn)$ complexity. In addition, we prove that the initial tree structure can be chosen flexibly, since tree geometry does not constrain the richness of the approximation up to the number of edge weights. This proof suggests a fast and recursive algorithm for computing the tree parameter basis set, which we find crucial to realising the efficiency gains at scale. Finally, we employ the tree-WSV algorithm to several single-cell RNA sequencing genomics datasets, demonstrating its scalability and utility for unsupervised cell-type clustering problems. These results poise unsupervised ground metric learning with TWD as a low-rank approximation of WSV with the potential for widespread application.
|
unsupervised learning, optimal transport, distance-based learning, clustering, trees, wasserstein distance
|
Unsupervised ground metric learning with tree-based optimal transport is computationally efficient yet still more accurate than alternative approximation appraoches, and scales favourably on genomics datasets
| 11,127
|
2411.07432
|
[
0.005549048539251089,
-0.10129018127918243,
0.03325953334569931,
0.026379365473985672,
0.1428091675043106,
-0.006099957041442394,
-0.014097468927502632,
0.0046896799467504025,
0.03209390491247177,
-0.029560338705778122,
0.011756309308111668,
-0.020717917010188103,
-0.04930165410041809,
0.09623686969280243,
-0.10374975204467773,
-0.015345419757068157,
0.11065074801445007,
0.04856078699231148,
-0.04007934406399727,
0.010333356447517872,
-0.010471915826201439,
0.06209094822406769,
0.031248105689883232,
0.004771220497786999,
0.08694814145565033,
0.0001305541954934597,
0.04668640345335007,
-0.006478017661720514,
-0.011917013674974442,
0.0006910608499310911,
0.0398055799305439,
-0.03984509035944939,
-0.005564318969845772,
-0.023469507694244385,
-0.07120334357023239,
0.11146662384271622,
0.03727582097053528,
0.07028084248304367,
-0.08391916751861572,
0.03987523540854454,
-0.01302857231348753,
0.08683230727910995,
0.04013175144791603,
0.09918974339962006,
-0.0030169093515723944,
0.08102284371852875,
-0.036188315600156784,
-0.017208963632583618,
-0.0018400290282443166,
0.007204117253422737,
-0.0027478805277496576,
-0.01414019800722599,
-0.06707378476858139,
0.10228326916694641,
0.003977430984377861,
-0.03215186670422554,
0.021478984504938126,
-0.02230500988662243,
0.0019251572666689754,
0.005689900368452072,
0.03537694364786148,
-0.09593424201011658,
-0.012533019296824932,
-0.06702303141355515,
0.005265016574412584,
-0.03634422644972801,
0.04391001537442207,
0.033480264246463776,
-0.011757920496165752,
0.005114628467708826,
-0.02785857394337654,
0.11441779136657715,
-0.12217847257852554,
0.046099547296762466,
-0.004871560260653496,
0.06903096288442612,
0.06556328386068344,
0.03796299919486046,
0.09701919555664062,
-0.06431880593299866,
0.028425714001059532,
0.06433160603046417,
0.01767786778509617,
-0.01915595307946205,
0.08166322112083435,
0.038258764892816544,
-0.04635579511523247,
0.07870296388864517,
0.03902319818735123,
-0.02376718632876873,
-0.03306574374437332,
-0.06446459144353867,
-0.06019704416394234,
-0.07734598219394684,
-0.0942685529589653,
-0.06554562598466873,
0.013315003365278244,
0.012865961529314518,
0.039717745035886765,
0.10502471029758453,
-0.011725308373570442,
0.04798831418156624,
-0.04897121340036392,
-0.06503324210643768,
0.03481324389576912,
0.027107300236821175,
0.0016085790703073144,
0.04263175278902054,
0.07366320490837097,
-0.018922151997685432,
0.08378207683563232,
-0.025432268157601357,
-0.0878719910979271,
0.06103485822677612,
-0.008965656161308289,
-0.003427698742598295,
0.08448922634124756,
-0.018178660422563553,
-0.05241944268345833,
0.04494692012667656,
-0.059786856174468994,
-0.03327262029051781,
0.04840848594903946,
0.040173616260290146,
0.06332814693450928,
-0.04686807096004486,
0.011092358268797398,
4.307609701505827e-33,
-0.010027158074080944,
0.05183122307062149,
0.040040623396635056,
-0.025677315890789032,
0.03298625349998474,
0.010833477601408958,
-0.037521444261074066,
0.033444296568632126,
0.013829081319272518,
0.05641857907176018,
-0.08695870637893677,
0.0018327528377994895,
0.030599422752857208,
-0.00594692537561059,
0.06385163217782974,
-0.025754554197192192,
-0.011559488251805305,
-0.013801639899611473,
-0.026037009432911873,
-0.05434542894363403,
0.08512316644191742,
-0.10205693542957306,
0.018566161394119263,
-0.026815567165613174,
-0.009644725359976292,
-0.07727938145399094,
-0.001261305995285511,
-0.022540049627423286,
0.0811106488108635,
-0.011383917182683945,
-0.0657210499048233,
-0.018439002335071564,
-0.040691111236810684,
0.052593592554330826,
0.015482855029404163,
-0.022142887115478516,
-0.041258733719587326,
-0.07262960076332092,
-0.00929891224950552,
-0.09871026128530502,
-0.010564330965280533,
0.010313354432582855,
0.008659638464450836,
-0.10426681488752365,
-0.05391409993171692,
0.0005100411362946033,
0.030024925246834755,
-0.06918235868215561,
0.0050489641726017,
-0.09774565696716309,
0.03296918049454689,
-0.02432176098227501,
0.0020808668341487646,
-0.03884442523121834,
-0.04865850508213043,
0.033850301057100296,
0.08504937589168549,
0.027123525738716125,
0.050316132605075836,
0.0785510241985321,
-0.022381870076060295,
0.023997383192181587,
0.04701612889766693,
-0.01872309111058712,
0.019380029290914536,
-0.009790833108127117,
0.04758448153734207,
0.04674094170331955,
0.007766289636492729,
-0.000025879975510179065,
0.0271455105394125,
0.006719415541738272,
-0.017760049551725388,
-0.059249039739370346,
-0.02794831246137619,
0.02315598353743553,
0.062325797975063324,
-0.039249468594789505,
-0.017141273245215416,
0.02052946574985981,
-0.053280241787433624,
0.017866063863039017,
-0.018707187846302986,
-0.0833042562007904,
-0.025014236569404602,
-0.037625085562467575,
0.017866669222712517,
-0.07407908141613007,
-0.10479853302240372,
-0.04557809606194496,
-0.024075865745544434,
0.06965509802103043,
-0.024732986465096474,
-0.04999816045165062,
0.04157394915819168,
-3.961148169964312e-33,
-0.04405837506055832,
0.04040980711579323,
0.06487922370433807,
0.031778208911418915,
-0.024302074685692787,
-0.016016172245144844,
-0.051785703748464584,
0.08906938135623932,
-0.025210164487361908,
-0.06064021587371826,
-0.05283506587147713,
-0.034198179841041565,
0.0522603914141655,
0.017773522064089775,
0.023648284375667572,
0.09364014863967896,
-0.03125636279582977,
0.05239549279212952,
-0.0077749392949044704,
0.004241885617375374,
0.05373717471957207,
0.08070214837789536,
-0.0524153932929039,
0.021124256774783134,
-0.046234987676143646,
0.017683546990156174,
0.03369317203760147,
0.045888375490903854,
-0.08597318828105927,
0.025843510404229164,
-0.011657381430268288,
-0.043593890964984894,
-0.09797615557909012,
-0.09989852458238602,
-0.04256828874349594,
0.05001292750239372,
-0.002532132901251316,
0.016671601682901382,
-0.03129101172089577,
0.13013234734535217,
0.0063616628758609295,
0.03864211589097977,
-0.05576556921005249,
0.009778586216270924,
-0.05322572961449623,
-0.04622874781489372,
-0.07531909644603729,
0.02644306793808937,
0.033935174345970154,
-0.02030981332063675,
-0.03693307563662529,
-0.021365435793995857,
-0.07060500234365463,
0.1053207665681839,
0.009645214304327965,
0.03158152475953102,
-0.03643420338630676,
0.03082648478448391,
0.031475894153118134,
-0.053682900965213776,
-0.0377223826944828,
-0.06019509211182594,
-0.06510742753744125,
0.02019161731004715,
0.01487291231751442,
-0.051383957266807556,
-0.012951277196407318,
-0.013460610993206501,
-0.07502710819244385,
0.045270927250385284,
-0.05093090608716011,
-0.03127550333738327,
0.004076051525771618,
0.01176503486931324,
-0.05549238249659538,
-0.01878354139626026,
0.10709910094738007,
0.005997573025524616,
-0.010867596603929996,
-0.004997538402676582,
0.023050466552376747,
0.010942966677248478,
0.008465706370770931,
0.07560136169195175,
0.11649419367313385,
0.028226740658283234,
0.06933820247650146,
-0.029034825041890144,
0.04378214851021767,
0.008194751106202602,
-0.028800267726182938,
0.048548273742198944,
0.048319507390260696,
0.027056878432631493,
-0.04287482798099518,
-3.980375495871158e-8,
-0.13190825283527374,
-0.032790202647447586,
0.048188574612140656,
-0.09975661337375641,
0.03183320164680481,
0.024128753691911697,
0.010823900811374187,
0.14499890804290771,
-0.09082698076963425,
0.01950216293334961,
0.017736095935106277,
-0.03424597159028053,
-0.09562413394451141,
0.047791607677936554,
-0.007046757731586695,
0.004905183333903551,
0.0026696170680224895,
0.046585358679294586,
-0.027312323451042175,
0.095132976770401,
-0.00656244158744812,
0.011828549206256866,
0.0662020817399025,
0.007688055746257305,
-0.015644315630197525,
-0.08141506463289261,
0.06880088895559311,
-0.010847209021449089,
0.06606701761484146,
-0.02687094919383526,
-0.06669852137565613,
0.057982731610536575,
0.016238482668995857,
-0.026631316170096397,
-0.02789197862148285,
0.08101051300764084,
-0.04555385932326317,
0.06678622961044312,
-0.08329793065786362,
0.03761601820588112,
-0.01847286894917488,
0.056854523718357086,
-0.055700965225696564,
0.015493442304432392,
0.002593828132376075,
0.07642904669046402,
-0.00048268501996062696,
0.024301301687955856,
-0.010922162793576717,
0.019317960366606712,
0.020526478067040443,
-0.047413382679224014,
0.005753338802605867,
0.028677649796009064,
0.058225709944963455,
-0.03673585504293442,
-0.07515013962984085,
-0.09148062765598297,
0.08229382336139679,
0.009068685583770275,
-0.00662445230409503,
-0.07004519551992416,
-0.0113868098706007,
-0.026143984869122505
] | 0
| 0
| 0
| 0
|
|
World Model on Million-Length Video And Language With Blockwise RingAttention
|
https://openreview.net/forum?id=HN8V0flwJF
|
[
"Hao Liu",
"Wilson Yan",
"Matei Zaharia",
"Pieter Abbeel"
] |
Poster
|
Enabling long-context understanding remains a key challenge in scaling existing sequence models -- a crucial component in developing generally intelligent models that can process and operate over long temporal horizons that potentially consist of millions of tokens. In this paper, we aim to address these challenges by providing a comprehensive exploration of the full development process for producing 1M context language models and video-language models, setting new benchmarks in language retrieval and new capabilities in long video understanding. We detail our long context data curation process, progressive context extension from 4K to 1M tokens, and present an efficient open-source implementation for scalable training on long sequences. Additionally, we open-source a family of 7B parameter models capable of processing long text documents and videos exceeding 1M tokens.
|
long context, world model, attention
| null | 11,124
|
2402.08268
|
[
0.004303147085011005,
-0.09511736035346985,
0.020246876403689384,
-0.04105392470955849,
0.09267888963222504,
0.0343233160674572,
-0.0406934916973114,
-0.006771037355065346,
0.018178200349211693,
-0.10721319913864136,
-0.01988064870238304,
-0.053278204053640366,
-0.012622677721083164,
0.06798344850540161,
-0.018217824399471283,
0.0009550047107040882,
0.020938193425536156,
0.08950156718492508,
-0.12212156504392624,
-0.058269284665584564,
0.027989616617560387,
0.008367259055376053,
0.06261562556028366,
-0.011110913939774036,
0.004957944620400667,
0.021651964634656906,
-0.00984637439250946,
-0.03155940771102905,
0.0935915857553482,
-0.03690493851900101,
0.06292546540498734,
0.04564041644334793,
0.045723218470811844,
0.07185322791337967,
-0.034630972892045975,
0.03853327035903931,
-0.0483752079308033,
-0.03188753128051758,
-0.05122550204396248,
-0.05656864121556282,
0.07087229192256927,
0.06413069367408752,
0.09243898838758469,
0.00945659913122654,
0.05310548096895218,
-0.012562842108309269,
0.02625826746225357,
-0.03615785390138626,
0.017952371388673782,
-0.010266199707984924,
-0.14427867531776428,
0.030220484361052513,
-0.0440022349357605,
0.07644163817167282,
-0.06231292337179184,
-0.029581252485513687,
-0.018673205748200417,
0.0035268987994641066,
0.003063098294660449,
0.006147280335426331,
-0.0335531048476696,
-0.11931987851858139,
-0.03639231622219086,
0.013933499343693256,
-0.0413304939866066,
-0.018065351992845535,
0.035148654133081436,
0.032257236540317535,
-0.0237937830388546,
0.007732544094324112,
-0.07953638583421707,
0.06931937485933304,
-0.039682820439338684,
0.1301521211862564,
-0.033026743680238724,
0.040252719074487686,
0.02050100639462471,
-0.039434146136045456,
0.03426128998398781,
-0.13059137761592865,
0.035644348710775375,
-0.05350532755255699,
0.15381252765655518,
-0.03501797094941139,
0.09322889149188995,
-0.004230517894029617,
0.024946678429841995,
0.028734970837831497,
-0.053447771817445755,
0.03370070829987526,
-0.05229828879237175,
-0.05334069952368736,
0.08558311313390732,
-0.03580240532755852,
0.02617661841213703,
0.02202342078089714,
0.012958494946360588,
-0.07670167833566666,
-0.06244165450334549,
0.09989988803863525,
0.0019454079447314143,
0.027421370148658752,
0.0045394585467875,
0.010949100367724895,
0.001459200750105083,
0.006500907242298126,
0.039161838591098785,
0.07197125256061554,
-0.0047940583899617195,
-0.041989974677562714,
0.0385160930454731,
0.04022214934229851,
-0.004842635244131088,
-0.05996792018413544,
0.03802701085805893,
-0.007719456218183041,
-0.03168207034468651,
-0.06535940617322922,
0.0025478776078671217,
0.08864599466323853,
0.00263881660066545,
-0.024567212909460068,
-0.08611129224300385,
-0.03905223682522774,
-0.07769821584224701,
-0.007854072377085686,
0.046546027064323425,
5.584694682229992e-34,
0.024028025567531586,
0.02472604624927044,
0.06395457684993744,
0.040074147284030914,
0.029606495052576065,
0.0183559600263834,
0.0209516528993845,
0.0699663981795311,
-0.06131723150610924,
-0.02274460345506668,
0.04540986195206642,
-0.017205119132995605,
-0.054693933576345444,
0.04961568862199783,
-0.002779521746560931,
-0.03573187068104744,
-0.036865800619125366,
0.04548873379826546,
0.00662919832393527,
0.04590709134936333,
0.04208708181977272,
-0.04410800710320473,
-0.002090056659653783,
0.033074911683797836,
0.013461865484714508,
-0.006849899422377348,
0.07438626140356064,
-0.0855172723531723,
0.0030104410834610462,
-0.0012311635073274374,
-0.11766605824232101,
0.008749393746256828,
0.01843641884624958,
0.02913585677742958,
0.09633028507232666,
-0.027576547116041183,
-0.033534932881593704,
-0.03502164036035538,
0.019181663170456886,
-0.07654856890439987,
-0.050504155457019806,
0.00989878736436367,
-0.008627834729850292,
-0.037906404584646225,
-0.1385800838470459,
-0.11468672007322311,
0.0426362007856369,
0.04845389723777771,
-0.06473060697317123,
0.046490300446748734,
0.11983103305101395,
0.028995223343372345,
-0.02882317081093788,
-0.058447930961847305,
0.011078773997724056,
0.008758212439715862,
0.07263188809156418,
0.013365410268306732,
0.017744550481438637,
0.05064250901341438,
0.024215705692768097,
-0.0014133467338979244,
0.0075020263902843,
0.11166644841432571,
0.026935381814837456,
0.005391815211623907,
-0.03326830640435219,
0.04971616342663765,
0.03693034127354622,
-0.01277440320700407,
-0.005853629671037197,
0.006710700690746307,
0.020790264010429382,
-0.004979792982339859,
0.0059822634793818,
0.016086572781205177,
0.01887676678597927,
-0.10789415240287781,
-0.025896253064274788,
0.0795917883515358,
0.01589147374033928,
-0.017462696880102158,
0.07305613905191422,
-0.050685860216617584,
0.013707065023481846,
-0.07233031839132309,
0.024723947048187256,
-0.11067884415388107,
0.017689431086182594,
-0.08294705301523209,
-0.021494265645742416,
-0.006297885440289974,
-0.012847254984080791,
-0.016686372458934784,
0.025291310623288155,
6.427974125700746e-34,
0.03420383110642433,
0.014937017112970352,
-0.02306419052183628,
0.034136589616537094,
0.017394084483385086,
-0.08697709441184998,
-0.011494201608002186,
0.1132025495171547,
-0.032434336841106415,
-0.05224606394767761,
0.02847699262201786,
-0.07322380691766739,
0.030671535059809685,
-0.01945677027106285,
0.021644853055477142,
-0.04764436185359955,
0.05243866890668869,
-0.00034408317878842354,
-0.02858494408428669,
0.028616556897759438,
0.008008508011698723,
0.006890800315886736,
-0.08879053592681885,
0.004126507788896561,
-0.06639094650745392,
0.03176788613200188,
-0.05347694829106331,
0.04879578948020935,
0.02417231723666191,
0.003487509675323963,
-0.039565227925777435,
-0.032406534999608994,
-0.016195373609662056,
-0.030455995351076126,
-0.08517765253782272,
0.011679576709866524,
0.08881766349077225,
0.002083159051835537,
-0.03433126211166382,
0.08246359974145889,
0.12075725942850113,
-0.012161731719970703,
-0.017272500321269035,
-0.003257918171584606,
-0.0031760120764374733,
-0.040388062596321106,
-0.06833431869745255,
0.027467627078294754,
-0.035087551921606064,
-0.05968364328145981,
-0.01997433975338936,
0.07514821738004684,
-0.05324028804898262,
0.005393521394580603,
0.002379904268309474,
-0.08393047004938126,
-0.05886876955628395,
-0.017604131251573563,
0.057359110563993454,
-0.05031794682145119,
-0.016630511730909348,
-0.029949035495519638,
-0.004774965345859528,
-0.029713187366724014,
0.06442269682884216,
-0.06625615060329437,
-0.002531446283683181,
0.0086787398904562,
-0.06628432869911194,
-0.020995916798710823,
0.029212428256869316,
-0.03244871646165848,
-0.055402252823114395,
0.060921043157577515,
-0.036679238080978394,
0.004611765034496784,
-0.004721983335912228,
-0.0671457052230835,
0.048630230128765106,
-0.03266469016671181,
-0.019946621730923653,
-0.03886035829782486,
0.03077836148440838,
0.025047406554222107,
0.11826595664024353,
0.0880875214934349,
0.0725671574473381,
0.08897233754396439,
0.027539143338799477,
0.07877369970083237,
-0.0185422170907259,
0.046573761850595474,
0.030940640717744827,
0.037923663854599,
-0.020083509385585785,
-3.2307429620459516e-8,
-0.09809128940105438,
0.04465826228260994,
-0.06721196323633194,
0.0038772926200181246,
-0.006805360782891512,
0.034509968012571335,
0.012772170826792717,
0.06686973571777344,
0.06153301149606705,
-0.023509236052632332,
0.037608250975608826,
-0.005444481503218412,
-0.025389451533555984,
-0.022074978798627853,
-0.06506971269845963,
0.08048661053180695,
0.00735508743673563,
0.023400072008371353,
0.0012506595812737942,
0.110826775431633,
0.06698044389486313,
0.07462695986032486,
0.0634288564324379,
-0.00559570174664259,
-0.012983893044292927,
-0.04382501170039177,
0.002312081167474389,
0.14480607211589813,
0.09231361001729965,
-0.1009172797203064,
-0.024805713444948196,
0.020305702462792397,
-0.023915406316518784,
-0.16347616910934448,
-0.046533968299627304,
0.031980354338884354,
0.0015289136208593845,
0.00944907683879137,
0.008913502097129822,
-0.01052234135568142,
0.05794445052742958,
0.023461870849132538,
-0.02309601567685604,
0.02170494943857193,
-0.016445647925138474,
0.03427474945783615,
0.014288485050201416,
-0.07398413866758347,
-0.016861949115991592,
-0.010396955534815788,
-0.02371881902217865,
-0.021360183134675026,
-0.0077510098926723,
0.031390853226184845,
0.06180606037378311,
0.03131642937660217,
0.023075658828020096,
0.046900052577257156,
0.024591051042079926,
0.04977044835686684,
-0.04884902387857437,
-0.016826527193188667,
-0.02725609391927719,
0.04416457191109657
] |
https://github.com/LargeWorldModel/LWM
| 7,269
| 82
| 0
| 18
|
Quamba: A Post-Training Quantization Recipe for Selective State Space Models
|
https://openreview.net/forum?id=mnna9LUg7P
|
[
"Hung-Yueh Chiang",
"Chi-Chih Chang",
"Natalia Frumkin",
"Kai-Chiang Wu",
"Diana Marculescu"
] |
Poster
|
State Space Models (SSMs) have emerged as an appealing alternative to Transformers for large language models, achieving state-of-the-art accuracy with constant memory complexity which allows for holding longer context lengths than attention-based networks. The superior computational efficiency of SSMs in long sequence modeling positions them favorably over Transformers in many scenarios. However, improving the efficiency of SSMs on request-intensive cloud-serving and resource-limited edge applications is still a formidable task. SSM quantization is a possible solution to this problem, making SSMs more suitable for wide deployment, while still maintaining their accuracy. Quantization is a common technique to reduce the model size and to utilize the low bit-width acceleration features on modern computing units, yet existing quantization techniques are poorly suited for SSMs. Most notably, SSMs have highly sensitive feature maps within the selective scan mechanism (i.e., linear recurrence) and massive outliers in the output activations which are not present in the output of token-mixing in the self-attention modules. To address this issue, we propose a static 8-bit per-tensor SSM quantization method which suppresses the maximum values of the input activations to the selective SSM for finer quantization precision and quantizes the output activations in an outlier-free space with Hadamard transform. Our 8-bit weight-activation quantized Mamba 2.8B SSM benefits from hardware acceleration and achieves a 1.72 $\times$ lower generation latency on an Nvidia Orin Nano 8G, with only a 0.9\% drop in average accuracy on zero-shot tasks. When quantizing Jamba, a 52B parameter SSM-style language model, we observe only a $1\%$ drop in accuracy, demonstrating that our SSM quantization method is both effective and scalable for large language models, which require appropriate compression techniques for deployment. The experiments demonstrate the effectiveness and practical applicability of our approach for deploying SSM-based models of all sizes on both cloud and edge platforms.
|
State Space Models, Model quantization
|
We propose a 8-bit static quantization method for selective state space models.
| 11,122
|
2410.13229
|
[
-0.004534534178674221,
-0.04762730747461319,
0.0031301698181778193,
0.03918488323688507,
-0.03987187519669533,
0.04993849992752075,
-0.02205265685915947,
-0.04448356479406357,
0.030465317890048027,
-0.05007422715425491,
-0.018175018951296806,
-0.06011934578418732,
-0.02126213349401951,
0.02195739932358265,
0.029040757566690445,
0.00938339438289404,
0.09785844385623932,
0.03069511614739895,
-0.03212885931134224,
-0.053450461477041245,
0.06453236937522888,
0.04445822536945343,
-0.05006125941872597,
0.008523434400558472,
0.039159368723630905,
0.026545310392975807,
0.0194743350148201,
-0.031086605042219162,
0.046244729310274124,
0.008669664151966572,
0.030601834878325462,
0.03065355122089386,
0.03185155242681503,
0.03871835768222809,
0.02556166984140873,
0.021930698305368423,
-0.09570780396461487,
-0.10625952482223511,
-0.031290993094444275,
-0.10262881219387054,
0.04635823890566826,
-0.012616490945219994,
-0.006652139127254486,
0.03650020435452461,
0.08214017748832703,
0.030691446736454964,
0.043633442372083664,
0.025955015793442726,
0.019811119884252548,
-0.053759586066007614,
-0.018236083909869194,
0.037555862218141556,
-0.06893694400787354,
0.12447050213813782,
-0.04458604380488396,
-0.038504961878061295,
0.06166552007198334,
0.06338817626237869,
-0.03777608647942543,
-0.01429192628711462,
-0.09131384640932083,
-0.08532516658306122,
0.05711665004491806,
-0.026844818145036697,
-0.006253632251173258,
0.004234558902680874,
0.04354412481188774,
0.04051588475704193,
-0.03185878321528435,
0.020509332418441772,
-0.029618453234434128,
0.009108543395996094,
-0.04570648819208145,
0.04375510290265083,
-0.019802549853920937,
0.042729590088129044,
0.05160767585039139,
0.0038396078161895275,
0.0906481221318245,
-0.031301528215408325,
-0.018286438658833504,
-0.022507578134536743,
0.0571599043905735,
0.018984083086252213,
0.060428887605667114,
-0.03915836662054062,
-0.04844537377357483,
0.008008002303540707,
0.015821073204278946,
0.008251793682575226,
-0.0014567181933671236,
-0.04327857121825218,
0.05065973848104477,
-0.03810638561844826,
0.09199139475822449,
0.02059037610888481,
0.004441861063241959,
0.007797003258019686,
-0.03578148037195206,
0.037862882018089294,
0.06021815165877342,
0.03486369550228119,
0.027037357911467552,
-0.03140493482351303,
0.031022807583212852,
0.014539423398673534,
-0.00406588613986969,
0.07782402634620667,
-0.0008431262103840709,
-0.08469270169734955,
-0.005036340095102787,
0.05530724301934242,
0.017877332866191864,
-0.04815010353922844,
0.020503390580415726,
-0.00016126288392115384,
0.001027388614602387,
-0.035529591143131256,
0.01036430150270462,
0.04373670369386673,
-0.03118159994482994,
-0.04320371896028519,
-0.028327956795692444,
0.04234261438250542,
-0.007470706943422556,
0.004826247226446867,
-0.12661387026309967,
2.1445841995966665e-33,
0.03297653794288635,
0.03219053894281387,
0.014913498423993587,
-0.03589197248220444,
0.04193660616874695,
-0.0003085185308009386,
0.07799836993217468,
0.043330904096364975,
-0.01916380226612091,
0.018354415893554688,
-0.04861259087920189,
0.0481591634452343,
0.002593273064121604,
0.0007958486676216125,
0.05977297201752663,
-0.04492504894733429,
-0.04132332280278206,
0.028420977294445038,
0.0025595054030418396,
-0.034702613949775696,
0.04350139573216438,
-0.03877796605229378,
0.013230311684310436,
0.006764689926058054,
0.062130801379680634,
0.017572320997714996,
0.10454369336366653,
0.002083346713334322,
-0.06935101002454758,
0.007988613098859787,
-0.10816159099340439,
-0.007633205968886614,
-0.04852983355522156,
-0.03440199792385101,
0.08780933171510696,
-0.05504896491765976,
0.009783229790627956,
-0.03134792298078537,
-0.022012941539287567,
-0.07559067755937576,
0.03292381018400192,
0.04340876638889313,
0.013103940524160862,
-0.01213237177580595,
-0.11312054097652435,
-0.05412627011537552,
-0.01891408860683441,
-0.07591850310564041,
0.013347689993679523,
0.009819497354328632,
0.09814688563346863,
-0.01716957613825798,
-0.08033741265535355,
-0.0017837604973465204,
0.05854889377951622,
-0.04705999046564102,
0.10115931928157806,
0.0642760768532753,
0.0557316355407238,
0.10029244422912598,
-0.11646567285060883,
-0.08616000413894653,
0.008350902236998081,
0.05703260377049446,
0.030059246346354485,
-0.0135112926363945,
-0.018219368532299995,
0.13335169851779938,
-0.05569867044687271,
0.01596050150692463,
-0.039926737546920776,
-0.01780969463288784,
0.03397335112094879,
-0.040964312851428986,
0.06729844957590103,
-0.08092700690031052,
0.09881081432104111,
-0.05905057489871979,
-0.04925581067800522,
-0.01950785703957081,
0.011647582054138184,
-0.001072116196155548,
-0.03134201839566231,
-0.007883795537054539,
-0.03845730423927307,
-0.06931377202272415,
0.030034439638257027,
-0.0312761627137661,
-0.0645660012960434,
-0.01599079556763172,
-0.015916908159852028,
-0.031483881175518036,
0.03497261926531792,
0.04345155507326126,
-0.13557952642440796,
-1.746548404425801e-33,
-0.06604485958814621,
0.011672533117234707,
-0.04967841878533363,
0.08318549394607544,
-0.04890451952815056,
-0.14155010879039764,
0.07232604175806046,
0.06597937643527985,
-0.0036737732589244843,
-0.028746677562594414,
-0.07009013742208481,
-0.05332816392183304,
0.10051216930150986,
0.035176198929548264,
0.017971118912100792,
0.05596736818552017,
-0.012690496630966663,
-0.019320189952850342,
-0.02961791679263115,
0.09775472432374954,
-0.03798350691795349,
0.04639740660786629,
-0.029569193720817566,
-0.05055101215839386,
-0.03537837043404579,
0.038819096982479095,
-0.04885796457529068,
0.0592239610850811,
0.04932992905378342,
-0.03249245882034302,
-0.020418871194124222,
-0.07129866629838943,
-0.07686144858598709,
-0.028156736865639687,
-0.04315986484289169,
0.003561814548447728,
0.06776472181081772,
0.051932189613580704,
0.012383456341922283,
0.09005524218082428,
0.07789100706577301,
-0.09595470875501633,
-0.09302708506584167,
0.11677863448858261,
-0.011481772176921368,
0.03424133360385895,
-0.06055159866809845,
0.04537343978881836,
-0.05938732996582985,
0.01431731041520834,
-0.052580393850803375,
0.044480521231889725,
-0.030800754204392433,
0.09927982836961746,
-0.04206272214651108,
-0.05912737548351288,
-0.06829112768173218,
0.019381465390324593,
-0.008527196012437344,
-0.01980479247868061,
-0.014443262480199337,
-0.07514249533414841,
0.08994555473327637,
-0.13696148991584778,
0.027329424396157265,
-0.05615473538637161,
-0.012155252508819103,
0.003681076457723975,
-0.015070383436977863,
-0.014661679975688457,
-0.027470121160149574,
-0.035033904016017914,
-0.03240750730037689,
0.013974525965750217,
-0.03884901851415634,
-0.03713984042406082,
0.034101080149412155,
-0.07292784750461578,
0.07844177633523941,
0.005425641313195229,
-0.08543122559785843,
0.09583049267530441,
-0.034673720598220825,
-0.013403864577412605,
0.022893017157912254,
0.08183229714632034,
0.06572602689266205,
0.06890667229890823,
-0.010077610611915588,
0.018823930993676186,
-0.06707300990819931,
0.05426321178674698,
0.006380193401128054,
0.10045435279607773,
-0.044771578162908554,
-4.626193828016767e-8,
-0.11142218112945557,
0.06292109191417694,
0.017535394057631493,
-0.013777690008282661,
0.09899202734231949,
-0.02811015024781227,
0.013103549368679523,
0.09173166751861572,
0.016012301668524742,
-0.022074704989790916,
0.05952141806483269,
-0.03261063992977142,
-0.07135459780693054,
-0.05768387392163277,
-0.018833132460713387,
0.07253344357013702,
-0.0010050387354567647,
-0.017025578767061234,
0.004720838740468025,
-0.007311674766242504,
-0.009443012066185474,
0.08398447185754776,
0.028567973524332047,
0.035092588514089584,
0.0007611415931023657,
-0.020174665376544,
0.05014612898230553,
0.08539171516895294,
0.05753441900014877,
0.007618020288646221,
-0.015206335112452507,
-0.0062170326709747314,
0.018702136352658272,
-0.014107904396951199,
-0.0026228874921798706,
0.03497022017836571,
0.015622814185917377,
-0.025822874158620834,
0.06042901426553726,
0.03185242414474487,
0.007354776840656996,
0.007614902686327696,
-0.08808507025241852,
0.06752616912126541,
-0.011676709167659283,
0.01091919094324112,
0.031186405569314957,
-0.03260137885808945,
0.01529594324529171,
0.00014151519280858338,
0.037093695253133774,
-0.044780172407627106,
-0.08759067952632904,
0.06384564936161041,
-0.004284895956516266,
0.049670275300741196,
0.056926269084215164,
-0.05715460330247879,
0.05723531171679497,
0.11196976900100708,
-0.01329693105071783,
-0.038838960230350494,
-0.06491182744503021,
0.012855559587478638
] |
https://github.com/enyac-group/quamba
| 40
| 0
| 0
| 0
|
Learning to engineer protein flexibility
|
https://openreview.net/forum?id=L238BAx0wP
|
[
"Petr Kouba",
"Joan Planas-Iglesias",
"Jiri Damborsky",
"Jiri Sedlar",
"Stanislav Mazurenko",
"Josef Sivic"
] |
Poster
|
Generative machine learning models are increasingly being used to design novel proteins. However, their major limitation is the inability to account for protein flexibility, a property crucial for protein function. Learning to engineer flexibility is difficult because the relevant data is scarce, heterogeneous, and costly to obtain using computational and experimental methods. Our contributions are three-fold. First, we perform a comprehensive comparison of methods for evaluating protein flexibility and identify relevant data for learning. Second, we overcome the data scarcity issue by leveraging a pre-trained protein language model. We design and train flexibility predictors utilizing either only sequential or both sequential and structural information on the input. Third, we introduce a method for fine-tuning a protein inverse folding model to make it steerable toward desired flexibility at specified regions. We demonstrate that our method Flexpert enables guidance of inverse folding models toward increased flexibility. This opens up a transformative possibility of engineering protein flexibility.
