title
stringlengths 14
154
| paper_url
stringlengths 42
42
| authors
sequencelengths 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
sequencelengths 384
384
|
|---|---|---|---|---|---|---|---|---|---|
Regretful Decisions under Label Noise | https://openreview.net/forum?id=7B9FCDoUzB | [
"Sujay Nagaraj",
"Yang Liu",
"Flavio Calmon",
"Berk Ustun"
] | Poster | Machine learning models are routinely used to support decisions that affect individuals – be it to screen a patient for a serious illness or to gauge their response to treatment. In these tasks, we are limited to learning models from datasets with noisy labels. In this paper, we study the instance-level impact of learning under label noise. We introduce a notion of regret for this regime which measures the number of unforeseen mistakes due to noisy labels. We show that standard approaches to learning under label noise can return models that perform well at a population level while subjecting individuals to a lottery of mistakes. We present a versatile approach to estimate the likelihood of mistakes at the individual level from a noisy dataset by training models over plausible realizations of datasets without label noise. This is supported by a comprehensive empirical study of label noise in clinical prediction tasks. Our results reveal how failure to anticipate mistakes can compromise model reliability and adoption, and demonstrate how we can address these challenges by anticipating and avoiding regretful decisions. | Uncertainty Quantification, Fairness, Model Multiplicity, Clinical Decision Support, Classification, Label Noise | null | 548 | null | [
-0.07392285764217377,
-0.013155557215213776,
0.013222488574683666,
0.05090199038386345,
0.045256827026605606,
-0.020541856065392494,
0.06856594234704971,
0.018650969490408897,
0.03816569969058037,
-0.08364991098642349,
-0.03630483150482178,
-0.043010011315345764,
0.07190313935279846,
-0.016090089455246925,
-0.07419419288635254,
-0.04823767766356468,
0.06844427436590195,
0.04855649545788765,
-0.0328245535492897,
0.024872176349163055,
-0.04837742820382118,
0.06173720955848694,
0.07205352187156677,
0.08322037756443024,
0.010963326320052147,
-0.04039164260029793,
-0.04234028235077858,
0.019541816785931587,
-0.05211110785603523,
0.013967209495604038,
0.028768988326191902,
-0.041969817131757736,
-0.00923213455826044,
-0.009017935022711754,
-0.0288024190813303,
-0.026704154908657074,
-0.08060446381568909,
0.03873346745967865,
0.0032599996775388718,
-0.0013089993735775352,
0.008183661848306656,
-0.004374041687697172,
0.009804152883589268,
0.07199308276176453,
0.03642166033387184,
-0.06433459371328354,
0.04484965652227402,
-0.052657101303339005,
0.004058328922837973,
0.060140710324048996,
-0.06660811603069305,
-0.09129153937101364,
-0.03463221341371536,
-0.0059442599304020405,
-0.022414112463593483,
0.03233559429645538,
0.07627853751182556,
0.03387732431292534,
-0.06937997788190842,
0.07058549672365189,
-0.015570544637739658,
-0.13224244117736816,
-0.03794349730014801,
-0.04931109398603439,
-0.023346351459622383,
0.052605677396059036,
-0.004417589399963617,
-0.0016173857729882002,
-0.025210801512002945,
0.08043526858091354,
-0.02275894582271576,
0.08546355366706848,
-0.016475461423397064,
0.09669015556573868,
-0.04689345136284828,
0.06507855653762817,
0.04388394579291344,
0.019289972260594368,
0.04286590591073036,
0.010502015240490437,
-0.043189797550439835,
0.0030747130513191223,
-0.023480094969272614,
0.003330904757604003,
0.16224339604377747,
-0.0570942685008049,
0.0748928040266037,
-0.002195211360231042,
0.023749396204948425,
-0.04358735308051109,
-0.0159157682210207,
-0.037139128893613815,
0.04598533362150192,
0.04632557928562164,
-0.03257626295089722,
0.0987878292798996,
-0.0365840382874012,
-0.08094221353530884,
-0.015037622302770615,
0.055908408015966415,
-0.06888345628976822,
0.02908831089735031,
0.016257639974355698,
-0.0498819574713707,
0.02731732279062271,
-0.016330905258655548,
-0.006953128147870302,
0.003479915903881192,
0.09570532292127609,
-0.05167241767048836,
-0.006500293966382742,
0.008658993057906628,
-0.016948649659752846,
0.013860472477972507,
0.017761779949069023,
-0.01252848468720913,
-0.0013695365050807595,
-0.029778381809592247,
-0.07259979844093323,
0.09907294809818268,
-0.059658482670784,
0.01967005804181099,
0.06317619234323502,
-0.023167965933680534,
-0.02637692727148533,
-0.0032237605191767216,
-0.07417123019695282,
-5.951993504178763e-34,
0.047861240804195404,
-0.03593509644269943,
0.057298678904771805,
0.002858958672732115,
0.06137719005346298,
-0.06944563239812851,
-0.12808910012245178,
-0.04563178867101669,
0.0759783461689949,
0.022862903773784637,
0.026464561000466347,
-0.03266400098800659,
0.007655712775886059,
0.0006464970065280795,
0.05953475832939148,
0.03413527458906174,
-0.05874335393309593,
0.10026036202907562,
-0.0349990539252758,
0.04326721653342247,
0.02539796195924282,
-0.06862456351518631,
0.018011637032032013,
-0.0834643691778183,
0.004735076799988747,
0.012352866120636463,
-0.051433589309453964,
0.06465405970811844,
0.08637021481990814,
0.0005269793327897787,
-0.05366955325007439,
-0.00412859907373786,
0.1416497379541397,
-0.007807889487594366,
-0.03305575251579285,
-0.03494260832667351,
-0.006804096978157759,
0.01793367601931095,
0.01713419333100319,
-0.017049921676516533,
0.0018600851763039827,
0.012680888175964355,
0.03431464359164238,
0.029033783823251724,
0.008085479028522968,
-0.018820878118276596,
0.0002663277555257082,
-0.08195757865905762,
-0.03362990543246269,
-0.05912249535322189,
0.00624640891328454,
0.002492462517693639,
-0.016582679003477097,
0.02671835571527481,
-0.0705472007393837,
0.051352884620428085,
0.0031224105041474104,
0.0100129758939147,
-0.06393176317214966,
-0.04231109097599983,
0.07177983969449997,
0.002537999302148819,
0.0021846673917025328,
-0.0006070495583117008,
-0.0567844994366169,
0.06784110516309738,
0.014386220835149288,
-0.025857072323560715,
-0.030423613265156746,
-0.10451944172382355,
-0.03308715671300888,
0.03715980052947998,
-0.0709771141409874,
-0.03749606013298035,
0.06177300214767456,
0.021978359669446945,
0.008952274918556213,
-0.061829887330532074,
0.033539578318595886,
-0.008196095004677773,
0.016400864347815514,
-0.09195637702941895,
-0.09047698229551315,
0.027763206511735916,
-0.0026278230361640453,
-0.020610103383660316,
0.02819252200424671,
-0.03850281983613968,
-0.05301011726260185,
0.028879327699542046,
-0.11543521285057068,
-0.03147512674331665,
-0.009618841111660004,
0.012031967751681805,
-0.00964661967009306,
-1.8909507222556186e-33,
-0.08777987211942673,
-0.0026012512389570475,
0.0329214408993721,
0.024192554876208305,
0.0363164022564888,
-0.0095050735399127,
0.01696174591779709,
-0.010080124251544476,
0.013651691377162933,
-0.061766333878040314,
-0.05470045283436775,
-0.019613122567534447,
0.04470217600464821,
0.01768001727759838,
-0.048923294991254807,
-0.03705940395593643,
-0.030731849372386932,
-0.025468213483691216,
0.012927633710205555,
0.01563844084739685,
0.07684770226478577,
0.05341895669698715,
-0.06200855225324631,
0.07297997176647186,
-0.057144854217767715,
0.05822408199310303,
-0.03024880401790142,
0.03784714639186859,
0.041387103497982025,
-0.08995412290096283,
-0.10454313457012177,
-0.0011160949943587184,
0.045423783361911774,
-0.0749470666050911,
0.03129710257053375,
0.059727322310209274,
0.10746511816978455,
-0.03695130720734596,
-0.06098399683833122,
0.06576170027256012,
0.05079421028494835,
0.013368484564125538,
-0.13982988893985748,
-0.030853116884827614,
0.005306576378643513,
-0.08680162578821182,
0.020611856132745743,
-0.035222455859184265,
0.017446573823690414,
0.002876750659197569,
-0.014307890087366104,
0.04299548268318176,
-0.08191409707069397,
0.06975580006837845,
-0.0477861687541008,
-0.10944841802120209,
-0.07068656384944916,
-0.05158497020602226,
0.06304524838924408,
0.09113962203264236,
-0.14786550402641296,
-0.005078073590993881,
-0.054343100637197495,
-0.01536340918391943,
-0.043141067028045654,
-0.007155891507863998,
-0.021024595946073532,
0.002320599975064397,
0.05028747022151947,
0.009436915628612041,
0.08152759075164795,
0.011410759761929512,
-0.05865858867764473,
0.0058858138509094715,
0.046903785318136215,
-0.0334121510386467,
-0.030712952837347984,
-0.03332962468266487,
-0.030297191813588142,
-0.031121965497732162,
0.10470891743898392,
-0.06878285109996796,
0.06943609565496445,
0.05792823061347008,
0.032401975244283676,
0.017573721706867218,
0.10026717185974121,
-0.004603368230164051,
-0.025067755952477455,
-0.018653051927685738,
-0.027772383764386177,
-0.022808589041233063,
-0.027000222355127335,
0.05025970935821533,
-0.09207356721162796,
-3.9983131472354216e-8,
-0.018823379650712013,
0.054515548050403595,
0.08230432122945786,
0.016427557915449142,
0.07476487755775452,
-0.005742989480495453,
-0.06643203645944595,
0.14136318862438202,
-0.03926375135779381,
-0.022282728925347328,
0.0670689195394516,
0.06408348679542542,
-0.06367840617895126,
-0.0398915596306324,
-0.03191263601183891,
0.08519398421049118,
-0.00924905575811863,
0.1081751361489296,
0.021254314109683037,
0.027596870437264442,
-0.00367987179197371,
0.009484254755079746,
0.05617402121424675,
-0.09897147119045258,
0.057960204780101776,
-0.09991377592086792,
0.005057790316641331,
0.10420754551887512,
-0.011897130869328976,
0.042746540158987045,
0.009224826470017433,
-0.0169871486723423,
0.035920411348342896,
0.050630681216716766,
-0.01511199027299881,
-0.019838042557239532,
0.01863454468548298,
-0.08772158622741699,
0.021493807435035706,
-0.00877021811902523,
0.018211811780929565,
0.06176508963108063,
-0.027891933917999268,
0.03441832587122917,
0.011687756516039371,
0.02342700958251953,
0.028654273599386215,
-0.0551014244556427,
0.010247900150716305,
-0.06377846747636795,
0.02320699952542782,
0.0039160349406301975,
0.009556114673614502,
0.05209296941757202,
0.06230393424630165,
0.018227532505989075,
0.0016600200906395912,
-0.043397098779678345,
-0.04769847169518471,
0.09027353674173355,
0.035785578191280365,
0.016702502965927124,
-0.013625502586364746,
-0.011232618242502213
] |
LaGeM: A Large Geometry Model for 3D Representation Learning and Diffusion | https://openreview.net/forum?id=72OSO38a2z | [
"Biao Zhang",
"Peter Wonka"
] | Poster | This paper introduces a novel hierarchical autoencoder that maps 3D models into a highly compressed latent space. The hierarchical autoencoder is specifically designed to tackle the challenges arising from large-scale datasets and generative modeling using diffusion. Different from previous approaches that only work on a regular image or volume grid, our hierarchical autoencoder operates on unordered sets of vectors. Each level of the autoencoder controls different geometric levels of detail. We show that the model can be used to represent a wide range of 3D models while faithfully representing high-resolution geometry details. The training of the new architecture takes 0.70x time and 0.58x memory compared to the baseline.
We also explore how the new representation can be used for generative modeling. Specifically, we propose a cascaded diffusion framework where each stage is conditioned on the previous stage. Our design extends existing cascaded designs for image and volume grids to vector sets. | diffusion, geometry, generative model, 3d | null | 546 | 2410.01295 | [
-0.03173578903079033,
-0.09264390915632248,
-0.002906598849222064,
0.0039376444183290005,
0.014351109974086285,
-0.013419413939118385,
-0.10293706506490707,
-0.06892397254705429,
0.005724309477955103,
-0.044164516031742096,
-0.018447233363986015,
-0.011023284867405891,
0.0038148509338498116,
0.11750934273004532,
-0.04505094140768051,
0.005147118121385574,
0.04277941957116127,
0.09586518257856369,
-0.10314038395881653,
-0.024875672534108162,
-0.05539727956056595,
-0.019019799306988716,
-0.03179114684462547,
0.018869200721383095,
0.09240689873695374,
0.05690260976552963,
0.11296862363815308,
-0.043069012463092804,
0.01398506574332714,
-0.03727266192436218,
0.07831135392189026,
-0.027304647490382195,
-0.04659860208630562,
0.05219375714659691,
-0.01469014585018158,
0.018901051953434944,
-0.06876865774393082,
-0.0023252919781953096,
-0.044306010007858276,
-0.04867814481258392,
0.027167247608304024,
0.060164786875247955,
0.010541316121816635,
0.047121644020080566,
0.0651683434844017,
0.012124997563660145,
-0.030956337228417397,
-0.0492425374686718,
0.03534022346138954,
-0.01908085308969021,
-0.030430760234594345,
-0.06950943917036057,
-0.043878935277462006,
0.02047996036708355,
-0.07131049036979675,
0.025638561695814133,
0.058463096618652344,
-0.03742111846804619,
0.007722924929112196,
0.015517176128923893,
-0.05021268501877785,
-0.010726089589297771,
-0.018012087792158127,
0.004938886966556311,
-0.047777507454156876,
-0.04333861917257309,
0.04173409193754196,
0.0533120296895504,
-0.028790410608053207,
0.030057813972234726,
0.03606583923101425,
-0.019798297435045242,
-0.06459842622280121,
-0.04283291473984718,
0.04403315484523773,
-0.025874130427837372,
0.0007771772798150778,
-0.08155997097492218,
0.020014986395835876,
-0.03232758864760399,
0.014953739941120148,
0.04341394826769829,
0.02247168868780136,
0.057864684611558914,
-0.006396336480975151,
0.013429868035018444,
-0.004893619567155838,
0.1022222489118576,
-0.06170845031738281,
-0.02167016640305519,
0.009851550683379173,
-0.009796639904379845,
-0.044938065111637115,
-0.019164519384503365,
-0.0220930315554142,
-0.01605408824980259,
0.09642612189054489,
-0.07477320730686188,
0.003649265505373478,
0.05966094508767128,
-0.0300787054002285,
-0.02191944606602192,
0.03936328366398811,
0.004247964359819889,
0.07129931449890137,
-0.035764239728450775,
0.09814418852329254,
0.08134192228317261,
-0.015862461179494858,
-0.033959414809942245,
0.011507009156048298,
0.0013514216989278793,
-0.026252584531903267,
-0.01266228873282671,
0.027163241058588028,
-0.027538228780031204,
0.01645330898463726,
-0.04024069383740425,
0.04636465385556221,
0.02726798690855503,
-0.03239915147423744,
0.0020864056423306465,
-0.07442794740200043,
0.030953356996178627,
0.00035868503618985415,
-0.060054730623960495,
-0.11028073728084564,
1.1272496749416407e-33,
-0.002700172597542405,
-0.004485905636101961,
0.02078244835138321,
0.07874662429094315,
0.10518603026866913,
0.05184771493077278,
0.03048393689095974,
-0.02231309935450554,
0.0536750890314579,
-0.008630653843283653,
-0.11028843373060226,
-0.047326087951660156,
-0.04981173947453499,
0.13743670284748077,
-0.000077620024967473,
-0.023463409394025803,
-0.043960701674222946,
0.045993972569704056,
-0.026517268270254135,
0.014545943588018417,
0.026048919185996056,
-0.022188590839505196,
-0.011464144103229046,
0.034490857273340225,
-0.06386757642030716,
0.016416117548942566,
0.05429525300860405,
-0.020458510145545006,
-0.032732658088207245,
-0.000529054319486022,
-0.07903928309679031,
-0.020801914855837822,
0.001741212559863925,
-0.005503619089722633,
0.03322279825806618,
0.04317096620798111,
-0.029465431347489357,
-0.02745622582733631,
-0.003104531206190586,
-0.06483444571495056,
0.07897432148456573,
0.04025422781705856,
0.011533932760357857,
-0.05041329935193062,
-0.13156527280807495,
0.023115133866667747,
0.0553760901093483,
0.045008257031440735,
-0.058415234088897705,
-0.015093481168150902,
0.13556548953056335,
0.05433568358421326,
-0.04123060032725334,
-0.04534459486603737,
-0.011949040926992893,
-0.03306248411536217,
0.005467492621392012,
-0.00797367375344038,
0.11717888712882996,
0.002567802555859089,
0.04814682900905609,
0.02682751603424549,
0.013312599621713161,
0.10026584565639496,
0.02608032152056694,
-0.04872746393084526,
0.07533342391252518,
0.06204294413328171,
0.0421869121491909,
0.018342716619372368,
-0.06103548780083656,
0.024151483550667763,
-0.02834874764084816,
-0.05214782804250717,
0.07374372333288193,
-0.022961143404245377,
0.03565547615289688,
-0.06709399819374084,
0.0008190061198547482,
0.02691054530441761,
-0.051229532808065414,
-0.0395713709294796,
0.029435740783810616,
-0.060702040791511536,
-0.013196521438658237,
-0.05069967359304428,
0.02579021081328392,
-0.04503148794174194,
0.0028772144578397274,
-0.004171926062554121,
-0.07657141983509064,
-0.12861335277557373,
0.033198852092027664,
-0.02906693145632744,
0.03136976435780525,
-1.9478422611125172e-33,
0.015737222507596016,
0.027911391109228134,
-0.013712039217352867,
0.0013239275431260467,
-0.011596850119531155,
-0.05852452293038368,
-0.045049238950014114,
0.07432526350021362,
-0.018763240426778793,
-0.04711684212088585,
-0.0076072243973612785,
0.04273870587348938,
0.04806682839989662,
-0.015306293033063412,
0.03259110823273659,
0.019099893048405647,
0.029580160975456238,
-0.03676660358905792,
-0.025283679366111755,
-0.007516270037740469,
-0.015756871551275253,
0.04257268086075783,
-0.06273642182350159,
0.041046589612960815,
-0.06597697734832764,
0.03246135264635086,
-0.06507192552089691,
0.04645603895187378,
0.021739302203059196,
0.07202780246734619,
0.006266864947974682,
-0.03414987027645111,
0.07146493345499039,
-0.03614240139722824,
-0.10574792325496674,
0.06560450792312622,
0.061432771384716034,
0.02432074211537838,
-0.03366681933403015,
-0.03673279657959938,
0.036966390907764435,
-0.1052030548453331,
-0.04884292185306549,
0.062397826462984085,
-0.014252028428018093,
0.025674544274806976,
-0.01894344761967659,
-0.002526775700971484,
0.01664450392127037,
0.0678524449467659,
-0.08565287292003632,
0.06517520546913147,
-0.022215619683265686,
-0.013791188597679138,
0.06806674599647522,
-0.05241754651069641,
-0.10009093582630157,
-0.037123892456293106,
0.10423222929239273,
-0.07651419937610626,
-0.02839738316833973,
-0.06849543750286102,
-0.01056844461709261,
-0.037429727613925934,
-0.03882956877350807,
-0.08636938035488129,
-0.07978300750255585,
-0.0812404602766037,
0.007192527875304222,
0.09501047432422638,
0.04127811640501022,
0.004256910644471645,
-0.046962566673755646,
0.048402611166238785,
-0.03480377048254013,
-0.12517084181308746,
0.03844791278243065,
0.03526201844215393,
0.0627632588148117,
-0.05192837864160538,
0.027642229571938515,
-0.003613249631598592,
0.03148628771305084,
0.07545164227485657,
0.06125950813293457,
0.02549356408417225,
0.03357808291912079,
-0.028073621913790703,
-0.008445159532129765,
0.07710366696119308,
0.05092517286539078,
0.02694273740053177,
-0.004874375648796558,
0.05319436639547348,
-0.03823133185505867,
-3.7702225341718076e-8,
-0.026265660300850868,
0.057337500154972076,
0.06260661035776138,
-0.05362014099955559,
-0.00950630009174347,
-0.06749340891838074,
0.07195121049880981,
0.16133205592632294,
-0.011753265745937824,
0.007979178801178932,
0.11442574858665466,
-0.08390827476978302,
-0.05727383866906166,
-0.011062937788665295,
0.06792516261339188,
0.09314412623643875,
0.029097246006131172,
-0.01669200509786606,
0.009245493449270725,
0.04340513423085213,
-0.056343190371990204,
0.07424542307853699,
0.01617751643061638,
-0.008003149181604385,
0.08399853855371475,
-0.019381124526262283,
-0.009964700788259506,
0.0889466255903244,
0.06731579452753067,
-0.06638562679290771,
0.0048655010759830475,
0.043332938104867935,
0.027721678838133812,
-0.018858516588807106,
-0.053309932351112366,
0.052160970866680145,
0.06707104295492172,
-0.02143539860844612,
-0.04073237627744675,
-0.15585121512413025,
-0.03338181599974632,
0.058429453521966934,
0.00788955483585596,
-0.015704980120062828,
-0.0019416512222960591,
0.0549471415579319,
0.0728975236415863,
-0.07456008344888687,
0.013604254461824894,
0.01047432143241167,
-0.02763219177722931,
-0.0036845507565885782,
-0.013627409934997559,
0.06329302489757538,
0.046966347843408585,
0.0737357810139656,
0.0019236784428358078,
-0.04529028758406639,
0.08646531403064728,
-0.031018288806080818,
-0.027580169960856438,
0.015561075881123543,
-0.04877450317144394,
0.01643257960677147
] |
InterLCM: Low-Quality Images as Intermediate States of Latent Consistency Models for Effective Blind Face Restoration | https://openreview.net/forum?id=rUxr9Ll5FQ | [
"Senmao Li",
"Kai Wang",
"Joost van de Weijer",
"Fahad Shahbaz Khan",
"Chun-Le Guo",
"Shiqi Yang",
"Yaxing Wang",
"jian Yang",
"Ming-Ming Cheng"
] | Poster | Diffusion priors have been used for blind face restoration (BFR) by fine-tuning diffusion models (DMs) on restoration datasets to recover low-quality images. However, the naive application of DMs presents several key limitations.
(i) The diffusion prior has inferior semantic consistency (e.g., ID, structure and color.), increasing the difficulty of optimizing the BFR model;
(ii) reliance on hundreds of denoising iterations, preventing the effective cooperation with perceptual losses, which is crucial for faithful restoration.
Observing that the latent consistency model (LCM) learns consistency noise-to-data mappings on the ODE-trajectory and therefore shows more semantic consistency in the subject identity, structural information and color preservation,
we propose $\textit{InterLCM}$ to leverage the LCM for its superior semantic consistency and efficiency to counter the above issues.
Treating low-quality images as the intermediate state of LCM, $\textit{InterLCM}$ achieves a balance between fidelity and quality by starting from earlier LCM steps.
LCM also allows the integration of perceptual loss during training, leading to improved restoration quality, particularly in real-world scenarios.
To mitigate structural and semantic uncertainties, $\textit{InterLCM}$ incorporates a Visual Module to extract visual features and a Spatial Encoder to capture spatial details, enhancing the fidelity of restored images.
Extensive experiments demonstrate that $\textit{InterLCM}$ outperforms existing approaches in both synthetic and real-world datasets while also achieving faster inference speed. Code and models will be publicly available. | diffusion model, face restoration | By regarding the lq image as intermediate state of the LCM, this paper propose the method InterLCM, along with extra conditions as visual embeddings and spatial embeddings, for efficient blind face restoration. | 536 | 2502.02215 | [
0.0028513646684587,
-0.10490510612726212,
0.08369795978069305,
0.03758266195654869,
0.03832045942544937,
-0.021324893459677696,
-0.05392187833786011,
-0.11699911952018738,
0.0316823311150074,
0.0023744218051433563,
0.02589493617415428,
0.02012973092496395,
0.09948474913835526,
-0.02331222966313362,
0.0038347954396158457,
0.010782919824123383,
0.03460569679737091,
0.11362142860889435,
-0.10700315237045288,
-0.04469223693013191,
-0.08917247503995895,
0.019812079146504402,
-0.024897737428545952,
-0.05234016105532646,
0.01774701103568077,
0.05944891273975372,
0.062442343682050705,
0.020731862634420395,
0.0586492083966732,
-0.07545558363199234,
-0.011593077331781387,
-0.016757426783442497,
-0.05373535677790642,
0.0031488430686295033,
-0.057638414204120636,
0.058393094688653946,
-0.03981459513306618,
0.015317539684474468,
-0.04494592547416687,
0.021620482206344604,
-0.031015193089842796,
0.0658072829246521,
-0.060807570815086365,
-0.012869978323578835,
0.05318176746368408,
-0.016720648854970932,
-0.013407310470938683,
0.005990522913634777,
0.008074458688497543,
-0.06279388070106506,
0.020092429593205452,
0.02066362276673317,
-0.05340125411748886,
0.050664499402046204,
-0.05320145934820175,
0.06959076225757599,
0.036799851804971695,
-0.03617601469159126,
-0.04196025803685188,
0.036503877490758896,
-0.04396403580904007,
-0.033378779888153076,
-0.04576314985752106,
0.013165595941245556,
0.03504333645105362,
0.06019110605120659,
0.05969461798667908,
-0.055744096636772156,
0.006770903244614601,
0.02730637788772583,
-0.03212163597345352,
0.013261892832815647,
-0.011111794039607048,
0.021233517676591873,
-0.0468837134540081,
0.011851592920720577,
0.051517289131879807,
-0.031037582084536552,
-0.002509321551769972,
-0.005322152748703957,
0.05419453978538513,
0.03962145373225212,
-0.0067205424420535564,
0.01851418800652027,
0.10720472037792206,
-0.05265725776553154,
-0.004479252267628908,
0.006282202433794737,
-0.06413068622350693,
0.0672546848654747,
-0.027998317033052444,
0.023841410875320435,
-0.04337760806083679,
0.03554220870137215,
0.005915807094424963,
0.03794987127184868,
0.11623906344175339,
0.0017914503114297986,
0.052720535546541214,
0.08256334811449051,
0.022681526839733124,
-0.013890042901039124,
-0.06729551404714584,
0.0064408681355416775,
0.03366176038980484,
-0.05804872885346413,
0.10862071812152863,
0.07703958451747894,
-0.019110988825559616,
0.023053575307130814,
0.03547266870737076,
0.012874141335487366,
0.00937210489064455,
0.061692338436841965,
0.05027146637439728,
-0.09215047210454941,
0.038644589483737946,
-0.09637763351202011,
0.07166307419538498,
-0.05912201106548309,
0.015200400725007057,
-0.008583144284784794,
0.050302356481552124,
-0.057408686727285385,
0.011426975019276142,
0.0029019848443567753,
0.01642519049346447,
3.5591445762641925e-34,
-0.0324489064514637,
0.012605274096131325,
0.03261713311076164,
-0.020134776830673218,
0.026708578690886497,
0.02319713495671749,
-0.040492452681064606,
-0.014045899733901024,
-0.07332362979650497,
-0.0020994313526898623,
-0.042123373597860336,
0.008277645334601402,
-0.14643949270248413,
0.046643298119306564,
-0.0481664277613163,
-0.04432426020503044,
0.011259189806878567,
0.05863143876194954,
-0.003312393557280302,
0.011426946148276329,
0.014187120832502842,
0.029796253889799118,
0.009964997880160809,
0.002404759405180812,
-0.02065316215157509,
0.004377624951303005,
0.11319247633218765,
-0.069753497838974,
0.0312254186719656,
-0.011203451082110405,
-0.16870397329330444,
0.069118931889534,
0.07070325314998627,
0.0013589176815003157,
0.02108553983271122,
-0.05285962298512459,
0.039587780833244324,
-0.04086298123002052,
-0.04555511102080345,
-0.041612833738327026,
-0.0324125811457634,
0.10795668512582779,
-0.03462991490960121,
-0.08505364507436752,
-0.03880089893937111,
0.008097806014120579,
0.041876278817653656,
0.0078250952064991,
-0.03999674320220947,
-0.0557384192943573,
0.005140428896993399,
0.022633913904428482,
-0.041598446667194366,
-0.07807056605815887,
-0.03748111054301262,
0.0003090830869041383,
0.04216828569769859,
-0.02556469291448593,
-0.005327451974153519,
-0.015388363040983677,
0.10452613979578018,
0.03440002351999283,
0.018455488607287407,
0.10201356559991837,
0.06383629143238068,
0.030968129634857178,
-0.059818465262651443,
0.010597681626677513,
-0.0996527224779129,
-0.07766375690698624,
0.017975885421037674,
0.03147770091891289,
-0.0445156991481781,
-0.03868545964360237,
0.0777568370103836,
-0.048310521990060806,
0.07251030206680298,
0.02138158306479454,
0.061159174889326096,
0.00798108521848917,
-0.028492704033851624,
-0.018497861921787262,
-0.10356405377388,
-0.0702376663684845,
-0.0756957083940506,
0.04218004271388054,
0.03752874955534935,
-0.05881252884864807,
-0.002772609004750848,
-0.10680975019931793,
0.0008645934867672622,
0.0011459288652986288,
0.03384825959801674,
-0.006692743860185146,
-0.04540599137544632,
-5.982624315567299e-34,
-0.015520126558840275,
0.014756040647625923,
0.041187867522239685,
0.13011349737644196,
-0.0361773818731308,
-0.08117017894983292,
0.010837149806320667,
0.04570741951465607,
0.1508704572916031,
-0.030506808310747147,
0.019755307585000992,
-0.009330634027719498,
-0.0544310137629509,
-0.05208387225866318,
0.018425628542900085,
-0.006174870766699314,
0.03776257112622261,
-0.04146634042263031,
-0.1025349497795105,
0.04244783893227577,
0.03402097523212433,
0.02767827920615673,
-0.06973917782306671,
0.05309342220425606,
-0.03602569177746773,
0.07768096029758453,
-0.0378587543964386,
0.0346810407936573,
0.014159570448100567,
-0.02311897464096546,
0.039446525275707245,
-0.025104044005274773,
0.04335017502307892,
-0.025475624948740005,
-0.05176132917404175,
0.04240316525101662,
-0.00003374011430423707,
-0.03603457286953926,
-0.11460430175065994,
0.08987743407487869,
0.03054661676287651,
-0.004897904582321644,
-0.10169440507888794,
0.010265304706990719,
0.05589636415243149,
-0.06649576872587204,
-0.0524926632642746,
-0.046943917870521545,
0.020105961710214615,
0.04564843699336052,
0.035513412207365036,
-0.006820785813033581,
-0.09638664126396179,
0.0866466537117958,
-0.014321768656373024,
-0.04490487650036812,
-0.0022411239333450794,
0.014060902409255505,
0.060853660106658936,
0.04931012541055679,
0.010334901511669159,
-0.041776884347200394,
-0.0671590268611908,
-0.025553934276103973,
-0.02002898044884205,
0.02292662300169468,
-0.03293287009000778,
0.039029061794281006,
-0.009199666790664196,
0.013706936500966549,
0.0010339391883462667,
0.0058745876885950565,
0.0025450100656598806,
-0.03338748961687088,
0.1254759132862091,
-0.050973210483789444,
-0.0015918499557301402,
0.03194264695048332,
-0.016240183264017105,
0.006211947184056044,
0.031286194920539856,
-0.05636066570878029,
0.021416043862700462,
0.04710701107978821,
-0.030842352658510208,
0.029583679512143135,
0.0377715528011322,
-0.10868919640779495,
0.023127295076847076,
-0.004180994816124439,
-0.02746829204261303,
0.011544700711965561,
-0.024039873853325844,
0.007566830608993769,
-0.11206954717636108,
-4.4424282918953395e-8,
-0.08657144755125046,
-0.0056664603762328625,
-0.03654209151864052,
-0.003643755568191409,
0.007353152614086866,
-0.16016079485416412,
0.025481052696704865,
0.01702185533940792,
0.01184572372585535,
-0.05168118700385094,
0.059918250888586044,
-0.045915402472019196,
-0.08204088360071182,
-0.0591733418405056,
0.006142542231827974,
0.06771604716777802,
0.03442782536149025,
0.024424387142062187,
0.027528848499059677,
0.04396447911858559,
-0.08650531619787216,
0.05696404352784157,
0.003136411774903536,
-0.01634611189365387,
0.08907133340835571,
0.01238219253718853,
0.005702260881662369,
0.07555918395519257,
0.03419191762804985,
0.02270706556737423,
0.04060443863272667,
0.011484102346003056,
0.08802533894777298,
-0.015075674280524254,
-0.027460109442472458,
0.0858841985464096,
0.10187498480081558,
-0.01768564060330391,
-0.05408506095409393,
-0.02645530365407467,
-0.04184090346097946,
0.03874091058969498,
-0.029108360409736633,
0.005395413842052221,
0.0391668826341629,
0.06166661158204079,
0.015096256509423256,
-0.021640349179506302,
0.02418977953493595,
-0.03189389035105705,
-0.008179978467524052,
0.017833370715379715,
0.023740140721201897,
0.07056951522827148,
0.0779140517115593,
-0.07531428337097168,
0.08244477212429047,
0.021351156756281853,
0.06189645454287529,
-0.011805858463048935,
0.06282658129930496,
-0.011667195707559586,
0.0491146594285965,
0.05069153755903244
] |
Generalization Bounds for Canonicalization: A Comparative Study with Group Averaging | https://openreview.net/forum?id=n0lXaskyk5 | [
"Behrooz Tahmasebi",
"Stefanie Jegelka"
] | Poster | Canonicalization, a popular method for generating invariant or equivariant function classes from arbitrary function sets, involves initial data projection onto a reduced input space subset, followed by applying any learning method to the projected dataset. Despite recent research on the expressive power and continuity of functions represented by canonicalization, its generalization capabilities remain less explored. This paper addresses this gap by theoretically examining the generalization benefits and sample complexity of canonicalization, comparing them with group averaging, another popular technique for creating invariant or equivariant function classes.
Our findings reveal two distinct regimes where canonicalization may outperform or underperform compared to group averaging, with precise quantification of this phase transition in terms of sample size, group action characteristics, and a newly introduced concept of alignment.
To the best of our knowledge, this study represents the first theoretical exploration of such behavior, offering insights into the relative effectiveness of canonicalization and group averaging under varying conditions. | invariances, group, symmetry, canonicalization | We investigate generalization bounds for canonicalization and provide a comparative analysis with group averaging. | 530 | null | [
-0.052762530744075775,
-0.06779380142688751,
0.04059327393770218,
-0.03885289281606674,
0.06478694826364517,
0.026297520846128464,
-0.03619232401251793,
-0.009778128936886787,
0.0831386148929596,
-0.011416810564696789,
0.039884213358163834,
0.03805171325802803,
0.01045036781579256,
-0.0324639230966568,
0.014526518061757088,
-0.034830689430236816,
-0.0682491809129715,
0.09253005683422089,
-0.09460511058568954,
-0.0018838349496945739,
0.04946571961045265,
0.038565050810575485,
0.006267031654715538,
0.01768931932747364,
0.05010945722460747,
-0.01566690392792225,
-0.030885331332683563,
-0.027238599956035614,
0.016761422157287598,
-0.0647975280880928,
-0.06002587825059891,
0.07035776972770691,
0.015868328511714935,
-0.03260202333331108,
-0.047262802720069885,
0.08665726333856583,
0.015763264149427414,
0.02805902436375618,
-0.004149169661104679,
0.012847923673689365,
-0.029082996770739555,
0.025011571124196053,
0.02227494865655899,
-0.011155810207128525,
0.05801234021782875,
0.11115112155675888,
0.09411328285932541,
-0.03207266703248024,
-0.0990690290927887,
0.052751824259757996,
0.012576629407703876,
0.06739640980958939,
-0.06743835657835007,
0.07219863682985306,
0.03430107235908508,
-0.06106552481651306,
-0.000768932281062007,
-0.07422783970832825,
0.040258947759866714,
-0.039006058126688004,
-0.06122751533985138,
-0.04674585908651352,
0.011771508492529392,
0.07160511612892151,
0.04004961624741554,
0.0030198772437870502,
0.06402537971735,
0.027811788022518158,
-0.041372478008270264,
0.051535241305828094,
-0.05364096537232399,
0.029427167028188705,
-0.05997120216488838,
0.0752730742096901,
0.033885177224874496,
0.10007938742637634,
-0.018553810194134712,
0.004381006583571434,
-0.04725664481520653,
-0.006438537500798702,
0.0644746869802475,
0.0596010722219944,
0.049417801201343536,
0.005902780219912529,
0.06395656615495682,
0.014257815666496754,
-0.023178182542324066,
0.02085064724087715,
0.0289326049387455,
0.05792880430817604,
-0.10774639248847961,
0.044386811554431915,
0.04383450001478195,
-0.047749798744916916,
0.023640092462301254,
-0.0549742765724659,
0.07222045958042145,
-0.0727635845541954,
0.1354067027568817,
0.10985169559717178,
-0.023852214217185974,
0.038934629410505295,
0.03001600317656994,
-0.039257049560546875,
0.015346182510256767,
-0.1155213862657547,
0.0158193688839674,
-0.032952796667814255,
0.005301977042108774,
-0.11518360674381256,
-0.011152539402246475,
-0.052094146609306335,
-0.021010087803006172,
-0.0266435407102108,
-0.04120095446705818,
-0.08996579051017761,
0.0569189190864563,
0.015133412554860115,
0.06932560354471207,
-0.05880466476082802,
0.08513513207435608,
-0.02603406086564064,
0.060909565538167953,
0.09571739286184311,
0.017154905945062637,
0.06069544702768326,
-0.11259107291698456,
5.483433925430314e-33,
0.03984115645289421,
0.0026798236649483442,
0.05294981598854065,
0.04615648090839386,
-0.025596898049116135,
-0.010997734032571316,
-0.0019284550799056888,
-0.025504417717456818,
-0.022848762571811676,
0.07747844606637955,
-0.0677918940782547,
0.057145632803440094,
0.0169514212757349,
0.038463082164525986,
-0.014605103991925716,
0.06504335254430771,
0.013456563465297222,
0.06250529736280441,
-0.03757590800523758,
-0.0445803664624691,
0.04536450281739235,
0.045760251581668854,
0.010312759317457676,
-0.06635738909244537,
-0.05843052268028259,
-0.005756737664341927,
-0.05939074978232384,
-0.038688771426677704,
-0.11006171256303787,
-0.03311574459075928,
-0.06654686480760574,
0.005631254054605961,
-0.021793721243739128,
0.0634283795952797,
0.022129762917757034,
-0.016669677570462227,
0.05720595270395279,
-0.09623077511787415,
0.009750347584486008,
-0.011531819589436054,
0.02401643618941307,
0.027302127331495285,
-0.027507292106747627,
-0.04947298765182495,
0.031266700476408005,
0.005239568650722504,
-0.011346060782670975,
0.037789732217788696,
-0.03981686383485794,
-0.017977945506572723,
-0.022815069183707237,
0.02866777963936329,
0.006136324722319841,
-0.04649699106812477,
-0.049906257539987564,
0.049467284232378006,
-0.01655239798128605,
0.03460723161697388,
-0.053184960037469864,
0.017032543197274208,
-0.06199203431606293,
0.024523215368390083,
-0.03087073378264904,
-0.05127163603901863,
-0.03365335240960121,
0.05286847800016403,
-0.026901597157120705,
0.004257981199771166,
0.15107767283916473,
0.010146263055503368,
0.06679438054561615,
0.027857935056090355,
-0.1250714659690857,
0.000830977747682482,
0.11686962097883224,
0.03235343471169472,
0.00816031452268362,
-0.07666170597076416,
-0.0633537545800209,
0.038839370012283325,
-0.04542429372668266,
0.02471740171313286,
-0.03931817039847374,
-0.05209161341190338,
-0.059135593473911285,
-0.08201885223388672,
0.05393150821328163,
-0.06771580129861832,
-0.0800914540886879,
-0.014916040934622288,
-0.10236720740795135,
0.034208592027425766,
-0.00188633578363806,
0.08979448676109314,
0.024750467389822006,
-6.832265571198902e-33,
-0.07222974300384521,
-0.009928413666784763,
0.09739618003368378,
0.09776334464550018,
0.020564448088407516,
-0.057779762893915176,
-0.0673844963312149,
-0.0065674856305122375,
0.023539137095212936,
-0.0038920969236642122,
0.03450574725866318,
-0.015811365097761154,
0.016562599688768387,
-0.04307801276445389,
0.011874436400830746,
0.05816101282835007,
-0.04063764587044716,
-0.015067336149513721,
0.06788865476846695,
0.08466149121522903,
0.04293287172913551,
0.011147791519761086,
0.0033833906054496765,
0.026645494624972343,
0.010824533179402351,
0.04367699846625328,
-0.022556666284799576,
-0.01714332029223442,
0.04057085141539574,
-0.01485177781432867,
-0.08636275678873062,
-0.03390996530652046,
-0.10851740092039108,
-0.008157662115991116,
-0.053633324801921844,
0.00865466333925724,
-0.035574398934841156,
0.06464812159538269,
-0.023833364248275757,
0.10230977088212967,
-0.032833971083164215,
0.05610254406929016,
-0.04204953834414482,
-0.036924030631780624,
0.06098014861345291,
-0.04665868729352951,
0.09486773610115051,
0.020378226414322853,
0.017818741500377655,
-0.016059214249253273,
-0.024026479572057724,
0.0737074762582779,
-0.0626605674624443,
-0.027345456182956696,
-0.01681709475815296,
-0.016224496066570282,
0.05724560469388962,
-0.0823080837726593,
-0.03577631339430809,
-0.014755399897694588,
0.009437810629606247,
-0.03843340650200844,
-0.01924973726272583,
-0.021647905930876732,
-0.041751209646463394,
-0.013561657629907131,
0.026224840432405472,
-0.028481507673859596,
0.02574019506573677,
0.067396380007267,
-0.023174550384283066,
-0.1014123186469078,
-0.025998026132583618,
0.0018538320437073708,
-0.013410098850727081,
-0.058749277144670486,
0.07303956151008606,
-0.08768778294324875,
-0.021868783980607986,
0.027598466724157333,
0.04092775657773018,
-0.0005435865023173392,
0.038940705358982086,
0.010719896294176579,
-0.055830225348472595,
0.12465450167655945,
0.058717694133520126,
0.03866242617368698,
0.060744013637304306,
0.007944764569401741,
-0.009807571768760681,
0.06307995319366455,
0.005747796967625618,
0.10418367385864258,
-0.10266139358282089,
-4.726798152887568e-8,
-0.039704348891973495,
0.00435228506103158,
-0.01753823831677437,
0.001387317548505962,
0.024393368512392044,
-0.034305982291698456,
-0.02490827813744545,
0.05196208506822586,
-0.019135834649205208,
-0.08481999486684799,
0.03601492941379547,
0.01407187432050705,
0.02048911713063717,
-0.1229097917675972,
0.03707109019160271,
0.022250724956393242,
-0.013271614909172058,
0.01960577629506588,
-0.03942115232348442,
0.03800409659743309,
-0.05793716013431549,
-0.018804458901286125,
-0.04832644388079643,
-0.10440381616353989,
-0.026015933603048325,
-0.13131453096866608,
-0.04668959602713585,
-0.009414593689143658,
0.027064841240644455,
0.0012745409039780498,
-0.04295755550265312,
0.06597774475812912,
-0.010595967061817646,
0.030280619859695435,
-0.007161388639360666,
-0.009048620238900185,
-0.04688919708132744,
-0.004713843110948801,
0.008884602226316929,
-0.02313142456114292,
0.00939851813018322,
0.016817297786474228,
-0.004387455992400646,
0.05161857232451439,
0.0024444942828267813,
0.020662829279899597,
0.02907724492251873,
-0.020666111260652542,
-0.010652238503098488,
-0.046141862869262695,
0.08633467555046082,
-0.0007632319466210902,
-0.03261517733335495,
0.01131858304142952,
0.06274878233671188,
0.08180899173021317,
-0.0015673728194087744,
-0.030621426180005074,
0.0430292934179306,
-0.05733591690659523,
-0.07087961584329605,
-0.011285815387964249,
-0.06580380350351334,
-0.05856049805879593
] |
VibeCheck: Discover and Quantify Qualitative Differences in Large Language Models | https://openreview.net/forum?id=acxHV6werE | [
"Lisa Dunlap",
"Krishna Mandal",
"Trevor Darrell",
"Jacob Steinhardt",
"Joseph E. Gonzalez"
] | Poster | Large language models (LLMs) often exhibit subtle yet distinctive characteristics in their outputs that users intuitively recognize, but struggle to quantify. These "vibes" -- such as tone, formatting, or writing style -- influence user preferences, yet traditional evaluations focus primarily on the singular vibe of correctness.
We introduce $\textbf{VibeCheck}$, a system for automatically comparing a pair of LLMs by discovering identifying traits of a model ("vibes") that are well-defined, differentiating, and user-aligned. VibeCheck iteratively discovers vibes from model outputs and then utilizes a panel of LLM judges to quantitatively measure the utility of each vibe.
We validate that the vibes generated by VibeCheck align with those found in human discovery and run VibeCheck on pairwise preference data from real-world user conversations with Llama-3-70b vs GPT-4. VibeCheck reveals that Llama has a friendly, funny, and somewhat controversial vibe. These vibes predict model identity with 80% accuracy and human preference with 61% accuracy. Lastly, we run VibeCheck on a variety of models and tasks, including summarization, math, and captioning to provide insight into differences in model behavior. VibeCheck discovers vibes like Command X prefers to add concrete intros and conclusions when summarizing in comparison to TNGL, Llama-405b often overexplains its thought process on math problems compared to GPT-4o, and GPT-4 prefers to focus on the mood and emotions of the scene when captioning compared to Gemini-1.5-Flash. | large language models, evaluation | Find interesting subjective properties that differentiate models | 529 | 2410.12851 | [
-0.02606983296573162,
-0.0978814885020256,
-0.0027525906916707754,
-0.006992179434746504,
-0.0038876512553542852,
-0.013018905185163021,
0.04402812570333481,
-0.053431279957294464,
0.036758311092853546,
-0.03652425482869148,
-0.0651492029428482,
-0.04520343616604805,
0.05774668604135513,
-0.00666886568069458,
0.04403391480445862,
0.03821616247296333,
0.13823920488357544,
-0.014113669283688068,
-0.027321865782141685,
0.021413613110780716,
-0.07622333616018295,
0.021549196913838387,
0.025117551907896996,
-0.0007243946311064065,
-0.0013736209366470575,
0.022181987762451172,
0.010517473332583904,
0.04845086857676506,
0.03987574949860573,
-0.039459485560655594,
0.033327288925647736,
0.11097265034914017,
-0.011063167825341225,
0.006401261314749718,
-0.08179488778114319,
0.026269612833857536,
0.001000237069092691,
0.025666961446404457,
0.0029501905664801598,
-0.096499964594841,
-0.08350402861833572,
-0.02469341643154621,
0.039767075330019,
0.03227146714925766,
0.04979411140084267,
-0.015011860989034176,
-0.04691353812813759,
0.024152914062142372,
-0.018089069053530693,
0.0007549357251264155,
-0.03675812482833862,
-0.10134388506412506,
0.030415918678045273,
0.017587387934327126,
-0.04101167246699333,
0.047870270907878876,
0.011132783256471157,
0.0679607093334198,
0.011702749878168106,
-0.01750721037387848,
-0.048477865755558014,
-0.1170399859547615,
-0.04562103748321533,
-0.004547101445496082,
-0.00916273146867752,
-0.004271253012120724,
-0.002396825235337019,
0.055147480219602585,
0.016623515635728836,
-0.023680103942751884,
-0.00775431701913476,
0.02068612538278103,
0.012150691822171211,
0.11082910001277924,
0.0011312870774418116,
0.11537424474954605,
0.07309840619564056,
-0.023497220128774643,
0.04028603807091713,
-0.11166562139987946,
-0.04102381318807602,
-0.006207064259797335,
-0.038671351969242096,
0.0017493508057668805,
0.0860646590590477,
0.006533407606184483,
-0.00002916730591095984,
0.029696980491280556,
-0.01749970018863678,
0.0662238597869873,
-0.019220096990466118,
-0.011579940095543861,
0.07208476960659027,
-0.08745462447404861,
0.03512018918991089,
0.05741975083947182,
-0.009361866861581802,
-0.030339675024151802,
0.007869009859859943,
0.07036814093589783,
-0.019215630367398262,
0.05376158654689789,
0.009937477298080921,
-0.025218266993761063,
0.041341692209243774,
-0.019311020150780678,
-0.020127424970269203,
-0.014422494918107986,
0.07121880352497101,
-0.06526006013154984,
-0.013450315222144127,
0.03515702486038208,
-0.0012021379079669714,
0.01376077439635992,
0.09335548430681229,
-0.021561788395047188,
0.0790814459323883,
-0.0282248817384243,
0.019669968634843826,
0.020591892302036285,
0.02317131869494915,
-0.003771870629861951,
0.060278501361608505,
-0.007289569359272718,
0.018973540514707565,
0.003844177583232522,
-0.04349535331130028,
3.7398554809502625e-33,
0.05704876035451889,
0.06785153597593307,
0.0520080141723156,
0.055829957127571106,
0.06639380007982254,
-0.01887129619717598,
-0.14439597725868225,
-0.016660865396261215,
-0.05949082225561142,
0.03849237412214279,
-0.061796098947525024,
0.07083270698785782,
-0.015360473655164242,
0.10806587338447571,
0.02180393598973751,
0.021782467141747475,
-0.0923953726887703,
-0.009073524735867977,
-0.05221978574991226,
-0.037696871906518936,
0.01453655119985342,
-0.02713843435049057,
0.030383193865418434,
-0.014903766103088856,
-0.01874162256717682,
0.007812409196048975,
0.09526263922452927,
0.011704841628670692,
-0.03207150474190712,
-0.00831734947860241,
-0.055791810154914856,
-0.07187888026237488,
0.086052805185318,
0.03790945187211037,
-0.005336802452802658,
-0.04118294268846512,
-0.009699249640107155,
-0.04696732386946678,
0.08118030428886414,
0.0058441259898245335,
0.016694428399205208,
0.003719947300851345,
-0.030558662489056587,
-0.004612326622009277,
-0.06885158270597458,
0.04267098754644394,
0.013583310879766941,
-0.017269253730773926,
-0.00530421594157815,
-0.023302461951971054,
0.05999774485826492,
-0.012823627330362797,
0.032218534499406815,
0.09012795984745026,
-0.11654476076364517,
-0.014210108667612076,
0.019350918009877205,
-0.03882057964801788,
0.01624050922691822,
-0.04306880384683609,
-0.0029118231032043695,
0.05784694105386734,
0.024906210601329803,
-0.10293542593717575,
0.07720421999692917,
-0.013879597187042236,
-0.04920530691742897,
0.0005828164867125452,
-0.005680850241333246,
0.04545672610402107,
-0.01879466138780117,
0.038590215146541595,
-0.0351049080491066,
-0.05325710028409958,
-0.030540814623236656,
-0.03846590965986252,
0.005108935758471489,
0.01382207777351141,
0.017547378316521645,
0.05436074733734131,
-0.0065110716968774796,
-0.0039512054063379765,
0.014793625101447105,
-0.0808977261185646,
-0.02486821636557579,
-0.06330176442861557,
0.04483013227581978,
-0.11741197109222412,
-0.055842772126197815,
0.012292305938899517,
-0.06559152901172638,
0.012406174093484879,
0.09096503257751465,
-0.00547010125592351,
-0.02680922858417034,
-4.2640333912368754e-33,
-0.11978641152381897,
-0.030932331457734108,
0.017463471740484238,
0.12035097926855087,
-0.043550048023462296,
-0.06513996422290802,
0.024435274302959442,
0.06364551931619644,
-0.08512984216213226,
-0.08944748342037201,
-0.006276783999055624,
-0.026410700753331184,
0.0745205357670784,
-0.019682222977280617,
0.04010630026459694,
-0.008691636845469475,
-0.04520180821418762,
-0.034647852182388306,
0.001410191529430449,
0.06126236915588379,
0.030061446130275726,
0.04242347925901413,
-0.12165287137031555,
0.035916540771722794,
-0.10809775441884995,
0.026045406237244606,
0.018412945792078972,
-0.0687054768204689,
0.013530648313462734,
-0.0016333521343767643,
-0.006957106292247772,
0.02345407009124756,
-0.036772314459085464,
-0.042923454195261,
-0.0027284862007945776,
-0.011997354216873646,
-0.007574775721877813,
0.02256891317665577,
0.00887356698513031,
0.11554864048957825,
-0.005500481463968754,
0.047955065965652466,
-0.09007351100444794,
-0.03752852976322174,
-0.05867401137948036,
0.00016480502381455153,
-0.09895115345716476,
0.04455678537487984,
-0.052080489695072174,
-0.055017393082380295,
0.06383444368839264,
0.01012061070650816,
0.017781879752874374,
0.08826160430908203,
0.005140709225088358,
-0.10424284636974335,
0.02341189607977867,
-0.007891328074038029,
0.0345984511077404,
0.07661312818527222,
-0.08821836113929749,
-0.05679456889629364,
0.001456362078897655,
-0.11051624268293381,
-0.001966057810932398,
-0.039286058396101,
0.024125566706061363,
-0.008243812248110771,
-0.018027761951088905,
-0.03581324964761734,
0.07067025452852249,
-0.04277823120355606,
-0.10189585387706757,
-0.028710875660181046,
0.019895706325769424,
-0.02089751325547695,
0.020668303593993187,
-0.05908697471022606,
-0.01574847474694252,
-0.04948892444372177,
-0.011875350959599018,
-0.006905185990035534,
0.050861697643995285,
-0.02512260712683201,
0.04543803259730339,
0.08004984259605408,
-0.03958696871995926,
0.11422649770975113,
-0.004456410650163889,
0.06081153824925423,
-0.03984318673610687,
0.0771874487400055,
-0.02305220626294613,
0.014764402061700821,
-0.047934334725141525,
-4.8356088910850303e-8,
-0.10571549087762833,
-0.06338150799274445,
0.01658673584461212,
0.07622133195400238,
-0.015495438128709793,
0.01401614211499691,
0.009534940123558044,
0.016564836725592613,
-0.0038634822703897953,
-0.005874125752598047,
0.03975215554237366,
0.0013313358649611473,
-0.12567132711410522,
-0.006354254204779863,
0.0363045334815979,
0.05889628455042839,
-0.039433930069208145,
0.0680035874247551,
-0.057854063808918,
-0.026620063930749893,
0.07878004759550095,
0.09418134391307831,
0.016984345391392708,
-0.0248115211725235,
0.04783651605248451,
-0.037586476653814316,
-0.04243173077702522,
0.0738748237490654,
0.03450242057442665,
0.061884570866823196,
0.009562008082866669,
0.02246629260480404,
-0.025964558124542236,
-0.09129946678876877,
-0.0002985941246151924,
0.04694021865725517,
-0.04035032168030739,
-0.049080945551395416,
0.09604407101869583,
0.05135642737150192,
0.08756119012832642,
0.07198905199766159,
-0.07578982412815094,
0.026926739141345024,
-0.06293229758739471,
-0.08347678929567337,
0.039487868547439575,
-0.09569631516933441,
-0.0009352532215416431,
0.028130846098065376,
0.03177744895219803,
-0.05111151933670044,
0.015355107374489307,
0.05234607681632042,
0.026637041941285133,
-0.04105571657419205,
-0.0877557247877121,
0.044441159814596176,
0.09104017168283463,
0.059917036443948746,
0.07062462717294693,
0.005726853851228952,
-0.0387108139693737,
-0.02540753223001957
] |
3D-SPATIAL MULTIMODAL MEMORY | https://openreview.net/forum?id=XYdstv3ySl | [
"Xueyan Zou",
"Yuchen Song",
"Ri-Zhao Qiu",
"Xuanbin Peng",
"Jianglong Ye",
"Sifei Liu",
"Xiaolong Wang"
] | Poster | We present 3D Spatial MultiModal Memory (M3), a multimodal memory system designed to retain information about medium-sized static scenes through video sources for visual perception. By integrating 3D Gaussian Splatting techniques with foundation models, M3 builds a multimodal memory capable of rendering feature representations across granularities, encompassing a wide range of knowledge. In our exploration, we identify two key challenges in previous works on feature splatting: (1) computational constraints in storing high-dimensional features for each Gaussian primitive, and (2) misalignment or information loss between distilled features and foundation model features. To address these challenges, we propose M3 with key components of principal scene components and Gaussian memory attention, enabling efficient training and inference. To validate M3, we conduct comprehensive quantitative evaluations of feature similarity and downstream tasks, as well as qualitative visualizations to highlight the pixel trace of Gaussian memory attention. Our approach encompasses a diverse range of foundation models, including vision-language models (VLMs), perception models, and large multimodal and language models (LMMs/LLMs). Furthermore, to demonstrate real-world applicability, we deploy M3’s feature field in indoor scenes on a quadruped robot. Notably, we claim that M3 is the first work to address the core compression challenges in 3D feature distillation. | Foundation Model, Gaussian Splatting, Large Multimodal Model, Robotics | null | 522 | 2503.16413 | [
0.07679544389247894,
-0.18398278951644897,
-0.011732655577361584,
0.00930735468864441,
0.12162255495786667,
0.02691524662077427,
-0.0270770862698555,
-0.02682400867342949,
-0.019277596846222878,
-0.08137977868318558,
0.0009746901341713965,
-0.07708626240491867,
0.01496442686766386,
0.13566185534000397,
-0.012785912491381168,
0.015199407003819942,
0.07892191410064697,
0.09905606508255005,
-0.1146644726395607,
0.002612985437735915,
0.10776900500059128,
-0.029505129903554916,
0.012414789758622646,
0.05061361938714981,
0.04838797450065613,
0.029461877420544624,
0.061894740909338,
-0.026357529684901237,
0.016614319756627083,
-0.012020712718367577,
0.02972417324781418,
0.007331441156566143,
-0.028564561158418655,
0.04363267123699188,
-0.021025752648711205,
0.0653410330414772,
-0.055721186101436615,
0.019180772826075554,
-0.045311540365219116,
-0.08075414597988129,
0.013528159819543362,
0.039058320224285126,
-0.023144172504544258,
0.053409140557050705,
0.06606624275445938,
0.11105947941541672,
0.017412614077329636,
-0.1017378494143486,
0.015712065622210503,
-0.02295861579477787,
-0.06264346092939377,
-0.030483832582831383,
-0.03629060462117195,
0.04362797364592552,
-0.03004492074251175,
0.033697087317705154,
0.030268749222159386,
-0.055004317313432693,
0.016257312148809433,
0.025346221402287483,
0.01729132980108261,
-0.09138477593660355,
0.04588316008448601,
-0.04417535290122032,
-0.012010264210402966,
-0.06774916499853134,
0.06402160972356796,
-0.05312873050570488,
-0.05570722743868828,
0.017200898379087448,
-0.09248071163892746,
0.03463241457939148,
-0.04321987181901932,
-0.03572819381952286,
-0.028410645201802254,
0.09386194497346878,
-0.020556746050715446,
-0.044977493584156036,
0.00022079478367231786,
-0.06936020404100418,
0.06936310976743698,
0.0307929627597332,
-0.012330296449363232,
0.02419210970401764,
0.12205033004283905,
0.012594180181622505,
0.019374320283532143,
0.07909397035837173,
-0.031631190329790115,
0.011509845033288002,
-0.02404623106122017,
-0.021800965070724487,
0.03052062727510929,
-0.05440615490078926,
0.01017744280397892,
-0.03225266560912132,
0.06263570487499237,
-0.10709347575902939,
-0.00526062585413456,
0.0010555378394201398,
0.034677889198064804,
0.026572497561573982,
0.054439399391412735,
0.011372546665370464,
0.10421746224164963,
-0.007807867601513863,
0.07845719903707504,
0.051138732582330704,
0.00893560890108347,
-0.03224393352866173,
0.029176736250519753,
0.027665140107274055,
0.012266716919839382,
0.021882731467485428,
0.006076024379581213,
0.005076720379292965,
0.017842972651124,
-0.03256830945611,
-0.004598935134708881,
0.026035426184535027,
0.00135506980586797,
-0.028659233823418617,
-0.050533704459667206,
-0.007687023375183344,
0.011806881986558437,
-0.06536764651536942,
-0.03114977292716503,
2.2526854151631688e-33,
0.03691738471388817,
0.0034263082779943943,
-0.0210859514772892,
0.07879827171564102,
0.028096089139580727,
-0.007926415652036667,
-0.022579003125429153,
0.044016528874635696,
-0.019148726016283035,
0.025412460789084435,
-0.0066610220819711685,
-0.010773122310638428,
-0.04015616327524185,
0.14618754386901855,
-0.045073311775922775,
0.021038273349404335,
-0.0724267065525055,
-0.0005473395576700568,
0.01308339461684227,
0.06659769266843796,
0.005662950221449137,
-0.0014795581810176373,
-0.018655145540833473,
-0.0013895357260480523,
0.0037631900049746037,
-0.04111596569418907,
-0.02778700739145279,
-0.06889531761407852,
-0.006030078511685133,
-0.028785524889826775,
-0.08631780743598938,
0.07543493807315826,
-0.03942957893013954,
0.0031613255850970745,
0.015428978018462658,
-0.09024924784898758,
-0.004121457692235708,
-0.11107641458511353,
0.021466972306370735,
-0.06451713293790817,
-0.05654093995690346,
0.021976426243782043,
-0.006408495828509331,
-0.015384398400783539,
-0.11166483163833618,
0.012883149087429047,
0.10395340621471405,
0.03166724368929863,
-0.02296825498342514,
-0.02489682286977768,
0.06492754817008972,
0.007021721452474594,
-0.07627101242542267,
-0.05731678381562233,
-0.02473670244216919,
-0.03983223810791969,
0.02592267282307148,
0.025720911100506783,
0.046370379626750946,
0.010444353334605694,
-0.06548253446817398,
0.08031021803617477,
0.02752712368965149,
0.08877216279506683,
-0.02225794456899166,
0.003006524872034788,
-0.030829371884465218,
0.12080413848161697,
0.038020361214876175,
0.054459426552057266,
-0.03635052219033241,
0.019262203946709633,
0.026089441031217575,
-0.030088480561971664,
-0.036142852157354355,
-0.008139302022755146,
0.0536063052713871,
-0.08574024587869644,
0.051605284214019775,
0.05138886719942093,
0.012377815321087837,
-0.05608449503779411,
0.002331690164282918,
-0.04241796210408211,
-0.05135628581047058,
-0.07154630124568939,
0.07375208288431168,
-0.11553778499364853,
-0.05369812622666359,
-0.07876769453287125,
-0.03866774961352348,
-0.04082396626472473,
0.04683160409331322,
-0.04645726829767227,
-0.07073401659727097,
-1.9710692948007963e-33,
0.01995798945426941,
0.029798073694109917,
-0.006793862208724022,
0.03014199063181877,
-0.003294240450486541,
-0.033025890588760376,
0.004334143362939358,
0.006663926877081394,
-0.02583901211619377,
-0.09702789038419724,
-0.0334557369351387,
0.011278335005044937,
0.019221311435103416,
0.05576309934258461,
-0.0043317824602127075,
0.02260541170835495,
0.008604269474744797,
-0.01877373270690441,
-0.054414261132478714,
-0.039790987968444824,
-0.014330616220831871,
0.002090725814923644,
-0.028061842545866966,
-0.04432171583175659,
-0.08467061817646027,
0.10279130935668945,
-0.06392994523048401,
-0.04018506407737732,
-0.044473886489868164,
-0.0016043753130361438,
-0.05552368611097336,
0.05186708644032478,
-0.008722788654267788,
-0.060567330569028854,
-0.06743891537189484,
0.011717868968844414,
0.07056445628404617,
-0.060726772993803024,
-0.050994694232940674,
0.02111819200217724,
0.0739995464682579,
-0.012156213633716106,
-0.017789166420698166,
0.026174429804086685,
-0.03460272029042244,
0.004631856456398964,
-0.0210138987749815,
0.0021522585302591324,
0.030476896092295647,
0.059790026396512985,
-0.04476495459675789,
0.09333345293998718,
-0.06838330626487732,
-0.0006965622887946665,
0.02217380702495575,
-0.10303141176700592,
-0.07862702012062073,
-0.0020449503790587187,
0.120456762611866,
-0.01014755293726921,
0.00030403025448322296,
-0.01722780615091324,
-0.0756460577249527,
-0.06565658003091812,
0.005454691592603922,
0.010958529077470303,
-0.0021244811359792948,
-0.04800451919436455,
0.003737247781828046,
0.039698220789432526,
0.009266826324164867,
-0.030006548389792442,
-0.043647490441799164,
-0.018672678619623184,
-0.0034821885637938976,
0.004257892724126577,
0.03941265121102333,
-0.011110200546681881,
0.0003752823977265507,
0.03649091720581055,
0.10080518573522568,
-0.03326481580734253,
0.05851256474852562,
0.1306440681219101,
0.07835572957992554,
0.09038800001144409,
-0.018774885684251785,
0.02649347297847271,
0.053192753344774246,
-0.037418119609355927,
-0.0400732047855854,
0.039389509707689285,
-0.0048825968988239765,
0.06525983661413193,
0.02668626420199871,
-3.752418464841867e-8,
-0.0755050778388977,
0.03838053345680237,
-0.006133985240012407,
-0.07842623442411423,
-0.03021506778895855,
-0.1129753589630127,
-0.0064889718778431416,
0.10039030015468597,
0.020675310865044594,
-0.0224218238145113,
0.04261450096964836,
-0.058239877223968506,
-0.0531192310154438,
-0.016060570254921913,
0.09739930182695389,
0.07344412803649902,
0.019288545474410057,
-0.006890204735100269,
0.03760269656777382,
0.021162014454603195,
-0.03018447943031788,
0.003315461566671729,
0.027469122782349586,
0.03678346052765846,
-0.00976236816495657,
-0.007081865333020687,
0.03879760950803757,
0.1445939689874649,
0.10044586658477783,
0.009026997722685337,
0.056358519941568375,
0.005528381559997797,
0.03512740135192871,
-0.05248039588332176,
-0.007504376582801342,
0.027629701420664787,
0.030414193868637085,
0.0010196021758019924,
0.030824068933725357,
-0.010295777581632137,
-0.00895233079791069,
0.01979675702750683,
-0.03135755658149719,
0.06514182686805725,
0.006217215675860643,
0.06371058523654938,
-0.01592288538813591,
-0.09763260930776596,
0.031550124287605286,
0.00898899044841528,
-0.024387376382946968,
0.04279036074876785,
-0.0013133177999407053,
0.13172629475593567,
0.07527343928813934,
-0.04323822259902954,
0.01743137463927269,
0.02243402600288391,
0.0648403912782669,
-0.0031384467147290707,
0.002453461056575179,
0.025436963886022568,
-0.05362117663025856,
-0.012755519710481167
] |
Identification of Intermittent Temporal Latent Process | https://openreview.net/forum?id=6Pz7afmsOp | [
"Yuke Li",
"Yujia Zheng",
"Guangyi Chen",
"Kun Zhang",
"Heng Huang"
] | Poster | Identifying time-delayed latent causal process is crucial for understanding temporal dynamics and enabling downstream reasoning. While recent methods have made progress in identifying latent time-delayed causal processes, they cannot address the dynamics in which the influence of some latent factors on both the subsequent latent states and the observed data can become inactive or irrelevant at different time steps. Therefore, we introduce intermittent temporal latent processes, where: (1) any subset of latent factors may be missing during nonlinear data generation at any time step, and (2) the active latent factors at each step are unknown. This framework encompasses both nonstationary and stationary transitions, accommodating changing or consistent active factors over time.