|
protein flexibility, protein flexibility prediction, protein design, protein sequence design, inverse folding
| null | 11,118
|
2412.18275
|
[
-0.08698678016662598,
-0.11282771825790405,
0.07358492165803909,
0.04306920990347862,
0.024916358292102814,
-0.008905072696506977,
-0.04735160246491432,
0.06201276555657387,
-0.02170698344707489,
-0.06573627889156342,
0.00901948194950819,
-0.023020252585411072,
-0.061932459473609924,
0.05775437876582146,
-0.0040998077020049095,
0.03074972704052925,
-0.017200179398059845,
0.04801459610462189,
-0.047817688435316086,
-0.07300744205713272,
0.05719401687383652,
0.02212127484381199,
0.00674850819632411,
-0.007730318699032068,
0.05579196661710739,
-0.0011021359823644161,
0.01744316890835762,
0.00032307993387803435,
-0.003280292497947812,
-0.06556876003742218,
0.04547039791941643,
-0.02921309694647789,
-0.06713642925024033,
-0.023030554875731468,
-0.00010636747174430639,
0.08284096419811249,
-0.013091037049889565,
-0.004035712219774723,
0.004936951212584972,
-0.03418586775660515,
0.0471121184527874,
0.02651132456958294,
0.054598115384578705,
0.04563354700803757,
0.03023526817560196,
0.0015597805613651872,
-0.028567740693688393,
-0.05803857371211052,
0.049569617956876755,
0.04490675777196884,
-0.002217062283307314,
-0.014414816163480282,
-0.03455400839447975,
0.029110996052622795,
0.0019980454817414284,
0.006238772999495268,
-0.08287716656923294,
0.023778919130563736,
-0.008763842284679413,
0.01566564291715622,
-0.004700508899986744,
-0.12286870926618576,
-0.03214668482542038,
-0.05119150131940842,
0.005381623283028603,
-0.06031002104282379,
0.029084865003824234,
0.08118437975645065,
-0.048845283687114716,
0.013099790550768375,
0.002867167117074132,
0.08341684937477112,
-0.07094937562942505,
0.03248613700270653,
0.11693089455366135,
0.028577139601111412,
-0.015450638718903065,
0.019540928304195404,
0.07399779558181763,
-0.01183246448636055,
-0.030673643574118614,
0.009540684521198273,
0.07330258935689926,
0.06453313678503036,
0.07781024277210236,
0.011728583835065365,
0.021357744932174683,
0.07266121357679367,
0.044534362852573395,
0.0181598961353302,
0.01813676953315735,
-0.06536813080310822,
0.06681060791015625,
-0.050748880952596664,
-0.07744651287794113,
0.0417749285697937,
0.008383049629628658,
-0.11104089766740799,
-0.05944118648767471,
0.046762652695178986,
-0.024340789765119553,
0.081276074051857,
0.10101506859064102,
-0.06345384567975998,
0.025955792516469955,
-0.08784647285938263,
-0.014781643636524677,
0.06764896959066391,
0.03084598295390606,
-0.022619586437940598,
0.0695958212018013,
0.03024989739060402,
0.07066875696182251,
-0.06070905551314354,
0.015683066099882126,
0.024534009397029877,
-0.11412574350833893,
-0.05696391314268112,
-0.011011993512511253,
0.04354223608970642,
-0.06242731213569641,
0.03149125352501869,
-0.02682328410446644,
0.06790423393249512,
-0.013587416149675846,
-0.03830612823367119,
-0.11332317441701889,
2.864617743994008e-34,
0.017339879646897316,
0.012304958887398243,
0.041376993060112,
0.09848421812057495,
0.03440704941749573,
-0.08010414987802505,
-0.041523851454257965,
-0.011829400435090065,
-0.05818173661828041,
0.02946888469159603,
-0.0038551539182662964,
0.01455676555633545,
-0.03292522951960564,
0.0550217404961586,
0.07118525356054306,
-0.009510725736618042,
0.011791540309786797,
0.11905542761087418,
-0.06285028904676437,
-0.022415513172745705,
-0.06282857060432434,
0.009878098964691162,
0.02944142557680607,
-0.024131517857313156,
0.03070647083222866,
-0.07056919485330582,
-0.042487286031246185,
-0.02395850419998169,
-0.07464504987001419,
0.004010173492133617,
-0.1353335678577423,
0.0221443809568882,
-0.041287634521722794,
0.017752032727003098,
0.03264015540480614,
-0.023239262402057648,
0.06519501656293869,
-0.07684945315122604,
-0.029828721657395363,
0.012049752287566662,
0.037124890834093094,
0.016862554475665092,
0.06939423829317093,
-0.018808258697390556,
-0.04050925746560097,
-0.019734637811779976,
0.08031099289655685,
0.009011035785079002,
-0.09727957844734192,
-0.051056139171123505,
0.048986028879880905,
0.027779338881373405,
0.04392024874687195,
-0.017702896147966385,
-0.030391937121748924,
0.029460906982421875,
-0.011840380728244781,
0.02126391790807247,
0.04671787470579147,
0.0673932358622551,
-0.004874423611909151,
0.026489781215786934,
-0.059122826904058456,
0.05196887254714966,
0.01730896718800068,
-0.009421377442777157,
-0.012247982434928417,
0.00978102721273899,
0.03444381430745125,
0.10700608044862747,
-0.06459172070026398,
0.0027065491303801537,
-0.007137177046388388,
-0.06919348984956741,
0.0020796500612050295,
0.02617204375565052,
-0.0006598561885766685,
-0.09076539427042007,
0.023529740050435066,
0.053977496922016144,
-0.032386697828769684,
0.034109584987163544,
-0.007160399574786425,
0.005066267680376768,
-0.041208479553461075,
-0.12880133092403412,
0.11205771565437317,
-0.06634734570980072,
-0.06808798015117645,
-0.026989586651325226,
-0.09780286252498627,
-0.038874536752700806,
0.003918392583727837,
-0.010370380245149136,
0.007588465232402086,
-5.028012240926731e-34,
0.011631647124886513,
-0.019624793902039528,
0.010454623028635979,
-0.019454475492239,
-0.03544146195054054,
0.006520192138850689,
0.03962584212422371,
0.07848820090293884,
-0.010501910001039505,
-0.11362560093402863,
0.01999199576675892,
-0.04115860536694527,
0.06198287755250931,
-0.0020734022837132215,
-0.042450327426195145,
0.09278511255979538,
-0.1089109256863594,
0.024729756638407707,
0.027199573814868927,
0.07172571867704391,
-0.05187378451228142,
0.07136812806129456,
-0.020395657047629356,
0.04840138182044029,
-0.034140899777412415,
0.07421676814556122,
-0.05006277933716774,
0.05652480199933052,
0.00513051263988018,
0.033858273178339005,
-0.11573368310928345,
-0.009295651689171791,
-0.027949435636401176,
-0.036227062344551086,
-0.10099178552627563,
0.06848517805337906,
-0.01919756643474102,
-0.0183134563267231,
0.13736921548843384,
0.00423565786331892,
0.04097508266568184,
0.028641894459724426,
-0.08474378287792206,
0.04320766031742096,
0.036632925271987915,
-0.027586141601204872,
0.026248982176184654,
0.018264364451169968,
0.006752682849764824,
0.00030155619606375694,
0.007734384853392839,
0.01869768649339676,
-0.018039638176560402,
-0.06727355718612671,
-0.09738606214523315,
-0.08536587655544281,
-0.035599611699581146,
-0.02289384976029396,
0.055888112634420395,
-0.0005043306155130267,
-0.076458640396595,
-0.05893633887171745,
0.07031174749135971,
-0.012881209142506123,
-0.026888055726885796,
0.009616817347705364,
0.023789746686816216,
-0.07048383355140686,
-0.0011400155490264297,
-0.050945840775966644,
0.03185759112238884,
0.02114984020590782,
0.09221567213535309,
0.06943152099847794,
0.08230197429656982,
-0.022509483620524406,
-0.020644070580601692,
-0.03841733932495117,
-0.02527421899139881,
-0.0431549958884716,
0.002589631825685501,
0.006328150629997253,
-0.04383470118045807,
0.04153051972389221,
0.08955790102481842,
0.06831788271665573,
0.06956562399864197,
0.09180374443531036,
0.08437000960111618,
0.03954127058386803,
-0.01576026901602745,
0.037611618638038635,
0.0033811344765126705,
0.15543724596500397,
0.023167815059423447,
-3.5418956656485534e-8,
-0.011049208231270313,
0.011360353790223598,
0.07586294412612915,
0.025907838717103004,
0.05797659978270531,
0.1006663367152214,
0.019545746967196465,
0.024784857407212257,
-0.023935062810778618,
-0.030249079689383507,
0.020586250349879265,
0.03615593910217285,
-0.04461535066366196,
-0.01650686375796795,
-0.016093827784061432,
0.08382076770067215,
-0.007015074137598276,
0.07170473784208298,
-0.04141625016927719,
0.009684444405138493,
0.017764395102858543,
0.039913009852170944,
0.02268998697400093,
-0.06001869589090347,
0.058110713958740234,
-0.12532973289489746,
-0.1304638534784317,
0.024084623903036118,
0.030438140034675598,
-0.014221073128283024,
-0.005806456785649061,
0.00527731841430068,
-0.015132784843444824,
0.009171763435006142,
-0.03410184010863304,
0.06078466400504112,
0.004997890908271074,
-0.06816209852695465,
-0.01952979527413845,
-0.011429594829678535,
0.023723537102341652,
0.08457806706428528,
-0.062268465757369995,
-0.013539265841245651,
-0.033857088536024094,
0.0038558603264391422,
0.046618979424238205,
-0.05542449653148651,
-0.004068587440997362,
0.02620627172291279,
0.028090713545680046,
-0.024803489446640015,
0.06041713431477547,
-0.06016340106725693,
-0.00956001691520214,
0.03059886209666729,
-0.02673603594303131,
0.007686924654990435,
-0.026177670806646347,
0.005375100765377283,
-0.07116445153951645,
-0.017926255241036415,
0.007381385192275047,
-0.012369037605822086
] |
https://github.com/KoubaPetr/Flexpert
| 16
| 0
| 0
| 0
|
ElasticTok: Adaptive Tokenization for Image and Video
|
https://openreview.net/forum?id=tFV5GrWOGm
|
[
"Wilson Yan",
"Volodymyr Mnih",
"Aleksandra Faust",
"Matei Zaharia",
"Pieter Abbeel",
"Hao Liu"
] |
Poster
|
Efficient video tokenization remains a key bottleneck in learning general purpose vision models that are capable of processing long video sequences. Prevailing approaches are restricted to encoding videos to a fixed number of tokens, where too few tokens will result in overly lossy encodings, and too many tokens will result in prohibitively long sequence lengths. In this work, we introduce ElasticTok, a method that conditions on prior frames to adaptively encode a frame into a variable number of tokens. To enable this in a computationally scalable way, we propose a masking technique that drops a random number of tokens at the end of each frames's token encoding. During inference, ElasticTok can dynamically allocate tokens when needed -- more complex data can leverage more tokens, while simpler data only needs a few tokens. Our empirical evaluations on images and video demonstrate the effectiveness of our approach in efficient token usage, paving the way for future development of more powerful multimodal models, world models, and agents. Video examples of using ElasticTok can be found on our website: http://largeworldmodel.github.io/elastictok
|
adaptive representation, adaptive tokenization, autoencoder
| null | 11,116
|
2410.08368
|
[
-0.005313071887940168,
-0.09210032969713211,
0.01770961843430996,
-0.023198554292321205,
0.12842939794063568,
-0.008396631106734276,
0.00022582654491998255,
-0.0678856372833252,
0.09795771539211273,
-0.06150637939572334,
-0.045580457895994186,
-0.03365898132324219,
-0.0016772508388385177,
0.04803483933210373,
0.009615267626941204,
-0.05289797484874725,
0.029760701581835747,
0.14604060351848602,
-0.14441511034965515,
-0.07236924767494202,
0.06235364452004433,
-0.08227495104074478,
0.027948183938860893,
-0.035157639533281326,
0.022588346153497696,
0.030407056212425232,
-0.011697559617459774,
-0.07125207781791687,
0.04393067955970764,
-0.037888187915086746,
0.05466228723526001,
-0.06265094876289368,
0.021016374230384827,
0.04210473597049713,
-0.029552901163697243,
0.043897707015275955,
-0.06503773480653763,
-0.009785771369934082,
-0.04129902645945549,
-0.023159721866250038,
-0.010419649071991444,
0.039237603545188904,
-0.008096586912870407,
-0.043556977063417435,
0.02990054339170456,
0.04446829482913017,
0.0969715341925621,
-0.07995786517858505,
0.00930975005030632,
0.001847389037720859,
-0.04628613963723183,
-0.05100006237626076,
-0.06282512098550797,
0.02021089382469654,
-0.0917736068367958,
-0.07645533978939056,
-0.013110260479152203,
-0.03381660208106041,
0.07925324887037277,
0.02647782675921917,
0.008138722740113735,
-0.09638950228691101,
0.0912780836224556,
0.05157889425754547,
-0.020809102803468704,
-0.051541805267333984,
0.041215766221284866,
0.010826067067682743,
0.034764211624860764,
0.0136112654581666,
-0.05121096223592758,
0.05963156744837761,
-0.06065993383526802,
0.06828805059194565,
-0.016414161771535873,
0.02893669158220291,
0.07199207693338394,
0.000831610057502985,
0.06949056684970856,
-0.09640007466077805,
0.07913797348737717,
-0.08198834210634232,
0.06202343851327896,
0.006729591637849808,
0.13989529013633728,
-0.016910336911678314,
-0.000860013475175947,
0.026616835966706276,
-0.019798366352915764,
0.045132558792829514,
-0.12266905605792999,
-0.005021284334361553,
0.12813572585582733,
-0.02104135788977146,
0.014546741731464863,
-0.051352933049201965,
-0.030642220750451088,
-0.04652317985892296,
-0.008327004499733448,
0.10578960180282593,
-0.037937019020318985,
-0.008533659391105175,
0.026509378105401993,
0.009492725133895874,
0.10181616991758347,
0.023403070867061615,
0.04246928170323372,
0.058047130703926086,
0.06493382155895233,
0.009981650859117508,
0.04022817686200142,
0.09452760219573975,
-0.02823800779879093,
0.03391275182366371,
-0.003574660047888756,
0.02564225159585476,
-0.02722771465778351,
-0.010204924270510674,
0.005229429807513952,
0.1115608662366867,
0.10774116963148117,
-0.06940687447786331,
-0.040189675986766815,
-0.012619814835488796,
-0.035641297698020935,
-0.022617336362600327,
-0.007505480665713549,
1.3748947432492183e-33,
-0.009706814773380756,
0.022249991074204445,
0.03987971320748329,
-0.015173042193055153,
0.05988285690546036,
0.025437429547309875,
-0.0023549094330519438,
0.040665529668331146,
-0.0634007677435875,
-0.04973169043660164,
0.005144997034221888,
-0.03521822392940521,
0.00113824219442904,
0.10427697747945786,
0.006917130202054977,
-0.07780817151069641,
-0.02924502268433571,
0.006554960738867521,
0.07486944645643234,
0.03381533920764923,
0.005067635793238878,
-0.05888819321990013,
-0.052273839712142944,
0.006359400227665901,
-0.020343920215964317,
-0.026635704562067986,
0.08884314447641373,
-0.10516706109046936,
0.06515800207853317,
0.016512449830770493,
-0.09507796913385391,
-0.008365594781935215,
0.027113867923617363,
0.0412420853972435,
0.06467676907777786,
-0.03727685660123825,
0.01685241609811783,
-0.019548965618014336,
-0.0019131307490170002,
-0.0324675552546978,
-0.027807822450995445,
0.025531576946377754,
-0.0682060644030571,
-0.032612890005111694,
-0.08762343227863312,
-0.05486825108528137,
0.05181785672903061,
0.0651634931564331,
-0.0523826964199543,
0.05307488515973091,
0.0991731807589531,
0.019995257258415222,
-0.021951377391815186,
-0.1329500377178192,
-0.020534442737698555,
-0.011425521224737167,
0.005157098639756441,
-0.02021673507988453,
-0.007386316079646349,
0.00097703339997679,
-0.03936658427119255,
-0.015393693931400776,
0.031205084174871445,
0.10513099282979965,
-0.01328649464994669,
-0.027091898024082184,
0.06923987716436386,
-0.0034832817036658525,
-0.004239145666360855,
0.006571033503860235,
-0.016197534278035164,
0.046661488711833954,
0.002645105356350541,
-0.05038434639573097,
-0.05131852626800537,
0.03390943631529808,
0.05295827239751816,
-0.02278534695506096,
-0.09302733093500137,
0.06143500655889511,
-0.02706117555499077,
-0.005011946894228458,
0.05779964476823807,
0.0027394911739975214,
-0.0331123024225235,
-0.050297778099775314,
0.010455158539116383,
-0.06765308976173401,
-0.02838950976729393,
-0.07374589890241623,
-0.024342354387044907,
-0.04265508055686951,
0.00587575463578105,
-0.017474623396992683,
0.0022409388329833746,
-1.0062662351515762e-33,
0.011356081813573837,
0.07912709563970566,
-0.018107157200574875,
0.04645831510424614,
-0.003455492900684476,
-0.05916707590222359,
0.011555520817637444,
0.036824434995651245,
-0.035900894552469254,
-0.07663185149431229,
-0.06124386936426163,
-0.05723700672388077,
-0.024835817515850067,
-0.020741930231451988,
-0.04214688390493393,
-0.02221705950796604,
0.008589129894971848,
-0.01330993790179491,
-0.0624079555273056,
-0.023591618984937668,
-0.005866051651537418,
0.04229551926255226,
-0.05631804093718529,
0.008349575102329254,
-0.049944646656513214,
0.10497923195362091,
-0.05407292768359184,
0.07037833333015442,
-0.0006432821392081678,
-0.026704825460910797,
-0.05076278746128082,
-0.01077226921916008,
0.029511304572224617,
-0.038245804607868195,
-0.049114909023046494,
0.02913830056786537,
0.06407459080219269,
0.06282401084899902,
-0.04584813117980957,
0.05343441292643547,
0.05063273012638092,
-0.029130375012755394,
-0.04211368039250374,
0.036396466195583344,
-0.03080153279006481,
-0.040299538522958755,
0.010489316657185555,
0.029928626492619514,
-0.01349854189902544,
0.03335485979914665,
-0.038072895258665085,
0.03166377544403076,
-0.10851481556892395,
0.008621209301054478,
-0.03592507913708687,
-0.04815046116709709,
-0.053902819752693176,
0.031904060393571854,
0.029566094279289246,
-0.0020479056984186172,
0.008336007595062256,
-0.04711342230439186,
-0.004753138870000839,
-0.00355550111271441,
0.020099405199289322,
-0.05519799143075943,
-0.050652727484703064,
-0.020791754126548767,
-0.08355310559272766,
-0.007291467394679785,
0.09397073835134506,
-0.041793230921030045,
-0.06541232019662857,
0.09955444931983948,
0.027710046619176865,
0.004643254913389683,
0.05886828526854515,
-0.06323254853487015,
-0.004803363233804703,
0.008573215454816818,
-0.04290610924363136,
-0.03163115307688713,
0.042568303644657135,
0.05656781792640686,
0.08493133634328842,
0.1390227973461151,
0.06242206320166588,
0.038569338619709015,
0.026514220982789993,
0.08696780353784561,
-0.06598091870546341,
0.04705841839313507,
0.0019404347985982895,
0.029083745554089546,
-0.007164222188293934,
-3.440836593426866e-8,
-0.05641418322920799,
0.027229594066739082,
-0.036897365003824234,
-0.03658472001552582,
-0.03902922570705414,
-0.004074334166944027,
0.001916360342875123,
0.036757294088602066,
0.032012760639190674,
-0.09464633464813232,
0.05117173120379448,
-0.011865074746310711,
-0.050021037459373474,
-0.04195767641067505,
-0.04626258462667465,
0.046430498361587524,
0.0005088032339699566,
0.021590882912278175,
-0.009386083111166954,
0.07481763511896133,
-0.07598824054002762,
0.05856873467564583,
-0.003865570528432727,
-0.07845064252614975,
0.004588301759213209,
-0.06186015531420708,
0.02269940823316574,
0.07192472368478775,
0.0505894236266613,
-0.014278693124651909,
-0.07438986748456955,
0.056842636317014694,
0.0378597192466259,
-0.07163474708795547,
-0.05371579900383949,
-0.001980349887162447,
-0.0077727618627250195,
-0.03155696764588356,
-0.046346068382263184,
-0.016406381502747536,
0.03510123863816261,
0.009234180673956871,
0.00895897950977087,
0.009757271967828274,
0.01968255266547203,
0.047444235533475876,
0.05492632836103439,
-0.0464552566409111,
0.015053454786539078,
-0.007635079324245453,
0.000913999683689326,
-0.04558584466576576,
-0.03175988793373108,
0.09789803624153137,
0.07401067763566971,
-0.00019059084297623485,
0.02506803534924984,
0.008030255325138569,
0.09084343910217285,
0.12883725762367249,
0.014637206681072712,
-0.009670604951679707,
-0.0466584637761116,
-0.05864263325929642
] | 0
| 0
| 0
| 0
|
|
Trajectory-Class-Aware Multi-Agent Reinforcement Learning
|
https://openreview.net/forum?id=uqe5HkjbT9
|
[
"Hyungho Na",
"Kwanghyeon Lee",
"Sumin Lee",
"Il-chul Moon"
] |
Poster
|
In the context of multi-agent reinforcement learning, *generalization* is a challenge to solve various tasks that may require different joint policies or coordination without relying on policies specialized for each task. We refer to this type of problem as a *multi-task*, and we train agents to be versatile in this multi-task setting through a single training process. To address this challenge, we introduce TRajectory-class-Aware Multi-Agent reinforcement learning (TRAMA). In TRAMA, agents recognize a task type by identifying the class of trajectories they are experiencing through partial observations, and the agents use this trajectory awareness or prediction as additional information for action policy. To this end, we introduce three primary objectives in TRAMA: (a) constructing a quantized latent space to generate trajectory embeddings that reflect key similarities among them; (b) conducting trajectory clustering using these trajectory embeddings; and (c) building a trajectory-class-aware policy. Specifically for (c), we introduce a trajectory-class predictor that performs agent-wise predictions on the trajectory class; and we design a trajectory-class representation model for each trajectory class. Each agent takes actions based on this trajectory-class representation along with its partial observation for task-aware execution. The proposed method is evaluated on various tasks, including multi-task problems built upon StarCraft II. Empirical results show further performance improvements over state-of-the-art baselines.
|
trajectory clustering, multi-agent reinforcement learning, trajectory-class-aware policy, multi-task
|
TRAMA enables agents to recognize task types by identifying the class of trajectories and to use this information for action policy.
| 11,114
|
2503.01440
|
[
-0.02352014370262623,
-0.14513911306858063,
-0.039922941476106644,
-0.00031777535332366824,
0.046831272542476654,
0.05861589312553406,
0.02589268982410431,
-0.07005250453948975,
0.016307152807712555,
-0.00589223625138402,
-0.0274044182151556,
-0.06801449507474899,
0.024389691650867462,
0.11101230978965759,
-0.007967549376189709,
0.013554089702665806,
0.061598509550094604,
0.06955499202013016,
-0.06789699196815491,
-0.033221691846847534,
0.057357218116521835,
0.008365797810256481,
0.09624425321817398,
0.02076808735728264,
-0.09013106673955917,
0.04697384312748909,
0.05384962260723114,
-0.07869245857000351,
-0.07059819996356964,
0.012665768153965473,
0.0968070700764656,
-0.04975445941090584,
-0.028482403606176376,
-0.02558917924761772,
-0.016078241169452667,
0.1134173646569252,
-0.10793283581733704,
0.01203090138733387,
-0.027120111510157585,
-0.026702018454670906,
0.04325035214424133,
-0.00824821088463068,
0.015724560245871544,
0.02816922590136528,
0.03047354146838188,
0.0009090850362554193,
-0.047332122921943665,
-0.024979082867503166,
0.02092459425330162,
0.004191011656075716,
0.01911887712776661,
-0.0631612166762352,
0.011344285681843758,
-0.00459264637902379,
0.009206556715071201,
-0.010397166013717651,
0.01316847000271082,
0.019358500838279724,
0.04361177235841751,
-0.08414220809936523,
-0.005399881862103939,
-0.09593567252159119,
0.05308578163385391,
-0.04716017469763756,
-0.06343597918748856,
-0.05867169424891472,
0.008553915657103062,
0.05954323709011078,
-0.007398995570838451,
0.03353777527809143,
-0.0590362548828125,
-0.002476123161613941,
-0.010499954223632812,
0.049668166786432266,
-0.015935488045215607,
0.05699712038040161,
0.006252402905374765,
-0.013359347358345985,
-0.03580857440829277,
-0.060669440776109695,
0.015157354064285755,
-0.012508195824921131,
-0.03012698143720627,
0.06622849404811859,
0.14162211120128632,
-0.018547778949141502,
-0.06446272134780884,
0.054034605622291565,
0.05191607400774956,
-0.03504359349608421,
-0.028005274012684822,
0.047897495329380035,
0.06798195838928223,
-0.03256146237254143,
-0.05808627977967262,
0.04369324445724487,
-0.05333375558257103,
-0.08470118790864944,
-0.004860200919210911,
0.05184440687298775,
-0.039017122238874435,
0.04860791936516762,
0.030032172799110413,
0.10154375433921814,
0.024850470945239067,
-0.011268009431660175,
-0.0290254857391119,
-0.020777292549610138,
0.019193090498447418,
-0.0528104230761528,
-0.0256735198199749,
-0.03716433048248291,
0.01175612211227417,
0.04423196613788605,
0.04777489975094795,
0.04239997640252113,
0.034381203353405,
0.045031122863292694,
-0.03605208173394203,
0.042431965470314026,
-0.028081703931093216,
-0.07523976266384125,
0.06331072002649307,
0.029545195400714874,
0.09166816622018814,
-0.06102544069290161,
-0.04796919226646423,
-2.5535651328706338e-33,
0.08138275146484375,
0.01369637344032526,
0.05738123133778572,
0.015940368175506592,
0.032094474881887436,
0.002612763550132513,
-0.020117362961173058,
-0.016857245936989784,
0.030919905751943588,
-0.007535973098129034,
0.04225212708115578,
0.01784089021384716,
-0.03515502065420151,
0.10392485558986664,
0.025618579238653183,
0.04060065373778343,
-0.06718006730079651,
0.020685650408267975,
-0.011845041066408157,
-0.044673480093479156,
0.00761762261390686,
0.042698003351688385,
-0.0437452532351017,
-0.02008718252182007,
-0.019409753382205963,
0.008442776277661324,
0.07819660007953644,
-0.05385034531354904,
0.02845866233110428,
-0.007984278723597527,
-0.02381700649857521,
0.020007280632853508,
-0.03757137805223465,
0.05106741562485695,
-0.0025687438901513815,
0.03868981823325157,
-0.06426017731428146,
-0.010111854411661625,
-0.027751902118325233,
0.03514845296740532,
-0.01519028376787901,
-0.051311951130628586,
0.08181262761354446,
-0.04220109060406685,
-0.0654798075556755,
-0.11046906560659409,
-0.03462131693959236,
-0.006062924861907959,
-0.05302780121564865,
-0.061645492911338806,
0.02663707733154297,
-0.041732966899871826,
0.04683871194720268,
-0.11720079928636551,
-0.03927390277385712,
-0.021535690873861313,
0.0018092661630362272,
0.010647677816450596,
0.021259471774101257,
-0.022651998326182365,
0.0122909527271986,
-0.06330560892820358,
-0.020255934447050095,
0.047732140868902206,
-0.039585813879966736,
0.07950978726148605,
-0.025066692382097244,
0.14161467552185059,
0.11301304399967194,
0.045255932956933975,
-0.014296377077698708,
0.05803905799984932,
0.04259902611374855,
-0.030817296355962753,
0.09140996634960175,
0.003093165345489979,
0.044097695499658585,
-0.1353781819343567,
-0.0021421920973807573,
0.034811146557331085,
-0.052071183919906616,
-0.08170809596776962,
-0.011323499493300915,
-0.002790643833577633,
0.023427506908774376,
-0.053571850061416626,
-0.005208214744925499,
-0.0613754466176033,
-0.009124376811087132,
0.02152714692056179,
-0.1181209608912468,
-0.004153781104832888,
-0.10248515009880066,
0.07957138866186142,
0.002412457950413227,
-1.0748354838644288e-33,
0.039937783032655716,
0.03996879979968071,
-0.03404571861028671,
0.004770809784531593,
-0.0034339996054768562,
-0.026641089469194412,
-0.04934865981340408,
-0.07122387737035751,
-0.0034226858988404274,
-0.03281348943710327,
-0.14991293847560883,
-0.008042178116738796,
0.06235801428556442,
-0.03634781390428543,
-0.0254440288990736,
0.007536744698882103,
0.04126053676009178,
0.006146184168756008,
-0.03922894224524498,
-0.01614133082330227,
-0.004264373332262039,
0.0248278621584177,
-0.06580641120672226,
0.02283286489546299,
0.024860665202140808,
-0.005887764506042004,
0.07230628281831741,
0.002336039673537016,
0.004918559454381466,
0.07203246653079987,
-0.03849434107542038,
-0.11335857212543488,
0.062048472464084625,
-0.04554148390889168,
-0.10117103159427643,
0.038137100636959076,
0.06450074911117554,
0.03284437954425812,
-0.05465574562549591,
0.08601610362529755,
0.06291454285383224,
-0.14484326541423798,
-0.013876578770577908,
0.055804960429668427,
-0.00803233589977026,
-0.012935529462993145,
-0.004311719909310341,
0.08869978785514832,
-0.016689062118530273,
0.04111244156956673,
-0.07326140999794006,
0.026642734184861183,
-0.08706358075141907,
-0.0038584680296480656,
0.03048635832965374,
-0.02703467756509781,
0.024803290143609047,
-0.006111402064561844,
0.021188395097851753,
-0.032617464661598206,
0.04459226503968239,
0.05071135610342026,
-0.007854470051825047,
0.05070117115974426,
0.013942220248281956,
-0.06903983652591705,
-0.02856687270104885,
-0.009946025907993317,
0.011616417206823826,
-0.00540144881233573,
0.025905663147568703,
0.05450401455163956,
-0.021694142371416092,
-0.046069271862506866,
0.05126839131116867,
-0.07654304802417755,
-0.020555568858981133,
-0.014043452218174934,
0.01902821660041809,
-0.021976059302687645,
0.025262432172894478,
0.020134665071964264,
0.05306579917669296,
0.0061446488834917545,
-0.008655435405671597,
0.06171902269124985,
0.0287301205098629,
0.14096948504447937,
0.049534060060977936,
0.018525760620832443,
-0.027531450614333153,
0.04417983442544937,
-0.05852562561631203,
0.0037401942536234856,
-0.11523093283176422,
-3.948632354422443e-8,
-0.0899805799126625,
0.07156577706336975,
0.049430519342422485,
-0.05128300562500954,
0.027784699574112892,
0.01716000586748123,
-0.014909173361957073,
0.048831816762685776,
0.031958796083927155,
0.0008385740802623332,
0.09773194044828415,
0.05005984008312225,
-0.019453728571534157,
-0.03950394317507744,
-0.041252244263887405,
0.007185475435107946,
-0.02440028265118599,
-0.06254944205284119,
-0.007804615423083305,
0.0674823597073555,
0.0013817717554047704,
-0.003951935563236475,
-0.012773244641721249,
-0.03898843750357628,
0.00247568660415709,
-0.030232645571231842,
-0.06685072183609009,
0.0577249638736248,
0.06396884471178055,
0.0186940785497427,
-0.011128532700240612,
-0.03793782740831375,
0.048519257456064224,
-0.039279401302337646,
-0.0345148965716362,
0.04793938249349594,
0.039096757769584656,
-0.027897074818611145,
0.05005315691232681,
-0.07727671414613724,
-0.010596148669719696,
0.07854007929563522,
-0.018632272258400917,
0.023632459342479706,
0.07330064475536346,
0.04060789570212364,
-0.07242428511381149,
-0.11345979571342468,
0.06513730436563492,
-0.020517755299806595,
-0.040930550545454025,
-0.020655017346143723,
0.015339884907007217,
0.09940198063850403,
0.05553719773888588,
0.01612754724919796,
-0.025575978681445122,
-0.09224830567836761,
0.1252591460943222,
-0.02774599939584732,
0.01099811214953661,
-0.03831818699836731,
-0.06605395674705505,
-0.002398950047791004
] |
https://github.com/aailab-kaist/trama
| 19
| 0
| 0
| 0
|
Revisit the Open Nature of Open Vocabulary Semantic Segmentation
|
https://openreview.net/forum?id=2vHIHrJAcI
|
[
"Qiming Huang",
"Han Hu",
"Jianbo Jiao"
] |
Poster
|
In Open Vocabulary Semantic Segmentation (OVS), we observe a consistent drop
in model performance as the query vocabulary set expands, especially when it
includes semantically similar and ambiguous vocabularies, such as ‘sofa’ and
‘couch’. The previous OVS evaluation protocol, however, does not account for
such ambiguity, as any mismatch between model-predicted and human-annotated
pairs is simply treated as incorrect on a pixel-wise basis. This contradicts the open
nature of OVS, where ambiguous categories may both be correct from an open-
world perspective. To address this, in this work, we study the open nature of OVS
and propose a mask-wise evaluation protocol that is based on matched and mis-
matched mask pairs between prediction and annotation respectively. Extensive
experimental evaluations show that the proposed mask-wise protocol provides a
more effective and reliable evaluation framework for OVS models compared to the
previous pixel-wise approach on the perspective of open-world. Moreover, analy-
sis of mismatched mask pairs reveals that a large amount of ambiguous categories
exist in commonly used OVS datasets. Interestingly, we find that reducing these
ambiguities during both training and inference enhances capabilities of OVS mod-
els. These findings and the new evaluation protocol encourage further exploration
of the open nature of OVS, as well as broader open-world challenges. Project page: https://qiming-huang.github.io/RevisitOVS/.
|
Open vocabulary segmentation, Evaluation
| null | 11,105
| null |
[
0.03645938262343407,
-0.038629837334156036,
-0.024153035134077072,
0.05779292434453964,
0.15018793940544128,
-0.0784716010093689,
-0.008696242235600948,
-0.03552829101681709,
0.016845587641000748,
-0.09624209254980087,
-0.0001793101109797135,
-0.1115751564502716,
0.03281985595822334,
0.07219789177179337,
0.030889956280589104,
-0.016495760530233383,
0.05384976789355278,
0.06323184818029404,
-0.08165716379880905,
0.013725660741329193,
0.10625544935464859,
0.08782085031270981,
0.057474490255117416,
-0.04277874901890755,
0.02493654191493988,
0.004831809084862471,
-0.0300204586237669,
-0.041584212332963943,
0.014059006236493587,
-0.029665973037481308,
0.06310810148715973,
0.05021358281373978,
0.053344178944826126,
0.10573460161685944,
0.025583390146493912,
-0.010072852484881878,
0.033218324184417725,
0.044746704399585724,
-0.040702659636735916,
-0.03425546735525131,
-0.03870001807808876,
-0.042901068925857544,
-0.015607699751853943,
0.06604506075382233,
0.03545830771327019,
0.02874411642551422,
-0.041091009974479675,
-0.06450838595628738,
-0.026470519602298737,
0.01619061641395092,
-0.0645599439740181,
-0.04506676644086838,
-0.08782665431499481,
0.06565148383378983,
-0.06736990064382553,
0.019417790696024895,
-0.018126264214515686,
0.012408731505274773,
0.025726471096277237,
0.0133952172473073,
-0.015598928555846214,
-0.08196032792329788,
-0.002921689534559846,
0.08791454881429672,
-0.028949400410056114,
-0.10748306661844254,
-0.018006034195423126,
-0.05669308081269264,
-0.005917876958847046,
-0.030026772990822792,
-0.011777089908719063,
0.07762057334184647,
0.050219081342220306,
-0.026098567992448807,
-0.05995143949985504,
0.04766421020030975,
0.05311664938926697,
-0.02929169312119484,
0.0729445070028305,
-0.08109525591135025,
0.052059073001146317,
-0.02588907815515995,
0.03955085575580597,
-0.005638261325657368,
0.09710471332073212,
0.026016077026724815,
-0.013450936414301395,
-0.014646152965724468,
-0.04102640599012375,
0.013398198410868645,
-0.07244015485048294,
-0.09048029035329819,
0.05909314751625061,
0.0006212349981069565,
0.06350328028202057,
-0.06441427767276764,
0.038683708757162094,
-0.09227669984102249,
0.04630117490887642,
0.09228663146495819,
-0.017591796815395355,
0.01055864430963993,
-0.02190314047038555,
-0.030448678880929947,
-0.03349270299077034,
0.044412706047296524,
0.03131502494215965,
-0.00022107490804046392,
0.07612533122301102,
-0.012655336409807205,
0.08696618676185608,
0.03285159543156624,
0.06875228881835938,
-0.12828706204891205,
-0.03250081092119217,
0.011144328862428665,
-0.005519252270460129,
0.004750942345708609,
0.09361038357019424,
0.043304819613695145,
-0.11544999480247498,
0.039736177772283554,
0.061547886580228806,
-0.025181634351611137,
0.007342630997300148,
-0.011289875954389572,
-0.03614697605371475,
3.0251686909747354e-33,
0.044275667518377304,
-0.027789723128080368,
0.02568832039833069,
-0.05517113208770752,
0.05188313126564026,
-0.020844776183366776,
-0.020762460306286812,
0.04394020140171051,
-0.06407961249351501,
-0.013595908880233765,
0.055800553411245346,
-0.04993818700313568,
-0.049188315868377686,
0.07990748435258865,
0.028613487258553505,
0.059396713972091675,
-0.03335531800985336,
0.03666467219591141,
-0.12282413244247437,
0.06291753798723221,
0.027684226632118225,
0.01258789747953415,
0.03207479417324066,
0.06399969011545181,
0.028377605602145195,
-0.04669087380170822,
-0.0029329380486160517,
-0.07774727046489716,
-0.004507139325141907,
0.0397956445813179,
-0.05020976811647415,
0.05419803038239479,
0.10610677301883698,
0.06822263449430466,
-0.008608347736299038,
-0.04026050865650177,
-0.021011531352996826,
-0.034450531005859375,
0.02733527682721615,
-0.09896526485681534,
-0.040216077119112015,
0.039803437888622284,
0.018887707963585854,
-0.07945340126752853,
-0.024758463725447655,
-0.008630241267383099,
0.02689356543123722,
0.07304270565509796,
-0.06877440959215164,
0.03715941309928894,
0.08871188759803772,
-0.010939862579107285,
-0.07639584690332413,
-0.010133044794201851,
-0.10297182202339172,
-0.012931955978274345,
0.014464445412158966,
0.018945537507534027,
0.008137383498251438,
0.057645101100206375,
-0.044804301112890244,
0.06851843744516373,
0.01389093603938818,
-0.0416547954082489,
-0.06100078672170639,
-0.018493903800845146,
0.05943349748849869,
0.04767123609781265,
-0.01686944253742695,
0.05114634335041046,
-0.060594186186790466,
0.011853757314383984,
-0.02753705531358719,
-0.0365799181163311,
0.04059003293514252,
0.02769884094595909,
-0.013352420181035995,
-0.009081458672881126,
-0.009051003493368626,
0.01579487882554531,
0.005503767170011997,
-0.011739245615899563,
0.03226836025714874,
-0.04945310950279236,
-0.11888588964939117,
0.00015961460303515196,
0.07427328079938889,
-0.03439611569046974,
0.03671024739742279,
0.09062343090772629,
-0.07780247926712036,
-0.011408479884266853,
-0.0538884736597538,
0.017115196213126183,
0.06092873960733414,
-2.8105616260809874e-33,
-0.021485593169927597,
0.02556188404560089,
-0.011815181002020836,
0.03366900235414505,
-0.05471114441752434,
-0.09842436760663986,
0.03601669892668724,
0.030869988724589348,
0.020611004903912544,
-0.06782622635364532,
-0.01471153274178505,
0.0015699021751061082,
-0.014008578844368458,
0.01547199860215187,
0.023734621703624725,
-0.046035103499889374,
-0.03352412208914757,
0.0434839241206646,
-0.022908944636583328,
0.17987872660160065,
0.03241533786058426,
0.0039024080615490675,
-0.10278210043907166,
0.026318384334445,
0.026893600821495056,
0.07569501549005508,
0.006991726346313953,
-0.0334346741437912,
-0.036108892410993576,
-0.017999378964304924,
-0.023066876456141472,
0.04281498119235039,
-0.040093012154102325,
-0.043732307851314545,
-0.016705580055713654,
0.017089279368519783,
-0.0029820662457495928,
-0.0776708796620369,
0.0026457211934030056,
0.01590137928724289,
-0.03525710850954056,
-0.05840388685464859,
-0.11616788059473038,
0.005245700944215059,
-0.10061021149158478,
-0.11075884103775024,
-0.0489102303981781,
0.06292758882045746,
-0.031899865716695786,
0.019994517788290977,
-0.07523268461227417,
0.0643293708562851,
-0.10274961590766907,
0.014436746947467327,
-0.03568733483552933,
-0.03783947601914406,
-0.08799304068088531,
-0.06424102187156677,
0.01871488057076931,
-0.0018663688097149134,
0.019026219844818115,
0.011221439577639103,
-0.05593835562467575,
0.06643164902925491,
0.04912998899817467,
-0.047560062259435654,
-0.034014344215393066,
0.013070417568087578,
-0.02220924384891987,
0.00015890129725448787,
-0.004518910311162472,
-0.05011231452226639,
0.03815368935465813,
-0.002794106723740697,
-0.0023436436895281076,
0.0098733464255929,
0.02348553203046322,
-0.05317762866616249,
0.015160325914621353,
0.028584053739905357,
0.0051824250258505344,
-0.051627375185489655,
0.10308709740638733,
0.04750560224056244,
0.044927772134542465,
0.10144086927175522,
0.013713056221604347,
0.045080337673425674,
0.005193465389311314,
0.05215388908982277,
-0.05549488589167595,
-0.05280112102627754,
0.021184880286455154,
0.06850910931825638,
-0.03886255994439125,
-4.611649018215758e-8,
-0.07798808068037033,
-0.02225595712661743,
-0.011007769033312798,
-0.02446030080318451,
-0.03565613925457001,
-0.05334152653813362,
-0.007201370317488909,
0.08621351420879364,
-0.012719638645648956,
0.06298535317182541,
0.08487957715988159,
0.026848079636693,
-0.03896961361169815,
-0.031774524599313736,
-0.023040801286697388,
0.0761614665389061,
0.045240938663482666,
0.06500717997550964,
0.012210031971335411,
0.012287210673093796,
-0.04329666122794151,
0.0016740950522944331,
0.016709625720977783,
0.02730748802423477,
0.03373977914452553,
-0.04031311348080635,
-0.06350389122962952,
0.10159727185964584,
0.06786338239908218,
-0.0436241589486599,
0.007691989652812481,
-0.0013804436894133687,
0.01401479821652174,
-0.007091687526553869,
0.04312986508011818,
0.06503543257713318,
-0.04892180860042572,
-0.025093207135796547,
0.04019160196185112,
0.03995414078235626,
0.09268845617771149,
-0.051741112023591995,
0.036403801292181015,
-0.027117028832435608,
-0.001367580029182136,
0.08314542472362518,
0.05217383801937103,
-0.03877105191349983,
0.006595219485461712,
-0.04814215376973152,
-0.05402018874883652,
-0.00577026279643178,
-0.02733386680483818,
0.05665018782019615,
0.05584648624062538,
-0.06221899762749672,
-0.01192690059542656,
0.03975818306207657,
0.07213498651981354,
0.052356064319610596,
0.01620396412909031,
-0.002338474616408348,
-0.03130662813782692,
-0.02151666022837162
] | 0
| 0
| 0
| 0
|
|
Robust Barycenter Estimation using Semi-Unbalanced Neural Optimal Transport
|
https://openreview.net/forum?id=CI5Cj0vktS
|
[
"Milena Gazdieva",
"Jaemoo Choi",
"Alexander Kolesov",
"Jaewoong Choi",
"Petr Mokrov",
"Alexander Korotin"
] |
Poster
|
Aggregating data from multiple sources can be formalized as an *Optimal Transport* (OT) barycenter problem, which seeks to compute the average of probability distributions with respect to OT discrepancies. However, in real-world scenarios, the presence of outliers and noise in the data measures can significantly hinder the performance of traditional statistical methods for estimating OT barycenters. To address this issue, we propose a novel scalable approach for estimating the *robust* continuous barycenter, leveraging the dual formulation of the *(semi-)unbalanced* OT problem. To the best of our knowledge, this paper is the first attempt to develop an algorithm for robust barycenters under the continuous distribution setup. Our method is framed as a $\min$-$\max$ optimization problem and is adaptable to *general* cost functions. We rigorously establish the theoretical underpinnings of the proposed method and demonstrate its robustness to outliers and class imbalance through a number of illustrative experiments. Our source code is publicly available at https://github.com/milenagazdieva/U-NOTBarycenters.
|
unbalanced optimal transport, barycenter, generative modeling
| null | 11,104
|
2410.03974
|
[
-0.0788370817899704,
-0.07886724174022675,
0.07633816450834274,
0.01108128298074007,
-0.030113421380519867,
-0.010377448052167892,
0.05024328455328941,
0.023990759626030922,
0.00801830179989338,
-0.02282867394387722,
-0.08137176185846329,
-0.021998345851898193,
0.07522106915712357,
-0.003454770427197218,
-0.12737104296684265,
-0.010812394320964813,
0.13148856163024902,
0.06268268078565598,
-0.09494071453809738,
-0.0020491962786763906,
-0.057016752660274506,
0.03416835889220238,
-0.030144300311803818,
-0.023075737059116364,
0.12399739772081375,
-0.032805394381284714,
-0.00567376334220171,
-0.06782715022563934,
-0.023482605814933777,
-0.03407411649823189,
0.023789923638105392,
0.009795327670872211,
-0.007235895376652479,
-0.07094170898199081,
-0.046805720776319504,
-0.0036635396536439657,
-0.05854156240820885,
0.012470453046262264,
0.061391644179821014,
0.07853225618600845,
0.031079262495040894,
0.03171011433005333,
0.013369257561862469,
0.0072211334481835365,
-0.07765555381774902,
-0.029437879100441933,
0.00374002568423748,
-0.06761851906776428,
0.03741651773452759,
0.037814654409885406,
-0.010554877109825611,
-0.026323918253183365,
-0.015353202819824219,
0.08723196387290955,
-0.06247578561306,
-0.03332066535949707,
0.02477489411830902,
-0.04919954389333725,
0.00654589245095849,
0.10299446433782578,
0.014785454608500004,
-0.08141385763883591,
0.0037640919908881187,
0.014723648317158222,
-0.014762653969228268,
-0.023607755079865456,
0.01937723159790039,
0.09354276955127716,
-0.025983303785324097,
0.10876424610614777,
0.010387519374489784,
0.06981431692838669,
-0.15393048524856567,
0.03674032539129257,
0.08303028345108032,
-0.004471045918762684,
0.08587108552455902,
0.005594396032392979,
0.055137235671281815,
-0.030248267576098442,
0.01650611124932766,
0.022878145799040794,
0.01762661710381508,
-0.025882761925458908,
0.12280067801475525,
-0.0084110451862216,
-0.04004741832613945,
0.09294775873422623,
0.05595729872584343,
0.016169864684343338,
-0.026073424145579338,
0.08937261998653412,
-0.0050000958144664764,
0.02313411980867386,
-0.018631046637892723,
0.023262029513716698,
-0.01800399459898472,
0.004244710318744183,
0.11341091245412827,
0.1117052361369133,
-0.01171787828207016,
0.09840963035821915,
0.10633572190999985,
-0.060727521777153015,
0.02494927868247032,
-0.07980640977621078,
0.062014102935791016,
0.00008279312896775082,
0.0010579961817711592,
-0.08412598818540573,
0.010508861392736435,
-0.02942945808172226,
0.049112580716609955,
-0.020317424088716507,
-0.04552433267235756,
-0.005172339268028736,
0.025505738332867622,
0.03927759826183319,
-0.044790927320718765,
0.03909212350845337,
-0.08308142423629761,
-0.03795192390680313,
0.03264090418815613,
0.05716366693377495,
0.0029713637195527554,
0.08580765128135681,
-0.02314097248017788,
2.8060394625791756e-33,
-0.038225799798965454,
-0.01412525400519371,
0.02902374230325222,
-0.09635864943265915,
0.009620673954486847,
-0.04234249144792557,
-0.05852218344807625,
-0.030770106241106987,
0.021032040938735008,
-0.018080173060297966,
-0.0373440645635128,
-0.022367892786860466,
-0.11303818970918655,
0.011218588799238205,
0.04744821786880493,
-0.017619576305150986,
0.0022288875188678503,
0.06598079204559326,
0.022892791777849197,
-0.1170879453420639,
0.028219128027558327,
-0.12084639072418213,
-0.02862909436225891,
-0.07888998091220856,
0.03651541471481323,
-0.02589264325797558,
0.030969394370913506,
-0.005667189136147499,
0.02992611937224865,
0.02233107015490532,
-0.043341800570487976,
0.02316858619451523,
-0.01634189859032631,
-0.011214927770197392,
0.007088284008204937,
-0.06865917891263962,
-0.014037230983376503,
0.0136668486520648,
-0.0437823161482811,
0.04647121950984001,
-0.047539278864860535,
0.04626811295747757,
0.019345737993717194,
-0.01925264298915863,
-0.015752464532852173,
-0.025610296055674553,
0.04079452529549599,
-0.06709548830986023,
0.008925361558794975,
-0.12404855340719223,
-0.054962627589702606,
-0.04663660004734993,
0.022133145481348038,
-0.0081412885338068,
-0.030901828780770302,
-0.0017605772009119391,
-0.04849909618496895,
0.012143013998866081,
-0.00514057744294405,
0.09998608380556107,
-0.008276723325252533,
0.06072571501135826,
0.0690462738275528,
-0.010226539336144924,
0.076519675552845,
-0.00301153096370399,
-0.019056659191846848,
0.007997201755642891,
0.0003290768654551357,
0.019430110231041908,
0.07159488648176193,
0.02629324235022068,
-0.036629777401685715,
-0.04013599455356598,
0.0646025687456131,
0.04410035163164139,
0.0691620260477066,
0.008570513688027859,
-0.029887989163398743,
-0.054508429020643234,
-0.07820868492126465,
0.0006237310008145869,
0.058327704668045044,
0.024931399151682854,
-0.009625970385968685,
0.038281071931123734,
0.03573618829250336,
-0.010255209170281887,
-0.10883501917123795,
0.04188769310712814,
-0.057401567697525024,
-0.004372591618448496,
-0.10498960316181183,
-0.026981836184859276,
0.04384220391511917,
-2.2695006781806893e-33,
-0.04555243253707886,
0.09197715669870377,
0.0332636833190918,
-0.005201821681112051,
-0.03206200152635574,
-0.0035145943984389305,
-0.035366617143154144,
0.03770097345113754,
0.013003671541810036,
0.022574059665203094,
-0.12319772690534592,
0.007725366856902838,
0.04360280558466911,
-0.09046899527311325,
0.024001698940992355,
-0.07906043529510498,
0.009692901745438576,
0.04638214781880379,
-0.020877033472061157,
0.014534064568579197,
0.00806385837495327,
0.035628542304039,
-0.08967211097478867,
0.032646819949150085,
-0.02566605806350708,
0.019923381507396698,
-0.0736674815416336,
0.10330221056938171,
-0.06901765614748001,
-0.002251986414194107,
-0.06229875236749649,
-0.03530649095773697,
-0.023712623864412308,
-0.06155477091670036,
-0.059365514665842056,
0.04725433140993118,
0.011872523464262486,
0.07187724113464355,
-0.04001229628920555,
0.06853073835372925,
0.016258062794804573,
0.03359328955411911,
-0.04653175175189972,
-0.029543235898017883,
0.029769061133265495,
0.0026902456302195787,
-0.030090469866991043,
-0.06920697540044785,
-0.0014909796882420778,
-0.012022223323583603,
-0.034971799701452255,
0.04894419386982918,
-0.06102033704519272,
0.09392515569925308,
0.0090786786749959,
0.01804511807858944,
0.04133571684360504,
-0.013235587626695633,
0.06338544934988022,
-0.00531124509871006,
-0.0668492391705513,
-0.007127194665372372,
-0.06757006049156189,
-0.011101920157670975,
0.011347807943820953,
0.026136105880141258,
-0.02215862274169922,
-0.023218587040901184,
-0.015497250482439995,
0.04983530938625336,
0.04271158576011658,
0.03361054137349129,
0.06216857209801674,
0.021767813712358475,
-0.012249419465661049,
0.03533235192298889,
0.04098127782344818,
-0.039577048271894455,
0.004217507783323526,
0.04271382838487625,
0.029165633022785187,
-0.006243466399610043,
0.07464531809091568,
0.05998174473643303,
0.037942156195640564,
-0.01827128417789936,
0.07784144580364227,
-0.08676394820213318,
0.036120232194662094,
-0.03376229852437973,
-0.021328900009393692,
-0.010956820100545883,
-0.014601732604205608,
-0.007045171223580837,
-0.05680966004729271,
-4.674206977028916e-8,
-0.05698809400200844,
0.01460642646998167,
-0.011084545403718948,
-0.0395672582089901,
0.08525920659303665,
-0.022641120478510857,
-0.040772005915641785,
0.06514890491962433,
-0.049614157527685165,
0.0009095367277041078,
-0.014476466923952103,
-0.025689713656902313,
-0.007342691533267498,
-0.008416423574090004,
-0.04512123018503189,
0.04962010681629181,
-0.057987749576568604,
0.07889577746391296,
-0.018545100465416908,
-0.04431825131177902,
-0.019924024119973183,
0.09464012086391449,
0.0653497502207756,
0.004559642169624567,
0.08224145323038101,
-0.04340608790516853,
0.058548156172037125,
0.0695551186800003,
0.14378543198108673,
-0.04495278373360634,
-0.052775535732507706,
0.012828770093619823,
0.024285634979605675,
0.02372477576136589,
-0.0003954386920668185,
0.09479927271604538,
-0.05625578388571739,
-0.03189441189169884,
-0.08621961623430252,
-0.006296228151768446,
0.018061524257063866,
0.00001890348721644841,
-0.015804294496774673,
-0.01133913453668356,
0.05158869922161102,
0.03682176023721695,
0.04594285786151886,
0.033283378928899765,
0.025918038561940193,
-0.030719883739948273,
0.036968909204006195,
-0.00975258368998766,
0.08268579095602036,
0.07458144426345825,
0.039245981723070145,
-0.015884816646575928,
-0.05718826875090599,
-0.07532431930303574,
-0.05612098425626755,
0.06989738345146179,
0.024706343188881874,
0.039119090884923935,
-0.12444175034761429,
-0.03742881119251251
] |
https://github.com/milenagazdieva/u-notbarycenters
| 7
| 0
| 0
| 0
|
Inverse decision-making using neural amortized Bayesian actors
|
https://openreview.net/forum?id=zxO4WuVGns
|
[
"Dominik Straub",
"Tobias F. Niehues",
"Jan Peters",
"Constantin A. Rothkopf"
] |
Poster
|
Bayesian observer and actor models have provided normative explanations for many behavioral phenomena in perception, sensorimotor control, and other areas of cognitive science and neuroscience. They attribute behavioral variability and biases to interpretable entities such as perceptual and motor uncertainty, prior beliefs, and behavioral costs. However, when extending these models to more naturalistic tasks with continuous actions, solving the Bayesian decision-making problem is often analytically intractable. Inverse decision-making, i.e. performing inference over the parameters of such models given behavioral data, is computationally even more difficult. Therefore, researchers typically constrain their models to easily tractable components, such as Gaussian distributions or quadratic cost functions, or resort to numerical approximations. To overcome these limitations, we amortize the Bayesian actor using a neural network trained on a wide range of parameter settings in an unsupervised fashion. Using the pre-trained neural network enables performing efficient gradient-based Bayesian inference of the Bayesian actor model's parameters. We show on synthetic data that the inferred posterior distributions are in close alignment with those obtained using analytical solutions where they exist. Where no analytical solution is available, we recover posterior distributions close to the ground truth. We then show how our method allows for principled model comparison and how it can be used to disentangle factors that may lead to unidentifiabilities between priors and costs. Finally, we apply our method to empirical data from three sensorimotor tasks and compare model fits with different cost functions to show that it can explain individuals' behavioral patterns.