Our work shows that under certain assumptions, the latent causal variables are block-wise identifiable. With further conditional independence assumption, each latent variable can even be recovered up to component-wise transformations.
Using this identification theory, we propose an unsupervised approach, InterLatent, to reliably uncover the representations of the intermittent temporal latent process. The experimental findings on both synthetic and real-world datasets verify our theoretical claims. | unsupervised representation learning | null | 521 | null | [
-0.10915150493383408,
-0.09619130194187164,
0.0060611991211771965,
0.08160693198442459,
0.1381988227367401,
0.014515058137476444,
-0.00581235671415925,
-0.08737292885780334,
0.0793958380818367,
-0.004157652147114277,
-0.010493490844964981,
-0.014071629382669926,
-0.009146877564489841,
0.06232576444745064,
-0.030085762962698936,
-0.01690409705042839,
-0.00656017055734992,
0.001536394702270627,
-0.06965623050928116,
-0.06226406618952751,
0.014491918496787548,
0.007402710150927305,
-0.026031190529465675,
0.06287466734647751,
0.05930018797516823,
0.03650735691189766,
0.05796276405453682,
-0.00671409023925662,
0.0005705278599634767,
-0.0012232887092977762,
0.03428700193762779,
0.08475978672504425,
-0.0947689637541771,
-0.06066171079874039,
0.008286934345960617,
0.0626932829618454,
0.03193221613764763,
0.07272684574127197,
-0.06756297498941422,
0.007108598481863737,
0.06995807588100433,
-0.0550747811794281,
-0.06629959493875504,
0.0679171159863472,
-0.014474894851446152,
-0.013079253025352955,
0.0065607656724750996,
-0.03826425224542618,
-0.10668911784887314,
0.024137983098626137,
-0.062340155243873596,
0.08399562537670135,
0.0067956717684865,
0.055336177349090576,
0.053536977618932724,
0.03466873988509178,
0.07648574560880661,
0.038285307586193085,
0.017951054498553276,
0.07086336612701416,
-0.04973124712705612,
-0.011172128841280937,
-0.0005793377058580518,
-0.04461246728897095,
0.03910273686051369,
0.08330083638429642,
-0.04339158162474632,
0.05261329934000969,
0.018864726647734642,
0.026541590690612793,
-0.025320807471871376,
0.04607339948415756,
-0.11515986174345016,
-0.013383443467319012,
0.03772608935832977,
-0.03615475818514824,
0.012346808798611164,
-0.006780907511711121,
0.029642866924405098,
-0.05143585801124573,
-0.0023913392797112465,
0.06853719800710678,
-0.10612000524997711,
0.02183545008301735,
0.06543514877557755,
0.018615901470184326,
0.07542823255062103,
0.08103232830762863,
0.0025652337353676558,
0.023391298949718475,
0.019223781302571297,
-0.03634585812687874,
0.06903351843357086,
0.022492993623018265,
0.1184835135936737,
0.002851298777386546,
0.04023580998182297,
-0.01116994395852089,
0.1338084638118744,
0.0662948414683342,
-0.018998825922608376,
0.0501214899122715,
0.026390651240944862,
0.10066638141870499,
0.04607584327459335,
-0.0933976024389267,
0.05964377894997597,
0.05927430838346481,
0.005445194896310568,
-0.030926721170544624,
-0.05415065586566925,
-0.028640663251280785,
0.0542418546974659,
0.011208160780370235,
0.0020688313525170088,
-0.0579574890434742,
0.010401581414043903,
-0.04413691535592079,
-0.011152323335409164,
-0.0370652861893177,
0.026579370722174644,
0.004200961906462908,
0.05520571768283844,
-0.02800898067653179,
0.0715160071849823,
0.013296214863657951,
-0.051709264516830444,
4.010972232707884e-33,
0.025978699326515198,
-0.04473131150007248,
0.01178552396595478,
-0.026130862534046173,
-0.04241570457816124,
-0.052778683602809906,
-0.056818507611751556,
-0.054228026419878006,
0.05112774670124054,
0.035458918660879135,
-0.02062917873263359,
0.025528261438012123,
-0.11129771173000336,
0.031147966161370277,
-0.003927501384168863,
-0.011107342317700386,
0.0034227657597512007,
0.06468567252159119,
0.06216448172926903,
-0.007017043419182301,
0.022271325811743736,
-0.008362027816474438,
-0.04689175263047218,
-0.00939171388745308,
0.011007965542376041,
-0.0558546744287014,
0.0023761678021401167,
-0.007658245973289013,
-0.032918497920036316,
0.003011573338881135,
-0.06790893524885178,
0.058452535420656204,
-0.01757808029651642,
0.015857955440878868,
-0.004330710973590612,
-0.07118513435125351,
-0.015896271914243698,
-0.12315904349088669,
0.010191354900598526,
-0.05595092102885246,
0.046496447175741196,
0.01039755530655384,
-0.016655581071972847,
-0.0036021703854203224,
-0.1026398316025734,
-0.10536137223243713,
0.044257696717977524,
-0.0017546549206599593,
-0.019379466772079468,
0.031695716083049774,
0.024631401523947716,
0.12433364242315292,
0.09287132322788239,
-0.16047927737236023,
-0.08179336786270142,
0.018501102924346924,
-0.008601579815149307,
-0.0190677959471941,
-0.003867354243993759,
0.022564399987459183,
-0.03380050137639046,
0.019124167039990425,
-0.02064645104110241,
0.0100374361500144,
-0.020109351724386215,
0.020163778215646744,
-0.04003233090043068,
0.022234462201595306,
0.0655449852347374,
0.020733769983053207,
0.009239595383405685,
0.006983001716434956,
-0.012908803299069405,
-0.029993930831551552,
0.06920284032821655,
-0.05218743905425072,
-0.04316658154129982,
-0.05655674636363983,
-0.02484554797410965,
-0.01075583416968584,
-0.01586017943918705,
-0.07664996385574341,
-0.007172805722802877,
0.027730518952012062,
-0.04315932095050812,
-0.004283376038074493,
0.02289002574980259,
-0.07108920812606812,
-0.10195058584213257,
-0.025515642017126083,
-0.048631198704242706,
-0.03790416195988655,
0.03444399684667587,
0.04964599013328552,
0.0013569716829806566,
-5.2728136230957767e-33,
-0.03569631651043892,
-0.05316925048828125,
0.01561502180993557,
-0.012344684451818466,
0.05444740876555443,
-0.010359209030866623,
-0.04877729341387749,
0.005905816797167063,
0.05021372437477112,
-0.030768800526857376,
0.05518486350774765,
-0.07328598201274872,
-0.0220992062240839,
0.02597114071249962,
0.024090943858027458,
0.06271356344223022,
0.06931518763303757,
0.006949150934815407,
0.08007537573575974,
-0.006566215306520462,
0.015392567031085491,
0.015077860094606876,
-0.052308276295661926,
-0.029656438156962395,
0.007360695395618677,
0.027734098955988884,
0.03236783295869827,
0.07094001770019531,
-0.0793292373418808,
-0.004141682293266058,
-0.05509182810783386,
0.005577176809310913,
-0.05611749738454819,
0.013520270586013794,
-0.046023812144994736,
0.04228992387652397,
0.008387171663343906,
-0.02769619971513748,
-0.06187872216105461,
-0.036659739911556244,
0.09099366515874863,
0.01869085058569908,
-0.007262325845658779,
0.009960110299289227,
-0.0641651600599289,
-0.003067876910790801,
-0.0003573692520149052,
0.14448384940624237,
0.0036446687299758196,
0.038984473794698715,
0.03176592290401459,
0.0883854329586029,
0.028077606111764908,
0.007398215588182211,
-0.0023566707968711853,
0.0057299756444990635,
0.09734845161437988,
-0.07074689120054245,
0.020527344197034836,
0.04303983598947525,
-0.043674763292074203,
-0.011622344143688679,
0.0435115247964859,
-0.06121540069580078,
0.002396225929260254,
-0.08705654740333557,
0.008283124305307865,
-0.07473460584878922,
0.053360093384981155,
-0.023463154211640358,
0.10921257734298706,
0.044854097068309784,
-0.08391668647527695,
-0.006491968408226967,
0.06294695287942886,
-0.022345975041389465,
-0.03665979579091072,
0.0046461778692901134,
-0.05141108110547066,
-0.03153151273727417,
0.01653537154197693,
-0.0010010629193857312,
0.050321754068136215,
-0.010342954657971859,
-0.15024566650390625,
0.046840354800224304,
0.03629829362034798,
-0.0028973217122256756,
0.06298800557851791,
-0.00337339099496603,
-0.0455334447324276,
-0.04087599366903305,
-0.13920938968658447,
0.09676025807857513,
-0.07318040728569031,
-4.219428362262079e-8,
-0.011218021623790264,
0.04873077571392059,
-0.022147631272673607,
-0.031848158687353134,
0.06290484219789505,
-0.061907071620225906,
0.007251459173858166,
-0.003301881719380617,
0.011448193341493607,
-0.05107234790921211,
0.15877826511859894,
-0.05387724190950394,
0.058111511170864105,
-0.08375968039035797,
0.015770135447382927,
-0.007415290456265211,
0.05317949876189232,
-0.014403614215552807,
-0.04814835265278816,
-0.01311886589974165,
0.03514489904046059,
0.02719927206635475,
-0.05698472261428833,
0.009098735637962818,
0.03283457085490227,
-0.05235246196389198,
-0.050339117646217346,
0.08515377342700958,
0.045177191495895386,
-0.015811607241630554,
-0.003943241201341152,
0.049101028591394424,
0.0002708967658691108,
-0.014737259596586227,
-0.002829412929713726,
0.0576871782541275,
0.058105163276195526,
-0.06725145131349564,
-0.08311479538679123,
-0.016763171181082726,
-0.029234737157821655,
0.010752180591225624,
-0.09539566934108734,
-0.0010388283990323544,
0.015341488644480705,
0.016219761222600937,
-0.08903505653142929,
-0.07853344082832336,
-0.007263546343892813,
0.005908890627324581,
-0.007749456912279129,
-0.01711888797581196,
0.031588632613420486,
0.0708756223320961,
0.10727322101593018,
0.014571748673915863,
0.05487809330224991,
-0.027969053015112877,
0.026319559663534164,
-0.03158465400338173,
0.030423829331994057,
0.023851072415709496,
-0.0050069959834218025,
0.018544286489486694
] |
Meissonic: Revitalizing Masked Generative Transformers for Efficient High-Resolution Text-to-Image Synthesis | https://openreview.net/forum?id=GJsuYHhAga | [
"Jinbin Bai",
"Tian Ye",
"Wei Chow",
"Enxin Song",
"Qing-Guo Chen",
"Xiangtai Li",
"Zhen Dong",
"Lei Zhu",
"Shuicheng YAN"
] | Poster | We present Meissonic, which elevates non-autoregressive text-to-image Masked Image Modeling (MIM) to a level comparable with state-of-the-art diffusion models like SDXL. By incorporating a comprehensive suite of architectural innovations, advanced positional encoding strategies, and optimized sampling conditions, Meissonic substantially improves MIM's performance and efficiency. Additionally, we leverage high-quality training data, integrate micro-conditions informed by human preference scores, and employ feature compression layers to further enhance image fidelity and resolution. Our model not only matches but often exceeds the performance of existing methods in generating high-quality, high-resolution images. Extensive experiments validate Meissonic’s capabilities, demonstrating its potential as a new standard in text-to-image synthesis. | Text-to-Image Synthesis, Masked Generative Transformers | We present Meissonic, a next generation text-to-image foundation model with non-autogressive masked image modeling method. | 507 | 2410.08261 | [
-0.02293975278735161,
-0.0023621676955372095,
0.03281985595822334,
0.06063271313905716,
0.02580075152218342,
-0.051075126975774765,
-0.06316331773996353,
-0.020599517971277237,
0.08614351600408554,
-0.042552705854177475,
-0.012179295532405376,
0.012634754180908203,
0.07290612161159515,
0.05401476472616196,
-0.022339964285492897,
0.025884805247187614,
0.04034165292978287,
0.0657413974404335,
-0.06115366891026497,
-0.04329517111182213,
0.07700932025909424,
-0.03988812118768692,
0.04446006566286087,
-0.08447153866291046,
0.08442336320877075,
0.029965557157993317,
-0.037359412759542465,
-0.06875527650117874,
0.005266451742500067,
-0.03543686866760254,
-0.009606597013771534,
0.03583674132823944,
0.0182240828871727,
-0.020867403596639633,
-0.0022070708218961954,
0.024961236864328384,
-0.019814595580101013,
0.022816691547632217,
0.004222034942358732,
-0.03424059599637985,
0.04379758611321449,
0.03713188320398331,
0.03924385458230972,
-0.009278003126382828,
0.023761563003063202,
-0.029966438189148903,
-0.007570227608084679,
-0.010915827937424183,
0.019881408661603928,
-0.0017674807459115982,
-0.044377248734235764,
-0.01451036985963583,
-0.05794897302985191,
0.06009859964251518,
-0.04672820493578911,
0.0647604912519455,
-0.0075814612209796906,
-0.015497218817472458,
0.02607611007988453,
0.0456869974732399,
-0.07552867382764816,
0.003968203440308571,
0.0018373921047896147,
0.013059662654995918,
-0.04642218351364136,
-0.018060415983200073,
0.019147837534546852,
0.0027014503721147776,
0.014859004877507687,
-0.03588220477104187,
-0.01785610243678093,
-0.032305795699357986,
-0.051365360617637634,
0.012898755259811878,
0.04244130849838257,
0.06308427453041077,
0.027275724336504936,
-0.07664336264133453,
0.08380809426307678,
-0.02830052562057972,
0.05052345618605614,
-0.013997114263474941,
0.0952395498752594,
0.023965641856193542,
0.037003204226493835,
0.049911513924598694,
0.03792504966259003,
0.07371608912944794,
-0.03406994789838791,
0.006229381542652845,
-0.08229901641607285,
-0.018557878211140633,
0.004315282683819532,
-0.0038787852972745895,
-0.03144355118274689,
0.041497159749269485,
0.0022784308530390263,
-0.04298156872391701,
-0.03173408657312393,
0.07896292209625244,
0.037942443042993546,
-0.05632186308503151,
-0.02063443511724472,
-0.008024001494050026,
-0.002228580880910158,
-0.06576637923717499,
0.11410792171955109,
0.0556938573718071,
-0.013265433721244335,
-0.04241582751274109,
0.08041615784168243,
0.0012299607042223215,
-0.03324282169342041,
-0.03180374205112457,
0.03767402470111847,
-0.03063053824007511,
0.028518885374069214,
-0.0019893343560397625,
0.010280321352183819,
0.003137306310236454,
-0.039858996868133545,
-0.004054021555930376,
-0.1147550418972969,
0.029964933171868324,
-0.009590284898877144,
-0.012701363302767277,
-0.00961750466376543,
1.333982488888314e-33,
-0.047455091029405594,
0.049305640161037445,
0.027134113013744354,
0.059766609221696854,
0.05861034616827965,
-0.03440258651971817,
0.02154468558728695,
-0.04798441007733345,
-0.02459549717605114,
-0.023023804649710655,
-0.037076566368341446,
-0.08349676430225372,
-0.0973699688911438,
0.156995952129364,
-0.0038667565677314997,
-0.05766315013170242,
0.008880152367055416,
0.03081442601978779,
0.018574567511677742,
0.0027460725978016853,
0.04244648665189743,
0.031860463321208954,
0.02630036137998104,
-0.05175416171550751,
-0.03738442808389664,
-0.0023023770190775394,
0.02036353200674057,
-0.11128351837396622,
0.028231164440512657,
-0.018758824095129967,
-0.13545161485671997,
0.0010092529701068997,
0.09410329163074493,
0.032636359333992004,
-0.0021650418639183044,
0.027090182527899742,
0.019021226093173027,
-0.054244380444288254,
-0.005761986132711172,
0.09490995854139328,
-0.003461069893091917,
0.08848755806684494,
-0.026188595220446587,
-0.061095770448446274,
-0.0772281140089035,
0.06028563529253006,
0.04003814235329628,
0.03229295834898949,
-0.044343430548906326,
-0.05454499274492264,
0.06186409294605255,
0.05781755596399307,
-0.031216930598020554,
-0.00902492180466652,
-0.00663950527086854,
-0.055089373141527176,
0.11070192605257034,
-0.059147510677576065,
0.08173483610153198,
-0.012381783686578274,
0.007572048809379339,
-0.018428783863782883,
-0.03355763480067253,
0.07891029119491577,
0.05340877175331116,
0.0023319863248616457,
0.01051341649144888,
-0.0835019052028656,
-0.07678058743476868,
-0.03893822059035301,
-0.002394636394456029,
-0.010043669492006302,
-0.03354142606258392,
-0.05142426863312721,
0.03160108998417854,
0.008689656853675842,
0.013268642127513885,
-0.02879285253584385,
0.01736648753285408,
0.02593853510916233,
-0.05146002024412155,
-0.029686661437153816,
0.02098383568227291,
-0.10311471670866013,
-0.009322965517640114,
-0.0066854627802968025,
-0.001691248849965632,
-0.06503333151340485,
0.026777589693665504,
-0.029767585918307304,
-0.008147264830768108,
-0.009540095925331116,
0.029486769810318947,
0.024795284494757652,
0.0067557464353740215,
4.65635949810511e-34,
-0.04770195111632347,
0.07224882394075394,
-0.03470497205853462,
0.041419316083192825,
-0.053007837384939194,
-0.004283340647816658,
0.022402189671993256,
0.01804981380701065,
0.05367898941040039,
-0.08216268569231033,
-0.026791971176862717,
0.007687138393521309,
-0.030804714187979698,
-0.06846129894256592,
-0.02574031613767147,
-0.07261199504137039,
0.14438678324222565,
0.014132311567664146,
-0.08954352140426636,
-0.004677467048168182,
-0.033730193972587585,
0.05953739956021309,
0.008651547133922577,
0.03612067550420761,
-0.0249791219830513,
0.07265186309814453,
-0.006647799164056778,
0.043385446071624756,
0.033792153000831604,
-0.036917634308338165,
-0.03842635080218315,
0.007019907236099243,
0.0846642255783081,
-0.035155970603227615,
-0.05345341935753822,
0.09122694283723831,
-0.013130578212440014,
0.04657397419214249,
-0.0373852476477623,
0.050951339304447174,
0.02409169264137745,
-0.0703732967376709,
-0.08491867035627365,
0.0671527087688446,
-0.04823208600282669,
-0.025625688955187798,
-0.042454496026039124,
-0.06435553729534149,
0.11600496619939804,
0.03741185739636421,
-0.002470406237989664,
0.06181414797902107,
-0.10332892835140228,
-0.005953706800937653,
-0.02568950317800045,
-0.10094665735960007,
-0.0941784605383873,
-0.07488876581192017,
0.11378884315490723,
-0.06781168282032013,
-0.03237375244498253,
-0.06301465630531311,
0.011188452132046223,
-0.1306421160697937,
-0.01269470527768135,
0.024061543866991997,
-0.04038780927658081,
-0.0029868469573557377,
-0.021768195554614067,
0.0943792387843132,
0.04169108346104622,
-0.02821457013487816,
0.017633508890867233,
0.06595093011856079,
0.02545415610074997,
-0.096341572701931,
0.024030735716223717,
0.030808957293629646,
-0.004494804888963699,
-0.05038457736372948,
0.020760435611009598,
0.017222806811332703,
-0.051114194095134735,
0.10123099386692047,
0.11394063383340836,
0.07470041513442993,
0.0018018049886450171,
-0.0050685349851846695,
0.014688430353999138,
0.09371785819530487,
-0.022202758118510246,
0.12276652455329895,
0.009036367759108543,
0.03765088692307472,
0.019152769818902016,
-4.0677388568610695e-8,
0.0029690652154386044,
0.03206890448927879,
-0.0752335637807846,
0.0354640819132328,
-0.06089016795158386,
-0.08992432802915573,
0.039587073028087616,
0.016347315162420273,
0.058859411627054214,
-0.023825760930776596,
0.07154693454504013,
-0.06797125935554504,
0.012567795813083649,
-0.0698792040348053,
0.024695666506886482,
0.0490197092294693,
0.06862585991621017,
0.06229546293616295,
0.037303753197193146,
-0.049138884991407394,
-0.08390351384878159,
0.07747643440961838,
-0.028372416272759438,
-0.0924304723739624,
0.06801754236221313,
-0.000884730601683259,
-0.0010475398739799857,
0.030509354546666145,
0.07123815268278122,
0.03550022095441818,
0.010561817325651646,
0.026231061667203903,
-0.025483090430498123,
0.03173722326755524,
-0.008890563622117043,
0.06027025356888771,
0.07954632490873337,
0.014560257084667683,
-0.09775924682617188,
-0.10200832039117813,
0.020557915791869164,
0.06852250546216965,
-0.0240858756005764,
0.010584743693470955,
0.00015512507525272667,
-0.03174975514411926,
0.10914259403944016,
-0.11310583353042603,
0.031333424150943756,
0.00841308943927288,
0.02691805176436901,
-0.05808107554912567,
-0.013144575990736485,
0.050394102931022644,
0.08180076628923416,
-0.06695272028446198,
0.07858853042125702,
0.024889525026082993,
0.041845038533210754,
0.09914471209049225,
0.045409075915813446,
0.019164694473147392,
0.00001547524516354315,
-0.010081782937049866
] |
AHA: A Vision-Language-Model for Detecting and Reasoning Over Failures in Robotic Manipulation | https://openreview.net/forum?id=JVkdSi7Ekg | [
"Jiafei Duan",
"Wilbert Pumacay",
"Nishanth Kumar",
"Yi Ru Wang",
"Shulin Tian",
"Wentao Yuan",
"Ranjay Krishna",
"Dieter Fox",
"Ajay Mandlekar",
"Yijie Guo"
] | Poster | Robotic manipulation in open-world settings requires not only task execution but also the ability to detect and learn from failures. While recent advances in vision-language models (VLMs) and large language models (LLMs) have improved robots' spatial reasoning and problem-solving abilities, they still struggle with failure recognition, limiting their real-world applicability. We introduce AHA, an open-source VLM designed to detect and reason about failures in robotic manipulation using natural language. By framing failure detection as a free-form reasoning task, AHA identifies failures and provides detailed, adaptable explanations across different robots, tasks, and environments. We fine-tuned AHA using FailGen, a scalable framework that generates the first large-scale dataset of robotic failure trajectories, the AHA dataset. FailGen achieves this by procedurally perturbing successful demonstrations from simulation. Despite being trained solely on the AHA dataset, AHA generalizes effectively to real-world failure datasets, robotic systems, and unseen tasks. It surpasses the second-best model (GPT-4o in-context learning) by 10.3% and exceeds the average performance of six compared models including five state-of-the-art VLMs by 35.3% across multiple metrics and datasets. We integrate AHA into three manipulation frameworks that utilize LLMs/VLMs for reinforcement learning, task and motion planning, and zero-shot trajectory generation. AHA’s failure feedback enhances these policies' performances by refining dense reward functions, optimizing task planning, and improving sub-task verification, boosting task success rates by an average of 21.4% across all three tasks compared to GPT-4 models. Project page: https://aha-vlm.github.io | Robotic Manipulation; Data Generation; Vision-Language-Model; Failure Reasoning; Failure Detection | A vision-language model for detecting and reasoning about failures in robotic manipulation, which can be used to improve many downstream robotic applications. | 506 | 2410.00371 | [
0.006365049630403519,
-0.09636519104242325,
0.06451303511857986,
0.01503585372120142,
0.13740257918834686,
-0.014384706504642963,
0.02535756677389145,
-0.00585470674559474,
-0.015074452385306358,
0.009025773964822292,
-0.013365217484533787,
-0.07410240173339844,
0.036246269941329956,
0.08646736294031143,
-0.005175537429749966,
0.020079005509614944,
0.09442735463380814,
0.04311996325850487,
-0.10846129059791565,
-0.00038385207881219685,
0.02912234701216221,
0.0649283155798912,
0.06800757348537445,
0.042564183473587036,
-0.04990510642528534,
0.03319409862160683,
0.021573541685938835,
-0.07986771315336227,
-0.05197209119796753,
-0.05653134360909462,
0.010089271701872349,
-0.048821158707141876,
-0.059918198734521866,
0.025999361649155617,
0.03211105242371559,
0.10577958077192307,
-0.057487666606903076,
0.01869751326739788,
-0.05067959800362587,
-0.11787789314985275,
-0.07454493641853333,
0.0009728866862133145,
0.07650776207447052,
0.011770369485020638,
0.04223785176873207,
0.0559086799621582,
0.02445163019001484,
-0.0155034763738513,
0.12297190725803375,
0.043996166437864304,
-0.11794081330299377,
-0.03190786764025688,
0.08557727187871933,
-0.027015123516321182,
-0.011568326503038406,
0.016367660835385323,
0.015356642194092274,
-0.01082087866961956,
0.039400823414325714,
0.0077628763392567635,
0.05760136991739273,
-0.057463619858026505,
-0.0009752719197422266,
-0.017117643728852272,
-0.051427483558654785,
0.003399259177967906,
0.03357536718249321,
-0.044801171869039536,
-0.05022323504090309,
0.026967836543917656,
-0.017118876799941063,
0.042050790041685104,
-0.03360548987984657,
0.04205765202641487,
-0.022960280999541283,
0.029684867709875107,
-0.021264487877488136,
-0.0819089412689209,
0.021897833794355392,
-0.0029800839256495237,
0.0231357179582119,
-0.04522435739636421,
-0.036132071167230606,
0.07227715849876404,
0.1218404769897461,
0.005166196264326572,
0.027288127690553665,
0.06373883038759232,
0.052335064858198166,
0.007760040462017059,
-0.01704745925962925,
-0.07703085988759995,
0.036390822380781174,
-0.0108187859877944,
0.0859416127204895,
-0.009288354776799679,
0.0037574481684714556,
-0.09316414594650269,
-0.02365824207663536,
0.03558781370520592,
-0.018491076305508614,
0.011436638422310352,
0.0090026268735528,
-0.023913675919175148,
0.0385357141494751,
-0.009032697416841984,
0.0030755307525396347,
-0.010834118351340294,
0.04969397559762001,
-0.04421624913811684,
0.034886717796325684,
-0.06316738575696945,
0.06994848698377609,
0.0400417223572731,
0.030032265931367874,
-0.021301131695508957,
-0.0286552794277668,
-0.03270105645060539,
-0.07394315302371979,
0.014511801302433014,
-0.015366032719612122,
0.015834804624319077,
0.031343974173069,
0.026307718828320503,
0.07230744510889053,
-0.03996257483959198,
-0.004800613038241863,
-1.0010643971426991e-33,
0.12016116082668304,
0.009237687103450298,
0.05972147732973099,
0.022468356415629387,
0.07333815842866898,
-0.020400142297148705,
-0.01824202947318554,
0.026737947016954422,
-0.015143731608986855,
-0.01085574273020029,
-0.007051910739392042,
-0.028084054589271545,
-0.07657109200954437,
0.013103866949677467,
0.02422224171459675,
0.039539143443107605,
0.02237364649772644,
0.0057170274667441845,
-0.10927507281303406,
0.046392783522605896,
-0.03522435948252678,
-0.03694947063922882,
-0.0032820894848555326,
-0.003822955535724759,
-0.016285495832562447,
0.028300179168581963,
0.026492616161704063,
-0.0028050176333636045,
0.016066284850239754,
-0.03939869627356529,
-0.11008916795253754,
-0.018380766734480858,
-0.032098568975925446,
0.10457754135131836,
-0.0026218576822429895,
-0.05257164314389229,
-0.048306096345186234,
-0.05474618822336197,
0.020551685243844986,
-0.037977710366249084,
-0.044713813811540604,
0.03927016630768776,
-0.003998076543211937,
-0.010492452420294285,
-0.036654695868492126,
-0.05910996347665787,
0.046806782484054565,
0.03693190589547157,
-0.03456326201558113,
-0.007253632880747318,
0.027783237397670746,
0.08169017732143402,
-0.008028618060052395,
-0.038668837398290634,
0.015149855054914951,
-0.01693541184067726,
0.0626186653971672,
0.060023777186870575,
0.03748762607574463,
-0.023501677438616753,
0.04786529764533043,
-0.010912287048995495,
-0.038002580404281616,
0.11742395162582397,
-0.00796095933765173,
-0.021644243970513344,
-0.012637095525860786,
0.06829221546649933,
0.07433893531560898,
0.030120469629764557,
-0.02236851118505001,
-0.01339032780379057,
0.006875777151435614,
-0.006183919031172991,
0.030346063897013664,
0.034241218119859695,
0.026713641360402107,
-0.09253515303134918,
0.07720177620649338,
0.03253884240984917,
0.019762298092246056,
-0.030549483373761177,
-0.03060825914144516,
0.01257944293320179,
-0.05890369787812233,
0.003631359664723277,
0.0559130422770977,
-0.11148373782634735,
-0.05302587151527405,
-0.04229072108864784,
-0.0718972459435463,
-0.13354457914829254,
-0.07500486075878143,
-0.013229858130216599,
-0.0016193152405321598,
-2.567602234376131e-34,
-0.02678140439093113,
-0.04579528048634529,
0.0440584197640419,
0.05827277526259422,
-0.08436886966228485,
-0.10657189786434174,
-0.024863600730895996,
0.01945088617503643,
0.030321337282657623,
-0.04215063154697418,
-0.04115188494324684,
-0.04291679337620735,
-0.019175484776496887,
0.05602540448307991,
0.014106035232543945,
-0.09440219402313232,
0.012935551814734936,
-0.03316623345017433,
0.0031824312172830105,
0.029032442718744278,
0.01849624328315258,
0.1332101672887802,
-0.07542645186185837,
-0.01387197058647871,
-0.06555119901895523,
0.060549914836883545,
0.01655864715576172,
-0.010875758714973927,
-0.036888204514980316,
0.023624978959560394,
0.02363552898168564,
-0.021545806899666786,
0.02106844075024128,
0.05307060852646828,
-0.031619131565093994,
0.043549880385398865,
0.004968713968992233,
-0.01726987585425377,
-0.0423184297978878,
0.027514640241861343,
0.07840664684772491,
-0.05068107694387436,
-0.04861665889620781,
0.036842864006757736,
-0.04130641743540764,
-0.005981389433145523,
-0.028606409206986427,
0.023996111005544662,
-0.04806265980005264,
0.015857836231589317,
-0.07127844542264938,
-0.00037655391497537494,
-0.07109121233224869,
-0.014648948796093464,
-0.0008389939903281629,
-0.024544905871152878,
-0.014507698826491833,
-0.018661191686987877,
0.018208937719464302,
-0.04652596637606621,
-0.09945477545261383,
0.012750992551445961,
0.04686897248029709,
0.07647201418876648,
-0.04617131128907204,
0.05784019082784653,
-0.0776086300611496,
0.03524181246757507,
0.038258183747529984,
0.04310591518878937,
0.12120987474918365,
0.029390936717391014,
0.021224213764071465,
-0.005247562658041716,
0.036712367087602615,
0.06028106063604355,
-0.04936212673783302,
-0.04144422337412834,
0.014074290171265602,
-0.04976842179894447,
0.03431355208158493,
-0.07617727667093277,
0.08031859248876572,
0.15013450384140015,
0.016396570950746536,
0.0807691216468811,
0.018218055367469788,
0.03623756021261215,
0.014600488357245922,
0.043393395841121674,
-0.07246088981628418,
0.028968358412384987,
-0.024856090545654297,
0.11577123403549194,
-0.07173459976911545,
-3.746361798562248e-8,
-0.12469460815191269,
0.07649118453264236,
0.02230590209364891,
-0.060998525470495224,
0.02274615690112114,
0.008427033200860023,
-0.016517549753189087,
0.05033440887928009,
-0.029282663017511368,
0.015825526788830757,
0.02368706278502941,
-0.04734356328845024,
-0.042798642069101334,
-0.020655060186982155,
-0.03425304591655731,
0.06724555790424347,
0.05972699448466301,
0.022433029487729073,
-0.008649945259094238,
0.021723784506320953,
0.01757487840950489,
-0.012970085255801678,
0.06275148689746857,
0.004096563905477524,
-0.03743865713477135,
-0.047906938940286636,
-0.10138362646102905,
0.08875436335802078,
0.02284093014895916,
-0.03290468081831932,
0.0423966683447361,
-0.044129058718681335,
0.08183278888463974,
0.005884191952645779,
0.01986066997051239,
0.049669988453388214,
-0.007714244537055492,
-0.05241427943110466,
-0.0054709031246602535,
-0.05470449477434158,
0.04198873043060303,
0.09893225878477097,
-0.015463919378817081,
0.033062636852264404,
-0.02298147603869438,
0.06287434697151184,
-0.03690385818481445,
-0.1831761747598648,
-0.014120285399258137,
-0.04256187379360199,
-0.08868465572595596,
-0.015878397971391678,
-0.06747012585401535,
0.13021664321422577,
0.021686090156435966,
0.03203742578625679,
0.008516300469636917,
-0.016976753249764442,
-0.006926349829882383,
0.0958486795425415,
-0.02159896306693554,
0.06809815764427185,
0.0065793609246611595,
0.027390317991375923
] |
Shapley-Guided Utility Learning for Effective Graph Inference Data Valuation | https://openreview.net/forum?id=8X74NZpARg | [
"Hongliang Chi",
"Qiong Wu",
"Zhengyi Zhou",
"Yao Ma"
] | Poster | Graph Neural Networks (GNNs) have demonstrated remarkable performance in various graph-based machine learning tasks, yet evaluating the importance of neighbors of testing nodes remains largely unexplored due to the challenge of assessing data importance without test labels. To address this gap, we propose Shapley-Guided Utility Learning (SGUL), a novel framework for graph inference data valuation. SGUL innovatively combines transferable data-specific and modelspecific features to approximate test accuracy without relying on ground truth labels. By incorporating Shapley values as a preprocessing step and using feature Shapley values as input, our method enables direct optimization of Shapley value prediction while reducing computational demands. SGUL overcomes key limitations of existing methods, including poor generalization to unseen test-time structures and indirect optimization. Experiments on diverse graph datasets demonstrate that SGUL consistently outperforms existing baselines in both inductive and transductive settings. SGUL offers an effective, efficient, and interpretable approach for quantifying the value of test-time neighbors. | Graph Learning, Data Valuation, Graph Neural Networks, Data-centric AI | null | 505 | 2503.18195 | [