|
Bayesian actor models, perception and action, cognitive science, Bayesian inference, inverse modeling
|
We developed an efficient Bayesian inference methods for priors, uncertainties, and costs from behavior by amortizing Bayesian actor models using neural networks.
| 11,101
|
2409.03710
|
[
-0.06334318965673447,
-0.0876641646027565,
0.05875694379210472,
0.024298720061779022,
-0.017387425526976585,
0.007541139144450426,
0.0524948313832283,
0.03602227941155434,
0.03379504755139351,
0.05346665158867836,
-0.028865326195955276,
-0.06388062238693237,
0.017204182222485542,
0.047182608395814896,
-0.11357881128787994,
-0.01241347100585699,
0.12498263269662857,
0.05650815740227699,
-0.09595741331577301,
0.01148550771176815,
-0.018579162657260895,
0.020849891006946564,
-0.016340231522917747,
-0.01095812302082777,
0.029366912320256233,
-0.029729539528489113,
0.07331713289022446,
-0.07100647687911987,
-0.08303818106651306,
0.012382755056023598,
0.120976023375988,
-0.04319380596280098,
-0.0057993377558887005,
-0.026686808094382286,
-0.07999270409345627,
0.06108672916889191,
-0.08787384629249573,
0.006019992288202047,
0.011933096684515476,
-0.02053108997642994,
0.04156380891799927,
0.03444946929812431,
0.016765693202614784,
0.060403965413570404,
0.09212948381900787,
-0.029026255011558533,
-0.009786438196897507,
-0.08250775188207626,
0.05163051187992096,
0.04447302967309952,
-0.061152804642915726,
0.01049413625150919,
-0.0003865252365358174,
-0.020843056961894035,
0.05296469107270241,
0.028309054672718048,
0.05333821102976799,
0.016937173902988434,
-0.014032067731022835,
0.025807570666074753,
-0.0016389978118240833,
-0.07456362992525101,
-0.05163215473294258,
-0.03325125575065613,
-0.02970568835735321,
0.09387192875146866,
-0.002901754342019558,
-0.02845989167690277,
-0.02605009637773037,
0.0911860391497612,
-0.009240969084203243,
0.10962709039449692,
-0.05929712951183319,
-0.047913193702697754,
-0.01739461161196232,
-0.05049828067421913,
0.04634829983115196,
0.01584862917661667,
0.039562396705150604,
-0.09282024949789047,
0.03415216878056526,
-0.04444144666194916,
-0.017211729660630226,
-0.002139077754691243,
0.11950522661209106,
0.01196775771677494,
-0.004443879704922438,
0.1389361321926117,
0.0727686956524849,
0.06554926931858063,
-0.08404380083084106,
-0.04772428795695305,
-0.045514289289712906,
-0.09278988093137741,
-0.009120765142142773,
0.06444020569324493,
-0.001346818869933486,
-0.04158509895205498,
0.023569554090499878,
0.05147335305809975,
0.0006721914978697896,
0.048829350620508194,
0.029611412435770035,
0.0015878414269536734,
-0.05708269774913788,
-0.006288188509643078,
-0.013034617528319359,
0.051882170140743256,
0.02118387259542942,
-0.01555370818823576,
-0.03884279355406761,
-0.050001922994852066,
0.03422626480460167,
0.006783264223486185,
-0.004941094201058149,
0.0331718809902668,
0.07571200281381607,
0.001213633338920772,
-0.040019840002059937,
0.0402955487370491,
-0.005354998633265495,
-0.03726150840520859,
0.0004243739240337163,
0.025365769863128662,
0.01890963688492775,
-0.029697230085730553,
-0.07068485766649246,
3.723060605412889e-33,
0.0013134408509358764,
-0.03357269614934921,
0.044425833970308304,
0.005932648666203022,
-0.009358584880828857,
0.015963885933160782,
-0.01039127353578806,
-0.022536715492606163,
0.03486301004886627,
-0.00757801765576005,
-0.008584477007389069,
0.03971126675605774,
-0.10855019837617874,
0.03209327161312103,
0.06370384246110916,
0.02421068400144577,
-0.05922219902276993,
0.03530023619532585,
-0.04207213222980499,
-0.08246134221553802,
0.00404268829151988,
-0.0130311269313097,
-0.04993123188614845,
-0.03566272184252739,
-0.03304465487599373,
-0.05563564971089363,
0.04919921234250069,
0.010096301324665546,
-0.07172070443630219,
-0.0023437722120434046,
-0.07880783826112747,
0.007450951263308525,
0.0026245093904435635,
0.013291136361658573,
0.02027192898094654,
-0.019405120983719826,
0.0046068038791418076,
-0.0061338674277067184,
0.048694200813770294,
-0.03799671307206154,
-0.03361426666378975,
0.04495341330766678,
0.028939224779605865,
0.05364886671304703,
-0.08637002110481262,
-0.11682520806789398,
0.016652865335345268,
-0.03134612366557121,
0.04562297835946083,
-0.04126840457320213,
0.0061591798439621925,
-0.009680615738034248,
0.11380055546760559,
-0.12802813947200775,
-0.07518181204795837,
-0.02076002210378647,
0.0070820883847773075,
0.02020607702434063,
-0.015204319730401039,
0.02705579623579979,
0.027966046705842018,
0.03736818954348564,
0.011440379545092583,
0.04782405123114586,
0.00009581386257195845,
0.0809205025434494,
-0.020898722112178802,
0.08325382322072983,
0.0628061592578888,
-0.0030318284407258034,
0.003703661495819688,
0.01730925217270851,
0.025314629077911377,
-0.07339393347501755,
0.05775132030248642,
0.036765437573194504,
0.04863394424319267,
-0.04343865066766739,
-0.0126584367826581,
-0.04120863229036331,
-0.03796939179301262,
-0.010900040157139301,
-0.05855894088745117,
0.05888958275318146,
-0.00521985674276948,
-0.027303848415613174,
0.05211954936385155,
0.01031055673956871,
-0.030054738745093346,
0.0008789738640189171,
-0.057908594608306885,
-0.02385142631828785,
-0.046941835433244705,
0.00644299341365695,
0.002302632899954915,
-3.569051376329984e-33,
-0.09037427604198456,
-0.011929795145988464,
0.00457787187770009,
0.05159154534339905,
-0.03732750564813614,
-0.02129683457314968,
-0.010325852781534195,
0.007355173584073782,
0.009680195711553097,
-0.07436871528625488,
-0.13840016722679138,
0.003451528260484338,
0.04641500860452652,
0.003060847520828247,
0.037031516432762146,
-0.018490899354219437,
-0.06723557412624359,
0.05656134709715843,
0.0031740942504256964,
-0.06859876215457916,
-0.0038782746996730566,
0.05889073386788368,
-0.06448143720626831,
-0.01906493492424488,
-0.05383753404021263,
0.07046788930892944,
-0.012660454958677292,
0.12882855534553528,
-0.01874697394669056,
0.011190610937774181,
-0.13116809725761414,
0.004961538128554821,
-0.04493795335292816,
-0.009542198851704597,
-0.09445597976446152,
0.11196821928024292,
0.052407555282115936,
0.009702716022729874,
-0.07445918023586273,
-0.044879134744405746,
-0.01790440082550049,
-0.044292110949754715,
-0.09008008986711502,
-0.04683980345726013,
0.04577440768480301,
-0.08922605216503143,
0.02584291808307171,
-0.0040966724045574665,
0.01691855490207672,
0.023338764905929565,
-0.013655426912009716,
0.068716861307621,
-0.05257256701588631,
0.02248474210500717,
-0.10907359421253204,
-0.03294346109032631,
-0.00970577634871006,
-0.013774543069303036,
0.1557181030511856,
0.01575937122106552,
-0.02702048048377037,
0.0021588155068457127,
0.09941164404153824,
-0.01197850238531828,
-0.06401804089546204,
0.10494761168956757,
0.008616907522082329,
0.0681956484913826,
0.04084010422229767,
-0.03453439846634865,
-0.015482629649341106,
-0.030206937342882156,
0.06313963979482651,
0.0199593435972929,
0.05848371610045433,
-0.026062259450554848,
-0.028555357828736305,
0.03802943974733353,
-0.0012991796247661114,
0.014465033076703548,
0.10382344573736191,
0.013508393429219723,
0.011463855393230915,
0.039397891610860825,
-0.02375599928200245,
0.02372129261493683,
-0.01252734661102295,
0.03668314218521118,
0.030205318704247475,
-0.04153783991932869,
-0.04952109605073929,
0.0602593757212162,
-0.010721213184297085,
0.008228779770433903,
-0.09261002391576767,
-4.43325625099078e-8,
-0.05302097275853157,
0.08433530479669571,
0.13527905941009521,
0.05025554075837135,
0.06050480157136917,
-0.013954129070043564,
0.006354041397571564,
0.036161188036203384,
0.0030991609673947096,
-0.05058472603559494,
0.07114668190479279,
0.012124263681471348,
0.06510535627603531,
-0.06127355992794037,
-0.04773866757750511,
0.040471721440553665,
0.02780861407518387,
0.025004031136631966,
-0.03181575983762741,
0.04644240438938141,
0.07148736715316772,
0.05445278808474541,
-0.015803828835487366,
-0.023188326507806778,
0.08129414916038513,
-0.09304165840148926,
-0.12205709517002106,
0.10185310244560242,
0.0007809072267264128,
0.08865008503198624,
0.025530528277158737,
0.0006111138500273228,
0.013876776210963726,
0.02492711693048477,
-0.0006240104557946324,
0.03379818797111511,
0.011982348747551441,
-0.05832589417695999,
-0.014633920043706894,
-0.08859070390462875,
-0.037921663373708725,
0.016400929540395737,
-0.018770115450024605,
0.09039163589477539,
0.065396748483181,
-0.0030897040851414204,
0.004592174664139748,
-0.04266732931137085,
0.08026684075593948,
0.033979155123233795,
0.037051133811473846,
0.02390572428703308,
-0.05642188712954521,
0.030951492488384247,
-0.03939131274819374,
-0.039125800132751465,
0.045937035232782364,
-0.02765776216983795,
-0.027333823963999748,
0.02487674355506897,
0.02217055857181549,
0.059031397104263306,
-0.07936495542526245,
0.004717194940894842
] |
https://github.com/rothkopflab/naba
| 0
| 0
| 0
| 0
|
Reframing Structure-Based Drug Design Model Evaluation via Metrics Correlated to Practical Needs
|
https://openreview.net/forum?id=RyWypcIMiE
|
[
"Bowen Gao",
"Haichuan Tan",
"Yanwen Huang",
"Minsi Ren",
"Xiao Huang",
"Wei-Ying Ma",
"Ya-Qin Zhang",
"Yanyan Lan"
] |
Poster
|
Recent advances in structure-based drug design (SBDD) have produced surprising results, with models often generating molecules that achieve better Vina docking scores than actual ligands. However, these results are frequently overly optimistic due to the limitations of docking score accuracy and the challenges of wet-lab validation. While generated molecules may demonstrate high QED (drug-likeness) and SA (synthetic accessibility) scores, they often lack true drug-like properties or synthesizability. To address these limitations, we propose a model-level evaluation framework that emphasizes practical metrics aligned with real-world applications. Inspired by recent findings on the utility of generated molecules in ligand-based virtual screening, our framework evaluates SBDD models by their ability to produce molecules that effectively retrieve active compounds from chemical libraries via similarity-based searches. This approach provides a direct indication of therapeutic potential, bridging the gap between theoretical performance and real-world utility. Our experiments reveal that while SBDD models may excel in theoretical metrics like Vina scores, they often fall short in these practical metrics. By introducing this new evaluation strategy, we aim to enhance the relevance and impact of SBDD models for pharmaceutical research and development.
|
Stucture-Based Drug Design, Model Evaluation, Benchmark
| null | 11,094
| null |
[
-0.07365566492080688,
-0.0693136602640152,
-0.07445905357599258,
-0.017485134303569794,
0.03326858580112457,
-0.02814740501344204,
-0.019294708967208862,
0.07475168257951736,
0.1038043349981308,
-0.03736328333616257,
-0.06869936734437943,
-0.035602010786533356,
0.05913618952035904,
0.09543865919113159,
-0.05379735678434372,
-0.0287660900503397,
0.04781436547636986,
0.041824813932180405,
-0.04039815068244934,
0.04395414888858795,
-0.0036250967532396317,
0.033845845609903336,
0.05053771287202835,
-0.014238113537430763,
-0.02909437008202076,
0.029321637004613876,
0.0052389358170330524,
-0.019661348313093185,
-0.028143418952822685,
-0.06812270730733871,
0.08168616890907288,
0.0951644703745842,
0.029914503917098045,
-0.124552883207798,
0.02743220143020153,
-0.006249960511922836,
-0.0361037403345108,
0.006480250041931868,
-0.016340410336852074,
-0.0811033621430397,
-0.02987423725426197,
-0.013138948939740658,
0.03294763341546059,
0.04173542559146881,
0.0384320467710495,
-0.1456487476825714,
-0.093967504799366,
-0.029877211898565292,
0.013672156259417534,
0.08085168898105621,
-0.04355618730187416,
-0.030577661469578743,
-0.038912661373615265,
0.024980170652270317,
-0.010989041067659855,
0.1122250184416771,
-0.02429419569671154,
-0.00010152754111913964,
-0.020765310153365135,
0.01367612648755312,
0.022414881736040115,
-0.05973174795508385,
-0.004735707305371761,
-0.0059326714836061,
0.04319480434060097,
0.05817221850156784,
0.026939602568745613,
-0.00965846050530672,
0.041511714458465576,
0.008125756867229939,
-0.1116170734167099,
0.05846385657787323,
-0.025452759116888046,
0.03135710209608078,
0.039004672318696976,
0.05660757049918175,
-0.019579065963625908,
0.0595976822078228,
0.048322614282369614,
-0.027696039527654648,
0.024254396557807922,
0.019837649539113045,
0.02477777563035488,
0.033196065574884415,
0.06820980459451675,
0.08528808504343033,
0.032191772013902664,
0.030018476769328117,
0.08701905608177185,
-0.013773800805211067,
0.03389870002865791,
0.02462800219655037,
-0.06121550500392914,
-0.06728730350732803,
-0.025948798283934593,
0.012902256101369858,
0.019499290734529495,
-0.02621995285153389,
0.039494361728429794,
0.01681864820420742,
-0.00047871784772723913,
0.0725017711520195,
-0.020976688712835312,
-0.08406778424978256,
0.07973145693540573,
-0.04221673682332039,
0.06499835103750229,
0.04422656074166298,
-0.0051081436686217785,
-0.056675080209970474,
-0.003821892663836479,
0.0194561630487442,
0.0863257423043251,
0.005572299472987652,
-0.0047590057365596294,
0.06140286475419998,
0.027065489441156387,
-0.02523915469646454,
0.01678149215877056,
-0.025097545236349106,
-0.012010661885142326,
0.039370570331811905,
0.05677947402000427,
-0.059738002717494965,
-0.04612001031637192,
-0.014996561221778393,
-0.12827099859714508,
1.5292209884648609e-33,
0.007814958691596985,
0.04388635605573654,
0.08475085347890854,
0.08963347971439362,
0.04580339416861534,
-0.12033367156982422,
-0.021443121135234833,
0.013810344971716404,
-0.10436055064201355,
0.04010818526148796,
-0.09068942815065384,
0.028035469353199005,
-0.027692127972841263,
0.060726866126060486,
0.09484799951314926,
0.03891255334019661,
-0.03849444538354874,
0.07718450576066971,
-0.08163407444953918,
-0.07981757819652557,
0.001877520582638681,
-0.05994541570544243,
0.005939861293882132,
-0.03346896916627884,
0.007000158540904522,
0.043955493718385696,
-0.01849035546183586,
0.08976109325885773,
-0.058473095297813416,
-0.00855867750942707,
-0.10344104468822479,
0.012154984287917614,
0.01161462813615799,
0.007698424626141787,
-0.010397120378911495,
-0.01791810430586338,
-0.040083229541778564,
-0.047819919884204865,
-0.0090482821688056,
-0.008827464655041695,
-0.004955037031322718,
0.027348974719643593,
0.003929695580154657,
-0.0011554037919268012,
-0.026981402188539505,
0.0144176185131073,
0.012419088743627071,
0.016315558925271034,
-0.062186453491449356,
-0.018854601308703423,
0.021470749750733376,
0.034099217504262924,
0.0054296813905239105,
0.034410905092954636,
-0.11408460885286331,
-0.03382277116179466,
-0.020791595801711082,
0.026630539447069168,
-0.017002621665596962,
0.11281255632638931,
-0.023375414311885834,
0.0671442523598671,
-0.04145578294992447,
0.01672467775642872,
-0.046941276639699936,
0.03415365517139435,
-0.04126504063606262,
-0.02549656294286251,
0.020386191084980965,
0.1004597470164299,
0.046088241040706635,
0.011248039081692696,
-0.027432752773165703,
-0.03859866410493851,
0.11744623631238937,
-0.05392483249306679,
0.05339089408516884,
-0.045671310275793076,
0.03859003633260727,
-0.010654797777533531,
-0.02732008323073387,
0.04697701707482338,
-0.05612373352050781,
-0.018241439014673233,
-0.05653641000390053,
-0.1816822588443756,
0.03652798756957054,
-0.0667991116642952,
-0.11994230002164841,
-0.0558793731033802,
-0.03931700810790062,
-0.021277470514178276,
-0.026035280898213387,
-0.05031731724739075,
0.019628461450338364,
-3.013130894138346e-33,
0.025647811591625214,
-0.03190263733267784,
0.08405707776546478,
-0.011939848773181438,
-0.012913293205201626,
-0.00571902422234416,
-0.014962582848966122,
-0.07576799392700195,
0.07928478717803955,
-0.05788498744368553,
0.009259075857698917,
0.025344515219330788,
0.01627802662551403,
-0.013810081407427788,
-0.02610069140791893,
0.043185941874980927,
-0.0406448058784008,
-0.12301523983478546,
0.0321335569024086,
0.0015856946120038629,
0.057598505169153214,
0.03666279837489128,
-0.010509145446121693,
0.038114193826913834,
-0.07042811065912247,
0.012630478478968143,
0.04018549248576164,
-0.049792446196079254,
0.008055930957198143,
0.03343270719051361,
-0.05128130316734314,
-0.03205415606498718,
-0.12582610547542572,
0.04637269675731659,
-0.024733873084187508,
0.0673552006483078,
0.036856114864349365,
-0.04227745905518532,
0.027675095945596695,
0.01634403131902218,
-0.017693806439638138,
0.014746598899364471,
-0.13704712688922882,
0.009401807561516762,
0.04651405289769173,
-0.0021268262062221766,
-0.035437941551208496,
0.01022423803806305,
0.05569099262356758,
-0.05785081908106804,
0.019488506019115448,
0.05078195780515671,
0.002255914267152548,
-0.008124088868498802,
-0.07132218033075333,
-0.06631334125995636,
-0.06071591377258301,
-0.04694994539022446,
0.040900908410549164,
0.05136699974536896,
-0.006698806770145893,
-0.0247663427144289,
-0.04048304259777069,
0.010948721319437027,
-0.09484860301017761,
-0.012663661502301693,
0.06498485803604126,
0.03684309870004654,
0.013282976113259792,
-0.04621979594230652,
0.0227640513330698,
-0.035003483295440674,
0.0676007941365242,
0.12439627200365067,
0.012662366032600403,
-0.0501662977039814,
0.04402796924114227,
-0.07215707749128342,
0.004456420429050922,
-0.00959415640681982,
0.017716778442263603,
0.0005507601890712976,
0.00386443710885942,
-0.009220462292432785,
0.05403493717312813,
0.06726665049791336,
0.05069085583090782,
0.033671073615550995,
0.00037950300611555576,
0.031069539487361908,
0.017180919647216797,
0.020721565932035446,
-0.06768148392438889,
0.12494145333766937,
0.031869206577539444,
-4.156209598704663e-8,
0.08030104637145996,
-0.006477976217865944,
0.013587002642452717,
0.006686828099191189,
-0.05626823753118515,
0.06851296126842499,
-0.05769689381122589,
0.007873782888054848,
-0.02798689901828766,
0.00975894182920456,
-0.01732454262673855,
0.02183784358203411,
-0.0906461849808693,
0.08214420080184937,
0.008748996071517467,
0.05996008217334747,
-0.024644263088703156,
0.0531931109726429,
-0.06220731511712074,
-0.019509345293045044,
-0.036441370844841,
0.10351502895355225,
0.026906464248895645,
-0.04829063639044762,
0.05053728446364403,
-0.08207115530967712,
-0.019519321620464325,
-0.05769374594092369,
0.015998806804418564,
0.02821740321815014,
0.09562906622886658,
0.019526666030287743,
-0.018787970766425133,
0.04666738957166672,
0.023998264223337173,
0.0416221059858799,
-0.04080579802393913,
-0.026244090870022774,
-0.02525697834789753,
0.08826610445976257,
-0.04579990729689598,
0.00717083690688014,
-0.12130739539861679,
-0.011233989149332047,
-0.016233304515480995,
-0.008012696169316769,
-0.012723425403237343,
0.014311657287180424,
-0.02263566479086876,
0.019064687192440033,
0.03642765432596207,
-0.043097712099552155,
-0.010338892228901386,
-0.07603500783443451,
0.00904901884496212,
0.02971423789858818,
-0.01695116236805916,
0.007215550635010004,
0.07173816859722137,
-0.10430670529603958,
0.05186581611633301,
-0.03174743801355362,
0.04574253037571907,
-0.019027668982744217
] | 0
| 0
| 0
| 0
|
|
Valid Conformal Prediction for Dynamic GNNs
|
https://openreview.net/forum?id=i3T0wvQDKg
|
[
"Ed Davis",
"Ian Gallagher",
"Daniel John Lawson",
"Patrick Rubin-Delanchy"
] |
Poster
|
Dynamic graphs provide a flexible data abstraction for modelling many sorts of real-world systems, such as transport, trade, and social networks. Graph neural networks (GNNs) are powerful tools allowing for different kinds of prediction and inference on these systems, but getting a handle on uncertainty, especially in dynamic settings, is a challenging problem.
In this work we propose to use a dynamic graph representation known in the tensor literature as the unfolding, to achieve valid prediction sets via conformal prediction. This representation, a simple graph, can be input to any standard GNN and does not require any modification to existing GNN architectures or conformal prediction routines.
One of our key contributions is a careful mathematical consideration of the different inference scenarios which can arise in a dynamic graph modelling context. For a range of practically relevant cases, we obtain valid prediction sets with almost no assumptions, even dispensing with exchangeability. In a more challenging scenario, which we call the semi-inductive regime, we achieve valid prediction under stronger assumptions, akin to stationarity.
We provide real data examples demonstrating validity, showing improved accuracy over baselines, and sign-posting different failure modes which can occur when those assumptions are violated.
|
Graph neural networks, Graph machine learning, conformal prediction
|
A method to perform valid conformal prediction on dynamic graphs using GNNs.
| 11,093
|
2405.19230
|
[
-0.11222484707832336,
-0.10488395392894745,
0.005394053179770708,
0.07415149360895157,
0.0055991956032812595,
-0.016706587746739388,
-0.04051465541124344,
-0.022412994876503944,
-0.0741192027926445,
-0.06786760687828064,
-0.050960931926965714,
-0.06331159919500351,
-0.003498666686937213,
0.12359504401683807,
-0.005173433106392622,
-0.008875834755599499,
0.03895879536867142,
0.03196084871888161,
-0.0611812099814415,
-0.03048020415008068,
-0.006636624690145254,
0.02650078758597374,
-0.10612548887729645,
0.03679905831813812,
0.05517612025141716,
0.012539424002170563,
0.013954922556877136,
-0.04803953319787979,
0.03247666731476784,
0.03113614208996296,
0.008029521442949772,
-0.030546363443136215,
-0.024461520835757256,
-0.00528758903965354,
-0.025718675926327705,
-0.01756541058421135,
-0.0392330028116703,
-0.03311781585216522,
0.07518938183784485,
0.012910719029605389,
0.058310240507125854,
0.001276675146073103,
0.015327331610023975,
0.08397934585809708,
0.09208668768405914,
0.044033847749233246,
0.0118904709815979,
0.04406590387225151,
-0.014819455333054066,
0.037636324763298035,
-0.025912176817655563,
-0.040295861661434174,
-0.042324308305978775,
0.011246868409216404,
0.007372279651463032,
0.020579876378178596,
-0.02130541391670704,
0.008073216304183006,
-0.00785195454955101,
-0.013221671804785728,
0.029531022533774376,
-0.12000598013401031,
0.011647522449493408,
-0.027154486626386642,
0.010739466175436974,
0.07193285971879959,
-0.020053284242749214,
0.06920210272073746,
-0.05357366055250168,
0.10429619997739792,
-0.02481716126203537,
0.040683045983314514,
-0.1793762892484665,
0.014226735569536686,
0.04641130194067955,
0.05880388617515564,
0.07673202455043793,
-0.017553895711898804,
-0.007706035859882832,
-0.0287307221442461,
0.01133477222174406,
0.031078318133950233,
0.032713450491428375,
-0.010630199685692787,
0.026570722460746765,
-0.037797711789608,
-0.04832712188363075,
0.052163317799568176,
0.07862415164709091,
0.004452388733625412,
-0.06459896266460419,
0.05366222932934761,
-0.006548776291310787,
0.009591856971383095,
0.051170818507671356,
0.07479286938905716,
0.020343879237771034,
-0.05721086263656616,
0.0338340662419796,
0.09771745651960373,
-0.035828847438097,
0.054853618144989014,
0.06578262150287628,
-0.06378674507141113,
0.027717847377061844,
0.0021660206839442253,
0.03122716210782528,
0.03330812230706215,
-0.017559215426445007,
-0.052990369498729706,
0.004248100332915783,
0.08336330205202103,
0.06154284626245499,
-0.02745182067155838,
0.02455340325832367,
-0.03634914010763168,
-0.023525569587945938,
-0.014626782387495041,
-0.020329291000962257,
0.061567604541778564,
-0.023049838840961456,
0.049863461405038834,
0.06995850056409836,
0.03803102672100067,
-0.022077688947319984,
0.0009509458905085921,
-0.08774055540561676,
1.582440944185363e-33,
-0.037014998495578766,
-0.026342710480093956,
0.09043212234973907,
0.00781503226608038,
0.056011445820331573,
0.059467680752277374,
-0.07180391997098923,
-0.01078653335571289,
0.019879043102264404,
-0.0019462055061012506,
-0.09055361896753311,
0.06517896056175232,
-0.005478494334965944,
0.0270310677587986,
0.053087685257196426,
-0.004012495279312134,
0.03964390978217125,
-0.03497382625937462,
0.016744188964366913,
-0.04104296490550041,
0.05320174619555473,
-0.05701303109526634,
-0.03190619498491287,
0.014962323941290379,
0.011044367216527462,
0.006746669299900532,
-0.018708251416683197,
0.019019730389118195,
0.012282737530767918,
-0.024472452700138092,
-0.08773597329854965,
-0.020667513832449913,
-0.022569943219423294,
0.04184968024492264,
-0.009386919438838959,
-0.06633205711841583,
-0.0406486839056015,
-0.05484553426504135,
-0.03698638454079628,
-0.07970468699932098,
-0.04189037159085274,
0.02866901084780693,
-0.08306749165058136,
0.003287282772362232,
-0.04627619683742523,
-0.050415024161338806,
0.02499384433031082,
0.01482807844877243,
-0.020538080483675003,
-0.0961812436580658,
-0.012802927754819393,
-0.011827961541712284,
-0.10886196792125702,
-0.06252098828554153,
-0.037404268980026245,
-0.01106659509241581,
0.012318463064730167,
0.06036655977368355,
0.021721212193369865,
0.019364511594176292,
0.017198054119944572,
0.028486721217632294,
0.00308224861510098,
-0.006227824371308088,
-0.005715719424188137,
0.07453171908855438,
-0.07220159471035004,
-0.000039352547901216894,
0.022450894117355347,
-0.05943869799375534,
0.07022083550691605,
0.083424411714077,
-0.08738651126623154,
-0.03377268463373184,
0.040922846645116806,
-0.01365656591951847,
-0.008442559279501438,
-0.019627388566732407,
-0.02004621922969818,
0.05741264671087265,
-0.09624022245407104,
-0.031571466475725174,
-0.037289585918188095,
0.0635981559753418,
-0.05132420361042023,
-0.04908408969640732,
0.05669323354959488,
-0.042886052280664444,
-0.061854343861341476,
-0.044648464769124985,
-0.1343652307987213,
-0.03708678483963013,
0.0727677270770073,
0.011643274687230587,
0.05525348335504532,
-2.517752595446879e-33,
-0.029996152967214584,
0.06460125744342804,
0.023109188303351402,
0.0426558181643486,
-0.030091552063822746,
-0.04907798022031784,
-0.031264230608940125,
0.045201633125543594,
0.02670850045979023,
0.05581766366958618,
-0.01909887045621872,
-0.03148965165019035,
0.048074766993522644,
-0.031025120988488197,
0.038416538387537,
-0.03822877258062363,
-0.02641802839934826,
0.034340012818574905,
-0.02048412151634693,
0.003378507448360324,
0.036753181368112564,
0.04682006314396858,
-0.13691368699073792,
0.06720935553312302,
0.04744410142302513,
0.045720040798187256,
-0.026038382202386856,
0.04774860292673111,
-0.022014163434505463,
0.03580417111515999,
-0.10296507179737091,
-0.05987446755170822,
0.010284243151545525,
0.02318098396062851,
0.055837009102106094,
0.03998810425400734,
0.02627454698085785,
-0.004411784000694752,
-0.0018784059211611748,
0.0054917773231863976,
0.040776271373033524,
0.005394045729190111,
-0.09161330759525299,
0.049695756286382675,
-0.018390940502285957,
0.015849323943257332,
-0.0023939695674926043,
0.09163116663694382,
-0.017241166904568672,
-0.01875758171081543,
0.01832648180425167,
0.07538650184869766,
-0.056282542645931244,
-0.018094802275300026,
-0.07381341606378555,
0.007998716086149216,
0.03587996959686279,
-0.015560421161353588,
0.09703109413385391,
-0.017162298783659935,
-0.048977747559547424,
-0.08347506076097488,
0.00954245962202549,
0.04049981012940407,
-0.02159678377211094,
-0.04018748551607132,
-0.05306680127978325,
0.014030734077095985,
0.0987231656908989,
0.05286410450935364,
0.032930608838796616,
-0.048486240208148956,
-0.004869427997618914,
0.008787505328655243,
0.05569099634885788,
-0.047631196677684784,
0.021969342604279518,
-0.022105256095528603,
0.013521897606551647,
-0.023097429424524307,
0.013012060895562172,
0.005573343951255083,
0.08276274800300598,
0.060862086713314056,
0.06537267565727234,
-0.025265077129006386,
0.019304055720567703,
0.0508553721010685,
0.10816297680139542,
0.11356893926858902,
-0.012928090989589691,
-0.01318376511335373,
-0.0805666372179985,
0.1135995015501976,
-0.05427505820989609,
-4.300108713550799e-8,
-0.17571690678596497,
0.056912414729595184,
0.0716705322265625,
0.043997619301080704,
0.035744186490774155,
-0.0008408306166529655,
0.15908069908618927,
-0.0314728282392025,
0.001297496259212494,
0.07132207602262497,
0.057766351848840714,
-0.0694546326994896,
-0.08621910959482193,
-0.04978039488196373,
-0.05423780158162117,
0.04026861488819122,
0.004822332877665758,
0.004815978463739157,
0.0016813768306747079,
0.034678827971220016,
-0.0007492061122320592,
-0.03318806365132332,
-0.017434952780604362,
0.03160040080547333,
0.04062369465827942,
-0.15563437342643738,
-0.07924254238605499,
0.08212418854236603,
-0.00788742583245039,
0.03993609547615051,
-0.019509635865688324,
0.0006701677339151502,
0.07464023679494858,
-0.0029973338823765516,
-0.06982559710741043,
0.0693274661898613,
0.04864129051566124,
-0.06123906746506691,
0.004389354959130287,
0.03963717445731163,
-0.051390066742897034,
0.08540480583906174,
-0.023662816733121872,
-0.02008276991546154,
0.019972186535596848,
-0.053866323083639145,
0.03324345871806145,
-0.008013738319277763,
0.10136116296052933,
-0.04922311753034592,
-0.015158988535404205,
-0.009038202464580536,
-0.04595744237303734,
0.057732827961444855,
0.07232010364532471,
0.00808659940958023,
-0.046534836292266846,
0.006297664251178503,
0.0019302982836961746,
0.031960006803274155,
-0.0232289619743824,
0.005575768183916807,
-0.059142544865608215,
-0.0485914908349514
] |
https://github.com/edwarddavis1/valid_conformal_for_dynamic_gnn
| 3
| 0
| 0
| 0
|
ContextGNN: Beyond Two-Tower Recommendation Systems
|
https://openreview.net/forum?id=nzOD1we8Z4
|
[
"Yiwen Yuan",
"Zecheng Zhang",
"Xinwei He",
"Akihiro Nitta",
"Weihua Hu",
"Manan Shah",
"Blaž Stojanovič",
"Shenyang Huang",
"Jan Eric Lenssen",
"Jure Leskovec",
"Matthias Fey"
] |
Poster
|
Recommendation systems predominantly utilize two-tower architectures, which evaluate user-item rankings through the inner product of their respective embeddings. However, one key limitation of two-tower models is that they learn a pair-agnostic representation of users and items. In contrast, pair-wise representations either scale poorly due to their quadratic complexity or are too restrictive on the candidate pairs to rank. To address these issues, we introduce Context-based Graph Neural Networks (ContextGNNs), a novel deep learning architecture for link prediction in recommendation systems. The method employs a pair-wise representation technique for familiar items situated within a user's local subgraph, while leveraging two-tower representations to facilitate the recommendation of exploratory items. A final network then predicts how to fuse both pair-wise and two-tower recommendations into a single ranking of items. We demonstrate that ContextGNN is able to adapt to different data characteristics and outperforms existing methods, both traditional and GNN-based, on a diverse set of practical recommendation tasks, improving performance by 20\% on average.
|
graph neural networks, recommendation, relational deep learning
| null | 11,092
|
2411.19513
|
[
-0.07415919005870819,
-0.10489354282617569,
0.013242773711681366,
0.06790351122617722,
0.026887966319918633,
0.023238129913806915,
-0.03430526331067085,
0.0414678230881691,
-0.0519648939371109,
-0.07818392664194107,
-0.042171090841293335,
-0.005019709002226591,
0.0639096349477768,
0.0862186998128891,
-0.03757599741220474,
0.0737096294760704,
0.11475927382707596,
0.041858088225126266,
-0.01984582468867302,
-0.025030475109815598,
0.032066527754068375,
-0.010792919434607029,
0.0428394116461277,
-0.05939373001456261,
0.047691360116004944,
0.029883287847042084,
-0.08587440848350525,
0.013762774877250195,
0.021946804597973824,
0.0071372659876942635,
0.14979025721549988,
0.054380953311920166,
-0.08175862580537796,
-0.0067051369696855545,
-0.09813565760850906,
0.03307623043656349,
-0.05827293172478676,
-0.024643126875162125,
0.04277179390192032,
-0.034434493631124496,
0.03682327643036842,
0.06663688272237778,
0.017236454412341118,
0.1199638769030571,
0.060555797070264816,
0.05890240892767906,
-0.0002444938581902534,
0.009836933575570583,
0.03738420829176903,
0.11200054734945297,
-0.028017640113830566,
0.025551291182637215,
-0.02546493336558342,
0.011102131567895412,
-0.0002332305011805147,
0.04000029340386391,
-0.06858902424573898,
0.05567704886198044,
-0.04345405101776123,
0.015552426688373089,
0.07926782220602036,
-0.09804350137710571,
-0.04286961257457733,
-0.07656038552522659,
-0.011479416862130165,
-0.032417815178632736,
-0.027629001066088676,
0.018531300127506256,
-0.08548831939697266,
0.04257514327764511,
0.0520259253680706,
0.04797739163041115,
-0.03245498612523079,
0.06316269189119339,
0.01125222910195589,
0.048824045807123184,
0.017637738958001137,
-0.06712296605110168,
-0.032927725464105606,
0.013434793800115585,
-0.04692232608795166,
0.028436895459890366,
0.0019844945054501295,
-0.08126750588417053,
0.11157455295324326,
-0.019715862348675728,
0.006834222935140133,
-0.05226358398795128,
-0.09587794542312622,
0.011395344510674477,
-0.0021981389727443457,
0.013811701908707619,
0.024795373901724815,
-0.05961599573493004,
0.07077319175004959,
0.0729333832859993,
0.02575812302529812,
-0.012330280616879463,
-0.04572909325361252,
0.05680680274963379,
-0.09344030916690826,
0.07950545847415924,
0.01735365204513073,
-0.07119191437959671,
0.043238528072834015,
0.010109972208738327,
-0.03581426292657852,
0.07243425399065018,
0.03627442941069603,
-0.08095826208591461,
-0.04900825023651123,
0.10328537225723267,
-0.03437304496765137,
0.006104473024606705,
0.000533735379576683,
-0.03286624699831009,
0.07000372558832169,
-0.03285804018378258,
0.06122337281703949,
0.0868799015879631,
-0.07594534009695053,
0.046939972788095474,
0.006619033869355917,
-0.06934037804603577,
-0.11707016080617905,
0.005023137200623751,
-0.02130129747092724,
1.5581076601103647e-34,
0.0658702552318573,
0.024338727816939354,
0.027764488011598587,
0.04355388879776001,
0.0489114448428154,
0.017711855471134186,
-0.10134560614824295,
0.01807314157485962,
-0.06013886258006096,
0.011462872847914696,
-0.01552009116858244,
0.02788696438074112,
-0.01086464524269104,
0.08960136771202087,
0.060171350836753845,
0.024410001933574677,
0.0063595250248909,
-0.041131649166345596,
0.0019191631581634283,
-0.1328735053539276,
0.0019372605020180345,
0.026414779946208,
-0.014960919506847858,
0.025344539433717728,
0.00004256630927557126,
-0.01817200519144535,
0.07729652523994446,
0.004772027488797903,
0.05125633627176285,
-0.04476405680179596,
-0.05834532529115677,
0.0250965915620327,
-0.015241374261677265,
0.011387436650693417,
-0.033386968076229095,
-0.05223671719431877,
-0.008584062568843365,
-0.03797830641269684,
0.056465595960617065,
-0.07753076404333115,
-0.007991345599293709,
0.01634800061583519,
-0.00568638788536191,
-0.011798870749771595,
-0.04095844551920891,
-0.08089540898799896,
-0.04233570769429207,
-0.053795307874679565,
-0.013020355254411697,
-0.05190334841609001,
0.03796140104532242,
-0.035990625619888306,
0.010026081465184689,
0.09515251219272614,
-0.012394815683364868,
-0.0781075656414032,
0.034783974289894104,
0.040281154215335846,
0.018731093034148216,
-0.0416995994746685,
0.06598478555679321,
0.006584412883967161,
0.02186358906328678,
0.005307326093316078,
-0.028697390109300613,
0.012501735240221024,
-0.0046183448284864426,
0.0691564604640007,
0.015950068831443787,
-0.008985545486211777,
0.0006520063034258783,
0.1287890523672104,
0.05361904203891754,
-0.04433219134807587,
-0.022504596039652824,
0.047821443527936935,
-0.04839716851711273,
-0.12341858446598053,
-0.03669121861457825,
-0.007968842051923275,
-0.03910283371806145,
-0.03148568049073219,
0.02738369069993496,
0.016144437715411186,
-0.06561478227376938,
0.014964689500629902,
-0.008587419055402279,
-0.08363027125597,
0.0028801809530705214,
-0.04923269525170326,
-0.050859466195106506,
0.032673679292201996,
-0.018242763355374336,
0.0022627769503742456,
0.048190563917160034,
-1.6779812220116118e-34,
0.02288137562572956,
0.05650990828871727,
0.07755272090435028,
0.012392118573188782,
0.06502991914749146,
-0.04748387634754181,
-0.05664438009262085,
-0.0520855076611042,
-0.07360686361789703,
0.02894359640777111,
-0.014935705810785294,
-0.06270487606525421,
0.07987014204263687,
-0.01890542358160019,
0.07335519790649414,
0.03642600029706955,
-0.04587921127676964,
0.03921310603618622,
-0.055625539273023605,
-0.06613090634346008,
0.05254395306110382,
0.04083690792322159,
-0.060265541076660156,
0.04393206909298897,
0.019640032202005386,
-0.029505914077162743,
0.05463438481092453,
-0.026749972254037857,
-0.04670857638120651,
-0.02047216333448887,
-0.009411747567355633,
-0.05917654186487198,
0.04549615830183029,
-0.03361998870968819,
0.042017947882413864,
0.07272583246231079,
-0.00815805047750473,
-0.05535338073968887,
-0.06519120931625366,
-0.017115993425250053,
0.04669053107500076,
0.013162015937268734,
-0.07271161675453186,
0.00005115533713251352,
-0.017224274575710297,
0.029009366407990456,
-0.05926508083939552,
0.010493138805031776,
-0.017949890345335007,
-0.0315827876329422,
-0.05280270427465439,
0.046133264899253845,
-0.05372973158955574,
-0.024329842999577522,
-0.04228488355875015,
-0.03670342639088631,
-0.0037421593442559242,
0.05609855800867081,
0.12671765685081482,
-0.007144749164581299,
0.0010789314983412623,
-0.0776074230670929,
-0.006929102819412947,
-0.05586613714694977,
-0.05876723676919937,
-0.05051777884364128,
0.03019815683364868,
0.016011981293559074,
-0.014830092899501324,
0.08840400725603104,
-0.001741539454087615,
0.01446046307682991,
-0.02501278556883335,
-0.007077380083501339,
-0.03271064907312393,
-0.0204631257802248,
0.06490295380353928,
-0.021725600585341454,
-0.01044295635074377,
-0.062359295785427094,
0.035072438418865204,
-0.0366022028028965,
0.007655456196516752,
-0.014455915428698063,
0.13117700815200806,
-0.0022465339861810207,
0.03588762879371643,
0.08485912531614304,
-0.010237489826977253,
0.0018478494603186846,
-0.022152023389935493,
-0.0019998541101813316,
-0.07505319267511368,
0.029067648574709892,
-0.02616768702864647,
-3.472560905493083e-8,
-0.11219097673892975,
0.016408078372478485,
-0.01750829443335533,
0.08905767649412155,
-0.025926712900400162,
0.010459027253091335,
0.024595968425273895,
0.16379916667938232,
0.027111617848277092,
0.0330958291888237,
0.037839729338884354,
-0.01879972033202648,
-0.07542896270751953,
-0.09755922108888626,
0.04348932206630707,
-0.003187109949067235,
0.07172993570566177,
-0.056132007390260696,
0.053034212440252304,
0.034088607877492905,
0.022621871903538704,
0.024909917265176773,
0.06032673642039299,
0.05907667055726051,
-0.03772279620170593,
-0.010904993861913681,
0.010125340893864632,
0.09459541738033295,
0.06618132442235947,
0.04649168625473976,
0.006983713712543249,
0.021323570981621742,
0.052249833941459656,
-0.07416796684265137,
-0.0714547410607338,
0.07044167816638947,
0.03811633586883545,
-0.00781899131834507,
-0.04500006139278412,
-0.08007345348596573,
0.04688084498047829,
0.014198795892298222,
-0.019876016303896904,
0.061557967215776443,
0.021897541359066963,
0.02561957575380802,
0.059034861624240875,
-0.012511946260929108,
0.06702343374490738,
-0.010729745961725712,
0.012772477231919765,
-0.06418512761592865,
0.049809928983449936,
-0.01943529210984707,
0.011822343803942204,
-0.02359025925397873,
-0.07129871100187302,
0.011857396923005581,
0.07473462074995041,
0.012821948155760765,
0.00894047413021326,
-0.028869938105344772,
-0.07831691950559616,
0.06223039701581001
] |
https://github.com/kumo-ai/ContextGNN
| 25
| 0
| 0
| 0
|
KOR-Bench: Benchmarking Language Models on Knowledge-Orthogonal Reasoning Tasks
|
https://openreview.net/forum?id=SVRRQ8goQo
|
[
"Kaijing Ma",
"Xeron Du",
"Yunran Wang",
"Haoran Zhang",
"ZhoufutuWen",
"Xingwei Qu",
"Jian Yang",
"Jiaheng Liu",
"minghao liu",
"Xiang Yue",
"Wenhao Huang",
"Ge Zhang"
] |
Poster
|
In this paper, we introduce Knowledge-Orthogonal Reasoning (KOR), a concept aimed at minimizing reliance on domain-specific knowledge, enabling more accurate evaluation of models' reasoning abilities in out-of-distribution settings.
Based on this concept, we propose the Knowledge-Orthogonal Reasoning Benchmark (KOR-Bench), encompassing five task categories: Operation, Logic, Cipher, Puzzle, and Counterfactual.
KOR-Bench emphasizes models' effectiveness in applying new rule descriptions to solve novel rule-driven questions. O1-Preview and O1-Mini achieve accuracies of 72.88\% and 70.16\%, surpassing Claude-3.5-Sonnet and GPT-4o (58.96\% and 58.00\%), highlighting the effectiveness of KOR-Bench.
We perform detailed analyses, identifying bottlenecks in the Cipher task with Stepwise Prompting, where two rounds of Self-Correction yield optimal results.
We evaluate performance across three integrated tasks, explore the impact of Tricks on the Puzzle task, and visualize rule-focused attention. Additionally, we conduct an ablation study on dataset size, benchmark correlations, and zero-shot and three-shot "only questions" experiments.
KOR-Bench aims to enhance reasoning evaluation and support further research in this area.
|
Reasoning; Knowledge-Orthogonal; Rule-Based
|
We introduce the concept of Knowledge Orthogonal Reasoning (KOR) and propose five types of rule-based reasoning tasks to construct a KOR-Bench to fully evaluate the intrinsic reasoning ability of the model.
| 11,089
| null |
[
-0.034098945558071136,
-0.01367304939776659,
-0.016909928992390633,
-0.0002373192401137203,
-0.0008152903174050152,
-0.023045020177960396,
-0.0021493590902537107,
0.00968465581536293,
0.021416543051600456,
0.03175470978021622,
-0.06000124290585518,
-0.01434990856796503,
0.05972294136881828,
0.10800066590309143,
-0.006297779735177755,
-0.03374107554554939,
0.064451202750206,
-0.0003991980047430843,
-0.10430960357189178,
-0.08694028854370117,
-0.008798380382359028,
-0.08793620765209198,
0.09322885423898697,
-0.0686761736869812,
0.03234505653381348,
0.003965174313634634,
-0.026068931445479393,
-0.030094828456640244,
-0.01684366911649704,
0.014616253785789013,
0.029367415234446526,
0.036626413464546204,
0.01774897612631321,
0.008742819540202618,
-0.030892960727214813,
0.012853331863880157,
0.02858983725309372,
0.05194677412509918,
0.010275661014020443,
-0.0012020780704915524,
-0.0687413439154625,
-0.07306623458862305,
-0.019332053139805794,
0.03090246208012104,
0.05106966570019722,
-0.0056145391426980495,
-0.01628139801323414,
0.009121253155171871,
-0.02175692841410637,
-0.056446779519319534,
-0.10713561624288559,
-0.027705118060112,
-0.04341341182589531,
-0.015732986852526665,
0.02525162510573864,
0.0012358129024505615,
-0.010260092094540596,
0.036243122071027756,
0.04755081981420517,
-0.03244822844862938,
-0.012581213377416134,
-0.1527039110660553,
-0.07633831351995468,
0.10321950167417526,
-0.056693315505981445,
0.009535618126392365,
0.07321728765964508,
-0.0394052155315876,
-0.047459445893764496,
0.026738939806818962,
-0.018993856385350227,
0.10264678299427032,
-0.009215760976076126,
0.06495775282382965,
-0.014742112718522549,
0.029721418395638466,
0.003623258089646697,
-0.03453366085886955,
-0.013326743617653847,
-0.10302695631980896,
0.0120866559445858,
0.0015010754577815533,
-0.017905231565237045,
0.05137460678815842,
0.1455077826976776,
-0.00848347693681717,
0.029370052739977837,
0.04888433218002319,
0.04531670734286308,
0.033078718930482864,
0.026413505896925926,
-0.08911638706922531,
0.0528971329331398,
0.0073019047267735004,
0.029594717547297478,
0.08611489087343216,
0.04328583925962448,
-0.011365565471351147,
-0.08325374126434326,
0.05970612168312073,
0.0163328368216753,
0.06369257718324661,
-0.037750594317913055,
-0.08668334782123566,
0.060754116624593735,
0.0012583325151354074,
0.031274132430553436,
0.05003334954380989,
0.05109034478664398,
-0.08515861630439758,
-0.0000466598394268658,
0.07968489825725555,
0.07503937184810638,
0.02046414092183113,
-0.028582962229847908,
-0.012805401347577572,
0.03938896954059601,
0.026515720412135124,
0.05590636283159256,
0.04785061627626419,
0.020797936245799065,
0.00916539877653122,
0.03711218759417534,
0.016949526965618134,
0.061577875167131424,
-0.0008499028044752777,
-0.02407311461865902,
-8.567552923414818e-34,
0.07537015527486801,
0.07658732682466507,
0.04382454231381416,
-0.021562961861491203,
0.03227563574910164,
-0.013578974641859531,
0.00011092317436123267,
0.09107322245836258,
-0.038688454777002335,
0.02295863814651966,
0.006935339421033859,
-0.03398331627249718,
-0.0417240671813488,
-0.01728646643459797,
0.0015579144237563014,
0.039699383080005646,
-0.04511232301592827,
0.010269350372254848,
-0.072180837392807,
-0.033393822610378265,
0.05957190319895744,
-0.05824601277709007,
0.01680007204413414,
0.017047850415110588,
-0.018893498927354813,
0.054430603981018066,
0.06419946253299713,
-0.08259295672178268,
-0.03645462542772293,
0.011329886503517628,
-0.10903842002153397,
-0.09377643465995789,
0.001049224054440856,
0.07886970788240433,
-0.01652677357196808,
-0.09115618467330933,
0.012732825241982937,
-0.14045406877994537,
0.07894772291183472,
0.0014678334118798375,
-0.06406359374523163,
-0.0107340756803751,
-0.030473198741674423,
0.04032750800251961,
-0.015108266845345497,
-0.017069805413484573,
-0.027448469772934914,
0.03269161283969879,
0.006312733981758356,
0.04563582316040993,
0.03196708485484123,
0.07977518439292908,
-0.03216572105884552,
-0.052791304886341095,
0.019828958436846733,
-0.012592853046953678,
-0.0008832604507915676,
0.01027783565223217,
0.06354081630706787,
0.14420552551746368,
-0.025118282064795494,
0.0490754134953022,
-0.023161161690950394,
0.05180424451828003,
0.009762008674442768,
-0.007073464337736368,
-0.10474762320518494,
0.03975905478000641,
0.059266090393066406,
0.034156959503889084,
-0.027819272130727768,
-0.02589517831802368,
0.10494720935821533,
0.008767880499362946,
0.0878225713968277,
0.04158077761530876,
0.010990342125296593,
-0.07223671674728394,
0.03950968384742737,
-0.06328378617763519,
0.023420387879014015,
0.03645353391766548,
-0.034176237881183624,
-0.0337890088558197,
-0.04564329981803894,
-0.030858390033245087,
0.022495312616229057,
-0.08217751979827881,
-0.049096450209617615,
-0.03519769385457039,
-0.011092963628470898,
-0.08748102933168411,
-0.029314514249563217,
-0.09084134548902512,
0.0023248177021741867,
9.894353600997088e-34,
-0.012426084838807583,
-0.04651902616024017,
-0.09728331863880157,
0.029550664126873016,
-0.02466638945043087,
-0.06693018972873688,
-0.02749720960855484,
0.035975635051727295,
0.05715841054916382,
-0.04794524610042572,
-0.01099927257746458,
-0.06737536936998367,
0.05315649136900902,
-0.013644703663885593,
0.05396076664328575,
0.009580355137586594,
-0.04505882412195206,
0.00267421524040401,
-0.0010229827603325248,
0.11511113494634628,
-0.047748543322086334,
0.06051556020975113,
-0.1324850469827652,
0.07343658804893494,
-0.04903360828757286,
0.05499757081270218,
-0.027776828035712242,
0.014164247550070286,
-0.024969525635242462,
0.045411620289087296,
0.018463633954524994,
-0.027335507795214653,
-0.005362093448638916,
-0.006945145316421986,
-0.06973836570978165,
0.01078493706882,
0.062256909906864166,
-0.05637176334857941,
-0.06749935448169708,
0.04533706605434418,
0.021802086383104324,
-0.03455708175897598,
-0.04580119252204895,
-0.08758901804685593,
-0.039148956537246704,
0.04586079344153404,
-0.08752766996622086,
-0.013758819550275803,
-0.015216922387480736,
-0.01838339865207672,
-0.003786808345466852,
0.0009498126455582678,
-0.004101332742720842,
0.04570102319121361,
-0.01895933970808983,
-0.05350955203175545,
-0.11732453852891922,
-0.013807062990963459,
0.03458743169903755,
0.0069125802256166935,
0.028792116791009903,
0.05017876625061035,
0.06820029020309448,
0.04317566007375717,
0.06042599305510521,
0.03772544488310814,
-0.023919308558106422,
0.04281911253929138,
0.02498524636030197,
0.013953461311757565,
-0.07760811597108841,
-0.095691978931427,
-0.008667253889143467,
-0.006639079190790653,
0.10984139144420624,
0.07619450986385345,
0.006215631030499935,
-0.10594785958528519,
-0.06781341135501862,
0.03437570109963417,
-0.033251818269491196,
-0.05170516297221184,
0.006842475384473801,
0.04699668660759926,
-0.022862352430820465,
0.07292057573795319,
0.07831910997629166,
0.03905661404132843,
0.04423120245337486,
0.02015923522412777,
0.000791249331086874,
0.02609308250248432,
0.010202683508396149,
0.0922594666481018,
0.004255961626768112,
-4.846210543973939e-8,
-0.04778341203927994,
0.030889401212334633,
-0.009030547924339771,
0.01995975524187088,
0.05915671959519386,
-0.05409339815378189,
-0.08984477072954178,
0.011296896263957024,
-0.05887221172451973,
-0.027734752744436264,
0.11499769985675812,
-0.010592455975711346,
-0.003438070882111788,
-0.04194946959614754,
0.012402987107634544,
0.06584721058607101,
-0.000276675884379074,
-0.018587112426757812,
-0.025956913828849792,
-0.059082165360450745,
0.07649920880794525,
0.008474537171423435,
-0.039159730076789856,
0.028587767854332924,
-0.009599799290299416,
0.013590010814368725,
0.0007139028748497367,
0.1329421103000641,
0.11839248985052109,
0.036434806883335114,
-0.04105040058493614,
-0.05259280651807785,
0.07971611618995667,
-0.0016398789593949914,
0.07320212572813034,
0.02759496308863163,
0.04509111866354942,
-0.0027023768052458763,
0.053892847150564194,
0.03285432979464531,
0.09296821802854538,
0.005368545651435852,
-0.055999111384153366,
0.05753689631819725,
0.038192883133888245,
0.05614437535405159,
-0.0801449716091156,
-0.039864376187324524,
0.01933887228369713,
-0.006046969909220934,
0.009777296334505081,
-0.04958347603678703,
-0.06188167259097099,
0.03949544206261635,
0.013625966385006905,
-0.0014367880066856742,
-0.07145579904317856,
-0.023594489321112633,
0.001669498742558062,
-0.03750287741422653,
0.05974361672997475,
0.08465822041034698,
-0.017468729987740517,
0.012579658068716526
] | 0
| 0
| 0
| 0
|
|
Gaussian Splatting Lucas-Kanade
|
https://openreview.net/forum?id=dkrEoT68by
|
[
"Liuyue Xie",
"Joel Julin",
"Koichiro Niinuma",
"Laszlo Attila Jeni"
] |
Poster
|
Gaussian Splatting and its dynamic extensions are effective for reconstructing 3D scenes from 2D images when there is significant camera movement to facilitate motion parallax and when scene objects remain relatively static. However, in many real-world scenarios, these conditions are not met. As a consequence, data-driven semantic and geometric priors have been favored as regularizers, despite their bias toward training data and their neglect of broader movement dynamics.
Departing from this practice, we propose a novel analytical approach that adapts the classical Lucas-Kanade method to dynamic Gaussian splatting. By leveraging the intrinsic properties of the forward warp field network, we derive an analytical velocity field that, through time integration, facilitates accurate scene flow computation. This enables the precise enforcement of motion constraints on warp fields, thus constraining both 2D motion and 3D positions of the Gaussians. Our method excels in reconstructing highly dynamic scenes with minimal camera movement, as demonstrated through experiments on both synthetic and real-world scenes.
|
Gaussian Splatting, regularization, novel view synthesis
| null | 11,083
| null |
[
0.04067426174879074,
-0.11047766357660294,
0.032852061092853546,
-0.037611719220876694,
0.08973702788352966,
-0.029946720227599144,
-0.0015955892158672214,
-0.04151232913136482,
0.02378813736140728,
-0.1040097251534462,
0.05353367701172829,
-0.006419655866920948,
-0.09241892397403717,
0.07068347930908203,
-0.0342322513461113,
-0.06142404302954674,
0.024985812604427338,
0.0886261835694313,
-0.09451128542423248,
0.029525702819228172,
0.0034536116290837526,
-0.09340578317642212,
0.025311294943094254,
0.029606880620121956,
0.0676766186952591,
0.04895439371466637,
0.027162078768014908,
-0.019864220172166824,
-0.0044335569255054,
-0.012990044429898262,
0.039227310568094254,
0.03632490336894989,
-0.10011634230613708,
0.019943229854106903,
-0.047651711851358414,
0.035620275884866714,
0.048399072140455246,
0.03832555562257767,
-0.01304285041987896,
0.0052033523097634315,
0.007909627631306648,
0.03568693622946739,
0.03300534188747406,
0.04041785001754761,
0.11107464879751205,
0.09172290563583374,
0.07459887117147446,
-0.0737890973687172,
0.01456262543797493,
0.040270376950502396,
-0.07604023814201355,
0.01667383313179016,
0.06173381581902504,
0.01393735222518444,
0.01293798815459013,
0.0889410674571991,
0.013401631265878677,
-0.08113769441843033,
0.06307978183031082,
0.004612496122717857,
0.03033614158630371,
-0.04147850349545479,
0.013166534714400768,
0.018909111618995667,
-0.06896022707223892,
-0.08516962081193924,
0.02168693020939827,
-0.0349096953868866,
0.005329296458512545,
0.07574720680713654,
-0.05370752513408661,
0.06885164231061935,
0.06795058399438858,
-0.025554701685905457,
-0.031225642189383507,
0.020552389323711395,
0.030721614137291908,
-0.059582747519016266,
-0.02219109982252121,
-0.08184581995010376,
0.11876882612705231,
-0.06773467361927032,
0.049755312502384186,
0.058065373450517654,
-0.012670861557126045,
0.06995946168899536,
-0.013201409950852394,
0.10335838049650192,
0.022579338401556015,
0.04960636794567108,
-0.029457256197929382,
-0.059449128806591034,
0.0015490304213017225,
-0.09072408080101013,
-0.007980633527040482,
-0.023857852444052696,
-0.027922557666897774,
-0.11386203020811081,
0.08634984493255615,
0.04621642827987671,
0.0034072890412062407,
-0.08401639014482498,
0.04355553165078163,
0.07640083134174347,
0.10675079375505447,
0.01593060977756977,
0.04574849456548691,
0.06972672045230865,
0.013596456497907639,
0.001625075819902122,
0.02736089564859867,
0.023898571729660034,
0.029670465737581253,
-0.030353805050253868,
-0.06870216131210327,
0.051160164177417755,
-0.013602344319224358,
0.002602088265120983,
-0.046591997146606445,
0.03709062933921814,
0.01247463934123516,
-0.03729254752397537,
-0.07100530713796616,
0.008971319533884525,
0.006887858733534813,
-0.06508416682481766,
-0.0122801773250103,
1.652420510612714e-33,
0.03716293349862099,
0.04514395073056221,
0.07142291218042374,
0.03258895128965378,
-0.024454304948449135,
0.02278241328895092,
-0.04063954949378967,
-0.0008018254884518683,
-0.03227212280035019,
0.02111099474132061,
-0.05169498920440674,
-0.08684616535902023,
-0.03128281608223915,
0.0314653106033802,
0.002104672137647867,
0.009494246914982796,
-0.021974919363856316,
-0.028973720967769623,
0.02264576591551304,
0.07470458745956421,
0.055202145129442215,
-0.02777339331805706,
-0.07913044840097427,
-0.02427545189857483,
-0.08975863456726074,
0.015826569870114326,
-0.01702481135725975,
-0.0012520658783614635,
0.04240744560956955,
-0.023323198780417442,
-0.02916168048977852,
0.004985487554222345,
-0.01343043614178896,
0.0088715311139822,
0.07597950100898743,
-0.03379911929368973,
0.004888572730123997,
-0.07087322324514389,
0.043083976954221725,
-0.031942449510097504,
-0.053174376487731934,
0.05368201807141304,
-0.047327399253845215,
-0.010152837261557579,
-0.09446880221366882,
-0.04293501377105713,
0.12301185727119446,
0.1119692474603653,
-0.004198783077299595,
0.08540678024291992,
0.05854666605591774,
-0.03286362066864967,
-0.018666261807084084,
-0.05039292201399803,
-0.012819034047424793,
-0.011479359120130539,
0.028701171278953552,
-0.01592758484184742,
0.022783188149333,
0.02712508663535118,
0.03030254691839218,
0.011978086084127426,
-0.03676130622625351,
-0.00926221627742052,
-0.04222223535180092,
0.010437448509037495,
-0.006058586295694113,
0.04260856658220291,
0.05826805159449577,
-0.010926919057965279,
-0.07404562085866928,
-0.014861291274428368,
0.005370252300053835,
-0.06613434851169586,
0.08370696008205414,
0.035139985382556915,
0.012176682241261005,
-0.03252774849534035,
-0.004566406365483999,
0.025085119530558586,
-0.14212433993816376,
-0.012521125376224518,
0.021485237404704094,
0.025523347780108452,
0.018935536965727806,
0.00945241842418909,
-0.0398687869310379,
0.004331364296376705,
-0.05272939056158066,
-0.12223310023546219,
-0.004229398909956217,
-0.06597992777824402,
-0.060212329030036926,
-0.04076164960861206,
-0.04571858420968056,
-1.4091608627531688e-33,
0.02578435093164444,
0.05694911628961563,
-0.029495205730199814,
0.03313807025551796,
-0.019178831949830055,
-0.01013294979929924,
0.02357763797044754,
0.002013302640989423,
0.07115649431943893,
-0.08003295958042145,
-0.043314024806022644,
0.00872434675693512,
0.0004025567031931132,
-0.01819019764661789,
0.02118196338415146,
0.011617372743785381,
-0.023854194208979607,
-0.06527017802000046,
-0.059212882071733475,
-0.04654089733958244,
0.034162238240242004,
0.05928578972816467,
-0.039996903389692307,
-0.02285386249423027,
-0.05883010849356651,
0.10671151429414749,
-0.020465010777115822,
0.08286149054765701,
-0.054631948471069336,
0.00037669448647648096,
-0.008978317491710186,
-0.0921817272901535,
-0.005290169268846512,
-0.039575908333063126,
-0.0699680969119072,
0.01251919474452734,
0.008822964504361153,
0.049078311771154404,
-0.0171549953520298,
-0.037806738168001175,
-0.052409347146749496,
0.025019371882081032,
-0.03745673596858978,
0.05027118697762489,
-0.04444335028529167,
0.04523523524403572,
-0.0009821467101573944,
0.011433848179876804,
0.015253985300660133,
0.0799185037612915,
-0.10246703028678894,
0.0759781002998352,
-0.09041191637516022,
-0.02227052114903927,
0.007682438008487225,
-0.0037738820537924767,
-0.05142488330602646,
-0.030282968655228615,
0.058699727058410645,
-0.01486289594322443,
-0.012121642008423805,
-0.014300466515123844,
-0.006012284196913242,
0.02623112127184868,
0.006775654386729002,
0.03435184061527252,
-0.050455182790756226,
-0.009575705975294113,
-0.033156778663396835,
0.014170086942613125,
0.021036546677350998,
0.09337162971496582,
-0.04343010485172272,
0.031013991683721542,
-0.01757037453353405,
-0.04369695484638214,
0.022073475643992424,
-0.018701205030083656,
0.0838736742734909,
0.035222649574279785,
0.11475423723459244,
-0.034184034913778305,
0.07954184710979462,
0.06956002861261368,
0.05231059342622757,
0.13752257823944092,
-0.06395644694566727,
-0.05829519405961037,
0.06171870976686478,
-0.005019556265324354,
0.027485432103276253,
-0.058426178991794586,
0.055581558495759964,
0.012496693059802055,
-0.020854996517300606,
-4.149785937102024e-8,
-0.09958866983652115,
0.028555594384670258,
-0.02506231889128685,
-0.04452512413263321,
-0.04068894684314728,
-0.07755439728498459,
0.029326148331165314,
0.09607268869876862,
0.010183162987232208,
-0.12669940292835236,
0.06219331547617912,
-0.028364982455968857,
0.0028646052815020084,
-0.015319488011300564,
0.0011249689850956202,
0.06609649211168289,
0.03596826270222664,
-0.051978591829538345,
-0.06560003012418747,
0.09813132882118225,
-0.06722113490104675,
-0.007098302710801363,
-0.02526864781975746,
0.07926157116889954,
0.02743859589099884,
-0.003492888528853655,
-0.04427908733487129,
-0.0031768183689564466,
0.13769599795341492,
-0.0791880264878273,
-0.0008378361235372722,
-0.006679767742753029,
0.03478593751788139,
0.027582380920648575,
-0.07084906101226807,
0.04233642295002937,
-0.05632113292813301,
0.028376275673508644,
-0.02079981192946434,
0.0017887799767777324,
-0.0050306497141718864,
0.034132346510887146,
0.002731297630816698,
0.025807064026594162,
0.031551294028759,
0.09987956285476685,
0.03337918594479561,
-0.09563340991735458,
0.035407327115535736,
-0.01151308510452509,
0.02165733277797699,
0.06384991854429245,
-0.04675930365920067,
0.10247593373060226,
0.06315657496452332,
-0.019259663298726082,
0.025740548968315125,
-0.012832940556108952,
0.00969433132559061,
0.0033211628906428814,
-0.0684683620929718,
0.041049472987651825,
-0.015207012183964252,
-0.02082781121134758
] | 0
| 0
| 0
| 0
|
|
Beyond Content Relevance: Evaluating Instruction Following in Retrieval Models
|
https://openreview.net/forum?id=OlRjxSuSwl
|
[
"Jianqun Zhou",
"Yuanlei Zheng",
"Wei Chen",
"Qianqian Zheng",
"Shang Zeyuan",
"Wei Zhang",
"Rui Meng",
"Xiaoyu Shen"
] |
Poster
|
Instruction-following capabilities in large language models (LLMs) have progressed significantly, enabling more complex user interactions through detailed prompts. However, retrieval systems have not matched these advances, most of them still relies on traditional lexical and semantic matching techniques that fail to fully capture user intent. Recent efforts have introduced instruction-aware retrieval models, but these primarily focus on intrinsic content relevance, which neglects the importance of customized preferences for broader document-level attributes. This study evaluates the instruction-following capabilities of various retrieval models beyond content relevance, including LLM-based dense retrieval and reranking models. We develop InfoSearch, a novel retrieval evaluation benchmark spanning six document-level attributes: Audience, Keyword, Format, Language, Length, and Source, and introduce novel metrics -- Strict Instruction Compliance Ratio (SICR) and Weighted Instruction Sensitivity Evaluation (WISE) to accurately assess the models' responsiveness to instructions. Our findings indicate that although fine-tuning models on instruction-aware retrieval datasets and increasing model size enhance performance, most models still fall short of instruction compliance. We release our dataset and code on https://github.com/EIT-NLP/InfoSearch.
|
LLM, Instruction-Following, Retrieval Model, Benchmark
|
We propose InFoSearch, a novel benchmark, along with an evaluation protocol to assess the depth of instruction-following capabilities in information retrieval (IR) models.
| 11,080
|
2410.23841
|
[
0.009505191817879677,
-0.06884631514549255,
0.053906772285699844,
0.031553130596876144,
0.004179263021796942,
0.009814102202653885,
-0.007548510562628508,
0.0941704586148262,
0.019334228709340096,
-0.04479159787297249,
0.0025236504152417183,
0.06444701552391052,
0.04714462533593178,
0.010587853379547596,
0.006191319786012173,
-0.05656248331069946,
0.12110111117362976,
0.005686592776328325,
-0.09509860724210739,
-0.04842366650700569,
0.08786721527576447,
0.127137690782547,
0.09456140547990799,
0.0067953793331980705,
-0.02149146981537342,
-0.06223157420754433,
-0.09332618117332458,
-0.02195913903415203,
0.061050720512866974,
-0.05461026355624199,
0.0020680634770542383,
0.061743807047605515,
0.019093601033091545,
0.11453300714492798,
-0.08644860237836838,
0.027471449226140976,
-0.022020017728209496,
-0.0033950319048017263,
0.02124144323170185,
-0.017889950424432755,
-0.0641799345612526,
-0.025701189413666725,
0.036626338958740234,
0.05262183025479317,
0.1003790870308876,
-0.0341397300362587,
-0.07156810909509659,
-0.06549827754497528,
-0.0291200689971447,
0.05814320594072342,
-0.06998617202043533,
0.02304011955857277,
-0.0048781512305140495,
0.0032880627550184727,
0.000395853043301031,
0.017365820705890656,
0.018411079421639442,
0.029846712946891785,
-0.0312008336186409,
-0.10079716891050339,
-0.10867992788553238,
-0.11425409466028214,
-0.02647125907242298,
0.01759677566587925,
0.012312864884734154,
-0.041887860745191574,
0.029998553916811943,
-0.04107178747653961,
-0.008918353356420994,
-0.0060545033775269985,
-0.016904721036553383,
0.009193012490868568,
-0.009297734126448631,
0.12564320862293243,
-0.0375552624464035,
0.057935114949941635,
0.04792255908250809,
-0.01115418877452612,
-0.0250768531113863,
-0.013160542584955692,
-0.01585197262465954,
-0.0470736101269722,
0.014967760071158409,
0.06308712065219879,
0.019282031804323196,
-0.05762379243969917,
0.087922103703022,
-0.034573618322610855,
0.061714667826890945,
-0.011405258439481258,
0.05496390908956528,
-0.045043643563985825,
0.018487660214304924,
-0.02172175981104374,
0.03142142295837402,
0.034439247101545334,
0.04288151115179062,
-0.11074350774288177,
-0.03166799992322922,
-0.007406708784401417,
0.04341603070497513,
0.10150362551212311,
-0.028629474341869354,
-0.08744529634714127,
-0.05689350515604019,
0.047250159084796906,
0.04991987720131874,
0.12280664592981339,
-0.057329412549734116,
-0.09422899782657623,
0.031814705580472946,
0.04681349918246269,
-0.009705190546810627,
-0.03471184894442558,
0.059376999735832214,
-0.03308491408824921,
0.037240542471408844,
-0.04371862858533859,
0.03562154248356819,
0.07648567110300064,
0.007737191393971443,
0.08672136068344116,
0.02471020258963108,
-0.061093855649232864,
-0.04493457078933716,
-0.023697784170508385,
-0.05684243142604828,
4.7588805134185645e-34,
0.06367955356836319,
0.06409352272748947,
-0.03447042033076286,
0.05869416147470474,
0.02032369002699852,
-0.033543214201927185,
0.03336507827043533,
0.09904662519693375,
-0.11713337898254395,
-0.020195448771119118,
-0.013258661143481731,
0.1019008606672287,
-0.019687848165631294,
0.026661578565835953,
0.06715290993452072,
0.05155668035149574,
-0.08066851645708084,
0.04573219642043114,
-0.0589541494846344,
0.029752347618341446,
0.09304345399141312,
-0.0352364182472229,
0.048804476857185364,
-0.029061932116746902,
0.03625010326504707,
-0.018918277695775032,
-0.028573907911777496,
-0.06564341485500336,
-0.08445770293474197,
-0.01191095169633627,
-0.04775328189134598,
-0.04101855680346489,
0.02708251215517521,
-0.055259596556425095,
0.039588045328855515,
-0.0714060366153717,
-0.03457529470324516,
-0.0172387957572937,
0.058271341025829315,
-0.08125880360603333,
-0.0169285349547863,
-0.009282441809773445,
0.028110772371292114,
-0.03875987231731415,
-0.07750061899423599,
-0.08176866918802261,
-0.04579491540789604,
0.010800939053297043,
0.015731966122984886,
0.00703867431730032,
0.07347007095813751,
0.04475584626197815,
-0.02342710644006729,
-0.005948466248810291,
0.02945907600224018,
0.05632125958800316,
0.008691372349858284,
0.050723619759082794,
-0.03283994272351265,
0.05505678802728653,
0.013963775709271431,
0.10408167541027069,
0.04611432924866676,
0.02993142604827881,
-0.011199630796909332,
-0.025452814996242523,
-0.026939045637845993,
0.07498060166835785,
0.0776563361287117,
0.0368933267891407,
-0.028671957552433014,
-0.05012671276926994,
0.0419049970805645,
-0.014148653484880924,
0.0472724512219429,
-0.009000134654343128,
-0.00801483541727066,
-0.12503941357135773,
0.04896451160311699,
-0.05791499465703964,
0.04216710478067398,
-0.09288901835680008,
0.038548488169908524,
-0.07415501028299332,
-0.058352019637823105,
-0.02877659909427166,
0.041061341762542725,
-0.030349595472216606,
0.007320270407944918,
-0.004133902955800295,
0.09160251915454865,
0.020543133839964867,
-0.0619080476462841,
-0.014563885517418385,
0.007747975178062916,
4.233846420801271e-34,
0.057811278849840164,
-0.08037188649177551,
0.01722808927297592,
0.009808982722461224,
0.019729750230908394,
-0.030454501509666443,
-0.006865287199616432,
0.04168194532394409,
-0.06159370020031929,
-0.11461454629898071,
-0.04052048549056053,
0.021806295961141586,
0.01367112435400486,
0.019495712593197823,
0.005006885156035423,
0.03233988583087921,
-0.012832733802497387,
0.013952377252280712,
0.031021131202578545,
0.06663458794355392,
-0.042217522859573364,
0.021328693255782127,
-0.06968619674444199,
0.031274862587451935,
0.014598334208130836,
0.04037248715758324,
-0.04589967057108879,
0.028336746618151665,
-0.010196167044341564,
-0.01339574996381998,
0.02509021945297718,
-0.019775284454226494,
0.008033151738345623,
-0.020635580644011497,
-0.026368925347924232,
-0.0070678917691111565,
0.030447598546743393,
-0.02673022262752056,
-0.05827968567609787,
0.1481708288192749,
0.06691223382949829,
0.0009332576883025467,
-0.014937139116227627,
0.006285383366048336,
-0.02468986250460148,
0.02300354279577732,
-0.04911557957530022,
-0.012094937264919281,
-0.013107401318848133,
-0.03573836758732796,
0.007037173956632614,
-0.02817676216363907,
-0.00042597862193360925,
0.029256928712129593,
-0.0008992572547867894,
-0.06380824744701385,
-0.06176573783159256,
-0.0729684978723526,
0.005322189535945654,
0.035103484988212585,
-0.0011691662948578596,
0.0106987738981843,
0.00034457319998182356,
-0.025165781378746033,
0.001854793168604374,
-0.07777829468250275,
-0.016327373683452606,
0.007577067706733942,
-0.10828118771314621,
0.0003512415278237313,
0.03251131623983383,
-0.10188522189855576,
0.07577928155660629,
-0.09157196432352066,
0.012206259183585644,
0.033390164375305176,
0.033525388687849045,
-0.0174220260232687,
-0.06418515741825104,
-0.013909829780459404,
-0.028664998710155487,
0.044254451990127563,
-0.021595675498247147,
0.058453913778066635,
0.0353587307035923,
0.0946856290102005,
-0.00612594885751605,
0.12764675915241241,
-0.00733198830857873,
-0.007487358525395393,
0.007799867540597916,
-0.012821021489799023,
0.01059561874717474,
-0.013448217883706093,
-0.046386901289224625,
-3.8824079950927626e-8,
-0.05094948038458824,
0.03427313268184662,
-0.07106135785579681,
0.12769390642642975,
-0.0016494363080710173,
0.042642101645469666,
-0.03584565222263336,
0.062474336475133896,
-0.06881595402956009,
0.017359362915158272,
0.06317383795976639,
-0.03773004561662674,
-0.013851365074515343,
-0.020388931035995483,
0.08310399204492569,
0.09245947003364563,
-0.02247045561671257,
0.019651103764772415,
-0.005790540482848883,
-0.03426401689648628,
0.006493572145700455,
0.0178397074341774,
0.027809007093310356,
0.000883029424585402,
-0.009213042445480824,
0.0874028429389,
0.023693986237049103,
0.08284646272659302,
0.09993260353803635,
-0.0029877647757530212,
0.03483257442712784,
-0.016707714647054672,
-0.043651778250932693,
-0.04034796357154846,
0.02547222375869751,
0.03846548870205879,
-0.007103690877556801,
-0.07445130497217178,
0.003440095577389002,
0.04221046715974808,
0.048060063272714615,
0.005216223653405905,
0.018346669152379036,
0.09793911129236221,
0.076230987906456,
0.027172207832336426,
-0.0585395023226738,
-0.010706149972975254,
0.059733156114816666,
-0.026149533689022064,
-0.010640082880854607,
-0.13696424663066864,
-0.06314688175916672,
0.03810384124517441,
0.04410643130540848,
0.02944048121571541,
-0.05400363355875015,
0.030825208872556686,
0.036205511540174484,
-0.03174698352813721,
0.06361018866300583,
0.010966943576931953,
-0.02154836803674698,
0.022116780281066895
] |
https://github.com/EIT-NLP/InfoSearch
| 5
| 0
| 0
| 0
|
RobustKV: Defending Large Language Models against Jailbreak Attacks via KV Eviction
|
https://openreview.net/forum?id=L5godAOC2z
|
[
"Tanqiu Jiang",
"Zian Wang",
"Jiacheng Liang",
"Changjiang Li",
"Yuhui Wang",
"Ting Wang"
] |
Poster
|
Jailbreak attacks circumvent LLMs' built-in safeguards by concealing harmful queries within adversarial prompts. While most existing defenses attempt to mitigate the effects of adversarial prompts, they often prove inadequate as adversarial prompts can take arbitrary, adaptive forms. This paper introduces RobustKV, a novel jailbreak defense that takes a fundamentally different approach by selectively removing critical tokens of harmful queries from key-value (KV) caches. Intuitively, for an adversarial prompt to be effective, its tokens must achieve sufficient `importance' (measured by attention scores), which consequently lowers the importance of tokens in the concealed harmful query. Therefore, by carefully evicting the KVs of low-ranked tokens, RobustKV minimizes the harmful query's presence in the KV cache, thus preventing the LLM from generating informative responses. Extensive evaluation using benchmark datasets and models demonstrates that RobustKV effectively counters state-of-the-art jailbreak attacks while maintaining the LLM's performance on benign queries. Notably, RobustKV creates an interesting effectiveness-evasiveness dilemma for the adversary, leading to its robustness against adaptive attacks.{(Warning: This paper contains potentially harmful content generated by LLMs.)}
|
Jailbreak Attack, Large Language Model, KV cache optimization
|
We have designed a special KV cache compression policy that can help LLMs defend against Jailbreak Attacks.
| 11,079
|
2410.19937
|
[
-0.05859909579157829,
-0.0067218909971416,
-0.032692909240722656,
0.0487281009554863,
0.014037441462278366,
0.06291144341230392,
0.048698633909225464,
-0.04772365838289261,
-0.011162709444761276,
-0.036673203110694885,
-0.04291772469878197,
-0.08605381101369858,
0.09186166524887085,
-0.06566635519266129,
-0.018659628927707672,
0.013770455494523048,
0.11432087421417236,
0.09126085788011551,
-0.09035710990428925,
-0.02771124430000782,
-0.019155554473400116,
0.025975171476602554,
0.04398645460605621,
-0.011173971928656101,
0.0061508105136454105,
-0.023787997663021088,
-0.09200078994035721,
-0.018068978562951088,
0.0043702321127057076,
0.021574245765805244,
-0.004756484646350145,
-0.022596484050154686,
0.009326057508587837,
0.08628244698047638,
-0.06844279170036316,
-0.03672739490866661,
-0.048584163188934326,
0.04066888988018036,
0.02979770302772522,
-0.14869114756584167,
-0.08521807938814163,
-0.01011406909674406,
-0.06491989642381668,
0.023081088438630104,
0.03231968730688095,
-0.028157325461506844,
-0.009772395715117455,
0.0029580669943243265,
-0.016806870698928833,
-0.0355830118060112,
-0.06707185506820679,
-0.02133370190858841,
0.011959869414567947,
0.08133900165557861,
-0.06754528731107712,
0.009128980338573456,
-0.03676309436559677,
0.07307115942239761,
0.04128838703036308,
0.06347241997718811,
-0.05374950170516968,
-0.012268987484276295,
0.021683769300580025,
0.03642046079039574,
-0.004221571609377861,
-0.06480593979358673,
0.03638371080160141,
0.042715396732091904,
0.03763852268457413,
0.0673011988401413,
0.03373611718416214,
0.08531562983989716,
-0.0493183434009552,
-0.017895326018333435,
-0.03770928457379341,
0.1388089507818222,
-0.08165274560451508,
-0.06021347641944885,
0.0219525545835495,
-0.059506867080926895,
0.04267776384949684,
-0.03229649364948273,
0.029663244262337685,
0.056330811232328415,
0.12169583141803741,
0.017067672684788704,
0.02520816959440708,
-0.044698137789964676,
0.057723596692085266,
0.037974197417497635,
-0.016271285712718964,
-0.01911410503089428,
0.10468072444200516,
0.02429034560918808,
0.06278795748949051,
0.012621687725186348,
-0.020762475207448006,
-0.04624788463115692,
-0.059467654675245285,
0.09650745987892151,
0.03495479002594948,
0.056785013526678085,
0.006788055412471294,
-0.06621012836694717,
0.06360860913991928,
0.01758934184908867,
0.04875963181257248,
-0.011923830024898052,
0.026822354644536972,
0.013186585158109665,
0.0020620054565370083,
0.08547241985797882,
0.04470520839095116,
-0.04506482556462288,
0.009928176179528236,
0.008098559454083443,
-0.015815623104572296,
0.01139142457395792,
0.01973673887550831,
0.044413186609745026,
0.06285130977630615,
0.044856272637844086,
-0.00817880965769291,
-0.021182511001825333,
0.018808357417583466,
-0.022283034399151802,
-0.023988287895917892,
3.4609007532665486e-33,
0.01568242721259594,
0.03194555640220642,
-0.03887583315372467,
-0.018038861453533173,
0.036906857043504715,
-0.037414561957120895,
0.004929831251502037,
0.08844300359487534,
-0.027733586728572845,
-0.00968550331890583,
0.00959650706499815,
-0.018648089841008186,
-0.002865471411496401,
0.0274820439517498,
0.07165013998746872,
0.06527487933635712,
-0.02105998620390892,
-0.014675559476017952,
-0.023783309385180473,
-0.03059605322778225,
0.0443606860935688,
-0.09834560006856918,
0.04793274402618408,
-0.015585022047162056,
0.06400857120752335,
0.01112140342593193,
0.02618616819381714,
-0.006187289021909237,
-0.02256542071700096,
0.057196542620658875,
-0.1429680585861206,
-0.08003496378660202,
-0.018400991335511208,
0.03959643840789795,
0.023196132853627205,
-0.030040010809898376,
0.002311669522896409,
-0.06438788771629333,
0.008845838718116283,
-0.02259235642850399,
-0.07670628279447556,
0.03229231759905815,
-0.01767793297767639,
0.003639588598161936,
-0.05085248500108719,
0.017495175823569298,
-0.08722491562366486,
-0.0016232035122811794,
-0.061956387013196945,
-0.03335965797305107,
-0.007430811412632465,
0.023767299950122833,
-0.030646780505776405,
-0.031038150191307068,
-0.09683291614055634,
-0.021449852734804153,
0.035839688032865524,
0.015111418440937996,
-0.06135282665491104,
0.09932009130716324,
-0.07259368896484375,
0.03137302026152611,
-0.009561656042933464,
0.02892306074500084,
-0.020055284723639488,
0.005127493292093277,
-0.01483314298093319,
0.04305048659443855,
0.011505034752190113,
0.022673364728689194,
-0.035791222006082535,
0.002421123441308737,
0.0050004334188997746,
-0.009535309858620167,
-0.08350250869989395,
0.003768304130062461,
0.05578933283686638,
-0.01597200147807598,
0.04304222762584686,
-0.05187877267599106,
0.07947272062301636,
-0.08123039454221725,
0.013800031505525112,
0.03908418118953705,
-0.06711623817682266,
-0.054946623742580414,
0.028995681554079056,
-0.03479374200105667,
-0.023568809032440186,
0.05200955644249916,
-0.006152689456939697,
0.0022454550489783287,
0.009091737680137157,
-0.004954732023179531,
-0.08355825394392014,
-4.2154410261677745e-33,
-0.04426117241382599,
-0.04894731566309929,
-0.09591507911682129,
0.08105036616325378,
-0.07980865985155106,
0.004784164484590292,
-0.07403679192066193,
0.04664861783385277,
-0.04168259724974632,
-0.059030357748270035,
-0.1202261671423912,
-0.012329175136983395,
0.10412082821130753,
-0.012778115458786488,
0.024436360225081444,
-0.006964801345020533,
-0.0004109225410502404,
0.031473129987716675,
-0.0738300234079361,
0.10816223174333572,
-0.001287639606744051,
0.11971849948167801,
-0.10952677577733994,
0.038054343312978745,
-0.04399542883038521,
-0.04254361614584923,
-0.09846131503582001,
0.041334982961416245,
0.026618868112564087,
0.007979713380336761,
0.010338383726775646,
-0.008043309673666954,
-0.02746756747364998,
-0.036472566425800323,
-0.05712130665779114,
0.006312084849923849,
0.05859190598130226,
-0.06592333316802979,
-0.0604085847735405,
0.13587281107902527,
0.11097034066915512,
-0.019931716844439507,
-0.07662242650985718,
0.00500867236405611,
-0.06371994316577911,
0.026100708171725273,
-0.023063572123646736,
-0.0034421770833432674,
0.00832180306315422,
-0.06314090639352798,
0.03985855355858803,
-0.016537023708224297,
-0.017979606986045837,
0.05875486880540848,
-0.040798965841531754,
-0.05656406655907631,
-0.05273784324526787,
-0.024309873580932617,
0.003969250246882439,
-0.00760966632515192,
-0.03929739445447922,
-0.034926626831293106,
-0.010081414133310318,
-0.05768294632434845,
0.08745412528514862,
-0.06408695876598358,
0.0049179005436599255,
0.005723598878830671,
0.018044820055365562,
-0.08266032487154007,
0.1178777888417244,
-0.05395151674747467,
-0.015772400423884392,
0.05281379818916321,
0.0042684548534452915,
0.028318943455815315,
-0.02228129841387272,
-0.10726498812437057,
-0.04193048179149628,
0.025012662634253502,
0.062226712703704834,
0.05046214535832405,
0.03735329955816269,
0.014272205531597137,
0.032374318689107895,
0.06252797693014145,
0.06605979800224304,
0.05989653989672661,
0.004797905683517456,
0.0487794429063797,
0.040001895278692245,
0.016272252425551414,
-0.11006806790828705,
0.08013910055160522,
0.0046236454509198666,
-3.961985939326951e-8,
-0.024034181609749794,
0.034401461482048035,
-0.061124224215745926,
0.05688237026333809,
0.06829974055290222,
-0.011122389696538448,
-0.07497157156467438,
0.09178024530410767,
-0.004183534067124128,
-0.05399025231599808,
0.061712153255939484,
-0.0071001616306602955,
-0.015673063695430756,
-0.042670685797929764,
-0.021114826202392578,
0.04368998482823372,
-0.025986624881625175,
0.04135692119598389,
-0.057994965463876724,
-0.058349642902612686,
-0.04408688470721245,
0.06284284591674805,
-0.003214291762560606,
-0.054962433874607086,
0.003324004588648677,
-0.052981045097112656,
-0.003007026854902506,
0.04002758488059044,
0.0488387830555439,
0.045527949929237366,
-0.022068452090024948,
-0.018692202866077423,
-0.00045202605542726815,
-0.01983697898685932,
0.01648523285984993,
0.1735985428094864,
-0.042896658182144165,
-0.07343384623527527,
0.025211215019226074,
0.027660489082336426,
-0.00022897476446814835,
0.022518418729305267,
-0.034043509513139725,
0.03443384915590286,
-0.05002788081765175,
0.024342216551303864,
0.014551732689142227,
-0.07658184319734573,
0.08927864581346512,
-0.04013873636722565,
0.032752882689237595,
-0.1052161380648613,
0.019134853035211563,
0.060152579098939896,
0.038293611258268356,
0.03646673634648323,
-0.01929711364209652,
-0.0023057793732732534,
0.028111077845096588,
0.04430891200900078,
0.10863471031188965,
-0.013620368205010891,
-0.01159190945327282,
0.003072837833315134
] | 0
| 0
| 0
| 0
|
|
From Probability to Counterfactuals: the Increasing Complexity of Satisfiability in Pearl's Causal Hierarchy
|
https://openreview.net/forum?id=rvvSSmGIFS
|
[
"Julian Dörfler",
"Benito van der Zander",
"Markus Bläser",
"Maciej Liskiewicz"
] |
Poster
|
The framework of Pearl's Causal Hierarchy (PCH) formalizes three types of reasoning: probabilistic (i.e. purely observational), interventional, and counterfactual, that reflect the progressive sophistication of human thought regarding causation. We investigate the computational complexity aspects of reasoning in this framework focusing mainly on satisfiability problems expressed in probabilistic and causal languages across the PCH. That is, given a system of formulas in the standard probabilistic and causal languages, does there exist a model satisfying the formulas?