-0.025866327807307243,
-0.006751636974513531,
0.01106105837970972,
0.05142531916499138,
0.03380893915891647,
-0.004523727111518383,
-0.06816749274730682,
-0.09411972016096115,
-0.07385436445474625,
0.012472067028284073,
-0.0054073394276201725,
-0.052755385637283325,
0.03652000427246094,
0.04561129957437515,
0.0208602212369442,
0.03136783838272095,
0.13478311896324158,
0.056719712913036346,
-0.08378905802965164,
-0.05838839337229729,
-0.010243605822324753,
0.01399354636669159,
0.007760804612189531,
-0.009512039832770824,
0.09555643051862717,
-0.05275757610797882,
-0.007628765422850847,
-0.002909208182245493,
0.029894143342971802,
-0.00550132105126977,
0.06049805134534836,
-0.03383044898509979,
-0.07917189598083496,
-0.028459571301937103,
0.013958427123725414,
0.049191609025001526,
-0.0069108703173696995,
-0.007539737969636917,
0.03317414969205856,
0.019915327429771423,
0.027231767773628235,
-0.05612470209598541,
-0.03715914487838745,
0.020984526723623276,
0.03349660709500313,
0.062172938138246536,
-0.007718922570347786,
-0.01880659908056259,
0.002354659605771303,
0.003603184362873435,
-0.0038131517358124256,
-0.024323701858520508,
-0.03005046397447586,
-0.0092067401856184,
-0.02094893343746662,
0.016380636021494865,
0.000034227392461616546,
-0.04030856117606163,
0.03359859064221382,
0.005827457644045353,
-0.05280313640832901,
-0.12712909281253815,
-0.032972343266010284,
-0.0006126343505457044,
0.006291789934039116,
0.003754593897610903,
0.007135295309126377,
0.009843613021075726,
-0.10113799571990967,
0.03961976245045662,
0.05404174327850342,
0.056624885648489,
-0.07648301869630814,
0.05326002091169357,
0.040297236293554306,
0.15252168476581573,
0.06357791274785995,
0.013895072974264622,
0.02205517329275608,
-0.048609659075737,
-0.05696873366832733,
0.018801897764205933,
-0.03163878247141838,
0.0657932385802269,
0.05090934783220291,
-0.0009031795198097825,
0.044054433703422546,
0.004314491990953684,
-0.057749394327402115,
0.057268206030130386,
-0.016782671213150024,
0.010700802318751812,
-0.03987329080700874,
-0.010262545198202133,
0.0006223040982149541,
0.04526780918240547,
-0.008633829653263092,
-0.10516560822725296,
-0.05416721850633621,
0.07829121500253677,
-0.06814548373222351,
0.04908835515379906,
0.08075307309627533,
-0.09954120218753815,
0.0647486075758934,
0.001495958655141294,
-0.009487006813287735,
0.006340119522064924,
0.0074157994240522385,
-0.061270516365766525,
0.027517329901456833,
0.08696842938661575,
-0.009768513962626457,
-0.007933538407087326,
0.020140625536441803,
0.003121513407677412,
-0.03063667006790638,
-0.05665071681141853,
0.014220742508769035,
0.012438280507922173,
-0.01952628791332245,
0.04573877900838852,
0.04223354905843735,
-0.01600872538983822,
-0.06459339708089828,
-0.04780546948313713,
-0.06777729094028473,
4.548435433210697e-33,
0.06897369772195816,
0.028168393298983574,
0.07723288238048553,
-0.08274204283952713,
0.04676226153969765,
-0.00816489476710558,
-0.10504758358001709,
0.08726680278778076,
-0.05143977701663971,
0.010002264752984047,
-0.042399823665618896,
0.021348822861909866,
0.0036076621618121862,
0.03752117604017258,
-0.0040774778462946415,
0.04076623544096947,
0.0020655053667724133,
0.02895377203822136,
-0.030813083052635193,
-0.07453400641679764,
0.07051742076873779,
-0.09415886551141739,
0.033506713807582855,
-0.008671285584568977,
0.015596767887473106,
-0.01651736907660961,
0.0011517637176439166,
0.04345479607582092,
0.03699295222759247,
0.003286303486675024,
-0.060304202139377594,
0.02422679401934147,
0.03583051636815071,
0.035634055733680725,
-0.028967680409550667,
-0.012891441583633423,
-0.007343188859522343,
-0.07017876952886581,
0.023797467350959778,
-0.011697430163621902,
0.001319856964983046,
-0.01849350333213806,
0.008890854194760323,
-0.014432065188884735,
-0.08408430963754654,
-0.041831791400909424,
-0.01037940289825201,
-0.11373938620090485,
-0.03247910365462303,
-0.03681928664445877,
0.00876679178327322,
-0.04160681739449501,
-0.025180233642458916,
0.012494176626205444,
-0.09025035053491592,
-0.052259597927331924,
0.06029355525970459,
0.009331349283456802,
-0.0367576964199543,
0.0036145111080259085,
0.026520298793911934,
0.018279526382684708,
0.029804805293679237,
0.000654751667752862,
-0.06880900263786316,
0.050011053681373596,
-0.015737514942884445,
0.03866036236286163,
0.05215156823396683,
-0.009655660018324852,
-0.025747759267687798,
0.07815098017454147,
-0.06833834946155548,
-0.0394183024764061,
0.07117607444524765,
0.001673962688073516,
0.00696357898414135,
-0.10149858146905899,
0.01996627449989319,
0.016090618446469307,
-0.026903552934527397,
-0.02462449111044407,
0.010281487368047237,
0.019114717841148376,
-0.004833424463868141,
-0.05226796492934227,
0.04104078933596611,
-0.04422293230891228,
-0.028354346752166748,
-0.08998826146125793,
-0.10884267836809158,
0.055603090673685074,
-0.05058267340064049,
-0.03842618316411972,
0.007707992102950811,
-4.715283631998474e-33,
-0.025941485539078712,
0.08173184096813202,
0.04197980836033821,
0.12842507660388947,
0.00590868666768074,
-0.07183176279067993,
-0.03165724501013756,
-0.0553194023668766,
-0.014759224839508533,
0.04149612411856651,
-0.0406465157866478,
-0.09664073586463928,
0.07174178212881088,
-0.07079189270734787,
0.011757144704461098,
0.023474637418985367,
-0.1050669327378273,
-0.0478472001850605,
0.0067846504971385,
0.028338957577943802,
-0.017718516290187836,
0.1325308233499527,
-0.05767173320055008,
0.011876433156430721,
-0.005297672003507614,
0.005089981481432915,
-0.06154145300388336,
0.00868268497288227,
0.042857419699430466,
0.0352477990090847,
-0.09662792831659317,
0.0039599123410880566,
-0.04183405265212059,
-0.09284473210573196,
0.03766167536377907,
0.0767255499958992,
0.10142000764608383,
-0.04193748161196709,
-0.02739258110523224,
0.15199998021125793,
0.055132023990154266,
-0.011641819961369038,
-0.13008122146129608,
0.03008132055401802,
0.016387585550546646,
0.05408278480172157,
-0.0012067347997799516,
0.019917570054531097,
0.010869808495044708,
-0.00098261923994869,
0.020121823996305466,
0.07201489061117172,
0.00792771764099598,
0.13116194307804108,
-0.03178461268544197,
-0.07260389626026154,
0.05413717031478882,
0.06947685033082962,
0.042005933821201324,
0.004924016073346138,
-0.09823250025510788,
-0.015432390384376049,
0.027467520907521248,
0.030663322657346725,
-0.03324268013238907,
0.0027660562191158533,
0.04073798283934593,
-0.020015118643641472,
0.01229733694344759,
-0.024716002866625786,
-0.04584244638681412,
-0.020642979070544243,
-0.0051180897280573845,
-0.048672568053007126,
0.006299244239926338,
0.030183110386133194,
0.014322654344141483,
-0.00008238791633630171,
-0.04340299963951111,
0.01055263914167881,
0.1152305155992508,
0.008467016741633415,
0.0004263499577064067,
0.0403151698410511,
0.08578495681285858,
0.0066438037902116776,
0.0509619377553463,
0.058799050748348236,
0.011565014719963074,
0.05630674958229065,
-0.06803664565086365,
0.05487443506717682,
-0.13687315583229065,
0.075613833963871,
-0.0796096995472908,
-4.0233029352521044e-8,
-0.09540844708681107,
0.04800913482904434,
0.019185081124305725,
0.10637926310300827,
0.01949012093245983,
0.008846867829561234,
0.07911328971385956,
0.0901879221200943,
-0.031082183122634888,
0.029019150882959366,
0.05313285440206528,
0.009731931611895561,
-0.0240541510283947,
-0.06402099132537842,
-0.006013091653585434,
0.046692099422216415,
0.02802428789436817,
-0.002033776603639126,
0.051650431007146835,
0.054951321333646774,
-0.00889340415596962,
-0.015767965465784073,
0.05700276792049408,
0.004542012698948383,
0.041037533432245255,
-0.10361632704734802,
-0.0409710593521595,
0.06054036691784859,
0.057822078466415405,
0.05237541347742081,
0.0403047651052475,
0.002889246679842472,
0.06125998497009277,
-0.07687956839799881,
0.03668953850865364,
0.13095712661743164,
-0.021826716139912605,
0.0072619980201125145,
0.03348007798194885,
0.049877870827913284,
-0.0549272783100605,
0.10949986428022385,
-0.09820818156003952,
-0.008224994875490665,
-0.0662156492471695,
-0.005457930266857147,
0.035351961851119995,
0.019454987719655037,
-0.012948990799486637,
-0.045668359845876694,
-0.0012751035392284393,
-0.05283191055059433,
0.03445715829730034,
-0.036164358258247375,
0.0009026576881296933,
0.0012548610102385283,
-0.01859409548342228,
-0.08336220681667328,
-0.005207733251154423,
0.03634599968791008,
0.042069192975759506,
-0.04709259793162346,
-0.07266130298376083,
-0.054785873740911484
] |
DreamDistribution: Learning Prompt Distribution for Diverse In-distribution Generation | https://openreview.net/forum?id=oQoQ4u6MQC | [
"Brian Nlong Zhao",
"Yuhang Xiao",
"Jiashu Xu",
"XINYANG JIANG",
"Yifan Yang",
"Dongsheng Li",
"Laurent Itti",
"Vibhav Vineet",
"Yunhao Ge"
] | Poster | The popularization of Text-to-Image (T2I) diffusion models enables the generation of high-quality images from text descriptions. However, generating diverse customized images with reference visual attributes remains challenging. This work focuses on personalizing T2I diffusion models at a more abstract concept or category level, adapting commonalities from a set of reference images while creating new instances with sufficient variations. We introduce a solution that allows a pretrained T2I diffusion model to learn a set of soft prompts, enabling the generation of novel images by sampling prompts from the learned distribution. These prompts offer text-guided editing capabilities and additional flexibility in controlling variation and mixing between multiple distributions. We also show the adaptability of the learned prompt distribution to other tasks, such as text-to-3D. Finally we demonstrate effectiveness of our approach through quantitative analysis including automatic evaluation and human assessment. | Generative Models, Image Generation, Personalized Generation | Generation of novel and diverse images/3Ds following a set of user input images trhough prompt distribution learning | 503 | null | [
0.04560105875134468,
-0.07424363493919373,
0.0386074036359787,
0.07254143059253693,
0.09227941930294037,
-0.03314390033483505,
-0.06472231447696686,
-0.022434549406170845,
0.10165257751941681,
-0.03112129308283329,
-0.012273474596440792,
-0.02231527864933014,
0.0641685426235199,
0.043142568320035934,
0.015385103411972523,
-0.012405306100845337,
0.06692589819431305,
0.02851460874080658,
-0.08628277480602264,
-0.018972983583807945,
0.023151051253080368,
-0.008580476976931095,
0.12719957530498505,
-0.01115817204117775,
0.04421335086226463,
0.0367431715130806,
0.0005924430442973971,
-0.028149884194135666,
0.04711248725652695,
-0.02301647514104843,
0.008518275804817677,
0.01740679144859314,
-0.03133309632539749,
0.03529556840658188,
-0.03280680626630783,
0.04963278770446777,
-0.10857050120830536,
0.12210065871477127,
0.019276456907391548,
-0.011912133544683456,
0.004027624148875475,
-0.009156429208815098,
0.004990991670638323,
0.04207532852888107,
0.053526345640420914,
-0.04709973558783531,
-0.04917100444436073,
-0.0662054568529129,
-0.0006735904607921839,
0.09135305136442184,
-0.05872294306755066,
-0.07367949187755585,
-0.06315908581018448,
0.01857764646410942,
-0.021589677780866623,
0.06246839463710785,
0.0394403412938118,
-0.012185291387140751,
0.0601406991481781,
-0.04674157500267029,
-0.056971192359924316,
-0.059835780411958694,
0.010923529043793678,
0.015070958994328976,
0.04733338579535484,
-0.0228591226041317,
-0.01585332490503788,
0.05407654121518135,
-0.018893703818321228,
-0.050402719527482986,
-0.04326643422245979,
0.043666772544384,
-0.023670580238103867,
0.036025695502758026,
0.04171469062566757,
0.051699891686439514,
-0.033568184822797775,
-0.07372298836708069,
-0.05266684293746948,
-0.03632539138197899,
0.027307836338877678,
0.005630556493997574,
0.07455160468816757,
-0.0066443439573049545,
0.017354285344481468,
0.012934328988194466,
0.02995433658361435,
0.04657259210944176,
-0.048048295080661774,
0.006318183150142431,
-0.05553867667913437,
-0.005813551601022482,
0.04526650905609131,
-0.03690658509731293,
-0.04297293722629547,
0.060587912797927856,
0.0039144642651081085,
-0.1754898577928543,
0.00010143602412426844,
0.021949781104922295,
0.024064894765615463,
-0.006502324249595404,
0.03317341208457947,
-0.04684782773256302,
0.013191631995141506,
-0.030085183680057526,
0.10036137700080872,
-0.06492168456315994,
0.0016567164566367865,
-0.03821844235062599,
0.04351480305194855,
0.0034672461915761232,
-0.03459843993186951,
-0.03931640088558197,
0.03521576523780823,
-0.07707870751619339,
0.027337873354554176,
-0.01923954300582409,
0.04049358889460564,
0.0502912737429142,
0.010019836947321892,
0.02404003217816353,
-0.05109024420380592,
-0.01915222592651844,
-0.051904574036598206,
-0.08154180645942688,
-0.02360723726451397,
6.877825252678378e-34,
0.08139046281576157,
0.00022725285089109093,
0.05501595512032509,
0.1497308760881424,
0.12525804340839386,
-0.04375148192048073,
-0.07573633641004562,
-0.027030108496546745,
-0.08442725986242294,
-0.09249722957611084,
0.03232505917549133,
0.014672240242362022,
-0.06265206634998322,
0.16757145524024963,
0.05965019389986992,
-0.007767929695546627,
-0.08171039074659348,
0.0921514481306076,
0.02737373858690262,
0.08329972624778748,
-0.05052850767970085,
-0.02435801364481449,
0.00883067399263382,
-0.04571318253874779,
-0.008022792637348175,
0.00617899838835001,
0.059770312160253525,
0.00502507109194994,
0.030927510932087898,
-0.0156486164778471,
-0.08881968259811401,
-0.019117753952741623,
0.04836651682853699,
-0.03248866647481918,
0.023140592500567436,
0.02894681692123413,
-0.010674682445824146,
-0.019876200705766678,
0.03525294363498688,
-0.01693285070359707,
0.02008199691772461,
0.03781963139772415,
0.009674027562141418,
0.0345882847905159,
-0.0778246745467186,
0.0710076317191124,
0.0593709759414196,
0.05037449300289154,
-0.03397480770945549,
-0.03042953833937645,
0.03367083892226219,
-0.02992221526801586,
-0.029651014134287834,
-0.04588776454329491,
-0.01729770191013813,
-0.015251695178449154,
0.04761901870369911,
-0.06969092041254044,
0.05040028318762779,
-0.01125863753259182,
0.04933573305606842,
0.07113408297300339,
-0.01779038831591606,
0.06113841384649277,
0.04854830354452133,
0.03221714869141579,
0.026096180081367493,
0.011213011108338833,
0.05396232008934021,
-0.005960619077086449,
-0.04596303775906563,
-0.020669078454375267,
-0.07705451548099518,
-0.05095991864800453,
0.01668372005224228,
-0.020453475415706635,
-0.0011858055368065834,
-0.03297223895788193,
0.03836463391780853,
0.02953115850687027,
-0.060622211545705795,
-0.05672767013311386,
-0.013284810818731785,
-0.08743298798799515,
-0.03483706712722778,
-0.019119413569569588,
0.000193145708180964,
-0.11682850867509842,
-0.07348854839801788,
-0.04436180740594864,
-0.05340297892689705,
-0.014757564291357994,
-0.07216077297925949,
0.04262432083487511,
0.03481559827923775,
-4.786949114464293e-34,
-0.031376224011182785,
0.04982294142246246,
-0.0152577580884099,
0.046569254249334335,
0.010517849586904049,
0.009637533687055111,
-0.0023912235628813505,
0.06553645431995392,
0.08537699282169342,
-0.03554181382060051,
-0.061119962483644485,
0.03083759918808937,
-0.05225744470953941,
-0.08503884077072144,
-0.041186846792697906,
-0.07852911949157715,
0.05977538228034973,
0.02688552811741829,
-0.06676452606916428,
0.014253012835979462,
0.017144015058875084,
0.049692802131175995,
-0.045652542263269424,
0.0021192433778196573,
-0.05740341544151306,
0.05800636485219002,
-0.026241952553391457,
0.07196978479623795,
0.009424387477338314,
-0.03264257684350014,
-0.05401334539055824,
-0.0464450903236866,
0.03478563204407692,
-0.018463870510458946,
-0.07611220329999924,
0.028971655294299126,
0.04517824947834015,
-0.030382074415683746,
-0.02020689845085144,
0.09976091235876083,
0.040281280875205994,
-0.06717189401388168,
-0.056281641125679016,
-0.003107907949015498,
-0.07750807702541351,
-0.05377916991710663,
-0.017313554883003235,
-0.07079921662807465,
0.08265902101993561,
0.11279217153787613,
-0.039582498371601105,
0.013421723619103432,
-0.0987614244222641,
0.014834517613053322,
-0.049958620220422745,
-0.12722277641296387,
-0.0637345090508461,
-0.004571577068418264,
0.09521406143903732,
-0.013309778645634651,
-0.05502263829112053,
-0.0039441343396902084,
0.040500953793525696,
-0.05698443204164505,
-0.0521831139922142,
-0.07596808671951294,
-0.09836313873529434,
-0.03525668755173683,
-0.013644594699144363,
0.051917776465415955,
0.023064715787768364,
-0.03950412943959236,
0.008341093547642231,
-0.017145246267318726,
0.046147942543029785,
-0.07794786989688873,
0.022633332759141922,
0.0013309677597135305,
0.012081200256943703,
-0.10222812741994858,
0.05116228386759758,
0.004252396523952484,
-0.006084701977670193,
0.09827012568712234,
0.08646782487630844,
0.07152359187602997,
0.03265959024429321,
0.04332788288593292,
0.038943126797676086,
0.03422556072473526,
-0.025544466450810432,
0.03702161833643913,
0.03181428089737892,
0.10132144391536713,
-0.07860246300697327,
-3.756525757125928e-8,
-0.03500240668654442,
0.015512452460825443,
0.0071660736575722694,
0.10024386644363403,
-0.02104107104241848,
-0.01889052800834179,
-0.007716455031186342,
0.05887473374605179,
-0.02388579025864601,
-0.017107918858528137,
0.018708057701587677,
0.01080025639384985,
-0.04373509809374809,
-0.03214360401034355,
0.0008074308279901743,
0.06763872504234314,
-0.01174470130354166,
0.058978356420993805,
0.03293760493397713,
-0.04108044132590294,
0.04182014241814613,
0.03550786152482033,
0.022842342033982277,
-0.07631188631057739,
0.05997420474886894,
0.041992463171482086,
-0.053571801632642746,
-0.01332845352590084,
-0.02654102072119713,
-0.054012276232242584,
0.08626437187194824,
-0.027721066027879715,
0.011012832634150982,
0.013378206640481949,
-0.01316282618790865,
0.03504233807325363,
-0.019825907424092293,
-0.02024109475314617,
-0.025599706918001175,
-0.01870323345065117,
0.013229809701442719,
0.0932905450463295,
0.03660915419459343,
0.036586444824934006,
-0.020078513771295547,
-0.008467614650726318,
0.05363643914461136,
-0.06894535571336746,
0.035963598638772964,
-0.0335991308093071,
-0.04736192524433136,
-0.07329293340444565,
0.01596985012292862,
0.04114207252860069,
0.04781736806035042,
-0.0162477046251297,
0.043380189687013626,
0.06857427209615707,
0.0776510164141655,
0.07467379420995712,
0.08426304906606674,
0.07406650483608246,
-0.03517717123031616,
0.003967627417296171
] |
Personality Alignment of Large Language Models | https://openreview.net/forum?id=0DZEs8NpUH | [
"Minjun Zhu",
"Yixuan Weng",
"Linyi Yang",
"Yue Zhang"
] | Poster | Aligning large language models (LLMs) typically aim to reflect general human values and behaviors, but they often fail to capture the unique characteristics and preferences of individual users. To address this gap, we introduce the concept of Personality Alignment. This approach tailors LLMs' responses and decisions to match the specific preferences of individual users or closely related groups. Inspired by psychometrics, we created the Personality Alignment with Personality Inventories (PAPI) dataset, which includes data from over 320,000 real subjects across multiple personality assessments - including both the Big Five Personality Factors and Dark Triad traits. This comprehensive dataset enables quantitative evaluation of LLMs' alignment capabilities across both positive and potentially problematic personality dimensions. Recognizing the challenges of personality alignments—such as limited personal data, diverse preferences, and scalability requirements—we developed an activation intervention optimization method. This method enhances LLMs' ability to efficiently align with individual behavioral preferences using minimal data and computational resources. Remarkably, our method, PAS, achieves superior performance while requiring only 1/5 of the optimization time compared to DPO, offering practical value for personality alignment. Our work paves the way for future AI systems to make decisions and reason in truly personality ways, enhancing the relevance and meaning of AI interactions for each user and advancing human-centered artificial intelligence. The dataset and code are released at https://github.com/zhu-minjun/PAlign. | Personality Alignment, Large language models, behavioral preferences of LM | We introduce Personality Alignment for language models, efficiently tailoring responses to individual user preferences, providing the‘ PAPI dataset with over 300K subjects and a practical, efficient alignment method. | 495 | 2408.11779 | [
-0.012729870155453682,
-0.053209468722343445,
0.04795575514435768,
0.03854500874876976,
-0.036875177174806595,
0.01483666431158781,
-0.008168857544660568,
0.042679332196712494,
-0.03446846827864647,
-0.06939651072025299,
0.013953063637018204,
-0.1398307979106903,
0.00367680168710649,
-0.0012389981420710683,
0.020605135709047318,
0.03260257467627525,
0.08833829313516617,
0.06092783063650131,
-0.03845188394188881,
-0.04124581068754196,
-0.06246049329638481,
0.05943228304386139,
0.01864776201546192,
-0.08675223588943481,
-0.04692672938108444,
0.0034165908582508564,
0.0460624173283577,
-0.05068081617355347,
-0.023333370685577393,
-0.012873116880655289,
0.059403590857982635,
0.04074493423104286,
0.026047272607684135,
0.06628768891096115,
-0.06830868124961853,
0.030386436730623245,
-0.11155904829502106,
0.031695351004600525,
0.0006771351327188313,
-0.057694531977176666,
-0.005507390480488539,
-0.027245469391345978,
0.037224043160676956,
0.011634619906544685,
-0.018309839069843292,
-0.031153257936239243,
-0.019767506048083305,
-0.024638572707772255,
-0.004785224795341492,
0.003341733943670988,
-0.1095917671918869,
-0.058156441897153854,
0.03560450300574303,
0.09277787804603577,
-0.01882706768810749,
0.051670677959918976,
0.08606647700071335,
0.09423639625310898,
0.05184438079595566,
0.00005547772525460459,
-0.025389930233359337,
-0.0681227296590805,
-0.0028414593543857336,
-0.012839846312999725,
0.03404250368475914,
0.07924710959196091,
-0.002907945541664958,
0.01323362160474062,
-0.07574835419654846,
0.03966877982020378,
-0.05228763818740845,
0.09083873778581619,
-0.004572705831378698,
0.0406140498816967,
-0.0006433865637518466,
0.03451554477214813,
0.018507052212953568,
-0.042013704776763916,
0.08994631469249725,
-0.006541375070810318,
-0.05439089983701706,
0.04581340774893761,
0.008935545571148396,
0.06854145228862762,
0.10107539594173431,
-0.015615737996995449,
-0.03286421298980713,
-0.024431569501757622,
0.015599801205098629,
0.02151121385395527,
-0.07978926599025726,
0.02422770857810974,
0.08878733217716217,
-0.048728492110967636,
0.017456425353884697,
0.08144241571426392,
-0.024667173624038696,
-0.09959542751312256,
-0.04186057671904564,
0.01925842836499214,
0.02264520339667797,
0.080039843916893,
-0.06664380431175232,
-0.035629622638225555,
-0.051688194274902344,
-0.06590896099805832,
0.013951604254543781,
-0.03912876918911934,
-0.008297332562506199,
-0.08441036939620972,
-0.024482231587171555,
-0.017203815281391144,
0.04838547855615616,
-0.02633102610707283,
0.13845427334308624,
-0.015191258862614632,
0.011186528950929642,
0.032214514911174774,
0.009297970682382584,
0.039326012134552,
0.022957276552915573,
0.021028460934758186,
0.08165361732244492,
0.04265313968062401,
0.00963370781391859,
0.04496157169342041,
-0.04462570324540138,
8.965709390427635e-34,
0.005682266782969236,
0.0844593420624733,
0.04079712554812431,
0.07829564809799194,
0.032701533287763596,
-0.03411431610584259,
-0.016525045037269592,
0.030321892350912094,
-0.08380387723445892,
-0.0055923606269061565,
-0.05560341104865074,
0.07990441471338272,
-0.08367782086133957,
0.13497400283813477,
0.038604654371738434,
0.047027338296175,
0.009940252639353275,
0.06490147113800049,
-0.06060340628027916,
-0.005894571077078581,
0.056281983852386475,
-0.022701997309923172,
0.0614469088613987,
-0.003658226691186428,
0.020945312455296516,
-0.019999613985419273,
0.08042780309915543,
-0.028691107407212257,
-0.03561205044388771,
-0.031179670244455338,
-0.1097087562084198,
0.02001280151307583,
0.050122473388910294,
0.032521944493055344,
0.04341775178909302,
-0.030926618725061417,
-0.041976045817136765,
-0.03329600393772125,
0.006137814838439226,
0.005244850646704435,
-0.0023892526514828205,
0.0746544674038887,
0.01573377475142479,
0.03264911472797394,
-0.09480731934309006,
-0.04786200821399689,
0.04189335182309151,
-0.040194783359766006,
0.027082985267043114,
-0.05280294269323349,
0.05009199678897858,
0.009455591440200806,
0.02837090939283371,
-0.00041628326289355755,
-0.08723653852939606,
0.003602182725444436,
-0.014298955909907818,
-0.0024962311144918203,
-0.05815066024661064,
0.003395575098693371,
0.05397742986679077,
-0.020348776131868362,
-0.014253553003072739,
-0.005122800823301077,
0.05878070741891861,
-0.042398788034915924,
-0.009019712917506695,
0.07393702864646912,
0.035321738570928574,
0.013442817144095898,
0.025226030498743057,
0.009002331644296646,
0.012908167205750942,
-0.033116623759269714,
-0.06960050761699677,
0.056605879217386246,
0.042911872267723083,
-0.09982237219810486,
0.005321684759110212,
-0.0003331136831548065,
0.012398172169923782,
0.04616992175579071,
-0.047926273196935654,
-0.12894606590270996,
-0.044267043471336365,
-0.1363077610731125,
0.03498910367488861,
-0.010457280091941357,
0.021252021193504333,
0.0034257040824741125,
-0.03957974165678024,
0.014479956589639187,
0.037029724568128586,
0.046532101929187775,
-0.04072435572743416,
-1.068048657027361e-33,
-0.06683150678873062,
-0.062224723398685455,
-0.013711800798773766,
-0.012831032276153564,
0.055372145026922226,
-0.04605759680271149,
0.06569809466600418,
0.029455995187163353,
0.031883593648672104,
-0.003651738865301013,
-0.05725295469164848,
-0.055588629096746445,
0.11206995695829391,
-0.03584273159503937,
0.04488348215818405,
-0.006505121476948261,
-0.03605937585234642,
0.08167703449726105,
0.022548027336597443,
0.008380845189094543,
0.09178957343101501,
0.10277511924505234,
-0.12727947533130646,
0.03509692847728729,
-0.04409347102046013,
0.01369280181825161,
0.003967670723795891,
-0.03662852570414543,
-0.010499046184122562,
0.0187382735311985,
-0.06732972711324692,
0.05487668886780739,
-0.11659842729568481,
-0.008025672286748886,
0.013396739028394222,
0.034889183938503265,
-0.0657029077410698,
-0.050625067204236984,
-0.025894204154610634,
0.07767922431230545,
0.0009879441931843758,
0.018571609631180763,
-0.06940637528896332,
-0.06864146888256073,
-0.018111174926161766,
0.043855126947164536,
0.018189894035458565,
-0.06053299456834793,
0.002647101180627942,
-0.026609396561980247,
-0.01749139465391636,
-0.004529599100351334,
-0.016142701730132103,
0.08084754645824432,
-0.014218175783753395,
-0.1002144068479538,
0.020914027467370033,
-0.04367228224873543,
0.04641379788517952,
-0.0011606086045503616,
-0.08547043055295944,
0.009751698933541775,
0.10933602601289749,
-0.030408950522542,
-0.06405426561832428,
-0.035289015620946884,
0.020252950489521027,
0.04220801219344139,
-0.07520294934511185,
-0.0865255743265152,
0.06128222867846489,
-0.08995702117681503,
0.017261603847146034,
0.015205904841423035,
-0.07523910701274872,
-0.0835513174533844,
-0.010618958622217178,
-0.03029004856944084,
-0.011233203113079071,
-0.08048257231712341,
0.019750630483031273,
0.03330788388848305,
-0.009754139930009842,
0.030731797218322754,
0.04094970226287842,
0.07423561811447144,
0.028301581740379333,
0.13455021381378174,
0.048240095376968384,
0.0659823790192604,
-0.053701888769865036,
0.030074456706643105,
-0.037267908453941345,
0.044805802404880524,
-0.10464496910572052,
-4.4198735338341066e-8,
-0.07445713877677917,
-0.03646950423717499,
-0.027975676581263542,
0.11670033633708954,
-0.020305994898080826,
-0.026821723207831383,
-0.0692889615893364,
0.022204842418432236,
0.026376141235232353,
0.014588185586035252,
0.009836999699473381,
0.04405035823583603,
-0.006193977314978838,
-0.07083375006914139,
-0.005371362902224064,
0.009505213238298893,
-0.02962707355618477,
0.03054550662636757,
0.018504386767745018,
0.011862660758197308,
0.031565286219120026,
0.04587521031498909,
-0.027146875858306885,
-0.029782501980662346,
0.02746349759399891,
-0.06128654256463051,
-0.06306862831115723,
0.06583885103464127,
-0.01689697615802288,
0.015706192702054977,
0.02528444491326809,
-0.021755389869213104,
0.04621398448944092,
-0.08425885438919067,
0.021510086953639984,
0.08977708965539932,
-0.025180647149682045,
-0.0058737993240356445,
0.009934162721037865,
0.04575805738568306,
0.05485481023788452,
0.07095319032669067,
-0.0036969934590160847,
0.052938953042030334,
-0.001166930771432817,
-0.027000807225704193,
-0.03739333152770996,
-0.04782867431640625,
0.07068079710006714,
0.03014555387198925,
-0.04787105321884155,
-0.05200875550508499,
0.02425302192568779,
0.0656728595495224,
0.07870468497276306,
-0.025893207639455795,
0.014771793968975544,
0.02552914060652256,
0.05668492987751961,
0.028422761708498,
0.10495021194219589,
0.004182775970548391,
-0.060885969549417496,
-0.007851865142583847
] |
CycleResearcher: Improving Automated Research via Automated Review | https://openreview.net/forum?id=bjcsVLoHYs | [
"Yixuan Weng",
"Minjun Zhu",
"Guangsheng Bao",
"Hongbo Zhang",
"Jindong Wang",
"Yue Zhang",
"Linyi Yang"