Our main contribution is to prove the exact computational complexities showing that languages allowing addition and marginalization (via the summation operator) yield NP^{PP}-, PSPACE-, and NEXP-complete satisfiability problems, depending on the level of the PCH. These are the first results to demonstrate a strictly increasing complexity across the PCH: from probabilistic to causal and counterfactual reasoning. On the other hand, in the case of full languages, i.e.~allowing addition, marginalization, and multiplication, we show that the satisfiability for the counterfactual level remains the same as for the probabilistic and causal levels, solving an open problem in the field.
|
complexity, causal reasoning, Pearl's Causal Hierarchy
|
What is the complexity of equation systems containing probabilistic and causal terms?
| 11,078
|
2405.07373
|
[
-0.019266271963715553,
-0.03544582054018974,
0.04923813417553902,
0.008036074228584766,
0.028927119448781013,
0.00878078117966652,
-0.012548643164336681,
0.015431631356477737,
0.03605763241648674,
0.08341673761606216,
-0.0400475412607193,
-0.0388028584420681,
0.059423983097076416,
0.08327051997184753,
-0.01260384451597929,
-0.03581620007753372,
0.009752817451953888,
-0.00517182657495141,
-0.05055970698595047,
-0.0727727860212326,
-0.011459102854132652,
-0.06843680143356323,
-0.04795320704579353,
-0.018414855003356934,
-0.06341277062892914,
-0.021161027252674103,
-0.02144962176680565,
-0.05933165177702904,
0.04290684685111046,
0.05149564519524574,
-0.0624539852142334,
0.006068709306418896,
-0.014499234035611153,
-0.05349147319793701,
0.041878484189510345,
0.03886142000555992,
0.011197321116924286,
0.007302568759769201,
-0.02693241275846958,
0.029389824718236923,
0.013755126856267452,
-0.02813691645860672,
-0.009272890165448189,
-0.03830568119883537,
0.07957981526851654,
0.01738034188747406,
0.006355128716677427,
0.005686996504664421,
-0.05593010038137436,
-0.07733456045389175,
-0.10487069934606552,
-0.0027078213170170784,
0.003173800650984049,
-0.013474813662469387,
0.08882129937410355,
-0.09627676010131836,
0.06772218644618988,
0.026869043707847595,
-0.045684922486543655,
-0.10749489068984985,
-0.08046304434537888,
-0.06287826597690582,
-0.09318950027227402,
-0.020160725340247154,
0.03616105020046234,
0.09015841782093048,
-0.0448935329914093,
0.04665244743227959,
-0.02832365781068802,
0.17375773191452026,
-0.02851160243153572,
0.05317340046167374,
-0.05485447496175766,
0.05646643415093422,
-0.006187379825860262,
-0.0007556738564744592,
0.052988335490226746,
-0.006419582758098841,
-0.017342839390039444,
-0.01527587790042162,
0.00800970010459423,
0.02530515380203724,
0.021435966715216637,
0.009974799118936062,
0.07289934158325195,
-0.02193908765912056,
-0.01622913032770157,
0.021106155589222908,
0.04125409200787544,
0.03657817468047142,
-0.021155407652258873,
-0.059403568506240845,
0.02670900709927082,
0.026689579710364342,
0.03681686520576477,
0.016218969598412514,
-0.033287372440099716,
0.008529851213097572,
0.062113963067531586,
0.04145640507340431,
0.03576594963669777,
0.021335285156965256,
0.01556468941271305,
-0.08859223127365112,
0.012592467479407787,
0.01474013738334179,
-0.07336314022541046,
-0.011270209215581417,
-0.0156342051923275,
-0.051329612731933594,
-0.02049906924366951,
-0.031851064413785934,
0.05385933071374893,
0.004742676392197609,
-0.09998096525669098,
-0.012726599350571632,
0.003393247490748763,
-0.01103257481008768,
0.06426122784614563,
-0.1390388160943985,
0.08559153974056244,
0.024719029664993286,
0.08304329961538315,
0.008778804913163185,
0.07881138473749161,
-0.06706919521093369,
-0.01719694770872593,
1.1335841201246343e-33,
0.0017141294665634632,
0.036862459033727646,
0.01541640143841505,
-0.0011520660482347012,
0.03478459641337395,
0.06062161922454834,
-0.03660159558057785,
0.02545342594385147,
0.020665761083364487,
0.02959739789366722,
0.053087908774614334,
0.05703243240714073,
-0.05411900579929352,
-0.023876408115029335,
0.09690310806035995,
0.05935320630669594,
-0.008833395317196846,
0.044571150094270706,
-0.051609911024570465,
-0.03887435793876648,
0.08537168800830841,
-0.01938464678823948,
-0.045868344604969025,
0.03585115075111389,
-0.027310842648148537,
-0.028454633429646492,
0.03585498780012131,
-0.04877552390098572,
-0.023279942572116852,
-0.011528801172971725,
-0.045610349625349045,
-0.061808474361896515,
0.0023726134095340967,
0.0784626454114914,
-0.018220534548163414,
-0.08464936167001724,
-0.04359467327594757,
-0.1058923751115799,
0.005452514626085758,
-0.04999079555273056,
-0.032236699014902115,
0.04490600526332855,
0.021025776863098145,
0.0525752529501915,
-0.05411095544695854,
-0.11047515273094177,
-0.001981448382139206,
0.010854130610823631,
-0.00719371298328042,
-0.04336808994412422,
0.05633746460080147,
0.00569086242467165,
0.06477613002061844,
-0.045535773038864136,
-0.020240994170308113,
0.028893694281578064,
-0.050322525203228,
0.08491233736276627,
-0.02470248192548752,
0.03375772386789322,
0.0033787901047617197,
-0.03084409050643444,
0.010040991939604282,
0.004428219050168991,
0.0035110274329781532,
0.08153767883777618,
-0.1350426822900772,
-0.006211153697222471,
0.13472416996955872,
0.007925054058432579,
-0.033072613179683685,
0.03721301630139351,
0.01036989688873291,
0.04796406999230385,
0.06542621552944183,
-0.06099367514252663,
-0.023544423282146454,
-0.04464677721261978,
0.01883295364677906,
0.00480202492326498,
-0.10176504403352737,
-0.009115722961723804,
-0.007243028841912746,
0.05359894409775734,
-0.06063520163297653,
-0.06374932825565338,
-0.050908997654914856,
0.021975692361593246,
-0.02412387542426586,
-0.11259908229112625,
0.009421064518392086,
-0.04021848738193512,
0.06699710339307785,
0.02553732506930828,
-0.03794326260685921,
-4.020292433541966e-33,
-0.09853584319353104,
-0.02569759264588356,
0.004830470308661461,
-0.012790415436029434,
-0.01772967539727688,
-0.04400894418358803,
0.034351058304309845,
-0.07200325280427933,
0.062191177159547806,
-0.08627443015575409,
0.009661526419222355,
-0.026104124262928963,
0.12489090859889984,
0.003867589170113206,
-0.04476869851350784,
0.013582441955804825,
-0.029445592314004898,
0.06293381005525589,
0.0024235425516963005,
0.033690162003040314,
0.08651863038539886,
0.02252124808728695,
0.013602842576801777,
-0.05021252483129501,
0.05556934326887131,
0.051579348742961884,
-0.044495537877082825,
0.03471124544739723,
0.017415110021829605,
0.11631278693675995,
-0.03504970669746399,
-0.011960402131080627,
-0.09445077180862427,
-0.014305655844509602,
-0.035360462963581085,
0.028298931196331978,
0.08295605331659317,
0.06850268691778183,
-0.0687912181019783,
0.017194120213389397,
0.045958153903484344,
-0.062367044389247894,
-0.017537785694003105,
-0.03672116622328758,
0.025618692860007286,
0.02594800852239132,
-0.10204735398292542,
0.059145331382751465,
0.020876582711935043,
-0.057059187442064285,
-0.0005646018544211984,
0.02825944684445858,
0.045222990214824677,
0.07905332744121552,
-0.014061550609767437,
-0.012525605037808418,
-0.014004326425492764,
-0.05045459792017937,
0.05578041449189186,
0.0011747912503778934,
-0.07836510986089706,
-0.013923512771725655,
0.0016164955450221896,
0.07584211975336075,
0.005662053357809782,
0.01249469444155693,
-0.048719990998506546,
-0.08631234616041183,
-0.007688713725656271,
0.007516198325902224,
0.0023005083203315735,
0.07490752637386322,
-0.10206174850463867,
-0.01842435821890831,
0.007390848360955715,
0.04813776910305023,
0.027723902836441994,
0.016750724986195564,
-0.04690413177013397,
-0.09505753219127655,
0.04004613682627678,
-0.019907847046852112,
0.08787006139755249,
-0.06396972388029099,
-0.016682809218764305,
0.00582771934568882,
0.0459551177918911,
0.02074911631643772,
0.058899085968732834,
0.03290008381009102,
-0.06098737567663193,
0.043231260031461716,
-0.001329355058260262,
0.050546612590551376,
-0.10647361725568771,
-4.717047019653364e-8,
0.047317538410425186,
-0.016957692801952362,
-0.04056369885802269,
0.00014762982027605176,
0.06333523988723755,
-0.0984819233417511,
-0.06267966330051422,
0.004605851601809263,
-0.03808745741844177,
-0.012869230471551418,
0.024338863790035248,
0.05611071735620499,
0.01988905481994152,
-0.045166015625,
-0.026756443083286285,
0.04236293584108353,
0.026378419250249863,
-0.04189317673444748,
-0.04566610977053642,
0.026429254561662674,
0.06919099390506744,
-0.0362505204975605,
-0.12484539300203323,
0.04922471195459366,
0.014526765793561935,
-0.001967398449778557,
0.004287854302674532,
0.032911766320466995,
0.020621299743652344,
-0.00010184908023802564,
0.04585403949022293,
0.010813973844051361,
0.00323549029417336,
0.014969403855502605,
0.05381033942103386,
0.038610439747571945,
-0.008186934515833855,
0.019931994378566742,
-0.05674167722463608,
-0.06196865066885948,
0.03773227334022522,
0.06787868589162827,
-0.09734472632408142,
0.08974147588014603,
0.18177947402000427,
-0.03565023094415665,
-0.08005517721176147,
-0.043029867112636566,
-0.03290156275033951,
0.016773004084825516,
-0.018056487664580345,
-0.015682199969887733,
-0.018272660672664642,
-0.037111252546310425,
-0.022073278203606606,
0.09130965173244476,
0.0061130765825510025,
0.014930607751011848,
-0.011669088155031204,
0.006033605895936489,
0.10977808386087418,
0.09011179208755493,
0.0831112265586853,
-0.03683094307780266
] | 0
| 0
| 0
| 0
|
|
High-Dimensional Bayesian Optimisation with Gaussian Process Prior Variational Autoencoders
|
https://openreview.net/forum?id=SIuD7CySb4
|
[
"Siddharth Ramchandran",
"Manuel Haussmann",
"Harri Lähdesmäki"
] |
Poster
|
Bayesian optimisation (BO) using a Gaussian process (GP)-based surrogate model is a powerful tool for solving black-box optimisation problems but does not scale well to high-dimensional data. Previous works have proposed to use variational autoencoders (VAEs) to project high-dimensional data onto a low-dimensional latent space and to implement BO in the inferred latent space. In this work, we propose a conditional generative model for efficient high-dimensional BO that uses a GP surrogate model together with GP prior VAEs. A GP prior VAE extends the standard VAE by conditioning the generative and inference model on auxiliary covariates, capturing complex correlations across samples with a GP. Our model incorporates the observed target quantity values as auxiliary covariates learning a structured latent space that is better suited for the GP-based BO surrogate model. It handles partially observed auxiliary covariates using a unifying probabilistic framework and can also incorporate additional auxiliary covariates that may be available in real-world applications. We demonstrate that our method improves upon existing latent space BO methods on simulated datasets as well as on commonly used benchmarks.
|
Variational autoencoders, Gaussian processes, Bayesian optimisation
|
In this work, we propose a conditional generative model for efficient high-dimensional BO that uses a GP surrogate model together with GP prior VAEs.
| 11,070
| null |
[
-0.06227191165089607,
-0.1680414229631424,
0.033178120851516724,
0.0030813172925263643,
0.002665838925167918,
0.021900607272982597,
0.014062777161598206,
-0.035740938037633896,
-0.03249022737145424,
-0.03574612736701965,
-0.025712549686431885,
-0.004345695022493601,
-0.008604587987065315,
0.051042091101408005,
-0.037530843168497086,
0.002199535956606269,
0.14731095731258392,
0.09593622386455536,
-0.0840080976486206,
-0.04960116371512413,
0.020293965935707092,
0.07309294492006302,
-0.01356634683907032,
-0.02291194722056389,
0.006788274273276329,
-0.0214932132512331,
0.10916969180107117,
0.0223771333694458,
0.005506847519427538,
0.024919087067246437,
0.09260839968919754,
0.018460428342223167,
0.0014053439954295754,
-0.024965772405266762,
-0.004609578289091587,
0.08211696892976761,
-0.10453537106513977,
-0.0038226405158638954,
0.029547546058893204,
0.009715833701193333,
-0.003600404830649495,
0.0038031048607081175,
-0.020543519407510757,
0.09704619646072388,
0.014818524941802025,
-0.06459817290306091,
0.02690018340945244,
-0.09954585880041122,
0.00930356327444315,
-0.007825469598174095,
0.04175443947315216,
0.016478434205055237,
-0.042911287397146225,
-0.015606269240379333,
0.021380120888352394,
-0.015478056855499744,
0.023372147232294083,
-0.04848252981901169,
0.03442399576306343,
0.018612613901495934,
-0.04054492712020874,
-0.10380441695451736,
0.015657709911465645,
-0.07859022170305252,
-0.019064312800765038,
0.009784460999071598,
-0.0014983282890170813,
0.0737047791481018,
-0.029223930090665817,
0.04915934056043625,
-0.05149855464696884,
0.10919993370771408,
-0.09587125480175018,
-0.0038764041382819414,
-0.026115646585822105,
0.03306508809328079,
0.005274096038192511,
-0.06391574442386627,
0.059230346232652664,
-0.07501380890607834,
0.04075312614440918,
0.01503139641135931,
-0.005123179405927658,
0.05111856013536453,
0.10560527443885803,
0.04461173340678215,
0.0373266264796257,
0.10258978605270386,
0.03738273307681084,
0.021633995696902275,
0.0006629120325669646,
-0.008786165155470371,
-0.00895217340439558,
-0.02447623759508133,
-0.02421603910624981,
-0.03885183483362198,
0.016118595376610756,
-0.11681009829044342,
0.03838196396827698,
0.025935061275959015,
0.023124586790800095,
0.06402378529310226,
0.06366129964590073,
0.016424043104052544,
0.007462700363248587,
-0.09132308512926102,
0.02308242954313755,
0.15440739691257477,
0.016388805583119392,
-0.050286684185266495,
-0.0020961265545338392,
0.04442162066698074,
0.06211908161640167,
0.011330408044159412,
-0.02808600850403309,
0.014764883555471897,
0.04620411619544029,
-0.037086158990859985,
-0.004564810544252396,
0.02756255306303501,
-0.017852526158094406,
0.007609516382217407,
-0.04146844521164894,
0.013954054564237595,
-0.03668704256415367,
-0.040498942136764526,
-0.04770554229617119,
5.0345872032741626e-33,
0.016265157610177994,
0.00477439584210515,
0.07085101306438446,
0.03481932729482651,
-0.038907941430807114,
0.03435944765806198,
-0.03793376684188843,
-0.011248581111431122,
-0.018125304952263832,
0.04909578338265419,
-0.039669640362262726,
0.021566158160567284,
-0.1499747335910797,
0.0822889655828476,
0.06031080335378647,
0.03886096179485321,
-0.10606594383716583,
0.05799880623817444,
0.005515018478035927,
-0.0046577248722314835,
0.025483157485723495,
0.03194180876016617,
-0.08915126323699951,
0.053664430975914,
-0.01081880647689104,
-0.004857352934777737,
0.05735569819808006,
-0.06610547006130219,
-0.04048692062497139,
-0.03173767030239105,
-0.12042398750782013,
-0.04576966166496277,
0.0009370793122798204,
0.0702100619673729,
-0.0027296694461256266,
-0.02312856912612915,
0.014765911735594273,
-0.02534010075032711,
-0.04410328343510628,
-0.015077095478773117,
0.028425147756934166,
0.00640459731221199,
0.09438152611255646,
-0.0011727497912943363,
-0.13718454539775848,
-0.10266363620758057,
0.07382600009441376,
-0.031573835760354996,
0.02317514456808567,
-0.03539702668786049,
0.053180817514657974,
-0.0027812861371785402,
0.029353050515055656,
-0.03958258777856827,
-0.10255264490842819,
-0.009987409226596355,
0.005930463783442974,
-0.001108815660700202,
0.0698799341917038,
0.018528958782553673,
0.008047561161220074,
0.04356091096997261,
0.012436176650226116,
0.01333765871822834,
-0.020910637453198433,
-0.010630358941853046,
-0.0008405546541325748,
-0.0004027410177513957,
0.050205279141664505,
0.07098446041345596,
-0.02321857586503029,
0.02953348681330681,
0.0021932555828243494,
-0.13129101693630219,
0.09066790342330933,
0.04451707750558853,
0.021312223747372627,
-0.048904988914728165,
0.012407355010509491,
-0.033198282122612,
-0.03518689423799515,
-0.0031209303997457027,
-0.000027508251150720753,
-0.004296002443879843,
0.0007610323955304921,
-0.03069405071437359,
-0.04734257236123085,
-0.06436926871538162,
-0.1011328175663948,
-0.03383677452802658,
-0.061762578785419464,
-0.011788684874773026,
-0.009827861562371254,
-0.02380736730992794,
0.0014576134271919727,
-3.758136254815472e-33,
-0.02106502652168274,
0.013438115827739239,
0.060509759932756424,
0.01370918843895197,
0.007257069926708937,
-0.027424991130828857,
0.021974531933665276,
-0.04535231366753578,
-0.03980092704296112,
-0.09573030471801758,
-0.05724748224020004,
-0.05465933308005333,
0.07681103050708771,
0.03383127227425575,
-0.008011242374777794,
0.017861761152744293,
-0.04024174436926842,
0.047150418162345886,
-0.010981801897287369,
-0.029016120359301567,
0.001191791263408959,
0.06732799857854843,
0.014828885905444622,
-0.0030084382742643356,
-0.023921804502606392,
0.03647925332188606,
0.02012508548796177,
0.019225649535655975,
-0.003882902441546321,
0.0590023472905159,
-0.030126575380563736,
0.010520203039050102,
-0.019693449139595032,
0.03212055191397667,
-0.1043279692530632,
0.08673816174268723,
0.07986875623464584,
0.01740402728319168,
-0.06596294790506363,
0.007909678854048252,
0.02986312098801136,
-0.053171269595623016,
-0.003862454555928707,
0.02719416841864586,
0.004097538068890572,
0.02894502505660057,
0.041632164269685745,
0.0014060093089938164,
0.0715453028678894,
0.039141930639743805,
-0.03950927406549454,
0.11026778072118759,
-0.08194253593683243,
0.031563565135002136,
-0.04289275407791138,
-0.11580011248588562,
0.011522665619850159,
-0.1733638495206833,
0.10044871270656586,
-0.0028289910405874252,
0.01179793756455183,
-0.04096407815814018,
0.06014077365398407,
-0.08038453012704849,
-0.05251287296414375,
-0.04044368118047714,
0.010386662557721138,
0.05157998576760292,
-0.006943227257579565,
-0.022667089477181435,
0.002751793945208192,
-0.045019421726465225,
0.06749983131885529,
0.06907237321138382,
0.02305399253964424,
-0.04689401015639305,
0.052096035331487656,
0.0008019867236725986,
-0.017184047028422356,
-0.017427679151296616,
0.06956532597541809,
-0.04310627654194832,
0.007747955154627562,
0.03854837268590927,
0.007329355459660292,
0.051455240696668625,
0.0017526682931929827,
-0.03662547841668129,
0.06510549038648605,
-0.013034623116254807,
-0.10604898631572723,
0.023367980495095253,
0.02656637132167816,
0.028797369450330734,
-0.11519165337085724,
-3.8351014808313266e-8,
-0.042274605482816696,
0.05382988974452019,
0.070581816136837,
0.004040221683681011,
0.033805835992097855,
-0.023400122299790382,
0.00227051367983222,
0.09047337621450424,
0.04242105782032013,
-0.038806647062301636,
0.053942617028951645,
-0.0637049451470375,
-0.004370037000626326,
-0.08795792609453201,
0.04229244589805603,
0.035611841827631,
0.01908012665808201,
-0.0028091236017644405,
0.007611524313688278,
-0.0131031209602952,
-0.0003105426440015435,
0.051602527499198914,
-0.01732017658650875,
-0.0020612236112356186,
0.07941078394651413,
-0.08525493741035461,
-0.019441602751612663,
0.05154312029480934,
0.0935148373246193,
-0.006175657734274864,
0.026573210954666138,
0.035837624222040176,
-0.017153024673461914,
-0.020656965672969818,
-0.02996773272752762,
0.038140669465065,
0.03963052108883858,
0.021385308355093002,
-0.007999791763722897,
-0.0579194501042366,
-0.05694674327969551,
0.017371300607919693,
-0.10052364319562912,
0.026592187583446503,
0.03643505275249481,
0.06067566201090813,
0.027438022196292877,
-0.08019528537988663,
0.09605836868286133,
0.07651686668395996,
0.05098807066679001,
-0.050372399389743805,
0.04493032395839691,
0.09158686548471451,
0.06177138909697533,
0.004802827723324299,
0.07129858434200287,
-0.014209229499101639,
0.016055844724178314,
-0.04497336596250534,
-0.022834567353129387,
0.03644375130534172,
-0.05344986915588379,
0.027812357991933823
] | 0
| 0
| 0
| 0
|
|
LIFe-GoM: Generalizable Human Rendering with Learned Iterative Feedback Over Multi-Resolution Gaussians-on-Mesh
|
https://openreview.net/forum?id=gY08Ou8EL7
|
[
"Jing Wen",
"Alex Schwing",
"Shenlong Wang"
] |
Poster
|
Generalizable rendering of an animatable human avatar from sparse inputs relies on data priors and inductive biases extracted from training on large data to avoid scene-specific optimization and to enable fast reconstruction. This raises two main challenges: First, unlike iterative gradient-based adjustment in scene-specific optimization, generalizable methods must reconstruct the human shape representation in a single pass at inference time.
Second, rendering is preferably computationally efficient yet of high resolution.
To address both challenges we augment the recently proposed
dual shape representation, which combines the benefits of a mesh and Gaussian points, in two ways.
To improve reconstruction, we propose an iterative feedback update framework, which successively improves the canonical human shape representation during reconstruction.
To achieve computationally efficient yet high-resolution rendering, we study a coupled-multi-resolution Gaussians-on-Mesh representation.
We evaluate the proposed approach on the challenging THuman2.0, XHuman and AIST++ data. Our approach reconstructs an animatable representation from sparse inputs in less than 1s, renders views with 95.1FPS at $1024 \times 1024$, and achieves PSNR/LPIPS*/FID of 24.65/110.82/51.27 on THuman2.0, outperforming the state-of-the-art in rendering quality.
|
Generalizable human rendering, error feedback, dual representation
| null | 11,067
| null |
[
-0.022190311923623085,
-0.04270777478814125,
0.041484907269477844,
-0.01364271342754364,
0.06477733701467514,
0.0006769225583411753,
-0.026721319183707237,
-0.03541531786322594,
0.0033458578400313854,
-0.0027647847309708595,
-0.004115793853998184,
-0.039235323667526245,
-0.0016623103292658925,
0.07170948386192322,
-0.01962966099381447,
-0.04344005882740021,
0.10282379388809204,
0.07269785553216934,
-0.07089013606309891,
0.04070521518588066,
-0.020109061151742935,
-0.0020475785713642836,
0.0009769333992153406,
-0.023824824020266533,
0.06604444980621338,
-0.040518585592508316,
0.057994723320007324,
-0.0675283893942833,
-0.008581506088376045,
-0.015463540330529213,
0.007247239351272583,
-0.0335211455821991,
-0.02786371111869812,
-0.012837590649724007,
-0.08882542699575424,
0.03457610309123993,
0.02140728197991848,
0.06749856472015381,
-0.03661039471626282,
-0.04988829791545868,
0.029184134677052498,
0.025047916918992996,
-0.0633198469877243,
-0.00014217154239304364,
0.0824081301689148,
0.04834531247615814,
-0.022169623523950577,
-0.10354551672935486,
0.04611565172672272,
0.07328271865844727,
0.033010028302669525,
0.03449266031384468,
0.01714123599231243,
-0.014025176875293255,
0.038659051060676575,
0.026658903807401657,
0.005729044787585735,
-0.08315248787403107,
0.020910650491714478,
-0.01946335658431053,
0.006324631627649069,
-0.04418618977069855,
0.03705104812979698,
0.03137007728219032,
0.04614584892988205,
-0.10322465002536774,
0.0366397500038147,
-0.07528039813041687,
0.010060502216219902,
0.04371151328086853,
-0.012198060750961304,
-0.014207134954631329,
-0.013466009870171547,
-0.07166339457035065,
0.028296951204538345,
0.0037786844186484814,
-0.0003458267019595951,
-0.052585843950510025,
0.018920814618468285,
-0.012321898713707924,
0.07593581825494766,
0.07217030972242355,
0.028436413034796715,
-0.004944252781569958,
-0.012851571664214134,
0.043411217629909515,
-0.03271832689642906,
0.11011107265949249,
-0.002981761936098337,
0.014897486194968224,
-0.0272870771586895,
-0.011990650556981564,
0.0053627523593604565,
-0.010942350141704082,
-0.04841412231326103,
-0.017259230837225914,
0.056206632405519485,
-0.04797673225402832,
0.02013104408979416,
0.07299827039241791,
0.024097561836242676,
-0.05152980610728264,
0.07586415112018585,
-0.011194956488907337,
0.04277607426047325,
-0.013183888979256153,
0.08330985903739929,
0.07152877748012543,
0.04119793698191643,
-0.0267544724047184,
0.09371919929981232,
-0.06979977339506149,
-0.0315975546836853,
0.09785416722297668,
0.016590317711234093,
-0.019288478419184685,
0.014025004580616951,
-0.05178317055106163,
-0.03689732030034065,
0.033158134669065475,
-0.009962155483663082,
-0.013410086743533611,
-0.00664960453286767,
0.06832952052354813,
0.02499828115105629,
-0.06634808331727982,
-0.055842239409685135,
6.354447872411774e-34,
0.00449709827080369,
0.07889597117900848,
0.11320356279611588,
0.02063247375190258,
0.04255697503685951,
0.04275037720799446,
0.012998010031878948,
0.029401622712612152,
-0.0076174670830369,
0.013350787572562695,
-0.07245282083749771,
-0.06038044020533562,
-0.0535724051296711,
0.14336121082305908,
-0.0288570374250412,
-0.00011845634435303509,
-0.06537136435508728,
0.005193411372601986,
0.005927398335188627,
0.07540073245763779,
0.041253864765167236,
-0.05756993964314461,
-0.010778002440929413,
-0.016642263159155846,
-0.043439462780952454,
0.009671409614384174,
0.05830937623977661,
-0.005812028422951698,
-0.04620717465877533,
-0.0316784530878067,
-0.08411630988121033,
-0.010968155227601528,
-0.007115979213267565,
0.027477942407131195,
0.012000715360045433,
-0.04690269008278847,
0.07775084674358368,
-0.026458408683538437,
-0.0579647421836853,
0.01920512691140175,
0.017362423241138458,
0.0710306391119957,
0.019391464069485664,
-0.0031157045159488916,
-0.09782581776380539,
0.02705167420208454,
0.07180853933095932,
0.07841178774833679,
-0.10354168713092804,
-0.004341238643974066,
0.06301040202379227,
0.010596786625683308,
-0.04546643793582916,
-0.051112622022628784,
-0.0678979679942131,
-0.016471145674586296,
-0.027336087077856064,
-0.04716825857758522,
0.047204211354255676,
-0.023817608132958412,
0.0191995557397604,
0.006810850463807583,
-0.07532228529453278,
0.04562905803322792,
-0.10852336883544922,
0.01688288524746895,
0.04957006871700287,
0.03918918967247009,
0.0075608291663229465,
0.06026149168610573,
-0.07389085739850998,
0.017664873972535133,
0.047357916831970215,
-0.09177730232477188,
0.043902091681957245,
0.011356944218277931,
0.016525283455848694,
-0.03521973639726639,
-0.011352132074534893,
0.09384212642908096,
-0.041263602674007416,
0.05865737795829773,
-0.06161613017320633,
-0.05000091716647148,
-0.032303012907505035,
-0.08698152005672455,
0.044899832457304,
-0.010422979481518269,
0.009775561280548573,
-0.056250739842653275,
-0.05336149409413338,
-0.07161310315132141,
-0.01818288117647171,
-0.054015595465898514,
-0.03182157874107361,
8.037520683919125e-34,
0.06265170872211456,
0.006922414526343346,
-0.036961786448955536,
-0.010888205841183662,
0.08069994300603867,
-0.055385153740644455,
-0.013421662151813507,
0.04006229341030121,
0.003238166216760874,
-0.09968915581703186,
-0.01300494372844696,
-0.004837313666939735,
0.059016160666942596,
0.010369433090090752,
0.03516721352934837,
0.059902459383010864,
-0.0552055798470974,
-0.016015706583857536,
-0.058591488748788834,
-0.056693386286497116,
-0.02829016000032425,
-0.017141349613666534,
-0.03689490258693695,
-0.03767102584242821,
-0.10411012172698975,
0.17608746886253357,
-0.02009708620607853,
0.04036518558859825,
0.030160261318087578,
0.06348372250795364,
-0.04021518677473068,
-0.048482514917850494,
-0.001380958128720522,
-0.005398021545261145,
-0.0603342242538929,
0.07306326180696487,
0.027304496616125107,
0.0066103143617510796,
-0.028906557708978653,
0.031576938927173615,
0.016017846763134003,
-0.04676814004778862,
-0.0181825440376997,
0.1203535944223404,
0.031848546117544174,
0.01614815555512905,
-0.03703818470239639,
-0.04033809155225754,
-0.01741299033164978,
0.08097444474697113,
-0.0504831001162529,
0.038102250546216965,
-0.056250203400850296,
0.024416597560048103,
-0.05217962712049484,
-0.1331140249967575,
-0.05285493656992912,
0.004901085048913956,
0.12060873210430145,
-0.016760455444455147,
-0.035706449300050735,
-0.07344814389944077,
-0.09331567585468292,
0.022617820650339127,
-0.020517636090517044,
-0.008718031458556652,
-0.014092043042182922,
-0.055512685328722,
-0.02103843353688717,
0.08533903956413269,
0.0012086270144209266,
-0.07663172483444214,
-0.010113612748682499,
0.026145514100790024,
0.030555350705981255,
-0.07671638578176498,
0.018862351775169373,
0.0017663209000602365,
0.09147287160158157,
0.017891325056552887,
0.05534059554338455,
-0.05933615565299988,
0.06165415048599243,
0.07279039919376373,
0.17756232619285583,
0.06135145202279091,
0.011317147873342037,
0.09016148000955582,
0.03027316927909851,
0.025614222511649132,
-0.03766375780105591,
0.0894869714975357,
-0.024947162717580795,
0.011006372980773449,
0.010575596243143082,
-3.848561647146198e-8,
-0.10052855312824249,
0.01073460653424263,
0.027598580345511436,
-0.00029629579512402415,
-0.019837118685245514,
-0.06577257812023163,
0.05214712396264076,
0.04179020971059799,
-0.007301128003746271,
-0.0891445130109787,
0.023604843765497208,
-0.06889353692531586,
-0.0004015347803942859,
-0.035815007984638214,
0.08107902854681015,
0.07843687385320663,
-0.04953673854470253,
0.03863940387964249,
-0.04107210412621498,
-0.02866324409842491,
-0.09183083474636078,
0.02087881602346897,
-0.06614799052476883,
-0.051329806447029114,
0.05112922564148903,
-0.04055079072713852,
0.00350516801699996,
0.0714816302061081,
0.04253624379634857,
-0.028677916154265404,
-0.06129937618970871,
0.04263778030872345,
0.005230473354458809,
-0.005147196352481842,
0.0037263683043420315,
-0.02937890589237213,
0.02163616009056568,
0.014460301958024502,
0.030757972970604897,
-0.09823868423700333,
-0.0661792904138565,
0.10440202802419662,
0.01716790348291397,
-0.025497261434793472,
0.06659795343875885,
0.07962995022535324,
0.07370616495609283,
-0.0397728830575943,
0.029727715998888016,
0.04061600938439369,
-0.01123893354088068,
-0.024659795686602592,
0.05415986105799675,
0.09493402391672134,
0.04234844818711281,
-0.017391275614500046,
0.024948518723249435,
0.008157214149832726,
0.0964203029870987,
0.03545166179537773,
-0.00012514842092059553,
0.0033729765564203262,
-0.0655076652765274,
0.005082794930785894
] | 0
| 0
| 0
| 0
|
|
NNsight and NDIF: Democratizing Access to Open-Weight Foundation Model Internals
|
https://openreview.net/forum?id=MxbEiFRf39
|
[
"Jaden Fried Fiotto-Kaufman",
"Alexander Russell Loftus",
"Eric Todd",
"Jannik Brinkmann",
"Koyena Pal",
"Dmitrii Troitskii",
"Michael Ripa",
"Adam Belfki",
"Can Rager",
"Caden Juang",
"Aaron Mueller",
"Samuel Marks",
"Arnab Sen Sharma",
"Francesca Lucchetti",
"Nikhil Prakash",
"Carla E. Brodley",
"Arjun Guha",
"Jonathan Bell",
"Byron C Wallace",
"David Bau"
] |
Poster
|
We introduce NNsight and NDIF, technologies that work in tandem to enable scientific study of the representations and computations learned by very large neural networks. NNsight is an open-source system that extends PyTorch to introduce deferred remote execution. The National Deep Inference Fabric (NDIF) is a scalable inference service that executes NNsight requests, allowing users to share GPU resources and pretrained models. These technologies are enabled by the Intervention Graph, an architecture developed to decouple experimental design from model runtime. Together, this framework provides transparent and efficient access to the internals of deep neural networks such as very large language models (LLMs) without imposing the cost or complexity of hosting customized models individually. We conduct a quantitative survey of the machine learning literature that reveals a growing gap in the study of the internals of large-scale AI. We demonstrate the design and use of our framework to address this gap by enabling a range of research methods on huge models. Finally, we conduct benchmarks to compare performance with previous approaches.
Code, documentation, and tutorials are available at https://nnsight.net/.
|
interpretability, safety, large language models, distributed inference, scalable infrastructure, deferred execution, computation graphs, resource sharing
|
We create a framework for separating experimental code and model runtime, enabling researchers to run experiments on very large models without local hosting.
| 11,061
|
2407.14561
|
[
-0.0864592120051384,
-0.14304159581661224,
0.02374127320945263,
0.06538641452789307,
0.07312007248401642,
0.028304489329457283,
-0.12377765774726868,
-0.021083375439047813,
-0.0038434830494225025,
-0.06814916431903839,
-0.09777893126010895,
-0.04423331469297409,
-0.04153306782245636,
0.04043534770607948,
-0.02326146513223648,
-0.0005200410960242152,
0.103078693151474,
0.041723862290382385,
-0.11417801678180695,
-0.07395271956920624,
-0.009517948143184185,
0.06981010735034943,
0.050073496997356415,
-0.05264892429113388,
0.06443340331315994,
-0.06162726879119873,
0.0073110624216496944,
-0.09497898817062378,
0.027657581493258476,
-0.020140111446380615,
0.06981492787599564,
-0.025957778096199036,
-0.020442545413970947,
0.019401252269744873,
0.033381301909685135,
0.027800166979432106,
-0.055135756731033325,
-0.02047547698020935,
-0.04009052366018295,
-0.05671786889433861,
0.026856955140829086,
-0.023029465228319168,
0.02133357897400856,
0.028701677918434143,
0.10211998224258423,
0.019858643412590027,
0.014422625303268433,
-0.016631001606583595,
0.04564443975687027,
0.04049668833613396,
-0.07691038399934769,
-0.03876183554530144,
0.030944427475333214,
0.07172470539808273,
0.0011969424085691571,
-0.011045352555811405,
0.041450224816799164,
-0.03247176483273506,
-0.048141054809093475,
0.010369781404733658,
0.018825441598892212,
-0.082645945250988,
-0.007371414918452501,
-0.01916382648050785,
-0.04740374907851219,
0.040817681699991226,
0.02985565736889839,
0.04567893221974373,
0.007174174766987562,
-0.036599475890398026,
0.037947043776512146,
0.10809088498353958,
-0.061333391815423965,
-0.006335162557661533,
0.0011849143775179982,
0.014698483981192112,
0.03081289865076542,
-0.02110207825899124,
0.12107580900192261,
-0.10405965894460678,
0.008708035573363304,
0.015754567459225655,
0.04125029221177101,
0.02960178814828396,
0.10314185917377472,
0.0011855775956064463,
0.01701495610177517,
0.04424931854009628,
0.0003656227490864694,
0.005192589946091175,
-0.05146193131804466,
-0.060692042112350464,
0.03043173998594284,
-0.023178830742836,
-0.0021483106538653374,
0.03243144974112511,
-0.030353544279932976,
-0.07431776076555252,
-0.09068751335144043,
0.0973251461982727,
-0.04460272565484047,
0.056342605501413345,
0.03818514943122864,
-0.043878234922885895,
-0.029948676005005836,
0.04607868567109108,
0.05005646497011185,
0.05391789227724075,
0.02534785494208336,
-0.019945036619901657,
0.03528538718819618,
0.05281364917755127,
0.02768435701727867,
0.01775331050157547,
0.04683680087327957,
-0.014650180004537106,
-0.012913752347230911,
-0.03794744610786438,
0.009029376320540905,
0.08879663050174713,
-0.06329856067895889,
0.05950060486793518,
-0.057625606656074524,
0.050801124423742294,
0.004009348805993795,
0.030728919431567192,
-0.09808360785245895,
2.3600962114695553e-33,
0.036507803946733475,
0.051536936312913895,
0.06545459479093552,
0.027827126905322075,
0.08376403898000717,
-0.02615976706147194,
0.03298827260732651,
0.07087171077728271,
-0.07890088856220245,
-0.022679276764392853,
-0.02235819213092327,
0.02929438091814518,
-0.03862043097615242,
0.1116589829325676,
0.032464563846588135,
-0.03213544934988022,
-0.01779532991349697,
0.0471956692636013,
-0.004076074343174696,
0.0021779413800686598,
0.07311943918466568,
-0.003714520949870348,
-0.03359303995966911,
0.02823326364159584,
-0.009936237707734108,
-0.020702414214611053,
0.059347592294216156,
-0.0211141649633646,
0.017592189833521843,
-0.001441414118744433,
-0.06370825320482254,
-0.040492381900548935,
-0.004027196671813726,
0.08097997307777405,
0.06315155327320099,
-0.045560166239738464,
0.003325063968077302,
-0.02272329106926918,
0.018037056550383568,
-0.025956299155950546,
-0.04647275060415268,
0.04069919139146805,
0.00508918147534132,
-0.06588800251483917,
-0.1506197303533554,
-0.096693255007267,
-0.0241928119212389,
-0.007409829646348953,
-0.027912497520446777,
-0.07012593746185303,
0.07873667776584625,
0.057566218078136444,
-0.04732994735240936,
-0.0402190238237381,
0.024519212543964386,
-0.0625218003988266,
0.03144874423742294,
0.05113526061177254,
0.04778745025396347,
0.029503220692276955,
0.011951812542974949,
0.005621267016977072,
-0.0005103604053147137,
0.08154677599668503,
0.0013024818617850542,
0.0334634855389595,
-0.061793483793735504,
0.08445829898118973,
0.037585750222206116,
-0.02958770841360092,
-0.06458287686109543,
0.08398166298866272,
0.01990428753197193,
-0.009302177466452122,
0.022823266685009003,
0.019196033477783203,
0.03373473882675171,
-0.10203182697296143,
-0.007367289159446955,
0.018047712743282318,
0.012550953775644302,
0.037284862250089645,
-0.02616852894425392,
-0.006924598012119532,
-0.03616392984986305,
-0.046794213354587555,
-0.014422386884689331,
-0.04640282690525055,
0.014392905868589878,
-0.020721139386296272,
-0.04368063807487488,
-0.08700139820575714,
-0.016157517209649086,
-0.048348747193813324,
-0.02871549129486084,
-1.5027033052782486e-33,
-0.02404225617647171,
0.022943731397390366,
-0.05825231224298477,
0.011439201422035694,
0.0016027424717321992,
-0.07344208657741547,
-0.044234585016965866,
0.03149271383881569,
-0.010031919926404953,
-0.06686114519834518,
-0.0073912725783884525,
-0.0195118747651577,
0.1007252037525177,
-0.0019475785084068775,
0.050769977271556854,
-0.04685842990875244,
-0.014871244318783283,
0.025245605036616325,
0.002294567646458745,
-0.009324351325631142,
-0.025513211265206337,
0.09102616459131241,
-0.06527473032474518,
-0.010928970761597157,
-0.029058802872896194,
0.019784631207585335,
-0.054986622184515,
0.029282573610544205,
-0.010398367419838905,
-0.01926085166633129,
-0.06427767872810364,
-0.015279040671885014,
-0.12058725208044052,
-0.005924300756305456,
0.023781534284353256,
0.10160929709672928,
0.1028677299618721,
-0.031100459396839142,
-0.06232905015349388,
-0.03254168480634689,
0.07625432312488556,
-0.06235768273472786,
-0.07095832377672195,
0.0064430152997374535,
-0.03722689300775528,
-0.0029280937742441893,
-0.15825891494750977,
0.014175204560160637,
-0.007874086499214172,
0.006900186184793711,
-0.03362278267741203,
0.022304514423012733,
0.0038723861798644066,
0.028361497446894646,
-0.025881517678499222,
-0.07373423874378204,
0.05002626031637192,
0.012776843272149563,
0.058953721076250076,
0.02432401105761528,
-0.06586920469999313,
-0.040153998881578445,
-0.028771132230758667,
-0.04862939193844795,
-0.03865921497344971,
-0.04875591769814491,
-0.04611416906118393,
0.05654638633131981,
0.012535876594483852,
0.00743730366230011,
0.01715518720448017,
-0.06221975013613701,
0.013484621420502663,
0.09565338492393494,
-0.06458309292793274,
0.016663474962115288,
0.05197308957576752,
-0.04777169227600098,
-0.012977438047528267,
0.007992557249963284,
0.03966175764799118,
-0.052894193679094315,
0.05100274458527565,
0.003706442192196846,
0.1648591011762619,
0.0306199062615633,
0.07190701365470886,
0.056793004274368286,
0.008437189273536205,
0.06530433148145676,
-0.07831219583749771,
0.0447874441742897,
-0.03323321416974068,
0.0634351447224617,
-0.0317261628806591,
-4.004862930173658e-8,
-0.04799764230847359,
0.059125643223524094,
0.048067688941955566,
0.0608261413872242,
-0.007313122972846031,
-0.06052365154027939,
0.00396294379606843,
0.06761767715215683,
-0.043924666941165924,
0.06081542745232582,
0.11887990683317184,
-0.014127877540886402,
0.026672931388020515,
-0.08492376655340195,
-0.0240045003592968,
0.11185148358345032,
0.05295176059007645,
-0.03314189240336418,
0.0175764299929142,
-0.05507083982229233,
0.024416441097855568,
0.06167962774634361,
0.026059744879603386,
0.022050099447369576,
0.025561166927218437,
-0.08094710856676102,
-0.006613182369619608,
0.05812239646911621,
0.017925594002008438,
0.007820157334208488,
0.00023510903702117503,
0.039010562002658844,
0.014919150620698929,
-0.07567013800144196,
0.030659139156341553,
0.10279326140880585,
-0.006337251514196396,
0.019891316071152687,
0.037230174988508224,
-0.05324868485331535,
0.009120586328208447,
0.05463851988315582,
-0.07805778086185455,
0.013937526382505894,
0.03567970171570778,
-0.042902614921331406,
-0.011886809952557087,
-0.08543233573436737,
0.018153872340917587,
0.01933244802057743,
-0.0010374531848356128,
-0.011818964965641499,
0.00827210582792759,
0.10305828601121902,
0.00042308305273763835,
0.06496243923902512,
-0.0355064831674099,
-0.031379226595163345,
-0.0015053601237013936,
0.12088090181350708,
0.006411139387637377,
0.02127779833972454,
-0.03872964158654213,
0.04742667078971863
] |
https://github.com/ndif-team/nnsight
| 552
| 0
| 0
| 0
|
Protecting against simultaneous data poisoning attacks
|
https://openreview.net/forum?id=rK0YJwL69S
|
[
"Neel Alex",
"Shoaib Ahmed Siddiqui",
"Amartya Sanyal",
"David Krueger"
] |
Poster
|
Current backdoor defense methods are evaluated against a single attack at a time. This is unrealistic, as powerful machine learning systems are trained on large datasets scraped from the internet, which may be attacked multiple times by one or more attackers. We demonstrate that multiple backdoors can be simultaneously installed in a single model through parallel data poisoning attacks without substantially degrading clean accuracy. Furthermore, we show that existing backdoor defense methods do not effectively defend against multiple simultaneous attacks. Finally, we leverage insights into the nature of backdoor attacks to develop a new defense, BaDLoss (**Ba**ckdoor **D**etection via **Loss** Dynamics), that is effective in the multi-attack setting. With minimal clean accuracy degradation, BaDLoss attains an average attack success rate in the multi-attack setting of 7.98% in CIFAR-10, 10.29% in GTSRB, and 19.17% in Imagenette, compared to the average of other defenses at 63.44%, 74.83%, and 41.74% respectively. BaDLoss scales to ImageNet-1k, reducing the average attack success rate from 88.57% to 15.61%.
|
backdoors, backdoor defenses, data poisoning
|
Multiple backdoor attacks can be effectively installed at the same time, existing defenses fail when that happens, and ours does well.
| 11,060
|
2408.13221
|
[
-0.05644974485039711,
-0.05789100006222725,
0.01833828166127205,
0.008069812320172787,
0.03848293796181679,
-0.053842294961214066,
-0.025998717173933983,
-0.036556143313646317,
0.01630820706486702,
0.02802061103284359,
-0.016240598633885384,
-0.006442912854254246,
0.11692984402179718,
-0.08567007631063461,
-0.042552728205919266,
0.02997053973376751,
-0.013451877981424332,
0.08992965519428253,
-0.05107172206044197,
-0.019764672964811325,
-0.11310691386461258,
0.002036063699051738,
0.003366742515936494,
-0.010003848932683468,
-0.02936629392206669,
-0.002035904908552766,
-0.033560000360012054,
-0.052118897438049316,
-0.0674639344215393,
-0.08141954988241196,
0.05978665128350258,
-0.10421554744243622,
-0.04669351875782013,
0.027654096484184265,
-0.06543735414743423,
0.009734184481203556,
-0.0393887497484684,
0.005348462611436844,
-0.020441528409719467,
-0.030267180874943733,
-0.020720619708299637,
0.0468367338180542,
-0.018914781510829926,
0.025123756378889084,
-0.0679916962981224,
-0.00837759580463171,
0.0474630706012249,
0.02185891754925251,
0.07812441885471344,
-0.09592839330434799,
-0.05856946483254433,
0.004088227637112141,
-0.040481388568878174,
0.0856490507721901,
0.006751811131834984,
-0.09585072100162506,
-0.04627269133925438,
0.05035242438316345,
-0.04058445990085602,
0.06700822710990906,
0.014841815456748009,
-0.03653145581483841,
-0.028909992426633835,
0.014940536580979824,
0.030472738668322563,
-0.008187559433281422,
0.03390919417142868,
0.042086780071258545,
0.039250265806913376,
0.04205743968486786,
0.046433933079242706,
0.04871508851647377,
-0.10909140110015869,
0.0013737044064328074,
0.024512778967618942,
0.11585001647472382,
-0.045400746166706085,
-0.01857016608119011,
0.02000677026808262,
-0.08118339627981186,
0.05069439485669136,
-0.008389313705265522,
0.03618045523762703,
-0.0034282247070223093,
0.13799656927585602,
-0.03500170633196831,
-0.041215334087610245,
-0.03080374002456665,
0.06555263698101044,
0.021368926391005516,
0.05982501804828644,
0.020269516855478287,
0.09891598671674728,
0.0659286230802536,
0.09473687410354614,
0.060181181877851486,
-0.0182980764657259,
-0.07627078145742416,
-0.03845910727977753,
0.07707220315933228,
-0.07279503345489502,
-0.029943449422717094,
-0.009036225266754627,
-0.04806340113282204,
0.1271699219942093,
0.004481066949665546,
0.05815187841653824,
-0.017787542194128036,
0.017136331647634506,
-0.09550939500331879,
-0.006495986599475145,
0.058077264577150345,
-0.013755344785749912,
-0.01205188874155283,
0.016744650900363922,
-0.022687247022986412,
-0.035988472402095795,
-0.004345052409917116,
-0.0052373381331563,
0.0709770917892456,
-0.024112747982144356,
-0.00972647126764059,
0.01255047507584095,
-0.006536765955388546,
0.00008532279753126204,
-0.08690048009157181,
-0.09806673228740692,
4.350037333388834e-33,
-0.02924175374209881,
0.028727011755108833,
-0.058138906955718994,
-0.050359707325696945,
0.034869659692049026,
-0.062149014323949814,
-0.023235732689499855,
0.07540275901556015,
-0.08752597868442535,
0.038991235196590424,
-0.059348829090595245,
-0.05820700526237488,
0.018612245097756386,
0.1020822823047638,
0.08395099639892578,
0.08051353693008423,
0.0442531481385231,
0.057839978486299515,
0.01951996237039566,
-0.03396916389465332,
0.06142045930027962,
-0.09109312295913696,
-0.009906725957989693,
-0.05445336550474167,
0.03260207921266556,
0.015763401985168457,
0.01801692508161068,
0.09864787012338638,
-0.012716041877865791,
0.029797188937664032,
-0.04975341260433197,
0.006454061716794968,
0.007915914990007877,
0.07302961498498917,
0.019230669364333153,
-0.018838148564100266,
-0.0409616157412529,
0.02671118639409542,
-0.022183706983923912,
0.03517618030309677,
-0.07608287036418915,
0.021621497347950935,
-0.050989195704460144,
-0.018529940396547318,
0.029421215876936913,
-0.041899681091308594,
-0.08690014481544495,
-0.005084599833935499,
-0.07708657532930374,
-0.05915314331650734,
0.05316074937582016,
-0.03825295716524124,
-0.004254275467246771,
0.026954777538776398,
-0.10042203962802887,
0.029152996838092804,
0.021557755768299103,
-0.0036840075626969337,
0.04410233348608017,
0.15600742399692535,
0.024167101830244064,
0.024982238188385963,
-0.061134662479162216,
0.01700037345290184,
-0.021636320278048515,
-0.014011320658028126,
0.03983098268508911,
0.034622516483068466,
0.0006590204429812729,
0.00395179120823741,
-0.013918980024755001,
0.0166927520185709,
-0.004546566400676966,
-0.058278005570173264,
-0.005941961891949177,
-0.05398013815283775,
0.10113421082496643,
-0.017649631947278976,
0.06149308755993843,
0.029610957950353622,
-0.004103748593479395,
-0.0073683676309883595,
0.04345989227294922,
0.000055983698985073715,
-0.09603388607501984,
0.011665494181215763,
-0.050981029868125916,
-0.07985791563987732,
-0.06015079468488693,
0.0810428112745285,
-0.08515244722366333,
-0.0708436369895935,
0.003911267966032028,
-0.04030356928706169,
-0.04026264324784279,
-5.124465501338035e-33,
-0.09150168299674988,
-0.0006084588821977377,
-0.03230620548129082,
0.07098633795976639,
-0.0702492743730545,
0.01423198077827692,
-0.04228966310620308,
0.03441499546170235,
-0.022263512015342712,
0.01410288643091917,
0.01800713501870632,
0.022358855232596397,
0.034105923026800156,
-0.0465632788836956,
0.04911816120147705,
-0.022476697340607643,
0.023533416911959648,
0.015093578025698662,
-0.05929190292954445,
-0.03268108516931534,
0.002061299979686737,
0.09619870781898499,
-0.05197804048657417,
-0.022805415093898773,
0.011741372756659985,
0.020259955897927284,
-0.041043590754270554,
0.11637647449970245,
0.019859297201037407,
0.04766321927309036,
-0.03603770583868027,
-0.02484232559800148,
0.02284080721437931,
0.017484232783317566,
-0.021464021876454353,
0.061180077493190765,
0.07846616953611374,
-0.007694434840232134,
-0.010479995049536228,
0.04957142472267151,
0.1034003496170044,
0.06920497119426727,
-0.09557337313890457,
0.013475789688527584,
-0.014072207733988762,
0.01578219048678875,
-0.058246295899152756,
0.06325092911720276,
0.028455717489123344,
0.004029815085232258,
0.03299685940146446,
-0.01764313317835331,
-0.025404566898941994,
0.07398582994937897,
-0.0062866234220564365,
-0.059811774641275406,
0.07381915301084518,
-0.001781180500984192,
0.009657236747443676,
0.03699512407183647,
-0.05665086209774017,
-0.08556995540857315,
-0.04511483013629913,
0.07479321211576462,
0.03569852188229561,
-0.032118480652570724,
-0.03325682505965233,
0.020180027931928635,
-0.03349453955888748,
0.0031384904868900776,
0.018129196017980576,
0.023031916469335556,
-0.0474802665412426,
0.005916536785662174,
-0.05849193409085274,
0.003820492420345545,
-0.04720734432339668,
-0.07696451246738434,
-0.025157736614346504,
0.08007419854402542,
0.01699824072420597,
-0.04660271853208542,
-0.02152992971241474,
0.05485440418124199,
0.06535718590021133,
0.021394239738583565,
0.06131786108016968,
0.0037655774503946304,
0.07930580526590347,
0.0060288794338703156,
-0.0826340988278389,
0.004541880916804075,
-0.00781985092908144,
0.04827263578772545,
0.005658404435962439,
-5.0570648113534844e-8,
0.0353853665292263,
0.06432205438613892,
0.04585402086377144,
0.040188781917095184,
0.06463742256164551,
0.007339849602431059,
-0.12306573987007141,
0.09947194159030914,
-0.007596404291689396,
-0.06177736446261406,
0.11217053979635239,
-0.00034770838101394475,
-0.05162592977285385,
-0.08824248611927032,
-0.023147333413362503,
0.0520101934671402,
-0.03540416061878204,
-0.07009740173816681,
-0.040012847632169724,
-0.02902315743267536,
-0.014232344925403595,
0.0015975290443748236,
0.016763173043727875,
0.004198029171675444,
0.07302582263946533,
-0.015734218060970306,
0.017374414950609207,
0.03886176273226738,
-0.012997597455978394,
0.04164528101682663,
-0.07759629935026169,
-0.07423572242259979,
0.02216072380542755,
0.03743593394756317,
-0.03742501139640808,
0.1327762007713318,
0.002249240642413497,
-0.03955261409282684,
-0.03364437818527222,
-0.001059100846759975,
-0.04263227805495262,
-0.03027442656457424,
0.028629828244447708,
-0.01017686165869236,
-0.014572296291589737,
-0.005795050412416458,
0.02523692138493061,
-0.02846730314195156,
0.04111899062991142,
-0.02219325490295887,
0.05765698105096817,
0.01539684273302555,
0.01324023213237524,
0.10853863507509232,
0.059453368186950684,
-0.014881092123687267,
-0.008001566864550114,
-0.10694136470556259,
0.039537444710731506,
0.1169930174946785,
0.0333283469080925,
-0.046964067965745926,
-0.03834616392850876,
-0.039767954498529434
] | 0
| 0
| 0
| 0
|
|
Lift Your Molecules: Molecular Graph Generation in Latent Euclidean Space
|
https://openreview.net/forum?id=uNomADvF3s
|
[
"Mohamed Amine Ketata",
"Nicholas Gao",
"Johanna Sommer",
"Tom Wollschläger",
"Stephan Günnemann"
] |
Poster
|
We introduce a new framework for 2D molecular graph generation using 3D molecule generative models. Our Synthetic Coordinate Embedding (SyCo) framework maps 2D molecular graphs to 3D Euclidean point clouds via synthetic coordinates and learns the inverse map using an E($n$)-Equivariant Graph Neural Network (EGNN). The induced point cloud-structured latent space is well-suited to apply existing 3D molecule generative models. This approach simplifies the graph generation problem into a point cloud generation problem followed by node and edge classification tasks, without relying on molecular fragments nor autoregressive decoding. Further, we propose a novel similarity-constrained optimization scheme for 3D diffusion models based on inpainting and guidance. As a concrete implementation of our framework, we develop EDM-SyCo based on the E(3) Equivariant Diffusion Model (EDM). EDM-SyCo achieves state-of-the-art performance in distribution learning of molecular graphs, outperforming the best non-autoregressive methods by more than 26\% on ZINC250K and 16\% on the GuacaMol dataset while improving conditional generation by up to 3.9 times.
|
Drug Design, Computational Biology, Molecule Generation, Graph Generation, Latent Diffusion Models
|
We propose a latent Euclidean space for molecular graph generation and demonstrate that a diffusion model on such a space achieves state-of-the-art performance on common molecular graph generation benchmarks.
| 11,047
|
2406.10513
|
[
-0.05669353902339935,
-0.09704924374818802,
0.01958990842103958,
-0.016790954396128654,
0.008594455197453499,
-0.02613314427435398,
-0.019591331481933594,
-0.05941024422645569,
0.032356761395931244,
0.01709723472595215,
0.017583312466740608,
0.014973973855376244,
-0.019602328538894653,
0.09151424467563629,
-0.018476082012057304,
0.011640388518571854,
0.01768866740167141,
0.03323954716324806,
-0.04635313153266907,
0.00756585830822587,
0.006680880207568407,
0.035580869764089584,
0.03435030207037926,
0.03952789679169655,
0.11237508058547974,
0.0349663607776165,
0.11655455827713013,
0.037825409322977066,
0.01002525258809328,
-0.0013497744221240282,
0.15126822888851166,
-0.026437511667609215,
-0.06460381299257278,
0.025549137964844704,
-0.009827152825891972,
0.04056750610470772,
-0.09971153736114502,
0.04963398352265358,
0.011725005693733692,
-0.02627188339829445,
0.03650752827525139,
0.04332410916686058,
0.0014644248876720667,
0.08715690672397614,
0.0900707095861435,
-0.03053097054362297,
-0.08050602674484253,
-0.00225801020860672,
0.08240653574466705,
0.0021634555887430906,
0.015481393784284592,
-0.0344984345138073,
-0.10315763205289841,
-0.011883657425642014,
-0.021169034764170647,
0.010449937544763088,
-0.022160522639751434,
-0.035346467047929764,
0.022416796535253525,
-0.03461828455328941,
0.06416012346744537,
-0.08651704341173172,
-0.006036900915205479,
-0.03009822964668274,
-0.015623968094587326,
-0.0040252478793263435,
0.0911124050617218,
0.045664407312870026,
-0.01677127368748188,
0.07328372448682785,
-0.05075271427631378,
0.04892968013882637,
-0.08791939914226532,
0.020016515627503395,
0.01526191458106041,
0.04152611643075943,
0.04591881483793259,
-0.016972847282886505,
-0.056069016456604004,
0.001875768881291151,
0.03394729644060135,
0.08565885573625565,
0.11405953019857407,
0.0669362023472786,
-0.03827828913927078,
0.044102977961301804,
-0.019378526136279106,
0.05213959887623787,
0.009835119359195232,
-0.005371378269046545,
-0.016800442710518837,
0.03216610848903656,
-0.06439053267240524,
-0.045724060386419296,
-0.043251074850559235,
0.0005774741293862462,
0.06451089680194855,
0.0000697276700520888,
0.05848206207156181,
0.05431564897298813,
0.030545709654688835,
-0.023567726835608482,
-0.03670608997344971,
-0.03725673630833626,
0.05013897642493248,
0.0027112853713333607,
0.03599628433585167,
0.07284892350435257,
-0.014375418424606323,
-0.03236978128552437,
-0.0314352884888649,
0.049073703587055206,
0.013469737023115158,
0.032260239124298096,
0.005655600689351559,
-0.018770499154925346,
0.02422228269279003,
-0.05135837197303772,
0.04826321452856064,
0.05834105610847473,
-0.11348623037338257,
0.0019944722298532724,
-0.07709340751171112,
0.0037171572912484407,
-0.042027782648801804,
0.009490756317973137,
-0.0780709832906723,
1.810265524699146e-33,
0.011434359475970268,
0.05639498308300972,
0.08182895928621292,
0.037529315799474716,
0.0854501873254776,
0.05419251695275307,
-0.04359632357954979,
-0.03606278821825981,
-0.04281153902411461,
-0.008324015885591507,
-0.1588205248117447,
-0.030718624591827393,
-0.02711358666419983,
0.06391575932502747,
-0.02231557108461857,
-0.016160693019628525,
0.009945962578058243,
-0.0016609119484201074,
0.02192203514277935,
-0.025450890883803368,
-0.024131976068019867,
-0.06743448227643967,
0.021452080458402634,
0.036579035222530365,
-0.06277655065059662,
0.017123889178037643,
-0.044314611703157425,
0.006309567950665951,
0.03374843671917915,
-0.02236974611878395,
-0.10456901043653488,
0.01070760190486908,
-0.02423160709440708,
0.012377370148897171,
0.009247547946870327,
-0.0034473505802452564,
-0.026791764423251152,
-0.03076598234474659,
0.05541081354022026,
-0.09460905194282532,
0.036556024104356766,
0.08333558589220047,
0.0392388254404068,
-0.0589480847120285,
-0.11765444278717041,
0.06364472955465317,
0.07301843166351318,
-0.03710601106286049,
-0.006626035086810589,
-0.05775175616145134,
0.005847547668963671,
0.028939763084053993,
0.026523804292082787,
-0.03221500664949417,
-0.019607210531830788,
-0.05008351057767868,
0.061485420912504196,
-0.0305999293923378,
0.0883290246129036,
0.04061377793550491,
0.0007978263311088085,
0.026829741895198822,
0.024368496611714363,
-0.0041814809665083885,
0.007445183582603931,
-0.028706472367048264,
-0.09989987313747406,
0.03174504265189171,
0.05089820176362991,
0.06473874300718307,
-0.0018610624829307199,
0.0868457779288292,
-0.009877011179924011,
-0.03635280579328537,
0.06637734174728394,
0.027612147852778435,
-0.035094406455755234,
-0.07584553956985474,
0.05050403252243996,
0.004441349767148495,
-0.09796129167079926,
-0.03266262635588646,
-0.033258672803640366,
-0.07714246213436127,
-0.06480558216571808,
-0.12916617095470428,
0.004702202044427395,
-0.06123041361570358,
-0.012480293400585651,
-0.05462947115302086,
-0.030142253264784813,
-0.09365109354257584,
-0.009115761145949364,
0.018559135496616364,
0.03271271288394928,
-2.224364817515983e-33,
-0.0010770885273814201,
0.0897267535328865,
0.00601925840601325,
0.031031938269734383,
0.024483777582645416,
-0.004195741843432188,
-0.05416920781135559,
0.028199072927236557,
0.024168115109205246,
-0.0015333229675889015,
-0.0007716332911513746,
-0.01695505902171135,
0.01329880952835083,
-0.027773115783929825,
0.058127135038375854,
-0.010638942942023277,
-0.03146906569600105,
-0.02546806074678898,
-0.006956766825169325,
0.00491951871663332,
-0.017798544839024544,
0.010576145723462105,
-0.04369747266173363,
-0.008831596933305264,
-0.03824019059538841,
0.07674062252044678,
0.02161874622106552,
0.02269519492983818,
-0.017266854643821716,
0.04895744472742081,
-0.02701464481651783,
-0.0004537099739536643,
-0.04213052615523338,
-0.011271582916378975,
-0.07926442474126816,
0.0888170525431633,
0.039906006306409836,
-0.0016835228307172656,
-0.029493428766727448,
0.0006888320785947144,
-0.029183950275182724,
-0.027692560106515884,
-0.10299400985240936,
0.077630415558815,
0.05219214782118797,
0.07126352190971375,
0.026137126609683037,
0.015540029853582382,
-0.001867191051132977,
-0.0026724166236817837,
-0.02065710350871086,
0.0960296168923378,
0.02345597743988037,
0.023638786748051643,
0.01967286691069603,
-0.07441266626119614,
-0.06972051411867142,
-0.012119164690375328,
0.09321153163909912,
-0.044421274214982986,
-0.057908326387405396,
-0.10785775631666183,
0.03639012202620506,
-0.004018016625195742,
-0.07391594350337982,
-0.06567791104316711,
-0.013458684086799622,
0.0506594143807888,
0.04108646512031555,
0.0411004014313221,
-0.015368258580565453,
-0.017083769664168358,
0.023288752883672714,
-0.004782682750374079,
0.053937509655952454,
-0.10774041712284088,
0.03920779749751091,
-0.015684179961681366,
-0.008296926505863667,
-0.05222668498754501,
0.03660706803202629,
0.003462870605289936,
-0.022427313029766083,
0.08026693761348724,
0.01439953874796629,
-0.016970951110124588,
0.07881319522857666,
0.01254665944725275,
-0.01065105851739645,
0.05987369641661644,
0.0190407857298851,
0.03946209326386452,
-0.03447458893060684,
0.10285794734954834,
0.022092608734965324,
-3.743087617635865e-8,
-0.07566395401954651,
-0.005349807441234589,
0.10608231276273727,
-0.056291744112968445,
0.017667749896645546,
0.038886815309524536,
0.11844062805175781,
0.14799341559410095,
-0.00153721880633384,
0.014557738788425922,
0.07718349993228912,
-0.062238018959760666,
-0.01830778457224369,
-0.01934056356549263,
-0.02203822322189808,
0.042724188417196274,
0.02221365086734295,
0.0050192903727293015,
0.023271732032299042,
0.04651787504553795,
-0.11533104628324509,
0.0471959114074707,
0.019445989280939102,
-0.0005604274338111281,
0.08692139387130737,
-0.07392366975545883,
-0.0219613965600729,
-0.01979454606771469,
0.032154958695173264,
0.015022117644548416,
-0.06161569431424141,
-0.006922428496181965,
0.06672107428312302,
0.05363817885518074,
-0.06785205006599426,
0.04683360829949379,
-0.014522254467010498,
-0.055976156145334244,
-0.12887431681156158,
-0.034016694873571396,
-0.060406047850847244,
0.06413228064775467,
-0.018442904576659203,
-0.037479761987924576,
0.009867832064628601,
0.09299308061599731,
0.11099337041378021,
0.057852547615766525,
0.022779114544391632,
-0.04735684394836426,
0.016886556521058083,
-0.03821276128292084,
-0.038854245096445084,
-0.036135148257017136,
0.016177453100681305,
0.024822738021612167,
-0.11131302267313004,
-0.04799532890319824,
0.0741916075348854,
-0.06636273860931396,
-0.060906607657670975,
-0.04002320021390915,
-0.06794045120477676,
-0.06175161153078079
] | 0
| 0
| 0
| 0
|
|
Neural Sampling from Boltzmann Densities: Fisher-Rao Curves in the Wasserstein Geometry
|
https://openreview.net/forum?id=TUvg5uwdeG
|
[
"Jannis Chemseddine",
"Christian Wald",
"Richard Duong",
"Gabriele Steidl"
] |
Poster
|
We deal with the task of sampling from an unnormalized Boltzmann density $\rho_D$
by learning a Boltzmann curve given by energies $f_t$ starting in a simple density $\rho_Z$.
First, we examine conditions under which Fisher-Rao flows are absolutely continuous in the Wasserstein geometry.
Second, we address specific interpolations $f_t$ and the learning of the related density/velocity pairs $(\rho_t,v_t)$.
It was numerically observed that the linear interpolation,
which requires only a parametrization of the velocity field $v_t$,
suffers from a "teleportation-of-mass" issue.
Using tools from the Wasserstein geometry,
we give an analytical example,
where we can precisely measure the explosion of the velocity field.
Inspired by Máté and Fleuret, who
parametrize both $f_t$ and $v_t$, we propose an
interpolation which parametrizes only $f_t$ and fixes an appropriate $v_t$.
This corresponds to
the Wasserstein gradient flow of the Kullback-Leibler divergence related to Langevin dynamics.
We demonstrate by numerical examples that our model provides a well-behaved flow field which successfully solves the above sampling task.
|
Sampling, Boltzmann densities, Fisher-Rao Curves, Wasserstein Gradient Flows, Diffusion, Interpolations
| null | 11,046
|
2410.03282
|
[
-0.136543869972229,
-0.06027977541089058,
-0.005783652886748314,
0.07291100919246674,
0.07791049033403397,
-0.004402771592140198,
-0.01799662970006466,
0.023661648854613304,
0.049708735197782516,
-0.02845614403486252,
-0.0068796975538134575,
-0.02708418294787407,
-0.08458270132541656,
0.008686581626534462,
-0.05028260871767998,
-0.04515138640999794,
0.033390481024980545,
0.0890687108039856,
-0.09098168462514877,
0.0862555280327797,
-0.0693746879696846,
0.05921393632888794,
-0.006095467135310173,
0.013267861679196358,
0.06397990137338638,
-0.07827970385551453,
0.05988354980945587,
0.02097192220389843,
-0.04872838035225868,
0.027916954830288887,
0.07723068445920944,
-0.03440024331212044,
-0.08693540841341019,
0.0406702421605587,
-0.004545979667454958,
0.045453958213329315,
-0.02272717095911503,
0.06761379539966583,
-0.01895206794142723,
0.07457324862480164,
0.006918312981724739,
0.03290028125047684,
0.07282660901546478,
0.006872357334941626,
0.09177295863628387,
0.04662493243813515,
0.09952837228775024,
-0.02733144909143448,
-0.060349997133016586,
0.02153059095144272,
0.00908735767006874,
0.03073354996740818,
-0.037909019738435745,
0.07408948987722397,
0.05019117891788483,
0.0033692324068397284,
0.004199377726763487,
0.010408353060483932,
-0.00042827354627661407,
0.0015912199160084128,
0.015391085296869278,
-0.02494955062866211,
-0.05849265307188034,
-0.008182408288121223,
-0.030201803892850876,
-0.039890777319669724,
0.028266379609704018,
0.04765079915523529,
0.024056635797023773,
0.0600433312356472,
0.002986380597576499,
0.11493437737226486,
-0.06914188712835312,
-0.020961932837963104,
0.0166422501206398,
0.0176442489027977,
-0.009518898092210293,
0.011880365200340748,
0.07190863788127899,
-0.04080142080783844,
0.0634102001786232,
-0.0739765465259552,
-0.031719665974378586,
-0.10843579471111298,
-0.007779906503856182,
0.08135787397623062,
-0.013281187042593956,
0.11647692322731018,
0.02691180445253849,
-0.01862802542746067,
-0.0011121750576421618,
0.04282911494374275,
-0.136962428689003,
-0.0474146269261837,
-0.02763911336660385,
-0.0417889766395092,
0.013133175671100616,
-0.031923629343509674,
0.0703170895576477,
0.031673625111579895,
-0.05415438860654831,
-0.018911994993686676,
-0.025682954117655754,
0.0501420721411705,
-0.009248080663383007,
-0.06446489691734314,
0.09750694036483765,
0.10901028662919998,
-0.016067858785390854,
0.023052573204040527,
0.019683213904500008,
-0.039446376264095306,
0.05447377637028694,
0.022353220731019974,
-0.02117592841386795,
-0.025262344628572464,
-0.003702615387737751,
0.038852814584970474,
-0.03536050394177437,
0.05793551728129387,
-0.06642169505357742,
-0.022735856473445892,
-0.06783538311719894,
0.07834907621145248,
0.018223511055111885,
0.023369329050183296,
0.026029188185930252,
2.2959254238451897e-33,
0.02556709386408329,
0.018327098339796066,
0.0029656977858394384,
0.022946955636143684,
0.025500081479549408,
0.0923333391547203,
-0.02272890694439411,
-0.041996218264102936,
0.022244097664952278,
0.005917952861636877,
-0.050855547189712524,
-0.022048285230994225,
-0.011820648796856403,
-0.015414656139910221,
-0.030086269602179527,
-0.04138787090778351,
-0.07746659964323044,
0.001613583299331367,
-0.0037434257101267576,
-0.00044362424523569643,
0.11944662034511566,
0.030927833169698715,
-0.035988811403512955,
-0.011086348444223404,
-0.08757629990577698,
0.052317310124635696,
-0.039643701165914536,
0.026036936789751053,
0.017576517537236214,
0.05673348903656006,
-0.0836118534207344,
-0.0369088239967823,
-0.05455336719751358,
-0.07083940505981445,
0.03243168815970421,
0.01113254576921463,
0.004465071950107813,
-0.04980368912220001,
-0.08744703978300095,
-0.06415532529354095,
0.04643746837973595,
0.033131349831819534,
-0.02243683487176895,
-0.015154802240431309,
-0.09123394638299942,
-0.052828092128038406,
0.06692862510681152,
0.047715429216623306,
-0.010468479245901108,
-0.09190721064805984,
-0.028095310553908348,
-0.020916668698191643,
0.0636027380824089,
-0.07927118241786957,
0.07924743741750717,
-0.007191525772213936,
0.03917146474123001,
0.013619779609143734,
0.011434992775321007,
0.09170249104499817,
-0.02504979819059372,
0.04021693021059036,
-0.01701181009411812,
-0.015211776830255985,
0.07844189554452896,
0.05142850801348686,
-0.04072955250740051,
0.008162480778992176,
0.10793791711330414,
-0.0332910493016243,
0.0402841791510582,
0.04937046393752098,
-0.010990535840392113,
-0.11395786702632904,
-0.027258245274424553,
0.01790524460375309,
-0.011213752441108227,
0.05056776478886604,
0.029229339212179184,
0.01349056139588356,
-0.05056324601173401,
-0.04937043413519859,
0.012208248488605022,
-0.008588753640651703,
0.0008193585672415793,
-0.03611757233738899,
-0.01629672572016716,
-0.07758677005767822,
-0.09580446779727936,
-0.05803642049431801,
-0.008104698732495308,
-0.01980026438832283,
-0.022239303216338158,
-0.06878059357404709,
0.0091335978358984,
-2.3715851993809047e-33,
-0.06976766884326935,
0.0318518728017807,
-0.03734419122338295,
0.08409643173217773,
-0.025239793583750725,
0.07597426325082779,
-0.035562366247177124,
-0.016804024577140808,
0.015531950630247593,
-0.12203656882047653,
-0.05599632486701012,
-0.024798208847641945,
-0.03846107795834541,
-0.0025540629867464304,
0.07398495078086853,
-0.034528449177742004,
-0.0008000124944373965,
-0.09587697684764862,
0.045510999858379364,
-0.062407489866018295,
0.010077854618430138,
-0.008619768545031548,
-0.020335832610726357,
0.019013606011867523,
-0.009172401390969753,
0.02368578128516674,
-0.014515082351863384,
0.07740776985883713,
-0.03479794040322304,
0.07915422320365906,
-0.009534798562526703,
-0.06732076406478882,
-0.034146226942539215,
-0.04128043353557587,
-0.09662161767482758,
0.17178528010845184,
0.10694198310375214,
0.07099707424640656,
-0.02152802236378193,
0.023949014022946358,
0.018776237964630127,
-0.00873954314738512,
-0.07295489311218262,
-0.018389007076621056,
-0.0419432558119297,
0.05482017621397972,
0.0131844412535429,
0.044471316039562225,
0.032099105417728424,
0.0019097222248092294,
-0.04733764007687569,
0.026098722591996193,
-0.07055702060461044,
0.07032528519630432,
-0.06298511475324631,
0.016230668872594833,
0.011632277630269527,
-0.06505178660154343,
0.07882608473300934,
-0.08728169649839401,
-0.0868004858493805,
-0.07854697853326797,
-0.0062389797531068325,
0.04259481653571129,
-0.01647687517106533,
0.006779080722481012,
-0.009884903207421303,
-0.05869687721133232,
0.019128993153572083,
0.013487784191966057,
0.02782929129898548,
-0.07091382890939713,
0.04909555986523628,
0.044729214161634445,
-0.03566700965166092,
-0.012827165424823761,
0.10319692641496658,
0.01622079871594906,
-0.028011711314320564,
-0.014312179759144783,
0.012143155559897423,
0.006598769221454859,
0.06321155279874802,
0.05675826221704483,
0.12839369475841522,
0.004092687275260687,
-0.04096544161438942,
-0.06587957590818405,
0.006143901962786913,
-0.04707081615924835,
0.01714213192462921,
0.03202619403600693,
0.047656890004873276,
0.03272629901766777,
-0.0163640845566988,
-4.651215945727927e-8,
-0.02624029666185379,
-0.056095920503139496,
-0.0022191123571246862,
-0.010621310211718082,
0.04469092935323715,
0.019343674182891846,
0.002954880241304636,
0.05889630317687988,
-0.013459342531859875,
-0.0024667801335453987,
0.09433815628290176,
0.0018568282248452306,
0.03490851819515228,
-0.01573970541357994,
0.05479462444782257,
0.0101577527821064,
0.009892192669212818,
-0.050909705460071564,
-0.018165988847613335,
0.04137276113033295,
-0.009263476356863976,
0.0309741348028183,
-0.05662776902318001,
-0.0753866583108902,
0.03449263423681259,
0.049882277846336365,
-0.09159533679485321,
-0.10759594291448593,
-0.01049380749464035,
-0.001217102399095893,
-0.016220131888985634,
-0.0017475758213549852,
0.06824222207069397,
0.0275966078042984,
-0.08114107698202133,
0.048297181725502014,
-0.04472118243575096,
0.09742624312639236,
-0.10249094665050507,
-0.02626773528754711,
0.016110735014081,
0.08836659044027328,
-0.031614288687705994,
0.003972394857555628,
0.07690797001123428,
0.06364164501428604,
-0.011052027344703674,
0.042199231684207916,
0.009140778332948685,
0.0447244755923748,
-0.06137736886739731,
0.015294436365365982,
0.027243133634328842,
-0.014164712280035019,
0.03926890715956688,
-0.04444119334220886,
-0.013059587217867374,
-0.0846448615193367,
0.035959601402282715,
0.014986940659582615,
-0.01883402280509472,
0.05563872307538986,
-0.057619500905275345,
-0.00013948199921287596
] | 0
| 0
| 0
| 0
|
|
GEVRM: Goal-Expressive Video Generation Model For Robust Visual Manipulation
|
https://openreview.net/forum?id=hPWWXpCaJ7
|
[
"Hongyin Zhang",
"Pengxiang Ding",
"Shangke Lyu",
"Ying Peng",
"Donglin Wang"
] |
Poster
|
With the rapid development of embodied artificial intelligence, significant progress has been made in vision-language-action (VLA) models for general robot decision-making. However, the majority of existing VLAs fail to account for the inevitable external perturbations encountered during deployment. These perturbations introduce unforeseen state information to the VLA, resulting in inaccurate actions and consequently, a significant decline in generalization performance. The classic internal model control (IMC) principle demonstrates that a closed-loop system with an internal model that includes external input signals can accurately track the reference input and effectively offset the disturbance. We propose a novel closed-loop VLA method GEVRM that integrates the IMC principle to enhance the robustness of robot visual manipulation. The text-guided video generation model in GEVRM can generate highly expressive future visual planning goals. Simultaneously, we evaluate perturbations by simulating responses, which are called internal embeddings and optimized through prototype contrastive learning. This allows the model to implicitly infer and distinguish perturbations from the external environment. The proposed GEVRM achieves state-of-the-art performance on both standard and perturbed CALVIN benchmarks and shows significant improvements in realistic robot tasks.
|
Robot Manipulation; Vision Language Action Model
|
We propose a novel closed-loop VLA method GEVRM that integrates the internal model control principle to enhance the robustness of robot visual manipulation.