] | Poster | The automation of scientific discovery has been a long-standing goal within the research community, driven by the potential to accelerate knowledge creation. While significant progress has been made using commercial large language models (LLMs) as research assistants or idea generators, the possibility of automating the entire research process with open-source LLMs remains largely unexplored. This paper explores the feasibility of using open-source post-trained LLMs as autonomous agents capable of performing the full cycle of automated research and review, from literature review and manuscript preparation to peer review and paper refinement. Our iterative preference training framework consists of CycleResearcher, which conducts research tasks, and CycleReviewer, which simulates the peer review process, providing iterative feedback via reinforcement learning. To train these models, we develop two new datasets, Review-5k and Research-14k, reflecting real-world machine learning research and peer review dynamics. Our results demonstrate that CycleReviewer achieves promising performance with a 26.89\% reduction in mean absolute error (MAE) compared to individual human reviewers in predicting paper scores, indicating the potential of LLMs to effectively assist expert-level research evaluation. In research, the papers generated by the CycleResearcher model achieved a score of 5.36 in simulated peer reviews, showing some competitiveness in terms of simulated review scores compared to the preprint level of 5.24 from human experts, while still having room for improvement compared to the accepted paper level of 5.69. This work represents a significant step toward fully automated scientific inquiry, providing ethical safeguards and exploring AI-driven research capabilities. The code, dataset and model weight are released at https://wengsyx.github.io/Researcher. | Large Language Models, Automation of Scientific Discovery, AI Scientist | We propose CycleResearcher and CycleReviewer, open-source LLMs automating the research and review cycle. | 489 | 2411.00816 | [
-0.10210920870304108,
-0.06211601942777634,
-0.062215302139520645,
0.0895879715681076,
0.013698048889636993,
0.008810922503471375,
-0.05741722509264946,
0.04329651966691017,
0.006503810174763203,
0.04710818827152252,
-0.06722766906023026,
0.010270098224282265,
0.09197430312633514,
0.015584606677293777,
-0.05392511188983917,
0.06157726049423218,
0.0669797733426094,
-0.04742763191461563,
-0.03310639411211014,
-0.04785234108567238,
0.030335647985339165,
0.08024344593286514,
0.09651787579059601,
-0.006937419064342976,
0.015323537401854992,
-0.014217148534953594,
-0.03439542278647423,
-0.058347202837467194,
-0.004978415556252003,
0.051940612494945526,
0.022091204300522804,
0.04265874996781349,
-0.01705026626586914,
0.02891310676932335,
-0.06068217381834984,
0.04686960577964783,
-0.0017734229331836104,
0.06083448976278305,
0.010231317952275276,
-0.03899611905217171,
-0.026463212445378304,
-0.03691178560256958,
-0.0045973691157996655,
-0.0004887549439445138,
0.09725946933031082,
-0.01590064913034439,
-0.010204356163740158,
-0.027365781366825104,
-0.007589699234813452,
0.0638214647769928,
-0.07087741792201996,
-0.09030894190073013,
0.05760391429066658,
-0.013986333273351192,
-0.02798692137002945,
0.044513676315546036,
0.043843794614076614,
0.02828182280063629,
-0.016492219641804695,
-0.08397237211465836,
-0.00919367279857397,
-0.11682813614606857,
-0.0915081724524498,
-0.04160081967711449,
-0.012530118227005005,
-0.027995267882943153,
-0.03051319718360901,
0.05214090645313263,
-0.05253703519701958,
-0.0027389400638639927,
-0.012056903913617134,
0.08953294903039932,
0.07528281956911087,
0.04649174213409424,
0.04405613988637924,
0.046449534595012665,
-0.03761938214302063,
0.006209248211234808,
0.07734978944063187,
-0.051139723509550095,
-0.029819976538419724,
-0.006569418124854565,
0.03573804348707199,
0.0515299029648304,
-0.029356233775615692,
-0.049046047031879425,
0.11203913390636444,
0.07177615910768509,
0.030710140243172646,
0.03898465633392334,
-0.0035647230688482523,
0.01846778765320778,
0.011899652890861034,
-0.0667647272348404,
-0.0009055190603248775,
0.07265143096446991,
0.00037004382465966046,
-0.10432624816894531,
0.026140326634049416,
0.0621810257434845,
0.05701702460646629,
0.09943290054798126,
0.024367274716496468,
-0.07546652853488922,
-0.0600888766348362,
0.0050184233114123344,
0.006104144733399153,
0.0315399095416069,
0.06599628925323486,
-0.0705091655254364,
-0.012690354138612747,
0.022374959662556648,
0.048510290682315826,
-0.0002701645134948194,
0.06998177617788315,
0.0017715425929054618,
0.03360969200730324,
-0.029306840151548386,
0.05484240502119064,
0.11200982332229614,
-0.04267993941903114,
-0.010783438570797443,
0.04950610175728798,
-0.040841687470674515,
0.067270927131176,
0.03562723845243454,
-0.03860388696193695,
2.8502780904464185e-34,
0.054925184696912766,
0.08215738832950592,
0.09789307415485382,
0.03152722865343094,
0.05212227255105972,
-0.050825122743844986,
0.0025985310785472393,
0.027689777314662933,
-0.11513343453407288,
-0.1274886131286621,
-0.007821273989975452,
0.0006467882194556296,
-0.04595181345939636,
0.045193031430244446,
0.009000767953693867,
-0.042658306658267975,
-0.008710305206477642,
0.00999082438647747,
-0.06129291653633118,
-0.0726381242275238,
0.05154813453555107,
-0.11674364656209946,
0.06139584630727768,
-0.05365827679634094,
-0.0011400392977520823,
-0.03306169807910919,
0.03707454726099968,
-0.021208183839917183,
0.004115586169064045,
0.010341549292206764,
-0.0654987096786499,
0.024554502218961716,
0.01642359420657158,
-0.01594153791666031,
-0.026788480579853058,
0.03175158053636551,
-0.022788168862462044,
-0.00907173939049244,
0.03744782879948616,
-0.029441311955451965,
0.008284227922558784,
-0.0047377427108585835,
0.032164305448532104,
0.01334783248603344,
-0.075480155646801,
-0.015667665749788284,
0.02460568957030773,
0.016359904780983925,
0.026570670306682587,
-0.013456998392939568,
0.08027010411024094,
0.028350697830319405,
-0.039178140461444855,
0.023598624393343925,
-0.00526257511228323,
0.10691295564174652,
-0.002326358575373888,
0.0022477407474070787,
0.019867226481437683,
-0.05185713618993759,
0.07906296849250793,
0.08380317687988281,
-0.013376131653785706,
0.07438316196203232,
-0.002778930589556694,
0.07091613858938217,
0.03483276814222336,
0.07305648177862167,
0.042439188808202744,
0.012920836918056011,
-0.048885852098464966,
-0.00801850762218237,
-0.016458338126540184,
-0.06671669334173203,
0.04101259633898735,
-0.013238001614809036,
0.007550888694822788,
-0.0796361193060875,
0.045438457280397415,
0.0020715424325317144,
0.025986570864915848,
-0.039226166903972626,
-0.05515078827738762,
-0.1246488019824028,
-0.05132909119129181,
-0.12313118577003479,
0.004419281613081694,
-0.024069206789135933,
0.012497232295572758,
-0.04620935767889023,
-0.021435879170894623,
-0.02767721749842167,
-0.0043560899794101715,
0.06634243577718735,
0.025976000353693962,
-3.4851377937415733e-34,
-0.10112093389034271,
-0.11028734594583511,
0.018436774611473083,
0.1473073959350586,
0.03535376861691475,
-0.046234581619501114,
-0.08885738253593445,
-0.03121371753513813,
0.035437870770692825,
-0.008831885643303394,
-0.03209248185157776,
-0.1117032989859581,
0.06526560336351395,
0.004532817285507917,
0.05019502341747284,
-0.03402958810329437,
-0.0516027994453907,
-0.008844458498060703,
0.018851013854146004,
0.07238925993442535,
0.018341699615120888,
0.07669105380773544,
-0.05329124629497528,
0.0648636445403099,
-0.0042658355087041855,
0.021293556317687035,
0.050840988755226135,
0.055584389716386795,
0.01693006232380867,
0.02186717838048935,
0.018867263570427895,
-0.01998187229037285,
-0.05089346691966057,
-0.01593211479485035,
-0.09481924772262573,
0.06217635050415993,
0.04187578707933426,
-0.06951390951871872,
-0.05034472048282623,
0.06708142906427383,
0.054563257843256,
-0.013907245360314846,
-0.10008829087018967,
-0.09047042578458786,
-0.05980505049228668,
-0.006065822672098875,
-0.08748206496238708,
0.002214987063780427,
-0.005345834419131279,
-0.03675505891442299,
0.013358253054320812,
0.05814262107014656,
-0.03316517174243927,
-0.01916429027915001,
-0.008037921972572803,
-0.09858310967683792,
0.012922168709337711,
0.01604982651770115,
0.0201252568513155,
0.04013897106051445,
-0.08626502752304077,
-0.012241058982908726,
-0.022014150395989418,
-0.009975776076316833,
-0.03081163763999939,
-0.12673905491828918,
0.011893094517290592,
0.07385442405939102,
-0.01795332133769989,
0.049902692437171936,
0.004183168988674879,
0.014756050892174244,
-0.004847106523811817,
-0.02298794686794281,
0.05946284905076027,
0.020180294290184975,
0.00436569657176733,
-0.09160562604665756,
-0.041636426001787186,
-0.05142839998006821,
-0.027641260996460915,
0.05407967418432236,
0.03605162352323532,
0.05220240354537964,
0.029732707887887955,
-0.0021627116948366165,
0.040589556097984314,
0.1330355405807495,
-0.02111489325761795,
-0.0048965211026370525,
-0.010823862627148628,
-0.06681831926107407,
-0.019782952964305878,
0.027477625757455826,
-0.02949906513094902,
-4.0175983428980544e-8,
-0.022875752300024033,
0.024074213579297066,
0.03630613908171654,
0.10418854653835297,
0.018434680998325348,
-0.07984267175197601,
0.002323938999325037,
0.11062750965356827,
-0.019153786823153496,
-0.050269123166799545,
0.04461786523461342,
-0.03382542356848717,
0.004945453256368637,
-0.05944843590259552,
-0.020202357321977615,
0.027688652276992798,
0.06656073033809662,
-0.005444609094411135,
-0.00013844785280525684,
-0.045629724860191345,
0.06882280856370926,
0.0013177560176700354,
0.056913331151008606,
-0.03168799355626106,
0.08910039067268372,
-0.07082434743642807,
-0.01321452483534813,
0.053487926721572876,
-0.024855632334947586,
-0.022329416126012802,
-0.009240558370947838,
-0.03390049934387207,
-0.05392125993967056,
-0.05936049297451973,
0.040987107902765274,
0.0607026107609272,
0.08467070013284683,
-0.041613541543483734,
-0.006983737926930189,
0.006720461882650852,
0.02385588362812996,
0.07997152954339981,
-0.014521169476211071,
-0.016723206266760826,
-0.08374930173158646,
0.007277169730514288,
-0.04725858196616173,
-0.05178067833185196,
0.02146989293396473,
-0.05316866189241409,
0.02719317190349102,
-0.049355648458004,
0.03741835430264473,
0.01585586555302143,
0.056918054819107056,
0.07644904404878616,
-0.024663561955094337,
-0.009627585299313068,
0.03058738075196743,
0.027141926810145378,
0.09540238976478577,
-0.018697064369916916,
-0.003949430771172047,
0.05883210152387619
] |
Towards Understanding Why Label Smoothing Degrades Selective Classification and How to Fix It | https://openreview.net/forum?id=6oWFn6fY4A | [
"Guoxuan Xia",
"Olivier Laurent",
"Gianni Franchi",
"Christos-Savvas Bouganis"
] | Poster | Label smoothing (LS) is a popular regularisation method for training neural networks as it is effective in improving test accuracy and is simple to implement. ''Hard'' one-hot labels are ''smoothed'' by uniformly distributing probability mass to other classes, reducing overfitting. Prior work has shown that in some cases *LS can degrade selective classification (SC)* -- where the aim is to reject misclassifications using a model's uncertainty. In this work, we first demonstrate empirically across an extended range of large-scale tasks and architectures that LS *consistently* degrades SC.
We then address a gap in existing knowledge, providing an *explanation* for this behaviour by analysing logit-level gradients: LS degrades the uncertainty rank ordering of correct vs incorrect predictions by regularising the max logit *more* when a prediction is likely to be correct, and *less* when it is likely to be wrong.
This elucidates previously reported experimental results where strong classifiers underperform in SC.
We then demonstrate the empirical effectiveness of post-hoc *logit normalisation* for recovering lost SC performance caused by LS. Furthermore, linking back to our gradient analysis, we again provide an explanation for why such normalisation is effective. | Uncertainty Estimation, Selective Classification, Label Smoothing | null | 487 | 2403.14715 | [
-0.06377942115068436,
-0.05181281268596649,
0.0577908419072628,
0.07970926910638809,
0.07423950731754303,
-0.026445450261235237,
0.050393883138895035,
-0.02017960697412491,
-0.04605019465088844,
-0.08285035938024521,
-0.08294069766998291,
-0.015439200215041637,
0.043387047946453094,
0.017945118248462677,
-0.10007321834564209,
-0.05808233469724655,
0.0791206955909729,
0.08168508857488632,
-0.11213003844022751,
0.019690068438649178,
-0.028767749667167664,
0.05039810389280319,
0.016271501779556274,
0.07427410036325455,
0.06853675842285156,
-0.025677163153886795,
-0.0438154898583889,
0.010891576297581196,
-0.03354763984680176,
-0.03278402239084244,
-0.03529560938477516,
0.012543345801532269,
-0.05667424201965332,
0.013125400058925152,
0.006878240965306759,
-0.03850102797150612,
-0.0524103045463562,
0.04085984081029892,
0.09721479564905167,
-0.00456091808155179,
-0.005494846496731043,
0.002950830152258277,
0.014948724769055843,
0.020610442385077477,
0.034994225949048996,
0.009360240772366524,
0.005924407858401537,
-0.1226024255156517,
-0.07317928224802017,
0.01454270351678133,
-0.02306780032813549,
-0.06413868069648743,
-0.006508514750748873,
0.019587252289056778,
-0.1498287171125412,
0.0683819055557251,
0.0549066923558712,
0.04491567239165306,
-0.0025925789959728718,
0.015555212274193764,
-0.07766219228506088,
-0.10196833312511444,
-0.0344340056180954,
-0.026927504688501358,
0.022622575983405113,
-0.033140163868665695,
0.03757120668888092,
-0.007846571505069733,
0.034254394471645355,
0.11089448630809784,
0.032900307327508926,
0.1411411017179489,
-0.009488428011536598,
0.0961640253663063,
0.036987315863370895,
0.028221676126122475,
0.0383424311876297,
0.035940155386924744,
0.03321429342031479,
-0.0616721585392952,
0.004183141048997641,
0.0018533491529524326,
-0.0003312458866275847,
0.0176815465092659,
0.15264634788036346,
-0.005324029829353094,
-0.018356846645474434,
-0.011217867955565453,
-0.010229385457932949,
-0.03692452237010002,
-0.010776234790682793,
-0.04908564314246178,
0.03034554235637188,
0.006826743949204683,
0.0029479367658495903,
0.038046881556510925,
-0.021445637568831444,
-0.049559302628040314,
-0.025779947638511658,
0.10305501520633698,
-0.0632958859205246,
0.05372883379459381,
0.0015334548661485314,
-0.040250327438116074,
0.0401381179690361,
0.004215477965772152,
0.045727942138910294,
0.002995119197294116,
0.0420650951564312,
-0.11166682839393616,
-0.02529088221490383,
-0.0005110269994474947,
-0.033519379794597626,
-0.024287011474370956,
0.03608955442905426,
0.05569146201014519,
0.051195040345191956,
-0.03311001881957054,
-0.08075951039791107,
0.053058404475450516,
-0.09998033195734024,
0.045094821602106094,
0.029388362541794777,
0.004688816610723734,
0.016331126913428307,
0.09107349067926407,
-0.03596727177500725,
4.245676210238862e-33,
0.04786377772688866,
-0.007250525057315826,
0.009007446467876434,
-0.0011567602632567286,
0.033808767795562744,
-0.042962878942489624,
-0.01930394396185875,
0.014537218026816845,
0.02316708117723465,
0.027138682082295418,
0.00006534918065881357,
0.04997419938445091,
-0.018906399607658386,
0.030519116669893265,
0.0496380589902401,
-0.01594318076968193,
-0.025281313806772232,
0.04700789973139763,
-0.0812673345208168,
-0.00511093670502305,
-0.0234543327242136,
-0.0458836667239666,
0.010202770121395588,
-0.13272975385189056,
-0.0011707398807629943,
0.02921823225915432,
0.0171069148927927,
0.06463686376810074,
0.01203596405684948,
0.020695624873042107,
-0.04110302776098251,
-0.0241790059953928,
0.04753157123923302,
0.021007103845477104,
0.03415246680378914,
-0.025297366082668304,
0.010051831603050232,
-0.007254354190081358,
0.023607339709997177,
0.030079474672675133,
-0.035038888454437256,
0.04655585065484047,
-0.0062944283708930016,
0.02254842035472393,
-0.007421087007969618,
-0.008484725840389729,
-0.023844709619879723,
-0.030264845117926598,
-0.050943221896886826,
-0.10110896825790405,
0.011662440374493599,
0.01670219376683235,
0.015183206647634506,
-0.03722449392080307,
-0.06033520773053169,
-0.05415372923016548,
0.09246498346328735,
0.027199646458029747,
-0.011833330616354942,
0.0034386625047773123,
0.023471113294363022,
0.03850869461894035,
-0.054104071110486984,
0.06187928095459938,
0.037815969437360764,
0.07542368024587631,
0.0348445288836956,
0.08023575693368912,
0.004909800831228495,
-0.06004810333251953,
-0.01572122611105442,
0.004720854107290506,
-0.018394265323877335,
-0.06457637250423431,
0.09569405019283295,
0.025462476536631584,
0.0841081291437149,
-0.019116220995783806,
0.011153962463140488,
-0.03195872902870178,
-0.007829008623957634,
-0.02196534350514412,
-0.007486481219530106,
-0.04750148206949234,
-0.041307706385850906,
-0.0748484879732132,
0.008610987104475498,
-0.06713629513978958,
0.020401500165462494,
0.05968662351369858,
-0.07086099684238434,
-0.03234894201159477,
-0.017857393249869347,
0.039364561438560486,
-0.04188089072704315,
-3.102324283071872e-33,
-0.10059936344623566,
0.05919738858938217,
-0.03457583487033844,
0.08258704841136932,
-0.06180840730667114,
0.013847410678863525,
-0.0009118656744249165,
0.038195449858903885,
-0.010786822065711021,
-0.054401885718107224,
0.024826304987072945,
-0.019029654562473297,
0.04234528914093971,
-0.03315650299191475,
0.015306385233998299,
-0.023203950375318527,
-0.06452593952417374,
0.015170391649007797,
0.04666882008314133,
0.026142071932554245,
0.07392684370279312,
0.09810106456279755,
-0.09772223234176636,
0.05676589161157608,
-0.02091064862906933,
0.06094885617494583,
-0.06363753229379654,
0.10225006937980652,
0.044741928577423096,
-0.09355746954679489,
-0.007933758199214935,
-0.049945782870054245,
0.025946373119950294,
-0.05810882896184921,
0.04363299906253815,
0.0623127743601799,
0.0013545291731134057,
0.034848298877477646,
-0.02760634571313858,
0.07712642103433609,
-0.021417444571852684,
-0.030280010774731636,
-0.08521492779254913,
-0.03841852769255638,
0.0124976122751832,
-0.033561691641807556,
-0.05797185003757477,
-0.05782006308436394,
-0.042338136583566666,
0.014991279691457748,
-0.017814133316278458,
0.019016249105334282,
-0.10514213144779205,
0.08522363752126694,
-0.0037827929481863976,
-0.03694191202521324,
-0.09024795889854431,
-0.016512908041477203,
-0.021806124597787857,
0.09635201096534729,
-0.06294762343168259,
0.044782884418964386,
-0.01102073211222887,
0.010084981098771095,
0.008441713638603687,
-0.05176514387130737,
-0.037556495517492294,
0.026699164882302284,
0.036867085844278336,
0.002707025036215782,
0.05481712147593498,
-0.0035311533138155937,
-0.024707036092877388,
0.026353273540735245,
-0.07775472849607468,
-0.026643158867955208,
0.03987163305282593,
-0.0244524534791708,
0.020624125376343727,
0.00739300949499011,
0.016427580267190933,
-0.018061457201838493,
0.017016487196087837,
0.08203409612178802,
0.09331011027097702,
0.07985369116067886,
0.022378502413630486,
0.048784881830215454,
0.05411628633737564,
-0.025634318590164185,
0.015296366065740585,
-0.05396837741136551,
0.019661208614706993,
0.017653068527579308,
-0.09139511734247208,
-5.129438207518433e-8,
-0.05732208490371704,
0.06275118887424469,
0.012388146482408047,
0.0726214125752449,
0.07987216860055923,
-0.04542027413845062,
-0.018520537763834,
0.11316176503896713,
-0.05181284621357918,
-0.013443184085190296,
0.041685063391923904,
-0.022407401353120804,
-0.010242384858429432,
-0.08490415662527084,
0.024243367835879326,
0.06090666353702545,
0.00448948610574007,
0.1125461608171463,
-0.0014116360107436776,
0.023577431216835976,
0.03567252680659294,
-0.004028341267257929,
0.059708721935749054,
-0.07696401327848434,
0.0341438353061676,
-0.12038157880306244,
-0.0020779857877641916,
0.12070326507091522,
0.10473978519439697,
0.023803096264600754,
-0.05155719444155693,
-0.003215743927285075,
0.006950229872018099,
-0.009825353510677814,
0.02033672295510769,
0.13051624596118927,
0.023388130590319633,
-0.07894197106361389,
-0.01027123536914587,
0.047483891248703,
-0.04139174148440361,
-0.017454691231250763,
-0.05098896101117134,
0.015953995287418365,
-0.018968621268868446,
-0.02940070442855358,
-0.012082262896001339,
-0.069224514067173,
0.019990338012576103,
-0.04238705709576607,
-0.02360622026026249,
-0.0055460250005126,
-0.021785516291856766,
0.07930838316679001,
0.0588892437517643,
0.005134913604706526,
-0.04091285169124603,
-0.09035767614841461,
-0.11099600046873093,
0.0910177081823349,
0.0445958636701107,
-0.016679996624588966,
-0.062140535563230515,
0.05534311383962631
] |
Towards Improving Exploration through Sibling Augmented GFlowNets | https://openreview.net/forum?id=HH4KWP8RP5 | [
"Kanika Madan",
"Alex Lamb",
"Emmanuel Bengio",
"Glen Berseth",
"Yoshua Bengio"
] | Poster | Exploration is a key factor for the success of an active learning agent, especially when dealing with sparse extrinsic terminal rewards and long trajectories. We introduce Sibling Augmented Generative Flow Networks (SA-GFN), a novel framework designed to enhance exploration and training efficiency of Generative Flow Networks (GFlowNets). SA-GFN uses a decoupled dual network architecture, comprising of a main Behavior Network and an exploratory Sibling Network, to enable a diverse exploration of the underlying distribution using intrinsic rewards. Inspired by the ideas on exploration from reinforcement learning, SA-GFN provides a general-purpose exploration and learning paradigm that integrates with multiple GFlowNet training objectives and is especially helpful for exploration over a wide range of sparse or low reward distributions and task structures. An extensive set of experiments across a diverse range of tasks, reward structures and trajectory lengths, along with a thorough set of ablations, demonstrate the superior performance of SA-GFN in terms of exploration efficacy and convergence speed as compared to the existing methods. In addition, SA-GFN's versatility and compatibility with different GFlowNet training objectives and intrinsic reward methods underscores its broad applicability in various problem domains. | Generative Models, Generative Flow Networks, Exploration | null | 482 | null | [
-0.05682329088449478,
-0.09391278773546219,
0.04915904998779297,
0.022845981642603874,
0.005204727407544851,
0.028772292658686638,
-0.0014720504404976964,
-0.07312621176242828,
0.02577640861272812,
-0.047145992517471313,
-0.03974808752536774,
-0.07272939383983612,
0.0025876997970044613,
0.014086098410189152,
-0.00048381369560956955,
0.006177304778248072,
0.1081523671746254,
0.041130729019641876,
-0.03019161894917488,
-0.04623004049062729,
0.021088773384690285,
-0.017399704083800316,
0.045005496591329575,
-0.002101621590554714,
0.003992043901234865,
0.058210186660289764,
-0.024834219366312027,
-0.08510564267635345,
0.03803044185042381,
-0.04247952625155449,
0.11394719779491425,
0.023024985566735268,
-0.01761295273900032,
0.009859712794423103,
-0.06495407223701477,
0.11640425026416779,
-0.10516909509897232,
0.07340314984321594,
0.0031676546204835176,
-0.06671886891126633,
0.03657049685716629,
0.03006862662732601,
0.05176189914345741,
-0.00917388591915369,
0.07072987407445908,
0.00797493476420641,
-0.06245513632893562,
-0.08702325075864792,
0.04208613187074661,
0.0214549507945776,
0.0315118208527565,
-0.08691924065351486,
0.023376723751425743,
0.047355614602565765,
0.0823659598827362,
0.0941130667924881,
-0.04596467688679695,
-0.05844765901565552,
-0.017393555492162704,
-0.006422476842999458,
-0.017913702875375748,
-0.10655254870653152,
-0.04052721709012985,
-0.02920721471309662,
-0.08632995188236237,
-0.057273369282484055,
0.0614074170589447,
0.04999646171927452,
0.03751722723245621,
-0.028575409203767776,
0.09848399460315704,
0.05236964300274849,
-0.024573981761932373,
-0.01803574152290821,
0.050887130200862885,
0.08396187424659729,
-0.004973004572093487,
-0.015193294733762741,
-0.004956838209182024,
0.028561241924762726,
0.034686021506786346,
0.04185450077056885,
0.008121214807033539,
0.017125077545642853,
0.01325176376849413,
-0.03873523697257042,
-0.01004775334149599,
0.05235373601317406,
0.02270028181374073,
0.016902215778827667,
-0.04420190677046776,
0.02705995924770832,
0.023292357102036476,
-0.03270601108670235,
0.01814558357000351,
0.019403520971536636,
-0.02819746546447277,
-0.10415944457054138,
-0.10205584764480591,
0.0703899934887886,
0.029676422476768494,
0.030559729784727097,
0.0009349264437332749,
0.05877969041466713,
-0.014472749084234238,
-0.021108312532305717,
0.05620134249329567,
0.029515285044908524,
0.10031759738922119,
-0.0652015209197998,
-0.05807822197675705,
-0.038332078605890274,
0.058035463094711304,
0.003224617103114724,
-0.08190441876649857,
0.025871330872178078,
-0.01450272835791111,
0.0011391970328986645,
-0.0026141530834138393,
0.1120777353644371,
-0.03929969295859337,
-0.06785570830106735,
-0.0019533741287887096,
-0.026814427226781845,
0.022715281695127487,
-0.055720169097185135,
-0.031171996146440506,
4.1586730978887045e-33,
0.08072185516357422,
-0.007213435601443052,
0.04848988726735115,
0.032411959022283554,
0.09087684005498886,
0.024840328842401505,
0.037807513028383255,
-0.021842842921614647,
-0.01581333577632904,
-0.051223259419202805,
-0.07591243088245392,
0.06815404444932938,
-0.07087341696023941,
0.14576870203018188,
0.013063083402812481,
-0.08773108571767807,
-0.07354047894477844,
0.04150537773966789,
0.05949309840798378,
-0.05852138623595238,
0.02853166125714779,
-0.1030341237783432,
-0.02874302864074707,
-0.03755013644695282,
0.011721109971404076,
-0.03033403493463993,
0.021466737613081932,
-0.0028452810365706682,
-0.02659374102950096,
0.01121576875448227,
-0.09641502797603607,
0.023642856627702713,
-0.02337709069252014,
0.002495607128366828,
-0.08777657896280289,
0.016056472435593605,
0.015985596925020218,
-0.03463612496852875,
0.06164617836475372,
0.006264230702072382,
-0.015557446517050266,
-0.01674402691423893,
0.01479774434119463,
0.0007296916446648538,
-0.12148015201091766,
-0.03982071578502655,
0.05700419098138809,
0.014697432518005371,
-0.033729616552591324,
0.056766096502542496,
0.07237312197685242,
-0.012081771157681942,
0.03488331288099289,
-0.05227323994040489,
0.0015318036312237382,
0.04201861843466759,
0.020948557183146477,
0.06248890981078148,
0.007809734437614679,
0.017472347244620323,
0.0788227841258049,
0.0011843136744573712,
-0.04043474793434143,
0.06420005112886429,
0.01656562276184559,
0.12453359365463257,
0.0040295179933309555,
0.030138421803712845,
0.11427581310272217,
-0.03030047006905079,
-0.054097048938274384,
0.029834885150194168,
0.025923017412424088,
-0.11407020688056946,
0.11223912984132767,
0.002247676020488143,
0.009495416656136513,
-0.04405437037348747,
0.0020161112770438194,
0.0017782944487407804,
-0.10279711335897446,
0.025329604744911194,
-0.02892228402197361,
0.05109279230237007,
-0.015985000878572464,
-0.03480346128344536,
-0.004721153527498245,
-0.0815926194190979,
0.007742417510598898,
0.027930699288845062,
-0.043747399002313614,
-0.024286342784762383,
0.008523139171302319,
0.023677656427025795,
0.03834282606840134,
-3.5260517929769046e-33,
0.000030272705771494657,
0.028042446821928024,
-0.03458863124251366,
0.006176486145704985,
0.039496611803770065,
0.01279874611645937,
0.00528090400621295,
-0.066380076110363,
-0.02095896378159523,
-0.026842797175049782,
-0.04537764564156532,
0.12929196655750275,
0.01107865385711193,
-0.056062985211610794,
0.03365076705813408,
-0.03316187858581543,
-0.03329960256814957,
0.03861246258020401,
0.062100350856781006,
-0.05127380043268204,
0.04366791248321533,
0.11527827382087708,
-0.14571283757686615,
-0.07191962748765945,
0.04064851626753807,
-0.003660943591967225,
0.0346197634935379,
0.07149874418973923,
0.046111103147268295,
0.054562561213970184,
-0.027289357036352158,
-0.0063595883548259735,
-0.028240293264389038,
-0.042271409183740616,
0.024824922904372215,
0.1336892545223236,
0.08212515711784363,
0.041079938411712646,
-0.11588206887245178,
0.047690216451883316,
0.031970396637916565,
-0.061147186905145645,
-0.055441372096538544,
0.04035573825240135,
-0.0032827064860612154,
0.034824568778276443,
0.0006158766918815672,
0.007287712302058935,
0.011406440287828445,
0.07324355840682983,
-0.008363722823560238,
0.06339377164840698,
-0.11898814886808395,
-0.021878894418478012,
-0.009793135337531567,
-0.07540775835514069,
0.03471231088042259,
-0.012123804539442062,
0.034207798540592194,
-0.05878163129091263,
-0.013943750411272049,
-0.018889065831899643,
-0.07220754027366638,
0.02456493116915226,
-0.07196030020713806,
0.027186334133148193,
-0.02546033449470997,
0.022188197821378708,
-0.000004724061000160873,
0.04391578212380409,
-0.059624191373586655,
0.06296585500240326,
0.0664345845580101,
0.00008589955541538075,
0.08892270177602768,
-0.008644980378448963,
0.012559724040329456,
-0.04222385212779045,
-0.03084816038608551,
-0.05111560598015785,
-0.026055628433823586,
0.06876811385154724,
-0.012680653482675552,
0.005644675809890032,
0.05164095386862755,
0.04335483908653259,
-0.02531333453953266,
0.04976426810026169,
0.029380973428487778,
0.014854981563985348,
-0.015189108438789845,
-0.0160163976252079,
0.028063639998435974,
0.015017945319414139,
-0.06777579337358475,
-4.36734346465073e-8,
-0.0849621519446373,
0.05512865260243416,
0.015444628894329071,
-0.061481915414333344,
0.034960728138685226,
0.036652348935604095,
0.003927834331989288,
0.038132667541503906,
-0.012687561102211475,
-0.06535889953374863,
0.07140691578388214,
-0.014839713461697102,
0.020672764629125595,
-0.07000009715557098,
0.034788236021995544,
-0.01852668821811676,
0.04250238090753555,
-0.040300093591213226,
-0.022471845149993896,
-0.013049651868641376,
0.023778637871146202,
-0.014456192962825298,
-0.014625594951212406,
-0.03359809145331383,
-0.03913787379860878,
-0.08758587390184402,
0.03431277722120285,
0.04021183401346207,
0.04517322778701782,
-0.03959248960018158,
0.024080486968159676,
-0.008842346258461475,
0.052720703184604645,
0.03572581335902214,
0.002297287806868553,
0.06557643413543701,
-0.013997337780892849,
-0.011222761124372482,
-0.019865207374095917,
0.0021578131709247828,
-0.04077807068824768,
0.026984065771102905,
0.03170724958181381,
-0.032226890325546265,
-0.13086307048797607,
0.03595823422074318,
-0.04666582867503166,
-0.0894513726234436,
0.04164281487464905,
0.02535046637058258,
-0.005790225695818663,
0.0013850515242666006,
-0.06986493617296219,
0.0910022184252739,
0.0894966647028923,
-0.04420312121510506,
-0.03291214630007744,
-0.08052608370780945,
0.04733358696103096,
0.06740908324718475,
-0.016127020120620728,
0.04222865402698517,
-0.060964908450841904,
-0.03675498440861702
] |
Compositional 4D Dynamic Scenes Understanding with Physics Priors for Video Question Answering | https://openreview.net/forum?id=6Vx28LSR7f | [
"Xingrui Wang",
"Wufei Ma",
"Angtian Wang",
"Shuo Chen",
"Adam Kortylewski",
"Alan Yuille"
] | Poster | For vision-language models (VLMs), understanding the dynamic properties of objects and their interactions in 3D scenes from videos is crucial for effective reasoning about high-level temporal and action semantics. Although humans are adept at understanding these properties by constructing 3D and temporal (4D) representations of the world, current video understanding models struggle to extract these dynamic semantics, arguably because these models use cross-frame reasoning without underlying knowledge of the 3D/4D scenes.
In this work, we introduce **DynSuperCLEVR**, the first video question answering dataset that focuses on language understanding of the dynamic properties of 3D objects. We concentrate on three physical concepts—*velocity*, *acceleration*, and *collisions*—within 4D scenes. We further generate three types of questions, including factual queries, future predictions, and counterfactual reasoning that involve different aspects of reasoning on these 4D dynamic properties.
To further demonstrate the importance of explicit scene representations in answering these 4D dynamics questions, we propose **NS-4DPhysics**, a **N**eural-**S**ymbolic VideoQA model integrating **Physics** prior for **4D** dynamic properties with explicit scene representation of videos.
Instead of answering the questions directly from the video text input, our method first estimates the 4D world states with a 3D generative model powered by a physical prior, and then uses neural symbolic reasoning to answer the questions based on the 4D world states.
Our evaluation on all three types of questions in DynSuperCLEVR shows that previous video question answering models and large multimodal models struggle with questions about 4D dynamics, while our NS-4DPhysics significantly outperforms previous state-of-the-art models. | Video question answering, Compositional reasoning, Physical scene understanding, 3D scene understanding | We introduce DynSuperCLEVR, a video question answering dataset focused on the dynamic properties of 3D objects. We propose NS-4DPhysics, which use 4D world states with a 3D generative model and uses neural symbolic reasoning to answer questions. | 479 | 2406.00622 | [
0.00001107115804188652,
-0.10614871978759766,
0.04081565886735916,
-0.011854488402605057,
0.05699126422405243,
0.0007584859849885106,
-0.014206483960151672,
-0.02211192436516285,
0.07836642116308212,
-0.06023736670613289,
-0.053681060671806335,
-0.13677465915679932,
-0.07153414189815521,
0.11511822789907455,
-0.025164460763335228,
-0.005596666596829891,
0.06613950431346893,
0.07087884098291397,
-0.1578776240348816,
0.013791014440357685,
0.006442544981837273,
-0.013939684256911278,
0.10055046528577805,
0.040384020656347275,
-0.00316207529976964,
0.022845806553959846,
0.024073798209428787,
-0.058636002242565155,
0.012302839197218418,
0.023487035185098648,
0.08485410362482071,
0.02270820364356041,
-0.022805223241448402,
0.04870885983109474,
-0.02214912511408329,
0.04991161450743675,
-0.003528234316036105,
0.012932397425174713,
-0.07822586596012115,
-0.03490111976861954,
-0.023503120988607407,
0.012526138685643673,
0.05929921194911003,
0.04031024128198624,
0.08319273591041565,
0.07790088653564453,
0.002177133923396468,
-0.041067931801080704,
0.0024030045606195927,
-0.0015283151296898723,
-0.07654082775115967,
-0.025176942348480225,
-0.01036466658115387,
-0.008214187808334827,
0.0002503113937564194,
0.04615333676338196,
0.048684608191251755,
-0.018599344417452812,
0.0455072857439518,
-0.0006512916879728436,
0.058422550559043884,
-0.08123590797185898,
0.05082353577017784,
0.043697718530893326,
0.0029798441100865602,
-0.021430686116218567,
0.008302274160087109,
0.0023732336703687906,
-0.04951066896319389,
0.02357923798263073,
-0.06719302386045456,
0.07048603147268295,
0.01377437636256218,
-0.03569399192929268,
0.00763965817168355,
0.0185701847076416,
0.01505323126912117,
-0.10473395884037018,
0.012795292772352695,
-0.07680822908878326,
0.07532637566328049,
-0.020683176815509796,
0.03687267750501633,
0.0315101183950901,
0.07740098983049393,
0.04521816596388817,
0.014940321445465088,
0.08631481975317001,
-0.05681845173239708,
-0.007680295966565609,
-0.07878964394330978,
-0.09840673953294754,
0.007028116844594479,
-0.0025114568416029215,
0.07305198907852173,
0.03222880885004997,
-0.010495183058083057,
-0.13018013536930084,
0.04151373356580734,
0.025390248745679855,
0.04025690257549286,
0.0247774887830019,
-0.020262792706489563,
0.045125026255846024,
0.05051013082265854,
0.015990935266017914,
0.04102233052253723,
-0.0010954451281577349,
-0.023476621136069298,
0.005604906473308802,
0.027412831783294678,
-0.0005080477567389607,
0.04968925565481186,
0.03411222621798515,
-0.046287860721349716,
-0.013797832652926445,
0.04879216104745865,
-0.008642499335110188,
-0.010489602573215961,
-0.06921570003032684,
-0.0033835386857390404,
0.025469066575169563,
0.02464260719716549,
0.023143837228417397,
0.07108572870492935,
-0.11149583011865616,
0.026786338537931442,
-1.801813352974749e-33,
0.04550940543413162,
-0.02127837762236595,
0.06539787352085114,
0.039869166910648346,
0.03926636278629303,
0.011995820328593254,
-0.017668385058641434,
0.057996369898319244,
0.012821105308830738,
0.00035039978683926165,
-0.053280264139175415,
-0.049466945230960846,
-0.028248002752661705,
0.07303225994110107,
-0.03351353853940964,
0.0062175048515200615,
-0.055792186409235,
-0.0023014035541564226,
-0.07227648049592972,
0.03355160728096962,
0.03398701548576355,
0.04844161495566368,
-0.03207637369632721,
0.04240185394883156,
-0.026258638128638268,
-0.008898002095520496,
0.04693865403532982,
-0.029715057462453842,
0.0005967799224890769,
-0.044491417706012726,
-0.14419987797737122,
-0.053573135286569595,
-0.06923981755971909,
-0.005953936371952295,
0.04794949293136597,
-0.001771321869455278,
-0.06051457300782204,
-0.0676252692937851,
-0.026941731572151184,
-0.031934622675180435,
-0.037356749176979065,
0.022882800549268723,
-0.07636905461549759,
-0.02615809626877308,
-0.09276007115840912,
-0.03470270708203316,
0.04176635295152664,
0.09118486940860748,
-0.11051034927368164,
0.011115179397165775,
0.052243657410144806,
0.018115295097231865,
-0.019775710999965668,
-0.15772747993469238,
-0.010961481370031834,
-0.03388253226876259,
0.045974548906087875,
-0.03769101947546005,
0.037121549248695374,
0.010730641894042492,
-0.02321482077240944,
0.029021067544817924,
-0.005481165833771229,
0.013250438496470451,
-0.052837859839200974,
0.028878269717097282,
-0.02007627859711647,
0.10518477857112885,
0.06608450412750244,
0.03522614762187004,
-0.024968095123767853,
0.02072511613368988,
0.04097547009587288,
-0.045638132840394974,
0.06784827262163162,
0.06576390564441681,
-0.0064980946481227875,
-0.082105353474617,
0.00851298589259386,
0.0719693973660469,
-0.04146706685423851,
-0.053179699927568436,
-0.03269325941801071,
0.009061657823622227,
-0.02861095778644085,
-0.02558687888085842,
0.059901922941207886,
-0.1169712021946907,
0.01660708151757717,
-0.08053535968065262,
-0.04563844949007034,
-0.12832914292812347,
-0.02102610655128956,
-0.042787160724401474,
0.011160413734614849,
-4.094958089831771e-34,
0.02092808112502098,
-0.015942407771945,
-0.10245657712221146,
0.027337556704878807,
-0.01711636781692505,
-0.05247543379664421,
-0.055216189473867416,
0.05921371281147003,
0.03768741711974144,
-0.07526720315217972,
-0.00624779611825943,
-0.030929923057556152,
-0.04152463376522064,
0.007908237166702747,
0.018520064651966095,
-0.0024131035897880793,
-0.05999051406979561,
-0.057672563940286636,
-0.01242018211632967,
0.06127246841788292,
0.027836386114358902,
0.07854710519313812,
-0.08381466567516327,
-0.025434840470552444,
-0.02907465025782585,
0.08217616379261017,
0.06903026252985,
-0.0041589136235415936,
-0.002559868386015296,
0.024270137771964073,
-0.027332328259944916,
-0.058838892728090286,
0.0024579239543527365,
0.0027257425244897604,
-0.1157434806227684,
0.035054177045822144,
0.06256560236215591,
-0.08942347019910812,
-0.05766846239566803,
0.006513276137411594,
0.05267229676246643,
-0.027671577408909798,
-0.046108536422252655,
0.05153830349445343,
-0.07792426645755768,
0.0029214429669082165,
-0.03862195461988449,
0.05972043052315712,
-0.016656246036291122,
0.05717465281486511,
-0.028003409504890442,
0.028326937928795815,
-0.050982438027858734,
0.018932407721877098,
0.015927055850625038,
-0.019939133897423744,
-0.06048380956053734,
-0.010844046249985695,
0.06434127688407898,
-0.001038881717249751,
-0.026168381795287132,
-0.010761409066617489,
0.015542596578598022,
0.02291390299797058,
0.011560955084860325,
-0.014707491733133793,
-0.10727152973413467,
-0.026638591662049294,
0.016728300601243973,
-0.01988804154098034,
0.03841683268547058,
-0.031137971207499504,
-0.01606982946395874,
-0.003820920828729868,
0.08112158626317978,
-0.03991745784878731,
0.007049637380987406,
-0.024720890447497368,
0.04427843540906906,
-0.009849147871136665,
-0.03305288031697273,
-0.06840875744819641,
0.14004835486412048,
0.1523260623216629,
0.0011950150365009904,
0.00583383534103632,
0.017371609807014465,
0.051614176481962204,
0.043055757880210876,
0.02812311053276062,
-0.010026526637375355,
0.02899157628417015,
0.01503577921539545,
0.10662420094013214,
-0.07452566176652908,
-3.588933239484504e-8,
-0.09632249921560287,
0.1054583489894867,
0.029880572110414505,
0.017722608521580696,
-0.06425764411687851,
-0.046224307268857956,
-0.026401188224554062,
0.09013628959655762,
-0.01906014233827591,
-0.036045558750629425,
0.09617388248443604,
-0.015430512838065624,
-0.020669106394052505,
-0.02764078415930271,
-0.011206163093447685,
0.10822676122188568,
0.013044092804193497,
-0.041900768876075745,
-0.002963864943012595,
0.026169193908572197,
0.03214230388402939,
0.021075012162327766,
-0.006589843425899744,
0.023288529366254807,
-0.016445478424429893,
0.06597597151994705,
-0.04586273431777954,
0.10128846764564514,
0.0799342691898346,
-0.03278873860836029,
0.03484161198139191,
0.0005132290534675121,
0.013454820960760117,
-0.06477836519479752,
-0.04216272756457329,
0.022935308516025543,
-0.005542943254113197,
-0.04020867496728897,
0.03892907127737999,
-0.009382142685353756,
-0.013624248094856739,
0.060602281242609024,
-0.02991887927055359,
0.07869179546833038,
0.031025489792227745,
0.02239920012652874,
-0.06213238462805748,
-0.08088761568069458,
0.02557392045855522,
0.0200483538210392,
-0.0873875841498375,
0.0007784849149174988,
-0.017407584935426712,
0.1446049064397812,
0.06115065887570381,
0.03193851560354233,
-0.008087551221251488,
0.008860450237989426,
-0.0043678744696080685,
0.005203303415328264,
0.00047043399536050856,
0.039052095264196396,
-0.003622551914304495,
-0.002093597548082471
] |
UGMathBench: A Diverse and Dynamic Benchmark for Undergraduate-Level Mathematical Reasoning with Large Language Models | https://openreview.net/forum?id=fovPyqPcKY | [
"Xin Xu",
"Jiaxin ZHANG",
"Tianhao Chen",
"Zitong Chao",
"Jishan Hu",
"Can Yang"
] | Poster | Large Language Models (LLMs) have made significant strides in mathematical reasoning, underscoring the need for a comprehensive and fair evaluation of their capabilities. However, existing benchmarks often fall short, either lacking extensive coverage of undergraduate-level mathematical problems or probably suffering from test-set contamination. To address these issues, we introduce UGMathBench, a diverse and dynamic benchmark specifically designed for evaluating undergraduate-level mathematical reasoning with LLMs. UGMathBench comprises 5,062 problems across 16 subjects and 111 topics, featuring 10 distinct answer types. Each problem includes three randomized versions, with additional versions planned for release as the leading open-source LLMs become saturated in UGMathBench. Furthermore, we propose two key metrics: effective accuracy (EAcc), which measures the percentage of correctly solved problems across all three versions, and reasoning gap ($\Delta$), which assesses reasoning robustness by calculating the difference between the average accuracy across all versions and EAcc. Our extensive evaluation of 23 leading LLMs reveals that the highest EAcc achieved is 56.3% by OpenAI-o1-mini, with large $\Delta$ values observed across different models. This highlights the need for future research aimed at developing "large reasoning models" with high EAcc and $\Delta = 0$. We anticipate that the release of UGMathBench, along with its detailed evaluation codes, will serve as a valuable resource to advance the development of LLMs in solving mathematical problems. | Math Reasoning, Undergraduate-Level Math problems, Benchmark | We introduce a diverse and dynamic Benchmark for undergraduate-level mathematical reasoning with large language models | 478 | 2501.13766 | [
-0.026002192869782448,
-0.027667591348290443,
0.012427561916410923,
0.031033633276820183,
0.07148151099681854,
-0.056251928210258484,
-0.01925753429532051,
0.06696568429470062,
-0.05081643909215927,
-0.014111651107668877,
-0.06826874613761902,
-0.06292460113763809,
0.03308236226439476,
0.060636233538389206,
0.023949824273586273,
0.038099225610494614,
0.0866074487566948,
0.03368847072124481,
-0.0578664094209671,
-0.0771615132689476,
-0.05617256835103035,
-0.015184837393462658,
0.08282942324876785,
-0.05335084721446037,
0.0670279785990715,
-0.03816935047507286,
-0.005388522520661354,
-0.05545118823647499,
0.0027614368591457605,
0.008899138309061527,
0.006838807370513678,
-0.00016210602188948542,
-0.020135344937443733,
0.005935541354119778,
0.0063440315425395966,
0.02737082540988922,
-0.04749477282166481,
-0.011910198256373405,
-0.021707894280552864,
0.03512948378920555,
-0.0948173925280571,
0.013085435144603252,
0.09398896247148514,
-0.030906233936548233,
0.07864135503768921,
-0.04111293703317642,
-0.04265526682138443,
0.03280356153845787,
-0.021792171522974968,
0.011873623356223106,
-0.07430437207221985,
-0.031180547550320625,
-0.08390502631664276,
0.011881696991622448,
-0.021633924916386604,
-0.030242519453167915,
-0.0021060805302113295,
-0.005113642197102308,
-0.021177954971790314,
-0.08004452288150787,
-0.08041055500507355,
-0.08006690442562103,
-0.03297790512442589,
-0.004343893378973007,
-0.019218649715185165,
-0.03756913170218468,
0.006022714078426361,
0.036936428397893906,
-0.01563589833676815,
0.12443483620882034,
-0.02430817298591137,
0.10145091265439987,
-0.04689773544669151,
0.10830867290496826,
0.035798076540231705,
0.04927479103207588,
0.051952771842479706,
0.005589161068201065,
0.06810778379440308,
-0.023861320689320564,
0.005416146945208311,
-0.03197266906499863,
-0.03165653720498085,
0.01885930635035038,
0.1518981158733368,
-0.014453770592808723,
0.020456023514270782,
0.030189525336027145,
0.027397949248552322,
-0.006312913727015257,
0.03048773854970932,
-0.05177390202879906,
0.04066424444317818,
0.10660439729690552,
0.0738040953874588,
0.03387856110930443,
0.06766287982463837,
-0.00803255196660757,
-0.04739149287343025,
0.09967272728681564,
0.009072624146938324,
0.1118900403380394,
-0.00043145864037796855,
-0.09029068797826767,
-0.022138599306344986,
0.00252651353366673,
0.12378347665071487,
0.02938184142112732,
0.032448090612888336,
-0.10090073943138123,
0.07395333796739578,
0.013432188890874386,
0.03431587666273117,
0.05375765636563301,
0.017828602343797684,
-0.0518447682261467,
0.091690793633461,
-0.03509730473160744,
0.02832181379199028,
0.023590220138430595,
-0.005895985756069422,
0.06548541784286499,
0.05403508245944977,
-0.01928812451660633,
0.054851505905389786,
0.026266366243362427,
-0.013236471451818943,
3.4977165016503334e-33,
0.09567185491323471,
0.06644505262374878,
0.024342842400074005,
0.08535295724868774,
0.026394100859761238,
0.03417444974184036,
0.03046455793082714,
0.08647417277097702,
-0.06894207000732422,
0.05843348428606987,
0.01049141027033329,
0.01375209167599678,
-0.08209636062383652,
0.02695663645863533,
0.10758482664823532,
-0.01374437753111124,
-0.04024936258792877,
0.01429017074406147,
-0.061985988169908524,
-0.032723814249038696,
0.04936308413743973,
-0.08923525363206863,
-0.008300386369228363,
-0.04435879364609718,
-0.0019875718280673027,
0.01586165465414524,
0.05595225468277931,
-0.02304925210773945,
-0.013478734530508518,
0.004869809839874506,
-0.087009496986866,
-0.06619652360677719,
-0.07331707328557968,
0.07342217862606049,
-0.019113998860120773,
-0.03831479325890541,
-0.039840202778577805,
-0.04762089625000954,
0.010714911855757236,
-0.06543136388063431,
-0.04736723750829697,
0.0423029288649559,
0.06076674908399582,
0.03184157609939575,
0.007727373391389847,
-0.020126458257436752,
-0.05896009877324104,
0.034639857709407806,
0.022504519671201706,
-0.08023303002119064,
-0.03859223797917366,
0.006659246049821377,
-0.03831985965371132,
-0.024450216442346573,
-0.011733738705515862,
-0.00762925622984767,
-0.008905448019504547,
0.04203382879495621,
0.0029698435682803392,
0.07680228352546692,
0.003264556871727109,
0.08546964824199677,
0.019131671637296677,
0.033098045736551285,
0.009233878925442696,
0.08884347975254059,
-0.06480955332517624,
0.07887071371078491,
0.07788453251123428,
0.033839140087366104,
-0.04248299077153206,
-0.06500498950481415,
0.013326476328074932,
-0.015034577809274197,
0.01089558843523264,
0.0016970743890851736,
0.05857911705970764,
-0.06043373420834541,
0.06502000242471695,
-0.03742613643407822,
0.05374840274453163,
0.00812853965908289,
-0.03184878081083298,
-0.05304647982120514,
-0.0730612501502037,
-0.05458558350801468,
0.03494211286306381,
-0.05710940062999725,
-0.01693183369934559,
-0.0393839031457901,
0.026524996384978294,
-0.060062989592552185,
0.01171062607318163,
-0.05719541385769844,
0.014301782473921776,
-3.0697810893113413e-33,
-0.055566783994436264,
-0.021140940487384796,
-0.06242340803146362,
0.08288577198982239,
-0.03467473015189171,
-0.09529386460781097,
-0.03505941852927208,
0.056903764605522156,
-0.021789226680994034,
-0.049304619431495667,
-0.009883350692689419,
-0.040994808077812195,
-0.0010170824825763702,
0.02176704630255699,
0.00999725516885519,
-0.08510537445545197,
-0.035366445779800415,
-0.011910675093531609,
-0.013891851529479027,
0.12097669392824173,
-0.03361836075782776,
0.07281884551048279,
-0.06057128682732582,
-0.009261978790163994,
-0.07448546588420868,
0.0558050163090229,
-0.005609499756246805,
0.02444808930158615,
-0.03315935656428337,
0.006541768554598093,
0.0064887735061347485,
-0.04054466262459755,
-0.10178955644369125,
-0.013886683620512486,
-0.07933826744556427,
-0.0078041525557637215,
0.03336392343044281,
-0.01549514103680849,
0.034062646329402924,
0.04027223214507103,
0.080256387591362,
-0.0967479720711708,
-0.020536750555038452,
-0.018400149419903755,
0.014986858703196049,
0.07774447649717331,
-0.08794528245925903,
0.038281697779893875,
-0.041412342339754105,
0.006367409601807594,
-0.06000729277729988,
-0.09061536192893982,
-0.0015973143745213747,
0.05140192061662674,
0.03904050216078758,
-0.024239933118224144,
-0.0616479218006134,
-0.046105291694402695,
-0.023558607324957848,
0.048824626952409744,
-0.05707617849111557,
0.09444113820791245,
0.035190705209970474,
0.0040444983169436455,
-0.008214824832975864,
-0.052727751433849335,
-0.0649399682879448,
-0.006063106935471296,
-0.05076923221349716,
0.04606432095170021,
-0.0195144172757864,
-0.0661126971244812,
-0.030263187363743782,
-0.006445551756769419,
0.04552360251545906,
0.07202643901109695,
0.03534339368343353,
0.021708866581320763,
-0.021529298275709152,
-0.12329670786857605,
0.06027203053236008,
-0.06244315207004547,
0.04719496890902519,
0.017451593652367592,
0.005280730314552784,
-0.051085878163576126,
0.0691843181848526,
0.09376861155033112,
0.06926558911800385,
0.10192815959453583,
-0.025560403242707253,
0.035312753170728683,
-0.02126375213265419,
-0.0050506144762039185,
-0.05736250802874565,
-4.438932776906768e-8,
-0.02193574234843254,
0.09579570591449738,
-0.021701056510210037,
0.035893864929676056,
-0.03468465432524681,
-0.04502326622605324,
-0.09795963764190674,
0.029711078852415085,
-0.027068402618169785,
0.03271016851067543,
0.05521421134471893,
-0.01832438074052334,
-0.07219358533620834,
-0.023237576708197594,
-0.04627631604671478,
0.14548711478710175,
-0.01671718619763851,
0.017632408067584038,
-0.029940949752926826,
-0.051753632724285126,
0.018608639016747475,
0.053491994738578796,
0.03036166913807392,
0.015417385846376419,
-0.034760843962430954,
-0.031204819679260254,
0.013553336262702942,
0.13709475100040436,
0.053825266659259796,
-0.044302746653556824,
-0.019980773329734802,
-0.003318448318168521,
0.02721022628247738,
-0.048546694219112396,
0.06288739293813705,
0.049980197101831436,
-0.0408153273165226,
-0.0012718780199065804,
0.051068343222141266,
-0.015005647204816341,
0.0218143742531538,
0.0219709575176239,
-0.044294573366642,
0.027008645236492157,
0.05496026948094368,
-0.02893199399113655,
-0.07283328473567963,
-0.02099788375198841,
-0.014099393971264362,
-0.011470588855445385,
-0.06067369133234024,
-0.055566199123859406,
-0.04449491947889328,
-0.028344793245196342,
0.018933217972517014,
0.06084860488772392,
-0.09294039011001587,
-0.018037734553217888,
-0.040142785757780075,
-0.012142329476773739,
0.0505913607776165,
0.02708279900252819,
-0.021241573616862297,
0.05642594397068024
] |
Can LLMs Solve Longer Math Word Problems Better? | https://openreview.net/forum?id=C9ju8QQSCv | [
"Xin Xu",
"Tong Xiao",
"Zitong Chao",
"Zhenya Huang",
"Can Yang",
"Yang Wang"
] | Poster | Math Word Problems (MWPs) play a vital role in assessing the capabilities of Large Language Models (LLMs), yet current research primarily focuses on questions with concise contexts. The impact of longer contexts on mathematical reasoning remains under-explored. This study pioneers the investigation of Context Length Generalizability (CoLeG), which refers to the ability of LLMs to solve MWPs with extended narratives. We introduce Extended Grade-School Math (E-GSM), a collection of MWPs featuring lengthy narratives, and propose two novel metrics to evaluate the efficacy and resilience of LLMs in tackling these problems. Our analysis of existing zero-shot prompting techniques with proprietary LLMs along with open-source LLMs reveals a general deficiency in CoLeG. To alleviate these issues, we propose tailored approaches for different categories of LLMs. For proprietary LLMs, we introduce a new instructional prompt designed to mitigate the impact of long contexts. For open-source LLMs, we develop a novel auxiliary task for fine-tuning to enhance CoLeG. Our comprehensive results demonstrate the effectiveness of our proposed methods, showing improved performance on E-GSM. Additionally, we conduct an in-depth analysis to differentiate the effects of semantic understanding and reasoning efficacy, showing that our methods improves the latter. We also establish the generalizability of our methods across several other MWP benchmarks. Our findings highlight the limitations of current LLMs and offer practical solutions correspondingly, paving the way for further exploration of model generalizability and training methodologies. | Large Language Models, Math Reasoning, Long Math Word Problems | We investigate the impact of long contexts on mathematical reasoning abilities of large language models. | 477 | 2405.14804 | [
0.0145066287368536,
0.036224424839019775,
0.07080629467964172,
-0.01872117631137371,
0.02764519862830639,
-0.034090083092451096,
-0.015938779339194298,
0.059258949011564255,
0.01938546635210514,
-0.016080062836408615,
-0.03623592108488083,
-0.03180652856826782,
0.05483701825141907,
0.046242304146289825,
0.1230541542172432,
0.04463133588433266,
0.0799475908279419,
0.06321731209754944,
-0.10410835593938828,
-0.03026464208960533,
0.029966799542307854,
-0.019477348774671555,
0.04697471112012863,
0.008023450151085854,
0.0025388384237885475,
-0.045913055539131165,
-0.04824864864349365,
-0.09913235902786255,
0.06306789815425873,
0.01790771633386612,
-0.022867856547236443,
0.0710533931851387,
0.01758570782840252,
0.03845454007387161,
0.002571913180872798,
0.04822700098156929,
-0.016554642468690872,
0.09366986900568008,
-0.04582037031650543,
-0.02330877259373665,
-0.08810718357563019,
-0.02516804449260235,
0.06596191227436066,
0.04223250597715378,
0.04353920370340347,
-0.06458775699138641,
-0.02718321606516838,
0.029559332877397537,
-0.0546562634408474,
-0.014539992436766624,
-0.08715175837278366,
-0.053557462990283966,
-0.0834706723690033,
-0.0034714634530246258,
0.021809084340929985,
-0.006652419921010733,
0.011850422248244286,
0.04897448420524597,
0.012779772281646729,
-0.06113503500819206,
-0.10480312258005142,
-0.081511490046978,
-0.09643522650003433,
0.049463558942079544,
-0.021660828962922096,
-0.007941368035972118,
0.00712854927405715,
0.08889815956354141,
-0.04168260470032692,
0.12574118375778198,
-0.03452164679765701,
0.07075735926628113,
-0.026789851486682892,
0.13265149295330048,
0.011677254922688007,
0.0672454684972763,
-0.018334360793232918,
-0.009170325472950935,
0.068192258477211,
0.007615399546921253,
0.0036430945619940758,
-0.04741581156849861,
-0.003153854748234153,
0.012247556820511818,
0.1019480749964714,
-0.002772295381873846,
0.033882055431604385,
0.037237875163555145,
0.03420872986316681,
0.019156599417328835,
-0.0013376690912991762,
-0.14897041022777557,
0.020841063931584358,
0.06525104492902756,
0.032702282071113586,
0.012990659102797508,
-0.012097354047000408,
-0.07962054014205933,
-0.046025898307561874,
0.09359024465084076,
0.013338950462639332,
0.07440457493066788,
0.019733300432562828,
-0.054251570254564285,
-0.0517144650220871,
0.026260050013661385,
0.0049096886068582535,
0.03979085385799408,
0.020119043067097664,
-0.12486013770103455,
0.06132153794169426,
0.050098203122615814,
0.02231447398662567,
0.007437918800860643,
0.0036888914182782173,
-0.0325290784239769,
0.059836495667696,
-0.06568504124879837,
0.060213446617126465,
-0.002522979862987995,
0.060559917241334915,
0.0689474418759346,
-0.008269434794783592,
-0.04484092816710472,
0.029905183240771294,
0.022662783041596413,
-0.03736918419599533,
1.7808489622408106e-33,
0.056201472878456116,
0.07706200331449509,
0.014795687049627304,
0.08712079375982285,
0.08842230588197708,
0.03668573871254921,
-0.00021834451763425022,
0.11714120209217072,
-0.05543774738907814,
0.07037779688835144,
0.07606734335422516,
0.04831288754940033,
-0.048005539923906326,
0.004677104763686657,
0.114191934466362,
0.014372305013239384,
-0.03620828688144684,
0.03137694299221039,
-0.04741619899868965,
-0.03518495708703995,
-0.020632212981581688,
-0.07899320870637894,
0.010470789857208729,
-0.017597565427422523,
0.03314954414963722,
0.05573029816150665,
0.07574037462472916,
-0.07086554914712906,
-0.00596682308241725,
-0.005441613029688597,
-0.08376118540763855,
-0.05309779569506645,
-0.04115447774529457,
0.05722641572356224,
0.003695591352880001,
-0.05580665171146393,
-0.019377414137125015,
-0.06854214519262314,
0.033585358411073685,
-0.016980178654193878,
-0.03477681428194046,
-0.008646944537758827,
0.028249332681298256,
0.0794733390212059,
0.010848493315279484,
-0.047204338014125824,
-0.008562620729207993,
0.03087049536406994,
-0.03668937832117081,
0.02157219685614109,
0.04146313667297363,
0.023308435454964638,
-0.09507021307945251,
-0.05230633541941643,
-0.012453119270503521,
0.030593855306506157,
-0.02855294942855835,
0.009160256013274193,
-0.037626735866069794,
0.024040939286351204,
0.06722663342952728,