| 11,042
|
2502.09268
|
[
0.05874928459525108,
-0.0976787880063057,
0.027675706893205643,
-0.002435004571452737,
0.043496545404195786,
0.02028876356780529,
0.004153543151915073,
0.015092686749994755,
0.05161315202713013,
-0.013227857649326324,
0.04159761220216751,
-0.040540676563978195,
0.009373473934829235,
0.04552580788731575,
0.014909151941537857,
0.00720862066373229,
0.10732652246952057,
0.036587368696928024,
-0.10105741024017334,
0.0371202751994133,
0.10277806967496872,
0.018941722810268402,
0.04881533607840538,
-0.003920965828001499,
-0.0294259712100029,
0.04822734370827675,
0.008131382055580616,
-0.022928614169359207,
0.023293638601899147,
-0.07094859331846237,
-0.023067209869623184,
-0.07495526969432831,
-0.023090707138180733,
-0.041401345282793045,
-0.071451336145401,
0.0671873539686203,
-0.04357670620083809,
-0.025804005563259125,
-0.06078346446156502,
-0.07984255999326706,
-0.08704420179128647,
0.027723871171474457,
0.05445728451013565,
-0.0180734284222126,
0.07225429266691208,
0.06711620092391968,
0.017601115629076958,
-0.11989505589008331,
-0.00011220440501347184,
0.0500011071562767,
-0.10241743177175522,
-0.039297886192798615,
0.03757458180189133,
-0.09374038875102997,
-0.01768539287149906,
0.03645014762878418,
0.047516342252492905,
-0.03278615325689316,
0.04523929953575134,
-0.03549632057547569,
0.04297276958823204,
-0.08338519185781479,
0.022078542038798332,
-0.04291563481092453,
-0.0484478659927845,
-0.07108533382415771,
0.1086675152182579,
-0.011696630157530308,
-0.027792181819677353,
0.01728675328195095,
-0.019035756587982178,
0.0054016150534152985,
0.03827722370624542,
-0.008821030147373676,
-0.04050479456782341,
0.059940360486507416,
-0.023988837376236916,
-0.02028319053351879,
0.09270306676626205,
-0.03796418756246567,
0.07207256555557251,
0.017458628863096237,
-0.07130623608827591,
0.061333026736974716,
0.06941276788711548,
0.08549509942531586,
0.00037402883754111826,
0.07628075033426285,
0.09896344691514969,
0.011919430457055569,
-0.021148113533854485,
-0.09010117501020432,
0.015987979248166084,
-0.04029931128025055,
0.00980058778077364,
0.012009118683636189,
0.015542148612439632,
-0.10945545881986618,
-0.04695429652929306,
0.03730103373527527,
0.029176069423556328,
0.032915931195020676,
0.019610095769166946,
0.01849157176911831,
0.030676549300551414,
-0.007121745031327009,
0.08128415048122406,
0.038161080330610275,
-0.02082837000489235,
-0.05965929478406906,
-0.004395546857267618,
-0.08435360342264175,
0.0524919256567955,
0.0738239735364914,
0.03796004876494408,
0.009138713590800762,
0.04512335732579231,
0.012289741076529026,
0.028337588533759117,
0.0025098174810409546,
0.05403565987944603,
-0.038557350635528564,
0.016125213354825974,
0.05379866436123848,
0.03506988286972046,
-0.025679651647806168,
-0.012563750147819519,
-1.1182812040578436e-33,
0.06637509912252426,
-0.030331052839756012,
0.042307619005441666,
0.09286729991436005,
0.06620960682630539,
0.012356788851320744,
0.021944846957921982,
0.018145063892006874,
-0.0728512778878212,
0.009831332601606846,
0.024388818070292473,
-0.08880529552698135,
-0.06577538698911667,
0.12972867488861084,
0.01951746828854084,
0.02735084854066372,
-0.01141568087041378,
0.023472707718610764,
-0.05477556586265564,
0.03385183960199356,
0.03451404720544815,
0.001761494087986648,
0.025851866230368614,
-0.09637156873941422,
-0.007677867077291012,
0.050374098122119904,
0.007013426627963781,
0.024041911587119102,
-0.03389640524983406,
-0.03569280356168747,
-0.046252019703388214,
0.044768527150154114,
0.04169490188360214,
0.03019941970705986,
0.02826230227947235,
-0.0007918919436633587,
-0.043315988034009933,
-0.058389924466609955,
0.0038937858771532774,
0.020426401868462563,
-0.02294389344751835,
0.07910893857479095,
-0.017842557281255722,
0.020378004759550095,
-0.07047606259584427,
-0.0020558717660605907,
0.027357343584299088,
0.09287817031145096,
-0.1079128086566925,
-0.04034360498189926,
0.0036133420653641224,
0.07740189880132675,
-0.0781305730342865,
-0.06140792742371559,
0.018169283866882324,
-0.041911717504262924,
0.04169799014925957,
0.06042822450399399,
-0.05882555991411209,
-0.014871619641780853,
-0.02650449424982071,
0.055371712893247604,
0.008510136976838112,
0.07671498507261276,
0.0077386037446558475,
0.041746027767658234,
-0.008596669882535934,
0.11146045476198196,
0.09551780670881271,
-0.004177979659289122,
-0.0073936874978244305,
0.009367118589580059,
-0.03328085318207741,
-0.11130961775779724,
0.008526311255991459,
-0.0019008176168426871,
-0.0061383298598229885,
-0.10960403084754944,
-0.011066023260354996,
-0.007047764956951141,
-0.03201252594590187,
0.00619605602696538,
-0.029538560658693314,
0.02652459591627121,
0.022258078679442406,
-0.037386972457170486,
0.029103169217705727,
-0.022380800917744637,
-0.023586641997098923,
-0.01935759000480175,
-0.06967606395483017,
-0.028267784044146538,
-0.017046868801116943,
0.02467108890414238,
-0.054803524166345596,
1.5716698113886067e-34,
0.01501015666872263,
0.05828629434108734,
0.011335154995322227,
0.008213651366531849,
-0.006088161841034889,
-0.03762055188417435,
0.054564908146858215,
-0.023864617571234703,
-0.0010362331522628665,
-0.07987041771411896,
-0.05719750374555588,
-0.03911777213215828,
-0.032569464296102524,
0.0759650319814682,
0.019848192110657692,
-0.017674895003437996,
0.022044215351343155,
-0.01195412129163742,
0.008084908127784729,
-0.041678253561258316,
-0.008003783412277699,
0.07020528614521027,
-0.01467074453830719,
-0.006634784862399101,
-0.040340591222047806,
0.03547372668981552,
-0.03225924074649811,
0.05620179697871208,
-0.029236789792776108,
-0.026364801451563835,
-0.027834385633468628,
0.004145060665905476,
-0.01150133553892374,
0.007120966445654631,
-0.007812676019966602,
0.007722774986177683,
0.05737219750881195,
-0.0429716520011425,
0.004842094145715237,
0.03441235423088074,
0.03755547106266022,
-0.05092683061957359,
-0.05843803659081459,
0.020986881107091904,
0.008777542039752007,
0.0178049486130476,
-0.051337894052267075,
0.014499813318252563,
-0.017458057031035423,
0.007155256811529398,
-0.09433754533529282,
0.005672790575772524,
-0.09632658213376999,
-0.031876374036073685,
-0.0816861242055893,
-0.05582854896783829,
-0.03258275240659714,
0.002270629396662116,
0.11841576546430588,
0.022029053419828415,
-0.05818624049425125,
0.04843182861804962,
0.01063686516135931,
-0.07002609223127365,
-0.0179870817810297,
0.11296524852514267,
-0.028565550222992897,
0.005366384983062744,
0.08030693233013153,
-0.028379470109939575,
0.05187557265162468,
-0.011654583737254143,
0.049084000289440155,
0.007384816650301218,
0.06782050430774689,
-0.01766999252140522,
-0.005069516599178314,
-0.0891479104757309,
0.03459055349230766,
-0.07872091233730316,
0.024324800819158554,
-0.01280551590025425,
0.037294529378414154,
0.08717828243970871,
0.009785352274775505,
0.11679810285568237,
-0.018298408016562462,
0.04617952182888985,
0.02391147054731846,
0.01760214939713478,
-0.044316865503787994,
0.05375637486577034,
-0.0038459666538983583,
0.1206236481666565,
-0.03172991797327995,
-4.42168293091072e-8,
-0.1375225931406021,
0.034487534314394,
0.03063168190419674,
0.024398304522037506,
0.017612988129258156,
-0.018541967496275902,
-0.027360739186406136,
-0.045383814722299576,
0.0612504668533802,
-0.04138981178402901,
0.028898440301418304,
-0.01947329007089138,
-0.006000427529215813,
-0.03405940905213356,
-0.05186304450035095,
0.03635318577289581,
0.012941209599375725,
0.013402750715613365,
0.018242739140987396,
-0.004662806633859873,
0.014441180974245071,
-0.03816470503807068,
0.0176275372505188,
0.0006466160411946476,
-0.0025108002591878176,
-0.10904857516288757,
-0.10951551795005798,
0.08577752858400345,
0.04121943935751915,
-0.027002647519111633,
0.08045640587806702,
0.022526545450091362,
0.06895710527896881,
0.05708568915724754,
0.012664197012782097,
-0.010648460127413273,
-0.04480712488293648,
-0.031682323664426804,
0.01253943145275116,
-0.04028402268886566,
0.03280235454440117,
0.10064549744129181,
0.009886256419122219,
0.00972585566341877,
-0.029399005696177483,
0.01986011117696762,
-0.002862753113731742,
-0.24962790310382843,
-0.010524030774831772,
-0.045985158532857895,
-0.019817838445305824,
-0.013199220411479473,
-0.019908936694264412,
0.12908414006233215,
0.05952957645058632,
-0.03889916464686394,
0.09676902741193771,
-0.06578127294778824,
-0.03235596418380737,
0.11437610536813736,
-0.03301813825964928,
0.010008740238845348,
-0.013633904047310352,
0.01116703450679779
] | 0
| 0
| 0
| 0
|
|
Revealing and Mitigating Over-Attention in Knowledge Editing
|
https://openreview.net/forum?id=4l3AH8Bhmt
|
[
"Pinzheng Wang",
"Zecheng Tang",
"Keyan Zhou",
"Juntao Li",
"Qiaoming Zhu",
"Min Zhang"
] |
Poster
|
Large Language Models~(LLMs) have demonstrated superior performance across a wide range of tasks, but they still exhibit undesirable errors due to incorrect knowledge learned from the training data. To avoid this, knowledge editing methods emerged to precisely edit the specific model knowledge via efficiently modifying a very small percentage of parameters. However, those methods can lead to the problem of **Specificity Failure**, where the existing knowledge and capabilities are severely degraded due to editing.
Our preliminary indicates that Specificity Failure primarily stems from the model's attention heads assigning excessive attention scores to entities related to the edited knowledge, thereby unduly focusing on specific snippets within the context, which we denote as the **Attention Drift** phenomenon.
To mitigate such Attention Drift issue, we introduce a simple yet effective method **S**elective **A**ttention **D**rift **R**estriction(**SADR**), which introduces an additional regularization term during the knowledge editing process to restrict changes in the attention weight distribution, thereby preventing undue focus on the edited entity.
Experiments on five frequently-used strong LLMs demonstrate the effectiveness of our method, where SADR can significantly mitigate Specificity Failure in the predominant knowledge editing tasks.
|
model editing, mechanistic interpretability, NLP, language models
|
We analyze the reasons behind specificity failure in knowledge editing and mitigate it with our method.
| 11,041
|
2502.14838
|
[
0.006062187720090151,
-0.020665572956204414,
0.02050190418958664,
0.08529723435640335,
0.07832641899585724,
0.04169348627328873,
-0.016441894695162773,
0.001721702516078949,
0.024421874433755875,
0.05180731043219566,
-0.01180962473154068,
0.0001980160450330004,
0.04475005343556404,
-0.022238528355956078,
-0.012394973076879978,
0.02437964640557766,
0.06860277056694031,
0.06376670300960541,
-0.14066214859485626,
-0.014858799986541271,
0.037383414804935455,
0.033100344240665436,
0.060632865875959396,
0.046023305505514145,
-0.007531246170401573,
-0.04959244281053543,
-0.04957649111747742,
-0.004749916959553957,
0.010033595375716686,
-0.030606161803007126,
-0.015594737604260445,
0.050458356738090515,
0.042175132781267166,
0.08209186792373657,
-0.054196495562791824,
0.07472161948680878,
-0.06730446219444275,
0.08412816375494003,
0.014835330657660961,
-0.07032069563865662,
-0.03335612267255783,
-0.04063773527741432,
-0.022998645901679993,
0.02120525948703289,
0.06411178410053253,
-0.036287643015384674,
-0.021693222224712372,
-0.013736838474869728,
-0.037848833948373795,
0.023828374221920967,
-0.08424843847751617,
-0.11839128285646439,
-0.037001196295022964,
-0.003725233022123575,
-0.017864610999822617,
-0.01120116375386715,
-0.007141303736716509,
0.019088704138994217,
0.021111665293574333,
-0.11167649924755096,
-0.049463607370853424,
-0.0348072424530983,
-0.02752593345940113,
-0.03998982161283493,
-0.015572838485240936,
-0.061034075915813446,
0.029147062450647354,
0.018967216834425926,
-0.04699788615107536,
0.04016384482383728,
0.037534840404987335,
0.0843668282032013,
-0.03643076866865158,
0.028036877512931824,
0.006202712655067444,
0.06001046299934387,
-0.0344926193356514,
-0.025617403909564018,
0.03162671625614166,
-0.06734464317560196,
0.028612734749913216,
0.011693052016198635,
0.04574136808514595,
-0.06177158281207085,
0.12417932599782944,
0.02692529931664467,
0.08068656921386719,
-0.008392765186727047,
0.04612145572900772,
0.018846608698368073,
-0.021803393959999084,
-0.11977523565292358,
0.08886291831731796,
-0.015834901481866837,
0.05246727913618088,
0.021472414955496788,
0.018435291945934296,
-0.0578477568924427,
-0.021934321150183678,
0.05613589659333229,
0.047373443841934204,
0.09448298066854477,
-0.06072913482785225,
-0.013735798187553883,
0.01419282890856266,
-0.03313387930393219,
0.03598842769861221,
0.06265848875045776,
-0.008516350761055946,
-0.12340021133422852,
0.02379881963133812,
0.052645597606897354,
0.011466220952570438,
-0.062105439603328705,
-0.005451367702335119,
-0.025348080322146416,
0.04926179349422455,
0.01110535953193903,
-0.05059255659580231,
0.06844034790992737,
0.012333950027823448,
0.0333692841231823,
0.010110558941960335,
0.023087861016392708,
0.010765104554593563,
-0.012701592408120632,
-0.06644558161497116,
1.7480447720001992e-33,
0.15871919691562653,
0.0854315236210823,
-0.04592190682888031,
0.0017211949452757835,
0.004114540759474039,
0.016353430226445198,
-0.04794390872120857,
0.021108049899339676,
-0.03372237831354141,
-0.02225601114332676,
0.018967414274811745,
-0.008420565165579319,
-0.02541937120258808,
0.05361992493271828,
0.050700657069683075,
0.014313440769910812,
0.037587765604257584,
0.07707233726978302,
-0.022611364722251892,
-0.015949921682476997,
0.008310846984386444,
-0.021779030561447144,
-0.02437601052224636,
-0.08450490236282349,
-0.05585680156946182,
0.05211591720581055,
0.053889911621809006,
-0.01776563748717308,
-0.03994785249233246,
0.022702276706695557,
-0.07942607998847961,
-0.04669598117470741,
0.04165666550397873,
0.07595809549093246,
-0.0304216630756855,
-0.09505204111337662,
-0.01109548844397068,
-0.10372041165828705,
0.08947203308343887,
-0.0306999571621418,
-0.02232481725513935,
0.043961454182863235,
-0.00899022538214922,
-0.0020721564069390297,
-0.060551390051841736,
0.005944801028817892,
0.010625486262142658,
0.05629124492406845,
-0.017605571076273918,
-0.013172898441553116,
0.08780922740697861,
0.08133380860090256,
-0.009887082502245903,
-0.018070783466100693,
-0.06325894594192505,
0.05759954825043678,
0.06082784757018089,
-0.014676184393465519,
0.024988193064928055,
0.05215374380350113,
0.016461703926324844,
0.07099567353725433,
-0.004264691844582558,
0.09136028587818146,
0.049572039395570755,
0.05985357612371445,
-0.02343798242509365,
0.05796242877840996,
0.04099288582801819,
-0.041962526738643646,
-0.07879003137350082,
-0.02836126647889614,
-0.010187545791268349,
-0.019853802397847176,
0.005127172451466322,
-0.05720360949635506,
0.05765872076153755,
-0.03902154415845871,
0.08517960458993912,
-0.036624494940042496,
0.016648681834340096,
-0.008454221300780773,
-0.032054461538791656,
-0.06192746385931969,
-0.038900021463632584,
-0.05400477722287178,
0.003501421771943569,
-0.060311462730169296,
-0.005345676559954882,
0.0009615389863029122,
0.04765830934047699,
-0.037710800766944885,
-0.08822385966777802,
-0.03417884185910225,
-0.020668914541602135,
-9.643354333027836e-34,
-0.021880246698856354,
-0.02524472586810589,
-0.07313001900911331,
0.06087120994925499,
-0.04040594771504402,
-0.07550598680973053,
-0.039231497794389725,
0.06075755134224892,
-0.051434475928545,
-0.1411219835281372,
-0.0004120351804886013,
-0.13011053204536438,
-0.009165340103209019,
-0.03269483149051666,
-0.01527741551399231,
-0.03180915489792824,
0.0156791303306818,
0.054267726838588715,
-0.04994741827249527,
0.08643786609172821,
-0.02177453227341175,
-0.0068636685609817505,
-0.0257297083735466,
0.0750475525856018,
0.019975390285253525,
0.038077253848314285,
-0.07221443951129913,
0.06865102052688599,
-0.0033269417472183704,
-0.0475597083568573,
-0.07410437613725662,
0.012053563259541988,
-0.057922907173633575,
-0.02135157398879528,
-0.05362970754504204,
0.03564136475324631,
0.018247174099087715,
-0.03248415142297745,
-0.040738869458436966,
0.08559923619031906,
0.07692483067512512,
0.0032134426292032003,
-0.015027866698801517,
0.016717765480279922,
-0.019663741812109947,
-0.00037508716923184693,
-0.12325939536094666,
-0.0347479023039341,
0.023497141897678375,
0.00793683435767889,
-0.07701379805803299,
-0.0710889920592308,
-0.04297599941492081,
0.052590057253837585,
-0.013933928683400154,
-0.028939276933670044,
0.001820195815525949,
-0.05927637219429016,
0.013070791028439999,
-0.037039708346128464,
-0.04363545775413513,
0.010075628757476807,
0.024973811581730843,
-0.013505232520401478,
0.06672737002372742,
-0.02022588811814785,
0.0014060825342312455,
-0.009408798068761826,
0.06379926204681396,
-0.025463880971074104,
0.08834158629179001,
-0.04156215116381645,
0.007309278938919306,
-0.03167305141687393,
0.02636677399277687,
-0.030624864622950554,
0.01783534698188305,
-0.10240274667739868,
-0.019612114876508713,
-0.0691194161772728,
-0.02670019119977951,
0.011242584325373173,
0.03999684005975723,
-0.015351984649896622,
-0.004125847481191158,
0.1270034909248352,
0.05907249078154564,
0.09207512438297272,
-0.008966735564172268,
-0.00714441342279315,
-0.05640757083892822,
-0.0032255984842777252,
-0.018145577982068062,
0.07352885603904724,
-0.06454890966415405,
-4.650140539297354e-8,
-0.08699315786361694,
0.04168180748820305,
-0.06611613184213638,
0.09122708439826965,
0.09271086752414703,
-0.10723821073770523,
-0.05545584112405777,
0.08559394627809525,
0.010340695269405842,
-0.050159718841314316,
0.04692038148641586,
0.01218459103256464,
-0.023260870948433876,
-0.06251321732997894,
-0.016112487763166428,
0.05236789211630821,
-0.008931228891015053,
0.11376957595348358,
-0.050767652690410614,
-0.04406256973743439,
0.011989922262728214,
0.06134217232465744,
0.04910788685083389,
-0.04763079062104225,
0.046835388988256454,
-0.04495849087834358,
-0.015138233080506325,
0.10129237174987793,
0.04785073548555374,
0.001932500279508531,
0.017128493636846542,
-0.059250302612781525,
0.008042540401220322,
0.06940273940563202,
0.02633996680378914,
0.07007896900177002,
0.06880351901054382,
-0.004588577896356583,
0.02664290741086006,
0.063306525349617,
0.09839100390672684,
-0.0021774659398943186,
-0.03107035905122757,
0.11185216158628464,
0.012126008048653603,
-0.04038209468126297,
-0.006223320495337248,
-0.06508757174015045,
0.026349039748311043,
-0.1087534949183464,
0.06962351500988007,
0.003471239935606718,
-0.004212884232401848,
0.007762970868498087,
0.02660463936626911,
0.08384765684604645,
0.02555612288415432,
0.0369182825088501,
-0.03491384536027908,
-0.012848513200879097,
0.07707241922616959,
-0.02327609620988369,
0.0009560199687257409,
0.020790552720427513
] |
https://github.com/PinzhengWang322/Reveal_Attention_Drift
| 0
| 0
| 0
| 0
|
A Generalist Hanabi Agent
|
https://openreview.net/forum?id=pCj2sLNoJq
|
[
"Arjun V Sudhakar",
"Hadi Nekoei",
"Mathieu Reymond",
"Miao Liu",
"Janarthanan Rajendran",
"Sarath Chandar"
] |
Poster
|
Traditional multi-agent reinforcement learning (MARL) systems can develop cooperative strategies through repeated interactions. However, these systems are unable to perform well on any other setting than the one they have been trained on, and struggle to successfully cooperate with unfamiliar collaborators. This is particularly visible in the Hanabi benchmark, a popular 2-to-5 player cooperative card-game which requires complex reasoning and precise assistance to other agents. Current MARL agents for Hanabi can only learn one specific game-setting (e.g., 2-player games), and play with the same algorithmic agents. This is in stark contrast to humans, who can quickly adjust their strategies to work with unfamiliar partners or situations. In this paper, we introduce Recurrent Replay Relevance Distributed DQN (R3D2), a generalist agent for Hanabi, designed to overcome these limitations. We reformulate the task using text, as language has been shown to improve transfer. We then propose a distributed MARL algorithm that copes with the resulting dynamic observation- and action-space. In doing so, our agent is the first that can play all game settings concurrently, and extend strategies learned from one setting to other ones. As a consequence, our agent also demonstrates the ability to collaborate with different algorithmic agents ---agents that are themselves unable to do so.
|
Multi-Agent Reinforcement Learning (MARL), Cooperative game, Multi Agent Text-based game
|
First MARL agent that can play different Hanabi settings and cooperate with other algorithmic agents
| 11,039
|
2503.14555
|
[
-0.046387869864702225,
-0.0474575012922287,
-0.022209739312529564,
-0.01436865795403719,
-0.07020656764507294,
0.05252458527684212,
0.010517366230487823,
-0.008752462454140186,
0.012606441043317318,
-0.038977958261966705,
-0.0321413092315197,
-0.018948692828416824,
0.028714798390865326,
0.017773427069187164,
0.023020893335342407,
-0.008401262573897839,
0.057871267199516296,
0.05854025483131409,
-0.02189723588526249,
-0.01043173111975193,
-0.03615539148449898,
-0.021015943959355354,
0.018227895721793175,
-0.010692102834582329,
0.001151538803242147,
-0.01001469511538744,
-0.009054887108504772,
-0.04991656541824341,
-0.014442386105656624,
-0.007059757597744465,
0.08176404982805252,
0.05823545530438423,
0.012491326779127121,
0.029920538887381554,
-0.08788255602121353,
0.0807112604379654,
-0.12934350967407227,
0.051301516592502594,
-0.004072803072631359,
-0.06498059630393982,
-0.01392408274114132,
0.011499183252453804,
0.0031691391486674547,
-0.03235861286520958,
0.06624380499124527,
0.0017183776944875717,
-0.03771582245826721,
0.009604853577911854,
0.07994093745946884,
-0.0019041260238736868,
-0.07809475064277649,
-0.030175799503922462,
0.04138711094856262,
0.04218544811010361,
0.07174499332904816,
0.04017040133476257,
-0.015111847780644894,
0.03547217696905136,
0.007446459028869867,
-0.11897214502096176,
0.010729125700891018,
-0.10677947103977203,
0.05608624964952469,
-0.01543580275028944,
-0.06056074798107147,
-0.052331630140542984,
-0.019406070932745934,
0.057770878076553345,
0.000393665861338377,
-0.03394608572125435,
-0.01877538487315178,
0.0514385886490345,
0.015867723152041435,
-0.03264069929718971,
-0.0014111610362306237,
0.020904894918203354,
-0.013897077180445194,
-0.03545456752181053,
0.05455820634961128,
0.029531437903642654,
-0.017314070835709572,
-0.02547762542963028,
-0.01430310495197773,
0.034246157854795456,
0.06349706649780273,
-0.03993445634841919,
-0.009994748048484325,
0.0008142610895447433,
0.090255968272686,
-0.07331868261098862,
-0.02497050352394581,
0.03437327966094017,
0.09849561005830765,
0.00006511664832942188,
-0.0460873544216156,
0.051190607249736786,
0.010032077319920063,
-0.051166314631700516,
-0.04260639846324921,
0.1335059553384781,
0.02114405296742916,
0.1067766323685646,
-0.0125839589163661,
-0.05019448325037956,
-0.0071932184509932995,
0.01233655959367752,
0.028429681435227394,
0.030851151794195175,
0.014467407017946243,
-0.11158662289381027,
-0.01233272347599268,
-0.02466016821563244,
-0.030693912878632545,
-0.014592207968235016,
0.010248305276036263,
0.05734463036060333,
0.03841446340084076,
0.028810329735279083,
-0.02819703333079815,
0.07604355365037918,
0.0116274980828166,
-0.03934949263930321,
0.03726506978273392,
0.013331929221749306,
0.08148974925279617,
0.0422523058950901,
-0.008481008931994438,
5.477018022162228e-34,
0.05815364792943001,
0.04457489773631096,
0.04042165353894234,
0.041557345539331436,
0.040111202746629715,
-0.021712781861424446,
0.05024704337120056,
-0.04672626033425331,
-0.022156305611133575,
-0.06520166248083115,
0.031825952231884,
0.06344584375619888,
-0.03757128491997719,
0.08364246785640717,
-0.01860901340842247,
0.025724155828356743,
-0.12673243880271912,
-0.010366570204496384,
0.02426012046635151,
-0.026877274736762047,
0.07134077697992325,
-0.0024688411504030228,
0.02869960479438305,
-0.06555164605379105,
-0.008253764361143112,
-0.0003098580345977098,
0.04955798387527466,
-0.09075234830379486,
0.11050651222467422,
-0.010250716470181942,
0.019822727888822556,
-0.036436889320611954,
-0.03249867260456085,
0.09431936591863632,
0.00319021288305521,
0.03638433665037155,
-0.07145486772060394,
-0.019759071990847588,
-0.029413362964987755,
0.019384188577532768,
-0.024302920326590538,
0.013390268199145794,
-0.024423260241746902,
-0.042746588587760925,
-0.05431000143289566,
-0.0704696848988533,
-0.0005197766004130244,
0.007793413940817118,
-0.08146141469478607,
-0.05202513560652733,
-0.036272089928388596,
0.02757960744202137,
-0.028674263507127762,
-0.04234599322080612,
0.0414598323404789,
-0.020134249702095985,
0.012395511381328106,
0.02241447940468788,
-0.040107931941747665,
0.05295269191265106,
0.03249400854110718,
-0.028883816674351692,
-0.016070740297436714,
0.13303309679031372,
0.012981818988919258,
0.12127291411161423,
0.030261335894465446,
0.10692005604505539,
0.09087259322404861,
-0.06062379851937294,
0.040154218673706055,
0.017898982390761375,
0.11600898206233978,
-0.016590949147939682,
0.05553022027015686,
-0.008182794786989689,
0.10422752797603607,
-0.09381760656833649,
0.0465940497815609,
0.0005929372273385525,
-0.03463345021009445,
-0.04307038336992264,
-0.1264805793762207,
0.021841349080204964,
-0.06557965278625488,
-0.05516737699508667,
-0.003491353476420045,
-0.0691031888127327,
-0.06397800892591476,
0.05379002168774605,
-0.07240699231624603,
0.018379924818873405,
-0.024959774687886238,
0.0020780228078365326,
0.05634794384241104,
-6.079345000943415e-34,
0.002909568604081869,
-0.017774730920791626,
-0.06434968113899231,
0.0010309390490874648,
0.07119875401258469,
-0.04898881912231445,
-0.0746985599398613,
-0.042381852865219116,
0.03526477888226509,
-0.030752548947930336,
-0.0804336741566658,
0.0583384744822979,
0.06634048372507095,
0.025151455774903297,
-0.021178122609853745,
-0.009037421084940434,
0.09006672352552414,
0.01142257358878851,
-0.036246296018362045,
-0.025854675099253654,
0.027868302538990974,
0.08926573395729065,
-0.09774339199066162,
0.02499491348862648,
0.08268765360116959,
0.05434374511241913,
0.008024876937270164,
0.043560657650232315,
0.03518398478627205,
0.03803333640098572,
0.00014358811313286424,
-0.05714742839336395,
-0.009193251840770245,
-0.012252495624125004,
-0.0466439425945282,
0.053306352347135544,
-0.023087313398718834,
0.06506811082363129,
-0.10487528890371323,
0.11616800725460052,
0.04266585037112236,
-0.11209419369697571,
-0.11377275735139847,
0.0351780466735363,
-0.02546544186770916,
0.04713030904531479,
-0.0485529862344265,
-0.025741130113601685,
-0.0013488439144566655,
0.010987775400280952,
-0.0055017354898154736,
-0.06089306250214577,
-0.03835214301943779,
-0.08347490429878235,
-0.052799519151449203,
0.013746359385550022,
-0.00603320449590683,
0.05873618274927139,
0.00546574080362916,
-0.013444557785987854,
0.025431130081415176,
-0.050306931138038635,
-0.04213159531354904,
0.06415673345327377,
-0.0460023432970047,
0.01946219801902771,
0.020081842318177223,
0.01216248981654644,
0.02470565214753151,
-0.05124212056398392,
-0.035982489585876465,
0.05080946162343025,
0.022845713421702385,
-0.02007436938583851,
0.06035354733467102,
0.06214022636413574,
-0.1321324110031128,
-0.02354230172932148,
0.00986617524176836,
-0.0420810729265213,
-0.05733703449368477,
0.008015232160687447,
0.038706135004758835,
0.06422977149486542,
0.05727703124284744,
0.04361196607351303,
-0.0013325088657438755,
0.12405973672866821,
0.02235737070441246,
-0.05056603252887726,
-0.0005274730501696467,
0.07169115543365479,
0.0015765001298859715,
0.004510064143687487,
-0.09069635719060898,
-5.0002771700974336e-8,
-0.051934439688920975,
0.0003955461143050343,
0.016562659293413162,
0.008225427940487862,
0.05909654498100281,
-0.0060125235468149185,
-0.027389388531446457,
-0.02872607856988907,
-0.04737205058336258,
0.020063983276486397,
0.025319162756204605,
0.07039988785982132,
0.020992955192923546,
-0.041076891124248505,
0.01160090696066618,
0.030244644731283188,
0.053113386034965515,
-0.1214255839586258,
-0.0609390027821064,
-0.032643917948007584,
0.035050686448812485,
-0.0326574482023716,
0.009706162847578526,
-0.04375435784459114,
-0.05096976086497307,
-0.010046680457890034,
-0.04052489250898361,
0.020224440842866898,
0.04461013525724411,
-0.03604235500097275,
0.10507649928331375,
-0.01415195781737566,
0.06286631524562836,
-0.024069782346487045,
0.008252211846411228,
0.04114260897040367,
0.07255663722753525,
-0.057281430810689926,
0.007933867163956165,
-0.006025983486324549,
-0.04695535823702812,
0.0683874785900116,
0.006248754914849997,
0.008616598322987556,
0.0061650630086660385,
-0.02174667827785015,
-0.0668761357665062,
-0.1432051658630371,
0.031240416690707207,
-0.10359734296798706,
-0.01359387207776308,
-0.007619831245392561,
-0.029373060911893845,
0.047282829880714417,
0.12948735058307648,
0.03964736685156822,
-0.006810589227825403,
-0.023633688688278198,
0.1293216496706009,
0.050882842391729355,
-0.018660202622413635,
0.03729858994483948,
-0.06278178840875626,
0.011030834168195724
] |
https://github.com/chandar-lab/r3d2-a-generalist-hanabi-agent
| 1
| 0
| 0
| 0
|
GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
|
https://openreview.net/forum?id=5RUM1aIdok
|
[
"Tao Feng",
"Yihang Sun",
"Jiaxuan You"
] |
Poster
|
The powerful capabilities of Large Language Models (LLMs) have led to their growing use in evaluating human-generated content, particularly in evaluating research ideas within academic settings. Existing solutions primarily rely on prompt-based LLM methods or fine-tuned lightweight language models for idea evaluation. However, these methods are often unstable and struggle to comprehend the complex semantic information embedded in the ideas, impeding their ability to perform high-quality evaluations. To address the above challenges, we propose $\texttt{GraphEval}$, a lightweight graph-based LLM framework for idea evaluation. Our insight is that a complex idea can be broken down into comprehensible viewpoint nodes using prompts from small LLMs. These viewpoint nodes can then be linked together through edges created from LLM-based relation extraction and/or BERT similarity scores. The created viewpoint-graph can be used to conveniently propagate scores across view-nodes to improve the robustness of the idea evaluations. In particular, we propose two lightweight graph-based methods for idea evaluation: (1) GraphEval-LP: a training-free label propagation algorithm that propagates evaluation scores from known view-nodes to unknown nodes; (2) GraphEval-GNN: a Graph Neural Networks (GNN) that is trained to predict the evaluation scores given the observed graph with minimal computation resources. Moreover, to overcome LLM's limitation in objectively assessing the novelty of ideas, we further propose a novelty detection model to GraphEval-GNN to enhance its capability in judging idea novelty. Experiments on two datasets show $\texttt{GraphEval}$ improves F1 scores by at least 14% with low computation and API costs. Additionally, $\texttt{GraphEval}$ can effectively detect plagiarized ideas.
|
Idea Evaluation, View-graph, Lightweight model, Label propagation, Graph prediction
|
GraphEval is a lightweight, graph-based LLM framework for idea evaluation, offering two methods: GraphEval-LP (training-free) and GraphEval-GNN (minimally trained).
| 11,021
|
2503.12600
|
[
-0.06978658586740494,
0.012415367178618908,
-0.03618476167321205,
-0.0019512836588546634,
0.11094982177019119,
-0.008512202650308609,
-0.03946381434798241,
0.056809376925230026,
-0.014041858725249767,
0.018980197608470917,
-0.04755963757634163,
-0.03393964841961861,
0.08845886588096619,
0.08199916779994965,
0.014235482551157475,
0.09978882223367691,
0.13233670592308044,
-0.01978546753525734,
-0.020183313637971878,
-0.012776892632246017,
0.0025919717736542225,
0.04625019058585167,
0.031677961349487305,
0.020677952095866203,
0.035811763256788254,
0.008404038846492767,
-0.004103278275579214,
0.006335485260933638,
0.06332243978977203,
0.040322206914424896,
0.06213732808828354,
0.03253414109349251,
-0.03706816956400871,
0.06765829771757126,
-0.06029617041349411,
0.09305236488580704,
-0.017350543290376663,
0.07717961817979813,
-0.0024101831950247288,
-0.03917824849486351,
-0.04677228629589081,
0.018946589902043343,
0.04082988202571869,
0.061964865773916245,
0.0537843219935894,
0.004578661639243364,
-0.10441955924034119,
-0.03697242960333824,
-0.030453704297542572,
0.07007798552513123,
-0.1406257003545761,
-0.0240072812885046,
-0.022660275921225548,
-0.04997829347848892,
-0.02326064743101597,
0.002224453492090106,
0.02708725817501545,
-0.050805624574422836,
-0.014807285740971565,
-0.07408838719129562,
-0.03483353555202484,
-0.05386897921562195,
0.0014392004813998938,
-0.011576427146792412,
-0.06540661305189133,
-0.04260461404919624,
-0.011950491927564144,
0.05425933748483658,
-0.06590298563241959,
0.05987828969955444,
0.043281663209199905,
0.05312882363796234,
-0.050692543387413025,
0.050380244851112366,
-0.06863773614168167,
0.045328136533498764,
-0.010963581502437592,
-0.022169403731822968,
-0.011674811132252216,
-0.06844373047351837,
-0.002917597768828273,
0.01753089390695095,
0.022592326626181602,
0.038888365030288696,
0.04417816922068596,
0.05915462225675583,
0.07359198480844498,
-0.015589741058647633,
-0.015884321182966232,
0.04391246289014816,
-0.07142769545316696,
-0.05388374626636505,
0.035904429852962494,
0.00923238042742014,
0.10332001000642776,
0.027293261140584946,
0.02111445926129818,
-0.0988883376121521,
-0.08093620091676712,
0.07284492254257202,
-0.028749866411089897,
0.10340944677591324,
0.0007047146209515631,
-0.17054089903831482,
-0.005880305077880621,
0.020991509780287743,
0.021110860630869865,
0.03925391286611557,
0.028561295941472054,
-0.05576346069574356,
0.03189638629555702,
0.10276446491479874,
0.007698095869272947,
0.022320307791233063,
0.05029236152768135,
-0.02758854068815708,
0.06781333684921265,
-0.080011285841465,
0.044263508170843124,
0.03466014564037323,
-0.0525054857134819,
0.03563925251364708,
0.029767487198114395,
-0.020841827616095543,
-0.0379810705780983,
-0.04787953197956085,
0.013938252814114094,
7.1107044589006965e-34,
0.10978778451681137,
0.07140234857797623,
0.03760417550802231,
0.047682661563158035,
0.08447965979576111,
-0.006583781912922859,
-0.030496889725327492,
0.06437667459249496,
-0.08863827586174011,
-0.019880814477801323,
-0.04582759365439415,
0.023173993453383446,
0.03987431898713112,
0.0436587855219841,
0.014969456009566784,
-0.002186667174100876,
-0.013658544048666954,
0.021370744332671165,
-0.05346984043717384,
-0.08437753468751907,
-0.022922640666365623,
-0.05182713270187378,
0.030835764482617378,
-0.010364645160734653,
0.02102747932076454,
0.05087326094508171,
0.06368338316679001,
-0.040829624980688095,
-0.04616565629839897,
-0.009962550364434719,
-0.0614546500146389,
-0.004838708322495222,
0.033090509474277496,
-0.020361823961138725,
-0.013477073051035404,
-0.03922325745224953,
-0.04468624293804169,
-0.054636601358652115,
0.05877377465367317,
-0.052821431308984756,
-0.040466539561748505,
-0.0007007843814790249,
-0.012444901280105114,
-0.03950311243534088,
-0.03716370835900307,
-0.02757859230041504,
-0.046907298266887665,
0.03234288841485977,
0.014876944944262505,
-0.04599074646830559,
0.09190775454044342,
-0.029978930950164795,
-0.007796985562890768,
0.060442544519901276,
-0.057849690318107605,
-0.01656859554350376,
0.03070521168410778,
0.05547260865569115,
0.018645010888576508,
0.006944023538380861,
0.03977566957473755,
0.09741068631410599,
-0.010115931741893291,
0.008375016041100025,
0.028224892914295197,
0.04396247863769531,
-0.08466372638940811,
0.022776363417506218,
0.08129466325044632,
-0.02388855628669262,
-0.021933797746896744,
0.016358476132154465,
-0.014381402172148228,
-0.0667368546128273,
-0.022155050188302994,
0.025112608447670937,
-0.030696479603648186,
-0.11089041829109192,
0.06404939293861389,
0.06520608812570572,
-0.01936304196715355,
-0.057991672307252884,
0.043841540813446045,
-0.04984508454799652,
0.0015026323962956667,
-0.04143685847520828,
0.07740829139947891,
-0.05771627277135849,
-0.00004784978955285624,
-0.07600253820419312,
-0.023114170879125595,
-0.09075622260570526,
-0.07312678545713425,
-0.040938831865787506,
0.026540100574493408,
-6.306571732377209e-34,
-0.013803180307149887,
0.04404544457793236,
0.0032730717211961746,
0.016939781606197357,
0.04223114252090454,
-0.0881122350692749,
0.005879340693354607,
0.04851530119776726,
-0.05883440375328064,
0.006156970281153917,
-0.06473211199045181,
-0.16125720739364624,
-0.018370918929576874,
-0.006948818452656269,
0.09469690918922424,
-0.014186234213411808,
-0.06300019472837448,
-0.032535530626773834,
-0.051464781165122986,
0.07201087474822998,
0.01420650351792574,
0.10860398411750793,
-0.09696239978075027,
0.01583903841674328,
0.048542704433202744,
0.0495840460062027,
0.056224748492240906,
0.028797434642910957,
-0.025620797649025917,
0.006485690362751484,
-0.01626708172261715,
-0.006168844643980265,
-0.020029276609420776,
-0.05995733290910721,
-0.07211434841156006,
0.04236557334661484,
0.07003139704465866,
-0.08389154076576233,
-0.02107987180352211,
0.010846979916095734,
0.0257725790143013,
-0.024612339213490486,
-0.06266103684902191,
-0.01460352260619402,
-0.08530505746603012,
0.045518096536397934,
-0.08966488391160965,
0.024877317249774933,
-0.037463653832674026,
0.006956616882234812,
-0.046548228710889816,
0.009696596302092075,
-0.02509358339011669,
0.02701714262366295,
-0.027013540267944336,
-0.17964325845241547,
0.009055214934051037,
0.0400393046438694,
0.004656012635678053,
-0.03665608912706375,
-0.022745486348867416,
0.009040376171469688,
0.037308622151613235,
0.0318538136780262,
0.017197292298078537,
-0.044075410813093185,
-0.028120100498199463,
-0.0030588707886636257,
-0.047031670808792114,
0.041722726076841354,
0.040844567120075226,
0.0016792946262285113,
-0.021248193457722664,
-0.032021645456552505,
0.06580564379692078,
0.055202677845954895,
0.04452894255518913,
0.011195015162229538,
-0.007827341556549072,
-0.1295202374458313,
0.07643095403909683,
-0.035033274441957474,
0.0408795028924942,
0.02279829978942871,
0.05788789317011833,
-0.020320424810051918,
0.03922315314412117,
0.13043220341205597,
0.006773096043616533,
0.06367043405771255,
-0.055741555988788605,
-0.03176841884851456,
-0.054234765470027924,
0.07817205786705017,
0.00231083482503891,
-3.966333039784331e-8,
-0.1323874592781067,
0.015383954159915447,
-0.0518893226981163,
0.09985627233982086,
-0.042366232722997665,
0.003788818372413516,
0.006201756652444601,
0.08478985726833344,
-0.006908286828547716,
0.020419713109731674,
0.013097106479108334,
-0.003249224042519927,
-0.07510627806186676,
-0.035517215728759766,
-0.03243843838572502,
0.02382231131196022,
-0.006528080906718969,
0.012794360518455505,
0.034476038068532944,
0.051001060754060745,
0.013447990640997887,
0.022574283182621002,
0.0008572862716391683,
0.04568209871649742,
0.044843580573797226,
-0.051580093801021576,
-0.032434552907943726,
0.0693877711892128,
0.028795776888728142,
-0.026375029236078262,
-0.0050474307499825954,
0.04937336966395378,
-0.022107768803834915,
0.0022015972062945366,
0.01800881326198578,
0.07619941234588623,
0.0010778163559734821,
0.013955756090581417,
0.014852399937808514,
0.06551217287778854,
0.016198448836803436,
0.06227268651127815,
-0.03905436769127846,
0.01338987797498703,
0.008728294633328915,
0.012132189236581326,
0.0014768815599381924,
-0.03151582553982735,
0.016109468415379524,
-0.03748601675033569,
-0.10708831250667572,
-0.08774464577436447,
-0.041339289397001266,
-0.07215097546577454,
-0.04232696071267128,
0.06549009680747986,
-0.07974613457918167,
0.012502217665314674,
0.008426654152572155,
0.04996991902589798,
0.045623768121004105,
0.031128769740462303,
0.03926042094826698,
0.03825043514370918
] | 0
| 0
| 0
| 0
|
|
Modeling Future Conversation Turns to Teach LLMs to Ask Clarifying Questions
|
https://openreview.net/forum?id=cwuSAR7EKd
|
[
"Michael JQ Zhang",
"W. Bradley Knox",
"Eunsol Choi"
] |
Poster
|
Large language models (LLMs) must often respond to highly ambiguous user requests. In such cases, the LLM's best response may be to ask a clarifying question to elicit more information. Existing LLMs often respond by presupposing a single interpretation of such ambiguous requests, frustrating users who intended a different interpretation. We speculate this is caused by current preference data labeling practice, where LLM responses are evaluated only on their prior contexts. To address this, we assign preference labels by simulating their expected outcomes in future turns. This allows LLMs to learn to ask clarifying questions when it can generate responses that are tailored to each user interpretation in future turns. On open-domain QA datasets with multiple annotations, we evaluate systems based on their ability to ask clarifying questions to recover each user's interpretation and expected answer. We compare systems trained using our proposed preference labeling methods against standard methods, which assign preferences based on only prior context. Our method achieves a 5% improvement in F1 measured against the answer set from different interpretations of each query, showing the value of modeling future conversation turns. We further demonstrate that our method can be used to train models to judiciously determine when to ask clarifying questions, directly answering the question when clarification is unnecessary. In our experiments, we find that our method achives a 3% improvement in accuracy of such judgments over existing methods.
|
Clarifying Questions, QA, Ambiguity, RLHF
|
We RLHF train LLMs to ask clarifying questions in response to ambiguous requests by assigning preferences based on their expected outcomes in future turns.