0.09566355496644974,
0.03105895034968853,
0.016626106575131416,
0.009087273851037025,
0.06679140776395798,
-0.12020737677812576,
0.043492939323186874,
0.10957341641187668,
-0.020051071420311928,
-0.02898523397743702,
-0.06100969761610031,
-0.03126497566699982,
-0.005163566675037146,
0.03386956453323364,
-0.048166584223508835,
0.03765788674354553,
-0.05260249972343445,
0.017573345452547073,
0.016167176887392998,
0.09043444693088531,
-0.0034474364947527647,
0.006429542787373066,
-0.014423316344618797,
-0.04874183610081673,
-0.08559368550777435,
0.019184326753020287,
-0.08918801695108414,
-0.03968741372227669,
-0.05286889523267746,
-0.036837734282016754,
-0.013441330753266811,
-0.06215815618634224,
0.016961248591542244,
0.003338854294270277,
-2.1240637416511557e-33,
-0.0367458276450634,
-0.05306337773799896,
-0.05345390737056732,
0.07265330851078033,
-0.02842579036951065,
-0.06757320463657379,
-0.0030702543444931507,
0.03880235552787781,
0.031098265200853348,
-0.09363848716020584,
-0.07804010808467865,
-0.002350639784708619,
0.007648764178156853,
-0.008671700954437256,
-0.0018817849922925234,
-0.05887843668460846,
-0.029045509174466133,
-0.024841878563165665,
-0.03155532851815224,
0.09218905121088028,
0.021824736148118973,
0.08825413882732391,
-0.10768651962280273,
0.028110018000006676,
-0.007656407076865435,
0.009765302762389183,
-0.07485111057758331,
-0.003325595986098051,
-0.04246942326426506,
0.056849926710128784,
-0.003215229604393244,
-0.06263232976198196,
-0.046165306121110916,
-0.038705527782440186,
-0.07774239778518677,
0.01836962066590786,
0.06962775439023972,
0.020055480301380157,
-0.019723230972886086,
0.038186296820640564,
0.08113830536603928,
-0.09602104127407074,
0.037923384457826614,
-0.025546744465827942,
-0.008356242440640926,
0.03404702618718147,
-0.07031712681055069,
-0.023439789190888405,
0.07137341797351837,
0.0468720979988575,
-0.09452980011701584,
-0.05176999792456627,
-0.011922786943614483,
0.005913225002586842,
-0.03708012402057648,
-0.057540107518434525,
-0.040969137102365494,
-0.10593642294406891,
0.013781103305518627,
0.001795304357074201,
-0.033117346465587616,
0.02504672296345234,
0.04978944733738899,
-0.06630347669124603,
0.0701412484049797,
-0.07182937860488892,
-0.06339844316244125,
0.04217822477221489,
-0.06098923087120056,
-0.030434511601924896,
-0.008290446363389492,
0.02351524494588375,
-0.05432559549808502,
-0.023107361048460007,
0.0036503993906080723,
0.10178765654563904,
-0.015354140661656857,
-0.08054974675178528,
-0.04019223898649216,
-0.07231778651475906,
0.06906572729349136,
-0.01043732464313507,
0.0023794586304575205,
0.029509106650948524,
-0.03927065059542656,
-0.00007146990537876263,
0.09234650433063507,
0.15116813778877258,
0.028082547709345818,
0.044342588633298874,
-0.0011182265589013696,
0.0606994591653347,
0.009890860877931118,
0.056363340467214584,
-0.07530516386032104,
-4.726602043092498e-8,
-0.049419619143009186,
0.04107619822025299,
-0.025621965527534485,
0.023848822340369225,
0.05264393240213394,
0.002938553225249052,
-0.051221396774053574,
0.03807898983359337,
-0.010023120790719986,
0.01810375042259693,
0.033751364797353745,
0.022692540660500526,
-0.07609477639198303,
-0.011992606334388256,
-0.02683170698583126,
0.04082771763205528,
-0.016926296055316925,
-0.01910887286067009,
-0.05186045169830322,
-0.00910700112581253,
0.07324671745300293,
0.056622106581926346,
-0.04147883877158165,
0.02083401195704937,
-0.03407075256109238,
-0.03793436288833618,
-0.016624776646494865,
0.12069007754325867,
0.07238680869340897,
-0.02682514302432537,
-0.006242384668439627,
0.03982431814074516,
-0.012066352181136608,
0.0067571913823485374,
-0.013140945695340633,
0.028655046597123146,
-0.03107842244207859,
0.03580398112535477,
0.0699760690331459,
-0.029193352907896042,
-0.030362309888005257,
-0.03015192411839962,
-0.019647488370537758,
0.02482905425131321,
-0.0001654073130339384,
0.018600856885313988,
-0.019565606489777565,
-0.08458399772644043,
-0.05282926186919212,
0.0316888689994812,
-0.004351045470684767,
-0.047128867357969284,
-0.052001889795064926,
-0.028410090133547783,
0.05022721365094185,
0.007924377918243408,
0.0074371532537043095,
0.0059692589566111565,
-0.03192858770489693,
-0.006222814787179232,
0.01770152524113655,
0.06299517303705215,
-0.032183751463890076,
0.05345321074128151
] |
Progressive Token Length Scaling in Transformer Encoders for Efficient Universal Segmentation | https://openreview.net/forum?id=dmzM5UdAq6 | [
"Abhishek Aich",
"Yumin Suh",
"Samuel Schulter",
"Manmohan Chandraker"
] | Poster | A powerful architecture for universal segmentation relies on transformers that encode multi-scale image features and decode object queries into mask predictions. With efficiency being a high priority for scaling such models, we observed that the state-of-the-art method Mask2Former uses \~50% of its compute only on the transformer encoder. This is due to the retention of a full-length token-level representation of all backbone feature scales at each encoder layer. With this observation, we propose a strategy termed PROgressive Token Length SCALing for Efficient transformer encoders (PRO-SCALE) that can be plugged-in to the Mask2Former segmentation architecture to significantly reduce the computational cost. The underlying principle of PRO-SCALE is: progressively scale the length of the tokens with the layers of the encoder. This allows PRO-SCALE to reduce computations by a large margin with minimal sacrifice in performance (\~52% GFLOPs reduction with no drop in performance on COCO dataset). We validate our framework on multiple public benchmarks. Our code will be publicly released. | universal segmentation; efficient transformers | We propose an efficient transformer encoder for universal segmentation architectures. | 476 | 2404.14657 | [
-0.014012573286890984,
-0.018281232565641403,
0.02817687764763832,
0.03289343789219856,
0.05138888955116272,
-0.05394526943564415,
-0.07692009210586548,
0.016370922327041626,
-0.007457379251718521,
-0.07492674142122269,
-0.016711117699742317,
-0.07293324172496796,
-0.05308876931667328,
0.045880287885665894,
-0.017711546272039413,
0.004196220077574253,
0.012780139222741127,
0.1792505383491516,
-0.13352030515670776,
-0.08843863010406494,
0.0209474079310894,
-0.05619146302342415,
-0.030545396730303764,
-0.0007201553671620786,
0.062300387769937515,
0.0019507957622408867,
-0.038221392780542374,
-0.10901571810245514,
0.062281426042318344,
0.00419255904853344,
-0.014588331803679466,
-0.039615169167518616,
0.055900316685438156,
0.05284808948636055,
-0.021841073408722878,
0.023882877081632614,
0.0064022717997431755,
-0.05435614287853241,
0.015493803657591343,
-0.08438793569803238,
0.014607938006520271,
0.0078002712689340115,
0.0034242100082337856,
0.015742214396595955,
0.0426008366048336,
0.03666236251592636,
0.06007823348045349,
-0.03849365562200546,
-0.008743534795939922,
-0.033764202147722244,
0.05245031788945198,
-0.026702474802732468,
-0.07526690512895584,
0.0741683766245842,
-0.03691418468952179,
-0.04510059952735901,
0.047612905502319336,
-0.031970955431461334,
0.03800574690103531,
0.040263671427965164,
-0.0745793804526329,
-0.04401431605219841,
0.03461503982543945,
0.04359455406665802,
-0.0747281089425087,
-0.03142978250980377,
0.09787055104970932,
-0.07048771530389786,
0.0020640140864998102,
0.005975294392555952,
-0.042442042380571365,
0.06445034593343735,
-0.029080448672175407,
0.0881110355257988,
0.054928023368120193,
0.013715995475649834,
0.06875631958246231,
-0.0028499988839030266,
0.03765236958861351,
-0.05321435257792473,
0.063898466527462,
-0.012775805778801441,
0.059143926948308945,
0.01807810179889202,
0.11895251274108887,
-0.033423393964767456,
0.015170260332524776,
0.014131685718894005,
-0.036881010979413986,
0.04079093039035797,
-0.08535895496606827,
-0.07714447379112244,
-0.007515855133533478,
-0.01918124035000801,
-0.036638516932725906,
-0.0790669322013855,
-0.0682314857840538,
-0.02933623269200325,
-0.026795612648129463,
0.05804511904716492,
-0.05191480740904808,
-0.06439417600631714,
0.06438829749822617,
-0.007798413280397654,
0.052087876945734024,
0.03223468363285065,
0.0713016465306282,
0.05443143472075462,
0.08270031213760376,
0.02144269458949566,
0.06633350253105164,
0.05701589211821556,
0.009935717098414898,
-0.02733248472213745,
0.07940548658370972,
0.0074913944117724895,
-0.03176921606063843,
0.0007775385165587068,
0.036242514848709106,
0.0870574563741684,
-0.008163630962371826,
0.0005136426188983023,
-0.08610323071479797,
0.021655870601534843,
-0.03090318664908409,
0.006863252259790897,
-0.045219048857688904,
1.9636041709891055e-33,
-0.05524882674217224,
0.054049719125032425,
0.05223694071173668,
-0.05150941386818886,
0.021610617637634277,
0.006671423092484474,
0.027630330994725227,
0.05333530530333519,
-0.03916116803884506,
-0.0030654529109597206,
-0.026202155277132988,
-0.006148303393274546,
-0.005793672986328602,
0.12379016727209091,
0.07289742678403854,
-0.05823233723640442,
0.02193828672170639,
0.037367891520261765,
-0.009249748662114143,
-0.04478277638554573,
0.0028142337687313557,
-0.05497888848185539,
-0.019981691613793373,
-0.023738719522953033,
0.009744053706526756,
-0.04757796227931976,
0.016870634630322456,
-0.09815192222595215,
-0.033241402357816696,
0.012067817151546478,
-0.11187636107206345,
-0.0033073648810386658,
0.03054494969546795,
0.022693274542689323,
0.025560278445482254,
-0.07611655443906784,
0.07813980430364609,
-0.040039848536252975,
-0.03586804121732712,
-0.001348851597867906,
0.015358995646238327,
0.04827921465039253,
-0.05029246583580971,
-0.049222346395254135,
-0.05633549019694328,
-0.06528133153915405,
0.0020520419348031282,
0.05392998829483986,
-0.06133127585053444,
0.0006681153317913413,
0.1307901293039322,
0.0249886866658926,
-0.07436735183000565,
-0.03025161102414131,
0.06849940866231918,
-0.054984237998723984,
0.04092066362500191,
-0.02835729531943798,
0.06211673095822334,
0.07869832962751389,
-0.03006465919315815,
-0.014016627334058285,
-0.04816718026995659,
0.0888199657201767,
-0.04529186710715294,
-0.02239234745502472,
0.010476470924913883,
0.0048750475980341434,
-0.06770192831754684,
0.01605777069926262,
-0.03821830078959465,
0.041010305285453796,
0.021260790526866913,
-0.04353947192430496,
0.002951610367745161,
-0.010311641730368137,
0.010791201144456863,
-0.03261302784085274,
-0.051090825349092484,
0.06983817368745804,
-0.08287347108125687,
0.08164002746343613,
-0.019617464393377304,
-0.01482853852212429,
-0.0017327592941001058,
-0.02186688408255577,
0.03337227925658226,
-0.038962941616773605,
-0.04335634410381317,
0.018060138449072838,
-0.04980519413948059,
-0.02844039909541607,
0.02384795807301998,
-0.013400580734014511,
-0.017473898828029633,
-1.8380835985113555e-33,
0.015275473706424236,
0.07111576944589615,
-0.0487167127430439,
0.06489379703998566,
-0.047269243746995926,
-0.09393343329429626,
-0.01746072620153427,
0.053142085671424866,
-0.03301924094557762,
-0.06646332889795303,
-0.03435423597693443,
-0.017019372433423996,
0.04154808819293976,
-0.04073433578014374,
0.05373803526163101,
-0.02780536748468876,
-0.09105075150728226,
-0.015292933210730553,
-0.011768145486712456,
0.012317528016865253,
0.054642096161842346,
0.08993040025234222,
-0.05223386362195015,
0.042971134185791016,
-0.08621212095022202,
0.04842095449566841,
-0.08796767145395279,
0.03841099888086319,
0.07398467510938644,
-0.03106001392006874,
-0.07582831382751465,
-0.0040195779874920845,
0.02120724879205227,
-0.04047037661075592,
-0.030772404745221138,
0.02466432750225067,
-0.0018662228249013424,
0.01873568631708622,
0.025076549500226974,
0.035466331988573074,
0.009678859263658524,
-0.04499625787138939,
-0.01218603365123272,
0.06516727060079575,
-0.02611694112420082,
-0.036000657826662064,
0.03871173411607742,
0.003233448602259159,
-0.030947888270020485,
0.0429750420153141,
-0.0453343465924263,
0.05882222205400467,
-0.13176028430461884,
0.01261054165661335,
-0.03843620792031288,
-0.021515758708119392,
-0.10916908830404282,
-0.03033621795475483,
0.04135642200708389,
-0.036114130169153214,
0.006422960199415684,
-0.043586067855358124,
0.029376449063420296,
-0.04234066605567932,
0.019375698640942574,
-0.015037975274026394,
-0.019363611936569214,
-0.013507618568837643,
0.0031709112226963043,
0.051561273634433746,
0.037078287452459335,
-0.06031942740082741,
-0.03644613176584244,
0.02507890947163105,
-0.08891359716653824,
0.0077897971495985985,
0.040914762765169144,
-0.07426289469003677,
0.06332802772521973,
0.05673537775874138,
-0.07087249308824539,
0.019970882683992386,
0.006091064307838678,
0.06388244032859802,
0.08706650137901306,
0.06806864589452744,
0.03887537494301796,
0.03358025476336479,
0.02797943726181984,
0.07088522613048553,
-0.05456466227769852,
0.05775615572929382,
0.05054360628128052,
0.056558433920145035,
-0.0004740919976029545,
-3.8078912467653936e-8,
-0.04474833235144615,
0.04944198951125145,
-0.04196066036820412,
-0.012883069925010204,
-0.031473416835069656,
-0.028507156297564507,
0.04188687726855278,
0.12902072072029114,
-0.03405391797423363,
0.045995648950338364,
0.10704716295003891,
-0.03341737389564514,
-0.045722704380750656,
0.02174452692270279,
0.04841264337301254,
0.06178071349859238,
0.0673886090517044,
-0.020465856418013573,
-0.022266695275902748,
-0.028191175311803818,
-0.10236267000436783,
0.07624407857656479,
-0.04110616818070412,
-0.08948653191328049,
-0.0790570080280304,
-0.0837753564119339,
0.02063426747918129,
0.14965209364891052,
0.03673401102423668,
-0.021069027483463287,
-0.07821095734834671,
0.04760494455695152,
0.01817505806684494,
0.0035305407363921404,
0.03129355609416962,
0.018623342737555504,
0.005394218023866415,
0.026030737906694412,
0.02731419913470745,
-0.010575806722044945,
0.10666783899068832,
0.006225730292499065,
0.012668372131884098,
0.02753559872508049,
-0.0182539951056242,
0.049655698239803314,
0.06500299274921417,
0.03308914974331856,
-0.026664089411497116,
0.025601083412766457,
0.05274127051234245,
-0.03718035668134689,
-0.07617976516485214,
0.09972440451383591,
0.05705917254090309,
0.022514648735523224,
-0.0352969653904438,
-0.02606508880853653,
0.04995877668261528,
0.07284118235111237,
-0.006504020653665066,
-0.010463044978678226,
-0.09571365267038345,
-0.04208137094974518
] |
Understanding Matrix Function Normalizations in Covariance Pooling through the Lens of Riemannian Geometry | https://openreview.net/forum?id=q1t0Lmvhty | [
"Ziheng Chen",
"Yue Song",
"Xiaojun Wu",
"Gaowen Liu",
"Nicu Sebe"
] | Poster | Global Covariance Pooling (GCP) has been demonstrated to improve the performance of Deep Neural Networks (DNNs) by exploiting second-order statistics of high-level representations. GCP typically performs classification of the covariance matrices by applying matrix function normalization, such as matrix logarithm or power, followed by a Euclidean classifier. However, covariance matrices inherently lie in a Riemannian manifold, known as the Symmetric Positive Definite (SPD) manifold. The current literature does not provide a satisfactory explanation of why Euclidean classifiers can be applied directly to Riemannian features after the normalization of the matrix power. To mitigate this gap, this paper provides a comprehensive and unified understanding of the matrix logarithm and power from a Riemannian geometry perspective. The underlying mechanism of matrix functions in GCP is interpreted from two perspectives: one based on tangent classifiers (Euclidean classifiers on the tangent space) and the other based on Riemannian classifiers. Via theoretical analysis and empirical validation through extensive experiments on fine-grained and large-scale visual classification datasets, we conclude that the working mechanism of the matrix functions should be attributed to the Riemannian classifiers they implicitly respect. The code is available at https://github.com/GitZH-Chen/RiemGCP.git. | Global covariance pooling, SPD manifolds, Representation Learning, Riemannian Manifolds | We theoretically and empirically explain how matrix function normalization (e.g., matrix square root) enables Euclidean classifiers to effectively work with Riemannian covariance features in Global Covariance Pooling by Riemannian geometry. | 466 | 2407.10484 | [
-0.051444582641124725,
-0.09850916266441345,
0.04363377019762993,
-0.048243194818496704,
0.01180548220872879,
-0.023347435519099236,
-0.02194845676422119,
-0.08089201897382736,
0.01580137573182583,
-0.05001169815659523,
0.013864288106560707,
0.037349481135606766,
-0.04429179057478905,
0.06597692519426346,
-0.06570650637149811,
-0.02283000946044922,
0.04137270525097847,
0.1353764533996582,
-0.1149170994758606,
0.023451989516615868,
0.07116834074258804,
0.025517478585243225,
-0.018107512965798378,
0.04448896273970604,
0.041734401136636734,
-0.001668811310082674,
0.08258958160877228,
-0.03462328761816025,
-0.04619012400507927,
0.008469185791909695,
0.00016934324230533093,
0.01571829244494438,
-0.02919311262667179,
-0.026199618354439735,
-0.0808081105351448,
0.0909019485116005,
0.05704408511519432,
0.0517009012401104,
0.055660586804151535,
0.03704551234841347,
-0.010952473618090153,
-0.006221159361302853,
0.052092719823122025,
0.10869578272104263,
0.11943353712558746,
0.05948084220290184,
0.03000679612159729,
-0.09052788466215134,
0.014688114635646343,
0.002809096360579133,
-0.021610639989376068,
0.04882566258311272,
-0.0699562057852745,
-0.017248710617423058,
-0.07526994496583939,
-0.010130543261766434,
0.04668772220611572,
0.03614006191492081,
0.028631797060370445,
0.049718789756298065,
0.001849469030275941,
-0.07660012692213058,
0.01818222366273403,
0.022849276661872864,
0.020047441124916077,
-0.031364694237709045,
-0.007022658362984657,
-0.008523504249751568,
-0.0023678173311054707,
0.003544697305187583,
-0.009003610350191593,
0.07239097356796265,
-0.07555942237377167,
-0.03512350097298622,
-0.0056088268756866455,
0.034567639231681824,
0.0072600687853991985,
0.02807566709816456,
0.029260452836751938,
-0.05531986430287361,
0.07026705145835876,
0.12595881521701813,
0.09090922772884369,
0.01721864379942417,
0.12532839179039001,
0.0367744080722332,
-0.019977664574980736,
0.050651803612709045,
0.001061120186932385,
-0.07315851002931595,
-0.01598389260470867,
-0.001111412886530161,
-0.06282150000333786,
-0.08366786688566208,
0.002406294923275709,
-0.059404537081718445,
-0.02608633227646351,
0.0031264147255569696,
-0.00981944426894188,
0.06524434685707092,
0.008038580417633057,
-0.08988755196332932,
-0.012065444141626358,
0.07224880158901215,
0.06202629581093788,
0.01288062334060669,
0.09777914732694626,
0.07003684341907501,
0.012117729522287846,
-0.07034718990325928,
-0.001903208321891725,
0.04772033169865608,
-0.07338568568229675,
-0.010499173775315285,
-0.0012807410676032305,
0.03338983282446861,
0.004640300292521715,
0.014692996628582478,
0.056907519698143005,
0.08281554281711578,
-0.07495737820863724,
0.00040600894135423005,
-0.014098453335464,
0.036488309502601624,
0.013237114064395428,
0.07089410722255707,
-0.04142531007528305,
4.5841002991566296e-33,
-0.038408417254686356,
0.019025886431336403,
0.05829198285937309,
-0.011698989197611809,
-0.03145557641983032,
0.043390482664108276,
-0.04911360889673233,
-0.06243178993463516,
0.022640978917479515,
0.06624240428209305,
-0.06838146597146988,
0.05594711750745773,
-0.003839155426248908,
0.053918611258268356,
-0.04452383518218994,
0.05951043218374252,
-0.017174284905195236,
-0.03946661204099655,
0.004139491356909275,
-0.10129640251398087,
0.04117324948310852,
0.007243743631988764,
0.0581653006374836,
-0.003115240717306733,
-0.06741739809513092,
-0.018846118822693825,
0.010815096087753773,
0.02471238560974598,
-0.07244017720222473,
0.0018465797184035182,
-0.022476881742477417,
-0.01223916094750166,
0.0001191706323879771,
-0.005307382438331842,
-0.02921932563185692,
-0.008113382384181023,
-0.04476262256503105,
-0.007745365146547556,
0.042919665575027466,
-0.05464356020092964,
-0.10386008024215698,
0.03137802705168724,
-0.004684379789978266,
-0.038744110614061356,
-0.10557558387517929,
0.02265637181699276,
0.06559911370277405,
0.038233961910009384,
-0.03424065187573433,
-0.08806538581848145,
-0.002053435891866684,
0.029700839892029762,
-0.06888379156589508,
-0.033009618520736694,
0.019803663715720177,
-0.019707972183823586,
0.060909997671842575,
0.018524328246712685,
-0.028286200016736984,
0.05140133947134018,
0.027177851647138596,
0.026635056361556053,
-0.026771603152155876,
-0.03306801989674568,
-0.0604470819234848,
-0.03377733379602432,
0.06756416708230972,
0.08127628266811371,
0.08641437441110611,
-0.006035175174474716,
0.0044646854512393475,
0.01804438978433609,
-0.0387490950524807,
0.021675264462828636,
-0.011086849495768547,
0.04256775230169296,
0.000043635231122607365,
-0.0649736300110817,
-0.026562629267573357,
0.06805610656738281,
-0.054597850888967514,
-0.004875265993177891,
-0.014539801515638828,
-0.00605866638943553,
-0.09857215732336044,
-0.010618734173476696,
-0.026298293843865395,
-0.10252891480922699,
-0.013266339898109436,
-0.009653247892856598,
-0.0769142135977745,
-0.0218211617320776,
-0.006113112438470125,
-0.015877926722168922,
-0.04206884652376175,
-4.762650176864858e-33,
-0.053187549114227295,
-0.0023039565421640873,
-0.06599560379981995,
0.010922780260443687,
-0.03560400381684303,
-0.041761159896850586,
-0.05132586136460304,
0.030558383092284203,
-0.0489104762673378,
0.02358075976371765,
0.10283730924129486,
-0.023419639095664024,
-0.035820212215185165,
-0.016059398651123047,
0.06375058740377426,
0.05752643570303917,
-0.07116223871707916,
0.052769582718610764,
-0.03561033681035042,
-0.04485153779387474,
0.0654032975435257,
0.10293009132146835,
0.004643295891582966,
-0.0065454053692519665,
-0.05608206242322922,
0.018166374415159225,
-0.06042877212166786,
0.013622047379612923,
0.027675796300172806,
-0.016270915046334267,
-0.08394134789705276,
-0.06036633253097534,
-0.09594274312257767,
0.043459147214889526,
0.019985318183898926,
0.006939879152923822,
-0.03717222437262535,
-0.06320274621248245,
-0.0037505922373384237,
0.05958349257707596,
-0.025881867855787277,
-0.005479495041072369,
0.010403376072645187,
-0.016922064125537872,
0.001501442864537239,
-0.001708217547275126,
0.040568046271800995,
0.03256494924426079,
-0.03750056028366089,
-0.023241885006427765,
-0.06655685603618622,
0.015434658154845238,
-0.029456282034516335,
0.05904247239232063,
-0.08501069247722626,
0.016673726961016655,
-0.013276965357363224,
0.014633639715611935,
0.03688659518957138,
0.033691905438899994,
-0.03519021347165108,
-0.030305827036499977,
-0.0687863677740097,
0.022800985723733902,
-0.05040154233574867,
-0.08959856629371643,
-0.02507711388170719,
0.037188634276390076,
0.06884843111038208,
0.06962373852729797,
-0.014387045055627823,
-0.016747118905186653,
-0.005664439406245947,
-0.009961609728634357,
-0.08952509611845016,
-0.04143025353550911,
0.03298035264015198,
-0.03558320179581642,
-0.0043429662473499775,
-0.03363044187426567,
0.0835370272397995,
-0.01602674461901188,
0.06939312815666199,
0.11519245058298111,
0.13517911732196808,
0.09661436080932617,
0.04185395687818527,
-0.06316448003053665,
0.11677033454179764,
-0.020338241010904312,
0.00788172334432602,
0.0033510015346109867,
-0.056700970977544785,
0.10701276361942291,
-0.04809897392988205,
-4.021672239673535e-8,
-0.09626608341932297,
0.006543952040374279,
0.008424690924584866,
-0.06154930964112282,
-0.0068405852653086185,
-0.005876677576452494,
-0.03187422454357147,
0.1332169771194458,
-0.06369602680206299,
0.08434810489416122,
0.022359445691108704,
-0.09007924795150757,
-0.08174080401659012,
-0.08671236038208008,
-0.04817032814025879,
0.06647945940494537,
0.04442012310028076,
0.0644603744149208,
0.05319341644644737,
-0.0038309460505843163,
0.04226747900247574,
-0.01012706384062767,
0.07110723108053207,
0.011385560035705566,
0.002633733209222555,
-0.08390741795301437,
-0.007791610434651375,
0.05113833770155907,
0.0648658275604248,
-0.04400761052966118,
0.001215134747326374,
-0.013962413184344769,
0.021736497059464455,
0.0012685841647908092,
0.049026455730199814,
0.08619534224271774,
-0.024400658905506134,
-0.004537970293313265,
-0.023663341999053955,
-0.012149393558502197,
-0.06684623658657074,
0.03146869316697121,
0.029925620183348656,
0.026573877781629562,
0.038381177932024,
0.06377343833446503,
0.0618535540997982,
-0.039952702820301056,
-0.010952005162835121,
0.053270336240530014,
0.04766415059566498,
0.009282034821808338,
0.011437924578785896,
0.10423748940229416,
0.07058311253786087,
-0.005744268652051687,
0.0016944357194006443,
-0.04778328910470009,
-0.0002788550336845219,
0.05457785725593567,
-0.029525456950068474,
0.051515303552150726,
-0.0692712590098381,
-0.12196184694766998
] |
Gyrogroup Batch Normalization | https://openreview.net/forum?id=d1NWq4PjJW | [
"Ziheng Chen",
"Yue Song",
"Xiaojun Wu",
"Nicu Sebe"
] | Poster | Several Riemannian manifolds in machine learning, such as Symmetric Positive Definite (SPD), Grassmann, spherical, and hyperbolic manifolds, have been proven to admit gyro structures, thus enabling a principled and effective extension of Euclidean Deep Neural Networks (DNNs) to manifolds. Inspired by this, this study introduces a general Riemannian Batch Normalization (RBN) framework on gyrogroups, termed GyroBN. We identify the least requirements to guarantee GyroBN with theoretical control over sample statistics, referred to as \textit{pseudo-reduction} and \textit{gyroisometric gyrations}, which are satisfied by all the existing gyrogroups in machine learning. Besides, our GyroBN incorporates several existing normalization methods, including the one on general Lie groups and different types of RBN on the non-group SPD geometry. Lastly, we instantiate our GyroBN on the Grassmannian and hyperbolic spaces. Experiments on the Grassmannian and hyperbolic networks demonstrate the effectiveness of our GyroBN. The code is available at https://github.com/GitZH-Chen/GyroBN.git. | Gyrovector Spaces, Riemannian Manifolds, Riemannian Batch Normalization, Grassmannian Manifolds, Hyperbolic Manifolds | This paper proposes a general framework for Riemannian Batch Normalization (RBN) over gyrogroups (GyroBN), which incorporate several existing RBN methods, and showcase our framework on the Grassmannian and hyperbolic geometries. | 461 | null | [
-0.11063361912965775,
-0.05906516686081886,
0.03742349147796631,
-0.07168938219547272,
-0.029176440089941025,
0.005679471418261528,
-0.027922896668314934,
-0.0911281406879425,
-0.046179261058568954,
-0.04988772049546242,
0.028155729174613953,
0.0331500843167305,
-0.06215191259980202,
-0.0015247311675921082,
-0.02675548754632473,
-0.030478911474347115,
0.0025880231987684965,
0.11633795499801636,
-0.10425935685634613,
-0.0035099233500659466,
-0.04450899735093117,
0.024626802653074265,
0.03580014035105705,