| 11,015
|
2410.13788
|
[
-0.04181741178035736,
-0.01675656996667385,
0.014885622076690197,
0.037778761237859726,
0.028025079518556595,
-0.005527140107005835,
0.04521966353058815,
0.06122599169611931,
0.034639716148376465,
-0.05167220160365105,
-0.07317834347486496,
-0.10932005941867828,
0.0027586352080106735,
0.04546558856964111,
0.06336382776498795,
0.04857373610138893,
0.10316341370344162,
-0.0012749420711770654,
-0.07899203151464462,
0.010152525268495083,
0.01946803368628025,
0.055946338921785355,
0.0605829618871212,
0.03347434848546982,
-0.04355223476886749,
-0.027674397453665733,
0.012432885356247425,
0.04668465629220009,
0.023837879300117493,
0.025377554818987846,
0.032574526965618134,
0.05693431198596954,
-0.08198229223489761,
0.06274501234292984,
-0.03783918544650078,
0.002958283992484212,
-0.07139688730239868,
0.012769650667905807,
-0.030322961509227753,
-0.05488020181655884,
-0.10177189856767654,
-0.012833639048039913,
-0.007930934429168701,
0.02073052152991295,
0.0656915083527565,
-0.041662320494651794,
-0.01848340965807438,
0.025559596717357635,
-0.02440325729548931,
0.018187109380960464,
-0.05312364920973778,
-0.06219680979847908,
-0.006253665778785944,
0.013067325577139854,
-0.027094120159745216,
0.051067057996988297,
-0.010784135200083256,
0.07707635313272476,
-0.0015623918734490871,
-0.0365741141140461,
-0.14485342800617218,
-0.18719729781150818,
-0.039240989834070206,
0.0557074137032032,
-0.04423828795552254,
-0.04844173043966293,
-0.0243837833404541,
0.05594705790281296,
-0.07754264771938324,
0.021377582103013992,
-0.029968513175845146,
0.10308825969696045,
0.029255619272589684,
0.029366903007030487,
-0.005926521960645914,
0.08508181571960449,
-0.02892734482884407,
0.05940237641334534,
0.0029347070958465338,
-0.010626569390296936,
-0.04541834443807602,
-0.01616424322128296,
-0.012812772765755653,
0.01885990798473358,
0.0942406952381134,
0.014506060630083084,
0.019137507304549217,
0.08234531432390213,
0.002772421110421419,
0.06012122333049774,
-0.045259613543748856,
-0.03970738500356674,
0.04475323483347893,
0.03531888499855995,
0.05047958344221115,
0.08638472110033035,
0.010193209163844585,
-0.10725527256727219,
-0.0024293053429573774,
0.06038461998105049,
0.064129538834095,
0.1460241824388504,
0.011125612072646618,
-0.10919907689094543,
-0.03960498049855232,
0.027268068864941597,
-0.03768029436469078,
0.0788043960928917,
0.016561347991228104,
-0.10898856073617935,
0.0006147634121589363,
0.03829112648963928,
0.010545829311013222,
-0.02808424085378647,
-0.008233224973082542,
-0.028588026762008667,
0.0900249257683754,
0.004620369058102369,
0.004557425621896982,
0.024323606863617897,
-0.05273069813847542,
0.0039313193410634995,
0.04148900881409645,
-0.009639807976782322,
0.04618046060204506,
-0.0020479639060795307,
-0.06562298536300659,
2.8586738751761743e-33,
0.08712756633758545,
0.014450926333665848,
0.03822577744722366,
0.04214509204030037,
0.030061611905694008,
0.0284847654402256,
-0.027034249156713486,
0.07099965214729309,
-0.04296557977795601,
-0.04453732818365097,
0.0718250572681427,
0.017095837742090225,
-0.03212369978427887,
0.007963678799569607,
0.03643258288502693,
0.022053349763154984,
-0.07880456745624542,
0.03367221727967262,
-0.04484758526086807,
-0.011937977746129036,
0.03036973439157009,
0.010109066031873226,
0.040595460683107376,
-0.022802336141467094,
0.07553859055042267,
0.0038850305136293173,
0.055297765880823135,
-0.001694946433417499,
-0.03573845699429512,
-0.007098812609910965,
-0.08056088536977768,
-0.0728912279009819,
0.045842695981264114,
-0.007724906783550978,
0.03136111795902252,
0.016279570758342743,
0.0134851960465312,
0.013124145567417145,
-0.023638762533664703,
-0.06733642518520355,
0.04801790043711662,
-0.024476291611790657,
0.0423923060297966,
0.030757654458284378,
-0.08003509044647217,
-0.09142179787158966,
-0.025782786309719086,
0.02780771255493164,
-0.008673814125359058,
-0.0003946750075556338,
0.03450697660446167,
0.0017828185809776187,
-0.032699503004550934,
-0.012805153615772724,
-0.003142760368064046,
-0.016053151339292526,
-0.041042082011699677,
0.008958568796515465,
-0.05352112650871277,
0.012456563301384449,
0.04033846780657768,
0.04167936369776726,
0.027970358729362488,
0.03918224945664406,
-0.0076142302714288235,
-0.007275131996721029,
-0.014227538369596004,
0.08855415880680084,
0.04896824061870575,
-0.03017665445804596,
-0.04314010217785835,
-0.003925548400729895,
-0.014255186542868614,
-0.02705245651304722,
0.01774013601243496,
0.062356773763895035,
0.00283535267226398,
-0.027245067059993744,
0.10594137012958527,
-0.03730678930878639,
0.09353627264499664,
-0.02288886345922947,
0.023362141102552414,
-0.01732460968196392,
-0.04017002880573273,
-0.06567461788654327,
0.062480852007865906,
-0.04658396542072296,
-0.02180466055870056,
-0.015345152467489243,
-0.026631837710738182,
-0.025629520416259766,
-0.03842199221253395,
0.05572890490293503,
0.022988228127360344,
-2.325682512323279e-33,
-0.03944171965122223,
-0.03298964351415634,
-0.044710177928209305,
0.09546107053756714,
0.020573893561959267,
-0.03811141103506088,
0.0011987286852672696,
0.0633503869175911,
0.016645554453134537,
-0.09416386485099792,
-0.0870022252202034,
-0.009720603004097939,
0.014276367612183094,
0.03713354468345642,
-0.006247640121728182,
-0.021981941536068916,
0.000427058432251215,
-0.04797007143497467,
0.024171117693185806,
0.10989143699407578,
-0.06877370923757553,
0.11381785571575165,
-0.12850095331668854,
-0.0020591544453054667,
-0.01090661808848381,
0.06529010832309723,
-0.004715091548860073,
0.026221126317977905,
-0.033964790403842926,
-0.017775883898139,
-0.027454152703285217,
-0.0855260044336319,
-0.010656299069523811,
-0.06040877103805542,
0.015627427026629448,
-0.004780081566423178,
0.06772066652774811,
-0.0627928376197815,
-0.028030024841427803,
0.11015622317790985,
0.0178983174264431,
-0.05775855854153633,
-0.09556013345718384,
0.027887333184480667,
-0.07871776074171066,
-0.02257835865020752,
-0.04520488157868385,
0.003514495212584734,
0.02188495732843876,
0.01606062613427639,
-0.004215281456708908,
-0.00948202796280384,
-0.006471928209066391,
0.042145490646362305,
-0.015467005781829357,
-0.1299600750207901,
-0.0004142690740991384,
-0.021905720233917236,
0.037123046815395355,
-0.001367081655189395,
-0.04680870845913887,
0.0013271498028188944,
0.04758231341838837,
-0.0720636323094368,
0.032892126590013504,
-0.07082699239253998,
-0.003175582503899932,
-0.018532145768404007,
-0.04333111643791199,
-0.05247333273291588,
0.0826178565621376,
-0.07163529843091965,
-0.03531944751739502,
-0.03344409912824631,
0.10249155014753342,
-0.009796665981411934,
0.02573791891336441,
-0.11219387501478195,
-0.007567374501377344,
-0.015295827761292458,
0.06223103404045105,
0.014199481345713139,
0.013126818463206291,
0.10961786657571793,
0.01114246342331171,
-0.013176549226045609,
0.05092195048928261,
0.1151856929063797,
-0.0020292068365961313,
0.04755251854658127,
0.032988548278808594,
0.002276315586641431,
-0.03212149441242218,
0.0870605930685997,
-0.07033602148294449,
-4.248270002449317e-8,
-0.07947243750095367,
0.009374074637889862,
-0.03154069557785988,
0.11585218459367752,
0.014026006683707237,
-0.055794429033994675,
-0.019301166757941246,
0.08155518025159836,
-0.04054802283644676,
-0.04664004594087601,
0.02956586331129074,
0.07949300110340118,
-0.01776754856109619,
-0.04917418956756592,
-0.0050963847897946835,
0.0716700479388237,
0.03171556070446968,
0.012287747114896774,
-0.012474083341658115,
-0.04388992860913277,
0.03520210459828377,
0.028376873582601547,
0.020810356363654137,
0.002838662825524807,
0.07389458268880844,
-0.03990029916167259,
-0.0447896271944046,
0.14764268696308136,
0.06587819755077362,
0.028506051748991013,
-0.012640217319130898,
-0.08235172182321548,
-0.058560896664857864,
-0.014039983972907066,
-0.015246565453708172,
0.048599984496831894,
-0.017651693895459175,
-0.07408131659030914,
0.08591563999652863,
-0.019730055704712868,
0.01393505185842514,
-0.0022007639054208994,
-0.03661557659506798,
0.05029525235295296,
-0.03768838569521904,
-0.03189224749803543,
-0.03501996770501137,
-0.08407632261514664,
-0.03487442061305046,
-0.08128812164068222,
-0.06570154428482056,
-0.0984339639544487,
-0.009086846373975277,
0.059364620596170425,
0.04875548556447029,
0.0067608291283249855,
-0.014622196555137634,
-0.014283095486462116,
-0.024125540629029274,
0.044001393020153046,
0.04843004420399666,
0.06877646595239639,
-0.053240228444337845,
0.06271521747112274
] | 0
| 0
| 0
| 0
|
|
Towards Robust and Parameter-Efficient Knowledge Unlearning for LLMs
|
https://openreview.net/forum?id=1ExfUpmIW4
|
[
"Sungmin Cha",
"Sungjun Cho",
"Dasol Hwang",
"Moontae Lee"
] |
Poster
|
Large Language Models (LLMs) have demonstrated strong reasoning and memorization capabilities via pretraining on massive textual corpora. However, this poses risk of privacy and copyright violations, highlighting the need for efficient machine unlearning methods that remove sensitive data without retraining from scratch. While Gradient Ascent (GA) is commonly used to unlearn by reducing the likelihood of generating unwanted content, it leads to unstable optimization and catastrophic forgetting of retrained knowledge. We find that combining GA with low-rank adaptation results in poor trade-offs between computational cost and generative performance. To address these challenges, we propose two novel techniques for robust and efficient unlearning for LLMs. First, we introduce Inverted Hinge Loss, which suppresses unwanted tokens while maintaining fluency by boosting the probability of the next most likely token. Second, we develop a data-adaptive initialization for LoRA adapters via low-rank approximation weighted with relative Fisher information, thereby focusing updates on parameters critical for removing targeted knowledge. Experiments on the Training Data Extraction Challenge dataset using GPT-Neo models as well as on the TOFU benchmark with Phi-1.5B and Llama2-7B models demonstrate that our approach effectively removes sensitive information while maintaining reasoning and generative capabilities with minimal impact. Our implementation can be found in https://github.com/csm9493/efficient-llm-unlearning.
|
Machine Unlearning, Large Language Models, Low-rank Adaptation
|
We propose a novel loss function and LoRA initialization method for robust and parameter-efficient LLM unlearning.
| 11,013
|
2408.06621
|
[
-0.09207317233085632,
-0.09050051122903824,
0.012505549006164074,
0.07574788480997086,
0.057439859956502914,
0.03512769564986229,
-0.012944268062710762,
-0.03397829458117485,
0.011220626533031464,
0.011102454736828804,
-0.02796573005616665,
0.01942070946097374,
0.055201757699251175,
-0.023229626938700676,
-0.02581634186208248,
0.04461831972002983,
0.06897488981485367,
0.14556705951690674,
-0.06815066933631897,
-0.07948426902294159,
0.011827880516648293,
0.0618685819208622,
0.04579034447669983,
0.07385189086198807,
0.08204732090234756,
-0.060611922293901443,
0.011850784532725811,
0.008276468142867088,
0.013019493781030178,
-0.0434543639421463,
0.02687537856400013,
0.022221574559807777,
0.0017846684204414487,
0.022556180134415627,
-0.04722807928919792,
0.04998554661870003,
-0.05151677504181862,
0.039176855236291885,
0.03821825608611107,
-0.002761007519438863,
0.014810006134212017,
-0.029187962412834167,
-0.03778551518917084,
-0.023449977859854698,
0.05606435611844063,
0.02580084465444088,
-0.01295472215861082,
-0.004680158570408821,
-0.03744550049304962,
0.0382477268576622,
-0.022787485271692276,
-0.005051245912909508,
-0.09634602814912796,
0.0035776912700384855,
-0.045010942965745926,
-0.05815745145082474,
0.024097705259919167,
0.008890509605407715,
0.0022837778087705374,
-0.006234067026525736,
-0.07681386917829514,
-0.16175153851509094,
-0.0029635627288371325,
-0.01128667313605547,
-0.056432321667671204,
-0.05306122452020645,
0.0769769549369812,
0.05830272659659386,
-0.03338274359703064,
0.0386846587061882,
-0.005958623718470335,
0.07999499887228012,
-0.09628485888242722,
0.058438729494810104,
-0.024351373314857483,
0.07513750344514847,
0.07206167280673981,
-0.03782619908452034,
0.05026398226618767,
-0.023543981835246086,
0.044175151735544205,
0.029711922630667686,
0.07872079312801361,
0.015602225437760353,
0.05632558465003967,
0.0201832614839077,
0.05039086192846298,
0.07263893634080887,
0.06129516661167145,
-0.010666915215551853,
-0.00928215030580759,
-0.03499740734696388,
0.040316514670848846,
0.03274861350655556,
0.018809951841831207,
0.005645961966365576,
0.016553809866309166,
-0.04919465258717537,
-0.023076869547367096,
0.10036634653806686,
0.0187233779579401,
0.0737592875957489,
0.004721906967461109,
-0.012248900718986988,
-0.03497176244854927,
0.0007525657420046628,
0.0269789882004261,
0.0010545143159106374,
0.05779442563652992,
-0.049399878829717636,
0.05742957815527916,
0.10834512859582901,
0.007727003190666437,
0.036737460643053055,
-0.030648179352283478,
-0.022988473996520042,
0.01386854238808155,
-0.03186025843024254,
-0.013577924109995365,
0.09350213408470154,
-0.05057414621114731,
-0.01645781844854355,
0.04332612454891205,
0.026203488931059837,
-0.07144547253847122,
-0.013172569684684277,
-0.07620826363563538,
1.5927110913917035e-33,
0.05296303331851959,
0.06291017681360245,
0.0021308213472366333,
-0.005384007468819618,
0.03324142470955849,
0.013018602505326271,
0.039798542857170105,
0.07708501815795898,
-0.09277176856994629,
-0.02279346063733101,
-0.034361328929662704,
0.019648829475045204,
-0.0905088409781456,
0.0391494557261467,
0.02011420764029026,
0.015495848841965199,
-0.018026288598775864,
-0.012694490142166615,
-0.06251055747270584,
-0.10207249969244003,
0.07757574319839478,
-0.0238987784832716,
0.02099272981286049,
-0.026679176837205887,
0.011080391705036163,
-0.011314325965940952,
0.05935010686516762,
-0.06062731519341469,
-0.03618941083550453,
0.019794009625911713,
-0.10389602184295654,
-0.11429086327552795,
0.06185910850763321,
0.04083537682890892,
0.028691072016954422,
-0.06344014406204224,
-0.04094665125012398,
-0.05782843753695488,
0.03730834275484085,
-0.011119787581264973,
-0.03948143869638443,
0.08633195608854294,
0.04592681676149368,
-0.026231566444039345,
-0.10643035173416138,
-0.04957374185323715,
0.051348403096199036,
0.004656041506677866,
0.012333444319665432,
-0.05526195839047432,
0.0030995032284408808,
0.031136881560087204,
-0.08229433000087738,
-0.0226917564868927,
-0.06477914750576019,
-0.04430653899908066,
-0.015350082889199257,
-0.010653793811798096,
0.0040138293989002705,
0.025004200637340546,
-0.03054426610469818,
0.10388287156820297,
0.03153351694345474,
0.11672735959291458,
0.027333833277225494,
0.0021118693985044956,
-0.031643711030483246,
0.04137926176190376,
0.05379337817430496,
-0.019426601007580757,
0.021796638146042824,
-0.0158942062407732,
-0.026351338252425194,
-0.059703219681978226,
0.007024073041975498,
-0.003984512761235237,
0.042908377945423126,
-0.057847343385219574,
0.04524897038936615,
0.02665257640182972,
0.06925643980503082,
-0.009160115383565426,
0.027588920667767525,
-0.059800125658512115,
-0.07393447309732437,
-0.10662718117237091,
0.07598451524972916,
-0.07308444380760193,
0.0017860373482108116,
0.016561154276132584,
-0.0238718893378973,
-0.08312494307756424,
-0.0331684947013855,
-0.04751623421907425,
0.005510373972356319,
-1.2031046009066914e-33,
-0.04936688393354416,
-0.003923073876649141,
-0.012832595966756344,
0.07621026784181595,
-0.013059498742222786,
-0.04173016920685768,
-0.024542247876524925,
0.08040706068277359,
-0.07317259162664413,
-0.07017847150564194,
-0.04542624205350876,
-0.040517158806324005,
0.002191318431869149,
-0.03986680507659912,
0.08382286131381989,
-0.02671585977077484,
-0.052474167197942734,
0.05528485029935837,
-0.0978296771645546,
0.08144877851009369,
-0.009625070728361607,
0.0642370730638504,
-0.11167357861995697,
0.050598230212926865,
0.04449574276804924,
-0.006108545698225498,
-0.03264562413096428,
0.08108212798833847,
-0.004774880595505238,
0.028322234749794006,
-0.030769219622015953,
0.037232134491205215,
-0.03166428208351135,
-0.02670619823038578,
-0.11093330383300781,
0.013152559287846088,
0.06998532265424728,
0.02792716957628727,
-0.024653885513544083,
0.10041504353284836,
0.058078963309526443,
-0.002916766796261072,
-0.12026777118444443,
-0.01989128440618515,
-0.01729586161673069,
-0.025816375389695168,
-0.10904330760240555,
-0.015472942031919956,
0.10359784960746765,
0.04933283105492592,
-0.011605612933635712,
0.000981030985713005,
-0.02263343520462513,
0.012652663514018059,
-0.008512694388628006,
-0.12401867657899857,
-0.01045805774629116,
-0.020358433946967125,
-0.02442227490246296,
-0.023372121155261993,
0.005470005329698324,
-0.0003925851488020271,
0.0011480756802484393,
-0.06388654559850693,
0.0005946964956820011,
-0.05169650912284851,
-0.01484494749456644,
0.051297690719366074,
-0.03307430073618889,
-0.009230402298271656,
-0.00026830582646653056,
-0.021746961399912834,
0.025028357282280922,
-0.06902385503053665,
-0.0061377570964396,
0.023372191935777664,
-0.053472504019737244,
-0.03022666834294796,
-0.0375172421336174,
-0.06403975188732147,
0.08588789403438568,
-0.0018702588276937604,
0.04086055979132652,
0.09307141602039337,
0.09269079566001892,
0.016853418201208115,
0.12470528483390808,
0.07252196222543716,
0.025166526436805725,
0.04455586522817612,
-0.02172773703932762,
-0.021758809685707092,
0.030863158404827118,
0.06686824560165405,
-0.060872551053762436,
-4.303443290609721e-8,
-0.06325619667768478,
-0.033106256276369095,
-0.05154982581734657,
0.08658777922391891,
0.01667470484972,
-0.028406159952282906,
-0.046250924468040466,
0.11961797624826431,
-0.032877493649721146,
-0.018881773576140404,
0.08201789110898972,
-0.00466562807559967,
-0.03218533471226692,
-0.06701453030109406,
0.0008002248359844089,
0.09598099440336227,
0.03246702253818512,
0.03118217922747135,
-0.004116402473300695,
0.01019224151968956,
-0.00038479434442706406,
0.05702178552746773,
0.024526264518499374,
0.02026008628308773,
0.014580181799829006,
-0.09136173874139786,
0.006674776319414377,
0.08791892975568771,
0.11410117149353027,
0.015174385160207748,
-0.08198077231645584,
-0.0698036402463913,
-0.003070301143452525,
0.0010083225788548589,
0.03388548642396927,
0.09833183139562607,
0.009156815707683563,
0.0005511904018931091,
-0.010371854528784752,
-0.045601993799209595,
0.062300119549036026,
0.07123517990112305,
-0.03222203254699707,
0.040121860802173615,
-0.017636405304074287,
-0.03403560817241669,
-0.05989263951778412,
-0.06385503709316254,
0.034247227013111115,
-0.02507953532040119,
0.03248991817235947,
-0.02214336395263672,
-0.012244900688529015,
0.060325488448143005,
0.06385552883148193,
0.02630346082150936,
-0.0235342588275671,
0.03850236162543297,
0.02594427764415741,
0.0004329123185016215,
0.04374120011925697,
-0.024441778659820557,
-0.010610919445753098,
-0.004215942695736885
] |
https://github.com/csm9493/efficient-llm-unlearning
| 3
| 0
| 0
| 0
|
What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
|
https://openreview.net/forum?id=Zk9guOl9NS
|
[
"Kunhao Zheng",
"Juliette Decugis",
"Jonas Gehring",
"Taco Cohen",
"benjamin negrevergne",
"Gabriel Synnaeve"
] |
Poster
|
Prompting techniques such as chain-of-thought have established themselves as a popular vehicle for improving the outputs of large language models (LLMs). For code generation, however, their exact mechanics and efficacy are under-explored using unified metrics and benchmarks. We thus investigate the effects of a wide range of prompting strategies with a focus on automatic re-prompting over multiple turns and computational requirements. After systematically decomposing reasoning, instruction, and execution feedback prompts, we conduct an extensive grid search on the competitive programming benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama 3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that consistently improve performance across all models with small and large sampling budgets. We then show how finetuning with such an optimal configuration allows models to internalize the induced reasoning process and obtain improvements in performance and scalability for multi-turn code generation.
|
Large language Models, Multi-turn Code Generation, Chain-of-Thought
|
We break down the main components of recent CoT and self-repair techniques for multi-turn code generation identifying the most promising ones across benchmarks and models.
| 11,007
|
2410.08105
|
[
-0.001369453384540975,
-0.042095210403203964,
0.00048716968740336597,
-0.0061189113184809685,
0.044807568192481995,
-0.0948522686958313,
-0.030556201934814453,
0.05948019027709961,
-0.013231884688138962,
-0.00019040366169065237,
-0.10227452218532562,
-0.07002118974924088,
0.03113393671810627,
-0.02863524667918682,
0.07942382991313934,
-0.04413054883480072,
0.05403212830424309,
-0.027291346341371536,
-0.02617599070072174,
-0.0545416995882988,
-0.011724003590643406,
-0.07230575382709503,
0.06036793068051338,
0.00030267355032265186,
0.001444967812858522,
-0.025593625381588936,
0.013996448367834091,
-0.04164009913802147,
0.013154267333447933,
0.04549259692430496,
-0.017827462404966354,
0.08635177463293076,
-0.0009695247863419354,
0.07107273489236832,
-0.003551857080310583,
0.04467146843671799,
-0.02853304333984852,
-0.013760828413069248,
-0.0004118111392017454,
-0.0053536491468548775,
-0.11628638952970505,
0.039993517100811005,
0.07018543779850006,
-0.0019003490451723337,
0.07783173769712448,
-0.059867024421691895,
-0.012667208909988403,
0.036263357847929,
-0.027242084965109825,
-0.031660549342632294,
-0.08170558512210846,
-0.05776219069957733,
-0.024281153455376625,
-0.08260944485664368,
0.001364065334200859,
-0.03913719207048416,
0.13314063847064972,
0.0236278735101223,
0.03385002911090851,
-0.0426952987909317,
-0.10465451329946518,
-0.11969392001628876,
-0.0963192880153656,
0.02010156586766243,
-0.019033251330256462,
-0.0015276978956535459,
0.03618495166301727,
0.03253835439682007,
-0.0403817743062973,
0.03190341591835022,
-0.04332466050982475,
0.07121389359235764,
-0.06610015034675598,
0.0748511254787445,
0.00823394674807787,
0.0739428848028183,
-0.00485169468447566,
0.021811053156852722,
-0.007012544199824333,
-0.07920584082603455,
-0.008456068113446236,
-0.004667612258344889,
-0.0017974409274756908,
0.04641468822956085,
0.07110317051410675,
-0.015254879370331764,
0.075038842856884,
0.06732700020074844,
0.10612425953149796,
0.0430028922855854,
0.013479863293468952,
-0.02939506806433201,
0.027455385774374008,
0.048222675919532776,
0.016197524964809418,
0.11066741496324539,
-0.0011346357641741633,
-0.0913756713271141,
-0.04615107923746109,
0.05592423677444458,
0.020954130217432976,
0.08944451063871384,
0.07970195263624191,
-0.10182492434978485,
0.01964571699500084,
0.007977969944477081,
0.05550755560398102,
0.05552341789007187,
-0.006568749435245991,
-0.08715607970952988,
0.0323331356048584,
0.08060114830732346,
0.06695344299077988,
-0.013377231545746326,
0.014678867533802986,
-0.013239012099802494,
0.025608615949749947,
-0.03615470230579376,
0.0006213887827470899,
0.04031416401267052,
-0.00985883828252554,
0.00018632994033396244,
-0.0086569394916296,
-0.02385118044912815,
0.06210467591881752,
-0.017544258385896683,
-0.08015996962785721,
2.049871818768282e-34,
0.0375216007232666,
0.034042179584503174,
-0.001036095549352467,
0.09524369239807129,
0.07427676767110825,
-0.016392186284065247,
0.09335907548666,
0.05487194284796715,
-0.06131306663155556,
-0.007164530921727419,
0.02263066917657852,
-0.0031522491481155157,
-0.04454248771071434,
0.037195052951574326,
0.08479546010494232,
0.002211117884144187,
-0.029153242707252502,
0.026654783636331558,
-0.08382094651460648,
-0.004795565735548735,
0.05155665799975395,
-0.019493242725729942,
0.009922314435243607,
-0.029341958463191986,
0.019679075106978416,
0.07199282199144363,
0.0069224294275045395,
-0.06277044117450714,
-0.04479847103357315,
0.015513756312429905,
-0.07707736641168594,
-0.07841051369905472,
-0.013079851865768433,
0.08532518893480301,
0.058990728110075,
-0.0160598773509264,
-0.021899940446019173,
-0.05781782045960426,
0.04900185018777847,
0.02279869094491005,
-0.006216680631041527,
0.022621456533670425,
0.023709235712885857,
0.005923340562731028,
-0.04442497342824936,
-0.049970757216215134,
-0.06575214117765427,
0.0037198136560618877,
-0.02697339653968811,
0.01032019592821598,
0.05817149952054024,
0.11906153708696365,
-0.007754158694297075,
0.055740516632795334,
-0.0440661795437336,
-0.00041113479528576136,
-0.0032607088796794415,
-0.04640527814626694,
0.051299627870321274,
0.11890516430139542,
-0.03130952641367912,
0.04109020531177521,
-0.01949637196958065,
0.06576095521450043,
0.06051727756857872,
0.009647348895668983,
-0.052706025540828705,
0.03425195440649986,
0.06469282507896423,
0.03111131489276886,
-0.019973361864686012,
-0.07146725058555603,
-0.06775102764368057,
-0.023090997710824013,
0.021888628602027893,
-0.0345056988298893,
0.029280025511980057,
-0.039755284786224365,
0.06384337693452835,
-0.08643072098493576,
0.024805467575788498,
0.050926003605127335,
-0.026437388733029366,
-0.03900555893778801,
0.023874424397945404,
-0.08141544461250305,
0.02524353377521038,
-0.07388612627983093,
-0.0747045949101448,
0.015104063786566257,
-0.044503312557935715,
-0.039778273552656174,
0.005518802907317877,
-0.05827287212014198,
-0.017925076186656952,
-8.80175364242685e-34,
0.022024424746632576,
-0.05224465951323509,
-0.02137892320752144,
0.0853244960308075,
-0.05857295170426369,
-0.1083577424287796,
0.007958796806633472,
-0.033853087574243546,
-0.012792321853339672,
-0.04269927740097046,
-0.03014388680458069,
-0.05416518822312355,
0.07098869979381561,
0.053161971271038055,
0.04079664126038551,
-0.08056991547346115,
-0.00037706116563640535,
0.0031836838461458683,
0.07352571934461594,
0.05752792954444885,
-0.01217048242688179,
0.08113794773817062,
-0.07101136445999146,
-0.02447086200118065,
-0.03629986569285393,
0.031670648604631424,
-0.0918165072798729,
0.02285866066813469,
0.020129388198256493,
-0.003233281197026372,
-0.042357705533504486,
0.050133589655160904,
-0.008793946355581284,
0.018601134419441223,
-0.028045855462551117,
0.037862807512283325,
0.06208053231239319,
0.03851843625307083,
-0.02590969018638134,
0.13833987712860107,
0.071508027613163,
-0.08152613788843155,
-0.006335448008030653,
0.038737427443265915,
-0.05800652876496315,
0.08085109293460846,
-0.10319840162992477,
0.0285835899412632,
-0.01043800450861454,
-0.02953815460205078,
-0.014550554566085339,
-0.006275573279708624,
0.028657471761107445,
0.026884909719228745,
-0.013235263526439667,
-0.09611888229846954,
-0.04136784002184868,
-0.03977867215871811,
0.009640507400035858,
0.0177568681538105,
-0.05128937214612961,
-0.004771366249769926,
0.10959366708993912,
-0.10035907477140427,
0.04585365206003189,
-0.031577374786138535,
-0.04561533406376839,
-0.0547136589884758,
-0.031530074775218964,
0.009908161126077175,
-0.02031308226287365,
-0.015885569155216217,
-0.0932781845331192,
0.003736797021701932,
-0.033981647342443466,
0.06204020977020264,
-0.027986790984869003,
-0.0811752900481224,
-0.005166004411876202,
-0.04548509418964386,
0.015962570905685425,
0.012663522735238075,
-0.004630010109394789,
0.0655754953622818,
-0.07473165541887283,
-0.023116540163755417,
0.048378244042396545,
0.10201989859342575,
-0.024509930983185768,
0.04125944525003433,
-0.010608435608446598,
0.03283575549721718,
0.0627962276339531,
0.05717024207115173,
-0.03214077278971672,
-4.429240973991e-8,
-0.0307218786329031,
0.04748896509408951,
-0.0031962417997419834,
0.0976516529917717,
0.06153367832303047,
-0.010650989599525928,
-0.04881235957145691,
0.0400274321436882,
-0.023501859977841377,
0.03485419973731041,
0.07920599728822708,
-0.014660445973277092,
-0.03802678361535072,
0.01108111348003149,
0.015329012647271156,
0.047852419316768646,
-0.09482672065496445,
-0.035379596054553986,
-0.06910417973995209,
-0.029446333646774292,
-0.0031050178222358227,
0.03738778457045555,
-0.012666570954024792,
0.03533702716231346,
0.008148636668920517,
-0.045376576483249664,
0.060226526111364365,
0.09641537070274353,
0.061242468655109406,
-0.002306088339537382,
0.0051962509751319885,
-0.039419326931238174,
-0.013909511268138885,
-0.0062525211833417416,
0.026752330362796783,
-0.013162219896912575,
-0.04298416152596474,
-0.0049573518335819244,
0.09935083240270615,
-0.08722826093435287,
0.050269417464733124,
0.0430283397436142,
-0.08996809273958206,
0.050716858357191086,
0.012936597689986229,
-0.04631751403212547,
-0.07262378931045532,
-0.06010613590478897,
-0.037555646151304245,
-0.05135398730635643,
-0.041154779493808746,
-0.026203300803899765,
-0.04739551618695259,
0.029092993587255478,
0.018470074981451035,
0.11977103352546692,
-0.09654813259840012,
0.012696894817054272,
0.028286363929510117,
0.05015585571527481,
0.05172597989439964,
0.05994325876235962,
-0.01345866359770298,
-0.007968236692249775
] | 0
| 0
| 0
| 0
|
|
RMB: Comprehensively benchmarking reward models in LLM alignment
|
https://openreview.net/forum?id=kmgrlG9TR0
|
[
"Enyu Zhou",
"Guodong Zheng",
"Binghai Wang",
"Zhiheng Xi",
"Shihan Dou",
"Rong Bao",
"Wei Shen",
"Limao Xiong",
"Jessica Fan",
"Yurong Mou",
"Rui Zheng",
"Tao Gui",
"Qi Zhang",
"Xuanjing Huang"
] |
Poster
|
Reward models (RMs) guide the alignment of large language models (LLMs), steering them toward behaviors preferred by humans. Evaluating RMs is the key to better aligning LLMs. However, the current evaluation of RMs may not directly correspond to their alignment performance due to the limited distribution of evaluation data and evaluation methods that are not closely related to alignment objectives. To address these limitations, we propose RMB, a comprehensive RM benchmark that covers over 49 real-world scenarios and includes both pairwise and Best-of-N (BoN) evaluations to better reflect the effectiveness of RMs in guiding alignment optimization.
We demonstrate a positive correlation between our benchmark and the downstream alignment task performance. Based on our benchmark, we conduct extensive analysis on the state-of-the-art RMs, revealing their generalization defects that were not discovered by previous benchmarks, and highlighting the potential of generative RMs. Furthermore, we delve into open questions in reward models, specifically examining the effectiveness of majority voting for the evaluation of reward models and analyzing the impact factors of generative RMs, including the influence of evaluation criteria and instructing methods. We will release our evaluation code and datasets upon publication.
|
LLM Alignment, reward model, evaluation
| null | 11,004
|
2410.09893
|
[
-0.09312594681978226,
-0.08462174236774445,
-0.0279090479016304,
-0.010207751765847206,
0.013247796334326267,
0.07206980139017105,
-0.07327572256326675,
0.023342732340097427,
0.0744352713227272,
0.009680984541773796,
-0.05424091964960098,
-0.11477069556713104,
0.04169880598783493,
0.047484979033470154,
0.019235994666814804,
-0.008595606312155724,
0.03295385092496872,
0.02321680635213852,
-0.05065588280558586,
-0.043512973934412,
0.016893871128559113,
0.0857369676232338,
0.047434624284505844,
0.010835313238203526,
0.037366654723882675,
-0.009473403915762901,
-0.016140567138791084,
-0.01802053116261959,
-0.015815164893865585,
-0.03625993803143501,
0.014532594010233879,
0.03469553217291832,
0.050576016306877136,
0.02584138698875904,
-0.034068766981363297,
0.06433732062578201,
-0.12155824154615402,
-0.0549410879611969,
0.0005677920416928828,
-0.03583073988556862,
-0.02779393084347248,
-0.025198837742209435,
0.058829888701438904,
-0.02139611355960369,
0.10067375004291534,
-0.04842233285307884,
0.0022027126979082823,
-0.02447090856730938,
-0.03699230030179024,
0.01426537986844778,
-0.08904159069061279,
-0.009177052415907383,
-0.012922636233270168,
0.09688077121973038,
-0.03975466638803482,
0.07656294107437134,
0.05037225782871246,
0.038968514651060104,
0.02665293961763382,
-0.05695508420467377,
-0.017327547073364258,
-0.1265561878681183,
-0.01714286021888256,
0.03789038211107254,
0.013532881624996662,
-0.08826640248298645,
-0.011239091865718365,
0.022406037896871567,
-0.1058369055390358,
0.03197132423520088,
-0.027717353776097298,
0.03417947515845299,
0.025445429608225822,
0.05374493449926376,
0.014644088223576546,
0.07522736489772797,
0.013052084483206272,
-0.013126185163855553,
0.06142614409327507,
-0.037914302200078964,
-0.022798337042331696,
0.036240171641111374,
0.06422866135835648,
-0.00009062533354153857,
0.128840833902359,
-0.007434782572090626,
0.045049868524074554,
0.007399819325655699,
0.037077926099300385,
0.04311946779489517,
-0.07786177098751068,
-0.0626588761806488,
0.028086891397833824,
0.0051266709342598915,
-0.01071280986070633,
0.07506536692380905,
-0.006839017383754253,
-0.050273314118385315,
-0.057140473276376724,
0.05051450803875923,
0.05254868417978287,
0.14215818047523499,
-0.05982290580868721,
-0.11854427307844162,
0.011267391964793205,
-0.025736210867762566,
-0.012049037031829357,
0.01297495886683464,
0.02725290134549141,
-0.14107927680015564,
0.0018201369093731046,
0.02287072129547596,
0.015826299786567688,
0.06384096294641495,
-0.0027206826489418745,
-0.033336091786623,
0.01463713962584734,
-0.026260893791913986,
0.010004461742937565,
0.023141296580433846,
-0.016872627660632133,
0.007377028930932283,
0.04159930348396301,
-0.027364207431674004,
-0.013258636929094791,
0.04909197986125946,
0.018923429772257805,
7.451262621078183e-34,
0.06879055500030518,
0.06597217917442322,
0.06099817529320717,
0.03598730266094208,
0.06949996948242188,
-0.04030781239271164,
-0.0022429502569139004,
0.04303208366036415,
-0.08959978073835373,
0.005656793247908354,
-0.057481665164232254,
0.03320494294166565,
-0.01628805510699749,
0.07420621067285538,
0.000411162938689813,
-0.0108623206615448,
-0.06435709446668625,
0.07192379236221313,
-0.10156171023845673,
0.017898017540574074,
0.06824937462806702,
0.056826237589120865,
0.042311035096645355,
-0.06666962057352066,
0.03206264227628708,
-0.009440327994525433,
0.008222680538892746,
-0.07732942700386047,
-0.09474670141935349,
-0.006513138767331839,
-0.08419393002986908,
-0.054320160299539566,
0.03904452919960022,
-0.011757588014006615,
0.025247881188988686,
-0.032375212758779526,
-0.019458945840597153,
-0.051467373967170715,
-0.012768205255270004,
-0.05729295685887337,
-0.033048126846551895,
0.03251161426305771,
-0.023313181474804878,
0.023772433400154114,
-0.0809599757194519,
-0.07041577994823456,
0.0009901509620249271,
-0.003861040575429797,
0.057502344250679016,
-0.03117392025887966,
0.03960288316011429,
0.03057762235403061,
0.004421777557581663,
-0.008563240058720112,
-0.03254921734333038,
0.05059955641627312,
0.0013252947246655822,
0.029797976836562157,
-0.025749696418642998,
0.02759198099374771,
0.004865459632128477,
-0.016907835379242897,
-0.005659862887114286,
-0.024673132225871086,
0.0003004283644258976,
0.009418115019798279,
0.004953187890350819,
0.07042369246482849,
0.02790549397468567,
0.025925273075699806,
0.030208919197320938,
-0.009297757409512997,
-0.03699635714292526,
-0.043348561972379684,
0.07019411772489548,
0.044320572167634964,
0.07203755527734756,
-0.11183301359415054,
0.051523394882678986,
-0.038614995777606964,
-0.00724415248259902,
0.0301648136228323,
0.0059991441667079926,
-0.10401074588298798,
-0.028164029121398926,
-0.0769488662481308,
0.04252159222960472,
-0.08532237261533737,
-0.013217197731137276,
0.06157855689525604,
-0.01111446600407362,
0.03484908863902092,
-0.02059139497578144,
0.022719938308000565,
0.04160580039024353,
-1.636167648173665e-33,
-0.04030630737543106,
0.00007498211925849319,
0.06154249981045723,
0.045293305069208145,
-0.01714090257883072,
-0.03377409651875496,
0.009748592041432858,
0.014484641142189503,
-0.04621211811900139,
-0.03161704167723656,
0.0017106333980336785,
-0.022542616352438927,
0.009286719374358654,
-0.022764606401324272,
0.04051406681537628,
0.022672303020954132,
0.011018747463822365,
0.09386160224676132,
0.020106393843889236,
0.03294569253921509,
0.08242849260568619,
0.1570475548505783,
-0.12742556631565094,
0.023042093962430954,
-0.0034418676514178514,
0.029918944463133812,
0.0306863896548748,
0.004498723894357681,
0.0004807898076251149,
-0.003674689680337906,
-0.015077115036547184,
0.018182339146733284,
-0.1613273173570633,
-0.011198376305401325,
-0.05360174551606178,
0.047550637274980545,
0.03996127098798752,
-0.09109719097614288,
-0.024264078587293625,
0.11991974711418152,
0.023860037326812744,
0.0007593539776280522,
-0.08819039911031723,
0.0037690065801143646,
-0.0028741001151502132,
-0.004123652819544077,
-0.0976674035191536,
-0.03551797568798065,
0.013791265897452831,
0.006649221759289503,
-0.06675251573324203,
0.02392364665865898,
-0.05196324735879898,
0.06417793780565262,
-0.053078677505254745,
-0.0633254274725914,
-0.1185738742351532,
-0.05202208831906319,
0.04706193134188652,
-0.03675850108265877,
-0.02433454431593418,
0.07300951331853867,
0.02767670899629593,
-0.06367290765047073,
0.00540387537330389,
-0.01327219232916832,
0.057455193251371384,
0.02322397008538246,
-0.04958357661962509,
0.0010034709703177214,
0.03447757661342621,
0.004139311611652374,
0.010546057485044003,
0.02553807757794857,
-0.013005882501602173,
0.0034989486448466778,
0.010486580431461334,
-0.043301597237586975,
-0.03655710816383362,
-0.13320469856262207,
0.049418799579143524,
0.07004749029874802,
0.021124687045812607,
-0.013502088375389576,
-0.011113099753856659,
0.08397922664880753,
0.1386362612247467,
0.10089555382728577,
-0.0069807954132556915,
0.035036806017160416,
-0.04506811127066612,
-0.02375645563006401,
0.03226923942565918,
0.04413401335477829,
-0.0418306402862072,
-4.1143081830341544e-8,
-0.08629500865936279,
0.03907739371061325,
-0.044894419610500336,
0.0769651010632515,
-0.08287752419710159,
-0.037160541862249374,
-0.05301637947559357,
0.08344972133636475,
0.028436413034796715,
-0.05556999891996384,
0.031223904341459274,
0.06654785573482513,
-0.06017690151929855,
-0.03492700681090355,
-0.04084324836730957,
0.051728203892707825,
-0.07095133513212204,
0.06393499672412872,
-0.012667570263147354,
-0.02494615688920021,
0.024318533018231392,
0.09500141441822052,
0.0123536866158247,
-0.007812848314642906,
-0.025444287806749344,
-0.06965659558773041,
-0.03369716554880142,
0.10249576717615128,
0.055991269648075104,
-0.058993566781282425,
0.011798621155321598,
0.0051262048073112965,
0.019063062965869904,
-0.03163880854845047,
0.030804965645074844,
0.05263160169124603,
0.017001107335090637,
-0.015562774613499641,
0.031827978789806366,
0.03747328743338585,
0.027676211670041084,
0.053427983075380325,
-0.036355096846818924,
0.032723598182201385,
-0.032187506556510925,
-0.003878700779750943,
-0.09416711330413818,
0.002486727200448513,
-0.012012176215648651,
-0.09629009664058685,
0.05501943826675415,
-0.08082335442304611,
-0.05888955667614937,
0.01455909013748169,
0.06046668812632561,
0.044330332428216934,
-0.017943499609827995,
-0.05002645403146744,
0.05871577188372612,
0.01993352733552456,
0.04562421888113022,
-0.009654374793171883,
-0.04528439790010452,
0.011971748434007168
] |
https://github.com/zhou-zoey/rmb-reward-model-benchmark
| 34
| 0
| 0
| 0
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.