0.06018604338169098,
0.039360951632261276,
-0.03244703263044357,
0.0795971006155014,
-0.012885194271802902,
-0.004946908447891474,
0.015876367688179016,
0.04703659564256668,
-0.015958962962031364,
-0.003903220174834132,
-0.06513813138008118,
-0.03325022757053375,
0.05746326968073845,
-0.05630930885672569,
0.0333150252699852,
0.01651141047477722,
0.028362607583403587,
0.051579549908638,
-0.006349914241582155,
0.0031984515953809023,
0.07130230218172073,
0.03127450868487358,
0.0820421352982521,
0.05008872225880623,
-0.0718383863568306,
0.020693475380539894,
-0.0002647757937666029,
-0.040137313306331635,
-0.002003774745389819,
-0.04974791407585144,
0.06361740082502365,
-0.024774592369794846,
0.009929710999131203,
0.02264920249581337,
-0.0076435282826423645,
-0.01920580491423607,
-0.038063131272792816,
0.02570759318768978,
-0.08190415799617767,
-0.0880032405257225,
-0.02475505881011486,
-0.017768308520317078,
0.021348435431718826,
0.019870124757289886,
0.007544702384620905,
0.03286440670490265,
0.02788860723376274,
0.042063288390636444,
0.043022990226745605,
-0.056572746485471725,
-0.006413253489881754,
-0.036346253007650375,
0.07710815221071243,
-0.007630100008100271,
0.0034108266700059175,
0.005235090386122465,
-0.04222354292869568,
0.022501375526189804,
0.06931816786527634,
0.14881430566310883,
-0.023865962401032448,
0.05479806661605835,
0.024206742644309998,
-0.051755812019109726,
0.04641261324286461,
-0.009504932910203934,
-0.08411044627428055,
-0.00838471669703722,
0.012157767079770565,
-0.06767582893371582,
-0.10620319843292236,
-0.046817347407341,
-0.017716726288199425,
0.03828320652246475,
0.03957158327102661,
0.008431760594248772,
0.11523590981960297,
0.004853798542171717,
-0.12099205702543259,
0.014357074163854122,
0.01823940873146057,
-0.012489586137235165,
0.020519176498055458,
0.09050145000219345,
0.025786593556404114,
0.02530832588672638,
-0.09360804408788681,
-0.03169861435890198,
0.07911934703588486,
-0.049854639917612076,
0.008900103159248829,
-0.04938778281211853,
-0.07711008936166763,
0.03109937347471714,
0.022338083013892174,
0.010302695445716381,
0.09058749675750732,
-0.0016498628538101912,
0.015529350377619267,
0.029881609603762627,
0.07100460678339005,
0.007364276796579361,
0.05409973859786987,
-0.0971825122833252,
4.221540739822588e-33,
-0.0005847599823027849,
0.03036266751587391,
0.03429297357797623,
-0.06110532954335213,
0.012040198780596256,
0.05658303573727608,
-0.10020006448030472,
-0.034133922308683395,
0.031703416258096695,
0.042570676654577255,
-0.04548434540629387,
0.027214333415031433,
-0.018327496945858,
0.050876013934612274,
-0.08417564630508423,
-0.006326343864202499,
0.006133303977549076,
0.006712283939123154,
0.012430478818714619,
-0.11938035488128662,
0.0653369128704071,
0.05573304742574692,
0.021374616771936417,
0.009261771105229855,
-0.10881307721138,
0.031160222366452217,
0.02477811463177204,
0.027590513229370117,
-0.06872373819351196,
0.010917321778833866,
0.019417526200413704,
-0.053818438202142715,
-0.07626932114362717,
0.04791243001818657,
0.024583274498581886,
-0.0035670152865350246,
-0.06133845075964928,
-0.05204429849982262,
0.02400396391749382,
-0.012395546771585941,
-0.04154063016176224,
0.037630386650562286,
0.02352343685925007,
-0.03722495958209038,
-0.04649033397436142,
0.012133434414863586,
0.05495710298418999,
0.048470303416252136,
0.020915508270263672,
-0.07482751458883286,
0.004173957742750645,
-0.0438438318669796,
-0.026739349588751793,
-0.0393187515437603,
0.006835402455180883,
-0.010658252984285355,
0.07468210160732269,
-0.012585840187966824,
0.010406799614429474,
0.04168776795268059,
0.05050555616617203,
0.050643276423215866,
0.02003380097448826,
-0.06534922122955322,
-0.04475048929452896,
0.0017836331389844418,
0.01486067846417427,
0.04796319082379341,
0.08701051771640778,
-0.06790414452552795,
0.0070462641306221485,
0.05386378988623619,
-0.022568650543689728,
0.06666257977485657,
0.034037452191114426,
0.020584523677825928,
0.06974496692419052,
-0.06682495027780533,
-0.06034445762634277,
-0.010850566439330578,
-0.08429104834794998,
0.032910846173763275,
-0.10042934119701385,
-0.04092862084507942,
-0.09397870302200317,
-0.00694319698959589,
-0.03693235293030739,
-0.11112067848443985,
-0.04297499358654022,
-0.03246407210826874,
-0.12666790187358856,
-0.004731209948658943,
-0.07094195485115051,
-0.0010753074893727899,
-0.048921532928943634,
-4.053223540438268e-33,
-0.042217057198286057,
0.07338421046733856,
-0.05647881701588631,
0.038932859897613525,
0.005514513701200485,
-0.00915428064763546,
0.04445623606443405,
0.028184406459331512,
-0.037094034254550934,
-0.021638955920934677,
0.03293604031205177,
0.032650284469127655,
0.0374842993915081,
-0.02246510609984398,
0.05832255631685257,
0.016966301947832108,
-0.039566583931446075,
-0.015073433518409729,
0.022931573912501335,
-0.01677054539322853,
0.00006678403588011861,
0.06754910945892334,
-0.10248678922653198,
0.023385552689433098,
-0.02734195441007614,
-0.007009650580585003,
-0.06919673085212708,
0.10716079920530319,
0.05995754152536392,
0.045792724937200546,
-0.04458671435713768,
-0.053265515714883804,
-0.10122013092041016,
0.01964215189218521,
0.03792431578040123,
0.06885548681020737,
-0.01496270764619112,
0.0855899304151535,
-0.06836877018213272,
-0.008396380580961704,
0.042050331830978394,
-0.01479480043053627,
-0.010756821371614933,
0.04460795968770981,
0.030891865491867065,
0.009508352726697922,
-0.027282238006591797,
-0.03723004087805748,
-0.06798256188631058,
-0.007569900248199701,
-0.06467358767986298,
0.003311336971819401,
0.0011583241866901517,
-0.008969705551862717,
-0.036001451313495636,
0.0008287837263196707,
0.007039110641926527,
-0.012015784159302711,
0.03178511559963226,
0.00410677632316947,
-0.05806616693735123,
-0.050858497619628906,
-0.006402704864740372,
0.022988667711615562,
-0.04192807152867317,
-0.05610448122024536,
0.012904345989227295,
0.012906890362501144,
0.020562363788485527,
0.07710910588502884,
0.01755969412624836,
-0.015481354668736458,
0.011397134512662888,
0.018919771537184715,
-0.08681417256593704,
-0.030172526836395264,
0.08433444052934647,
-0.03994866833090782,
-0.019477367401123047,
-0.04145973548293114,
0.10617846995592117,
-0.03995412588119507,
0.05494121089577675,
0.06356503814458847,
0.13269007205963135,
0.10054688155651093,
0.07385684549808502,
-0.014206910505890846,
0.08667895942926407,
-0.0035216128453612328,
0.04390506073832512,
0.03268248215317726,
0.00446274084970355,
0.14393547177314758,
-0.03977590054273605,
-4.0777290877258565e-8,
-0.011386171914637089,
0.008465582504868507,
0.0555577352643013,
-0.005934773478657007,
0.02711295522749424,
-0.048861049115657806,
-0.011572867631912231,
0.03513362631201744,
-0.09182589501142502,
0.01457879226654768,
0.05041683837771416,
-0.056291379034519196,
-0.07423602789640427,
-0.05919623747467995,
-0.03896784037351608,
0.10647911578416824,
-0.016491377726197243,
0.0646062046289444,
0.011904103681445122,
0.018994415178894997,
0.0021731615997850895,
0.06046878173947334,
-0.0013899458572268486,
-0.003950058948248625,
0.08548106998205185,
-0.08108925819396973,
-0.017722604796290398,
0.0010200368706136942,
0.03278932347893715,
-0.013729299418628216,
0.06402444839477539,
0.016871772706508636,
0.11239317804574966,
0.05544225499033928,
0.004334640689194202,
0.05402784422039986,
-0.09342905879020691,
0.07923571020364761,
0.006910910829901695,
-0.042225126177072525,
-0.07137726247310638,
0.062101881951093674,
0.002400004304945469,
-0.006719459313899279,
0.008893259800970554,
0.03610244020819664,
0.031729232519865036,
-0.100397028028965,
0.03170108050107956,
0.08504427224397659,
0.08386023342609406,
0.021061686798930168,
0.03296511247754097,
0.11594603955745697,
0.03618461638689041,
0.020521951839327812,
0.0109231723472476,
-0.02542426623404026,
0.009614914655685425,
0.00978214479982853,
-0.0420968197286129,
0.021987538784742355,
-0.09809540212154388,
-0.08248724043369293
] |
T-JEPA: Augmentation-Free Self-Supervised Learning for Tabular Data | https://openreview.net/forum?id=gx3LMRB15C | [
"Hugo Thimonier",
"José Lucas De Melo Costa",
"Fabrice Popineau",
"Arpad Rimmel",
"Bich-Liên DOAN"
] | Poster | Self-supervision is often used for pre-training to foster performance on a downstream task by constructing meaningful representations of samples. Self-supervised learning (SSL) generally involves generating different views of the same sample and thus requires data augmentations that are challenging to construct for tabular data. This constitutes one of the main challenges of self-supervision for structured data. In the present work, we propose a novel augmentation-free SSL method for tabular data. Our approach, T-JEPA, relies on a Joint Embedding Predictive Architecture (JEPA) and is akin to mask reconstruction in the latent space. It involves predicting the latent representation of one subset of features from the latent representation of a different subset within the same sample, thereby learning rich representations without augmentations. We use our method as a pre-training technique and train several deep classifiers on the obtained representation. Our experimental results demonstrate a substantial improvement in both classification and regression tasks, outperforming models trained directly on samples in their original data space. Moreover, T-JEPA enables some methods to consistently outperform or match the performance of traditional methods likes Gradient Boosted Decision Trees. To understand why, we extensively characterize the obtained representations and show that T-JEPA effectively identifies relevant features for downstream tasks without access to the labels. Additionally, we introduce regularization tokens, a novel regularization method critical for training of JEPA-based models on structured data. | Self-Supervised Learning, Tabular Data, Representation Learning | This work presents a novel augmentation-free self-supervised method for tabular data based on the joint-embedding predictive architecture framework. | 459 | null | [
-0.07450642436742783,
-0.08043394982814789,
0.05884239077568054,
0.07033953815698624,
0.04697884991765022,
0.06551312655210495,
-0.020781653001904488,
-0.09084826707839966,
0.0038974848575890064,
-0.08371317386627197,
-0.031192408874630928,
-0.06752358376979828,
0.013938674703240395,
0.054282985627651215,
-0.00921593513339758,
-0.047811638563871384,
0.12370198965072632,
0.07157345861196518,
-0.0908195972442627,
-0.0906030610203743,
0.023077916353940964,
0.027840755879878998,
0.04404543340206146,
-0.0314081646502018,
0.06862684339284897,
-0.004256442654877901,
0.02616489864885807,
-0.019695186987519264,
-0.01607542112469673,
-0.03437482565641403,
0.0476367250084877,
-0.027745099738240242,
-0.07312086224555969,
0.041221071034669876,
-0.06343290954828262,
-0.006416045129299164,
-0.09091454744338989,
0.09388913959264755,
-0.02815108373761177,
-0.025741316378116608,
-0.018487168475985527,
-0.058594245463609695,
-0.04028021916747093,
0.076676145195961,
0.011058847419917583,
-0.00634957617148757,
0.032068461179733276,
-0.10460332036018372,
-0.010159855708479881,
0.014129456132650375,
-0.07340739667415619,
0.03132527321577072,
-0.025619423016905785,
0.040255192667245865,
-0.06719165295362473,
0.013621371239423752,
0.051925450563430786,
0.009316233918070793,
0.02644224651157856,
0.06744977831840515,
0.042909637093544006,
-0.0732010081410408,
0.01319179404526949,
-0.037213753908872604,
-0.049232181161642075,
-0.011198857799172401,
-0.027367182075977325,
0.06464092433452606,
-0.052820391952991486,
0.0006975195137783885,
0.020421884953975677,
0.07720183581113815,
-0.060661107301712036,
0.06280255317687988,
0.023774608969688416,
-0.03406888246536255,
0.0015033818781375885,
-0.0035874429158866405,
0.10307393223047256,
-0.0005118675180710852,
-0.021672695875167847,
0.05341755971312523,
0.008681508712470531,
0.07127508521080017,
0.10765018314123154,
0.025362910702824593,
0.02831258997321129,
0.04691311717033386,
-0.04134047403931618,
0.029591407626867294,
-0.003739504609256983,
-0.0561530739068985,
-0.02687828056514263,
-0.006814455147832632,
0.009699706919491291,
-0.024023571982979774,
0.03347181901335716,
-0.09948208183050156,
-0.0075121489353477955,
0.06744537502527237,
-0.024394569918513298,
0.08451536297798157,
0.03365826606750488,
0.04489442706108093,
-0.010069629177451134,
-0.019035417586565018,
0.07383961230516434,
0.02330339327454567,
0.037321366369724274,
-0.09345438331365585,
0.06035097315907478,
-0.01170068047940731,
-0.02941652573645115,
-0.06163233518600464,
0.05892765149474144,
0.019545435905456543,
-0.036183055490255356,
-0.018948262557387352,
0.049755845218896866,
0.046472568064928055,
-0.03339029848575592,
0.000862030079588294,
0.02710493840277195,
0.09069550037384033,
0.018352365121245384,
0.01721513457596302,
-0.11180473864078522,
4.179942198798895e-33,
0.03267420828342438,
0.07215181738138199,
-0.0020575011149048805,
-0.015504362992942333,
0.02250370942056179,
0.007882317528128624,
-0.0294501893222332,
0.05449387803673744,
-0.04187912121415138,
0.02922968752682209,
-0.02413320355117321,
-0.06742104142904282,
-0.05393589660525322,
0.07426326721906662,
0.004999254830181599,
0.021232958883047104,
-0.11064158380031586,
0.024460280314087868,
-0.047334667295217514,
-0.020544255152344704,
-0.014845561236143112,
0.033941321074962616,
0.007565016858279705,
0.009098000824451447,
0.025173155590891838,
-0.01258358359336853,
0.058487337082624435,
-0.04755440726876259,
-0.013436298817396164,
0.011237242259085178,
-0.0656292736530304,
0.06464065611362457,
0.02990058623254299,
-0.011860423721373081,
0.02841506339609623,
-0.030226437374949455,
0.05138589069247246,
-0.04369083791971207,
0.06967990845441818,
0.01989605277776718,
0.043960362672805786,
0.0831637904047966,
0.004694866482168436,
-0.04286151006817818,
-0.05815812945365906,
-0.11562090367078781,
0.034247465431690216,
0.034751780331134796,
-0.010167140513658524,
-0.03286522999405861,
0.07349548488855362,
0.03622904792428017,
-0.0619179792702198,
-0.0742500051856041,
-0.09235157817602158,
0.022901246324181557,
0.007740343455225229,
0.06898393481969833,
0.0784820094704628,
0.05763476341962814,
-0.05036664009094238,
0.013984405435621738,
-0.029549559578299522,
0.05833527818322182,
-0.061152201145887375,
-0.022748803719878197,
0.07468630373477936,
0.04910105839371681,
0.06754589080810547,
-0.06402497738599777,
-0.04152209684252739,
0.02661271020770073,
-0.0212370827794075,
-0.09648391604423523,
0.02907237410545349,
-0.01073153130710125,
0.06094348058104515,
-0.08832940459251404,
0.017171379178762436,
-0.025291452184319496,
0.04210595786571503,
0.03841225430369377,
0.0019346512854099274,
-0.06834976375102997,
-0.08014421164989471,
-0.01887941174209118,
0.06590751558542252,
-0.10114693641662598,
0.034359823912382126,
0.0382477305829525,
-0.03359430283308029,
-0.01214636117219925,
-0.01921018399298191,
0.010625465773046017,
-0.017280377447605133,
-2.5823873359978347e-33,
0.009629398584365845,
0.04295504465699196,
-0.05414058268070221,
0.01427094079554081,
0.0497981496155262,
-0.06701239198446274,
0.008922113105654716,
0.03610903024673462,
0.05529792979359627,
-0.026099303737282753,
-0.014500579796731472,
-0.0850554034113884,
0.08388648927211761,
-0.019664794206619263,
0.011268398724496365,
0.04477538540959358,
-0.051845476031303406,
0.035246241837739944,
-0.07384279370307922,
-0.025689570233225822,
-0.023319890722632408,
0.08420921862125397,
0.05158185586333275,
0.09897856414318085,
-0.006553677376359701,
0.0331566259264946,
-0.0022455418948084116,
0.09288924932479858,
-0.005770743824541569,
-0.013926386833190918,
-0.025626571848988533,
-0.024065803736448288,
-0.008144529536366463,
-0.05388893932104111,
-0.052971359342336655,
0.06562568247318268,
-0.01106447633355856,
-0.015571043826639652,
-0.01092316024005413,
-0.008215080015361309,
0.055340077728033066,
0.02533807046711445,
-0.06166766211390495,
0.013053355738520622,
-0.08738784492015839,
-0.08467043936252594,
-0.021275758743286133,
0.04727107286453247,
0.05815228074789047,
0.0402267761528492,
-0.12079622596502304,
0.017320148646831512,
-0.1320158690214157,
-0.009232763200998306,
-0.04424165189266205,
-0.06261588633060455,
0.03611956909298897,
0.0024231497664004564,
0.031936127692461014,
0.019851841032505035,
-0.0688193291425705,
-0.0008419144432991743,
0.10567732155323029,
-0.051483675837516785,
0.0812535509467125,
-0.1143338605761528,
-0.04965612664818764,
0.042696356773376465,
0.060809288173913956,
-0.0002409826556686312,
-0.0030912812799215317,
-0.016643645241856575,
-0.009117682464420795,
0.02201768010854721,
0.04374688118696213,
-0.07079377770423889,
0.011210804805159569,
0.021143289282917976,
-0.019480934366583824,
-0.04095115140080452,
0.017619024962186813,
0.006596030201762915,
-0.006543596740812063,
-0.015259578824043274,
0.08579335361719131,
0.09027424454689026,
0.07253994047641754,
0.06978863477706909,
-0.004084778018295765,
0.0053141009993851185,
-0.076463982462883,
0.003146533388644457,
-0.10850708931684494,
0.07574080675840378,
-0.00499790208414197,
-3.83664264802519e-8,
-0.07993621379137039,
0.012855518609285355,
0.006559792906045914,
0.08739689737558365,
0.004758020862936974,
-0.1050737127661705,
0.08988528698682785,
0.12091784179210663,
-0.03448163717985153,
-0.03166733682155609,
0.1042257696390152,
0.007729111239314079,
-0.03549346700310707,
-0.06965933740139008,
0.0050254035741090775,
0.03622429445385933,
0.014309617690742016,
0.0228032935410738,
-0.013130491599440575,
0.025485031306743622,
0.03055264987051487,
0.022978322580456734,
0.06623606383800507,
-0.04489544779062271,
0.05066608265042305,
-0.032632723450660706,
-0.05862290784716606,
0.10737119615077972,
0.04307538643479347,
0.009224706329405308,
-0.03078700788319111,
-0.009364400058984756,
0.05072416365146637,
-0.07951833307743073,
0.025901511311531067,
0.07364550977945328,
0.06864916533231735,
-0.026255235075950623,
-0.021610181778669357,
-0.023008350282907486,
-0.020076002925634384,
0.050399575382471085,
-0.03218429163098335,
-0.005635838024318218,
-0.011468637734651566,
0.008653704077005386,
0.0056600733660161495,
-0.04616101086139679,
0.04252948611974716,
-0.05534011498093605,
0.0016398652223870158,
-0.10131041705608368,
0.036183085292577744,
0.0744590312242508,
0.039730750024318695,
0.03650420159101486,
-0.002022738568484783,
0.019863981753587723,
0.08333480358123779,
0.06467242538928986,
-0.033090125769376755,
0.007312594912946224,
0.0014631173107773066,
0.008137276396155357
] |
Steering Masked Discrete Diffusion Models via Discrete Denoising Posterior Prediction | https://openreview.net/forum?id=Ombm8S40zN | [
"Jarrid Rector-Brooks",
"Mohsin Hasan",
"Zhangzhi Peng",
"Cheng-Hao Liu",
"Sarthak Mittal",
"Nouha Dziri",
"Michael M. Bronstein",
"Pranam Chatterjee",
"Alexander Tong",
"Joey Bose"
] | Poster | Generative modeling of discrete data underlies important applications spanning text-based agents like ChatGPT to the design of the very building blocks of life in protein sequences. However, application domains need to exert control over the generated data by steering the generative process—typically via RLHF—to satisfy a specified property, reward, or affinity metric. In this paper, we study the problem of steering Masked Diffusion Models (MDMs), a recent class of discrete diffusion models that offer a compelling alternative to traditional autoregressive models. We introduce Discrete Denoising Posterior Prediction (DDPP), a novel framework that casts the task of steering pretrained MDMs as a problem of probabilistic inference by learning to sample from a target Bayesian posterior. Our DDPP framework leads to a family of three novel objectives that are all simulation-free, and thus scalable while applying to general non-differentiable reward functions. Empirically, we instantiate DDPP by steering MDMs to perform class-conditional pixel-level image modeling, RLHF-based alignment of MDMs using text based rewards, and finetuning protein language models to generate more diverse secondary structures and shorter proteins. We substantiate our designs via wet-lab validation, where we observe transient expression of reward-optimized protein sequences. | Discrete diffusion models, language modeling, probabilistic inference | We fine-tune discrete diffusion models by solving a probabilistic inference task. | 453 | 2410.08134 | [
-0.09584029018878937,
-0.18060165643692017,
0.02630792371928692,
-0.012453820556402206,
0.0012435225071385503,
-0.01072003785520792,
0.03382701054215431,
-0.02080925926566124,
0.09295175224542618,
-0.010351547971367836,
-0.01154651865363121,
-0.07637975364923477,
0.022488314658403397,
0.0725078135728836,
-0.016326723620295525,
0.0014161539729684591,
0.08344416320323944,
0.02174457721412182,
-0.05448949709534645,
-0.0657682865858078,
0.038859300315380096,
0.06546945124864578,
0.02379441261291504,
-0.012456896714866161,
-0.009355058893561363,
0.05159160494804382,
0.06225655600428581,
-0.07505397498607635,
0.030213836580514908,
-0.005623634438961744,
0.02030697837471962,
-0.03980113938450813,
-0.01115262508392334,
-0.028101321309804916,
-0.018215613439679146,
0.08425088971853256,
-0.1250796765089035,
0.010894904844462872,
-0.001205738983117044,
-0.024021776393055916,
0.02273249626159668,
0.01838330179452896,
0.003111785976216197,
0.02336527779698372,
0.07420343905687332,
-0.07520047575235367,
-0.03960990160703659,
-0.005546820815652609,
0.027951449155807495,
-0.002203151350840926,
-0.03053857944905758,
-0.020752202719449997,
-0.019908763468265533,
0.05944741144776344,
-0.021838385611772537,
0.027524909004569054,
0.04962339624762535,
-0.02418929897248745,
-0.017063813284039497,
-0.03269603103399277,
-0.10213776677846909,
-0.0960528701543808,
-0.03681858628988266,
-0.04738442227244377,
0.004279240965843201,
-0.020672455430030823,
0.0009145360090769827,
0.1250225454568863,
0.009117715060710907,
0.07632763683795929,
-0.04496939107775688,
0.04647523909807205,
-0.04369468614459038,
0.0016680652042850852,
-0.005227214191108942,
0.009238759987056255,
-0.028834642842411995,
-0.05151616781949997,
0.05307353660464287,
-0.06952042132616043,
0.027975214645266533,
0.056422844529151917,
0.12197549641132355,
0.0033825973514467478,
0.08856654167175293,
0.013187418691813946,
0.019783416762948036,
0.09153906255960464,
0.06144581362605095,
0.012782059609889984,
-0.08018947392702103,
0.0301156435161829,
-0.0387314110994339,
0.020643752068281174,
-0.06372436881065369,
0.05875786766409874,
-0.05133157595992088,
-0.04350484535098076,
-0.003522081533446908,
0.08632464706897736,
-0.0031554531306028366,
0.06877164542675018,
-0.03447483852505684,
-0.05101897940039635,
0.017788467928767204,
-0.05611749365925789,
-0.01187829114496708,
0.03675388917326927,
0.0707702562212944,
-0.03413941711187363,
-0.015318429097533226,
0.050049565732479095,
0.0454704724252224,
0.05949697270989418,
0.015658816322684288,
-0.05182048678398132,
0.07137705385684967,
-0.0917869582772255,
-0.020666200667619705,
0.03957493230700493,
0.012706526555120945,
-0.028944453224539757,
0.0014983045402914286,
-0.028215721249580383,
-0.011115153320133686,
-0.0026126480661332607,
-0.06853456795215607,
1.788432370171553e-33,
0.08550062030553818,
-0.07279324531555176,
0.06873035430908203,
0.0759289562702179,
0.0387188196182251,
-0.0014771269634366035,
-0.026927810162305832,
-0.008764415048062801,
-0.03276371210813522,
-0.019006293267011642,
-0.005543834064155817,
0.00281341141089797,
-0.1305457353591919,
0.07375811785459518,
0.06663218885660172,
-0.036410316824913025,
-0.05240122973918915,
0.11355002224445343,
0.05092313513159752,
-0.04399624839425087,
-0.025492552667856216,
0.039306577295064926,
-0.05133818835020065,
-0.013767991214990616,
0.02502903714776039,
0.06567288190126419,
-0.00174788071308285,
-0.09078623354434967,
-0.009922927245497704,
0.02289850078523159,
-0.15062128007411957,
-0.019896183162927628,
-0.01685314252972603,
-0.00871268566697836,
0.0033876467496156693,
0.02480785734951496,
-0.03276030346751213,
-0.07374422252178192,
-0.04784191772341728,
-0.04601845517754555,
0.019520120695233345,
-0.024805154651403427,
0.04675294831395149,
-0.031685683876276016,
-0.1508103907108307,
-0.05213770642876625,
0.03489479422569275,
-0.02774740569293499,
-0.06515263020992279,
-0.04768357053399086,
0.11880737543106079,
-0.033626753836870193,
0.017782820388674736,
-0.09149698913097382,
-0.07232927531003952,
0.004647950641810894,
-0.03606131672859192,
0.016026169061660767,
0.0376216396689415,
-0.0033230690751224756,
0.00964408740401268,
0.0007403159979730844,
-0.008515343070030212,
0.0476422980427742,
0.026738960295915604,
0.0222140122205019,
-0.05251315236091614,
0.020513102412223816,
0.007155807688832283,
0.041755348443984985,
0.009462535381317139,
0.049346406012773514,
-0.005052151158452034,
-0.0940437912940979,
0.1117711141705513,
-0.000514045066665858,
0.02807144820690155,
-0.023284098133444786,
0.012264848686754704,
0.03601779788732529,
-0.030327267944812775,
-0.03344941884279251,
-0.06792820990085602,
-0.04370725154876709,
0.0046125538647174835,
-0.052379533648490906,
0.07079358398914337,
-0.0626954510807991,
-0.04110538214445114,
-0.07139578461647034,
-0.05523330345749855,
0.0091610262170434,
-0.021157683804631233,
0.07170955836772919,
0.01217428408563137,
-2.1994437862655734e-33,
-0.07909737527370453,
0.06917773187160492,
0.04474319890141487,
0.029914774000644684,
-0.009625825099647045,
-0.04034341499209404,
0.012062209658324718,
0.04694805666804314,
0.07391177862882614,
-0.08578082919120789,
-0.04228324815630913,
0.00800358783453703,
0.012002083472907543,
0.003745192661881447,
-0.013635092414915562,
-0.03488917648792267,
-0.022287076339125633,
0.04961659386754036,
-0.06657534092664719,
0.018479375168681145,
-0.0023242849856615067,
0.06652529537677765,
-0.13775719702243805,
0.056282803416252136,
0.00041520525701344013,
0.08451927453279495,
-0.005276740062981844,
0.11736441403627396,
0.025011982768774033,
0.017975835129618645,
-0.0418069064617157,
0.05788201466202736,
-0.06450077891349792,
-0.05787474289536476,
-0.11423995345830917,
0.08971329778432846,
0.046427447348833084,
0.019994450733065605,
-0.06388211250305176,
0.06675306707620621,
0.032117798924446106,
-0.023318573832511902,
-0.0913504958152771,
0.0254567451775074,
0.029243240132927895,
-0.04800121858716011,
0.003117367159575224,
0.011490939185023308,
0.07202193886041641,
0.001857449533417821,
-0.005725907627493143,
0.05709010362625122,
-0.03723687678575516,
0.04653570428490639,
-0.11097656190395355,
-0.08068682998418808,
-0.10144039988517761,
-0.046303775161504745,
0.10345502197742462,
-0.032611045986413956,
-0.05807669088244438,
-0.002446449361741543,
0.046096134930849075,
0.0027260787319391966,
-0.06251367926597595,
-0.06048371642827988,
-0.0019998541101813316,
0.008761168457567692,
-0.031576015055179596,
0.04427594318985939,
-0.006407288834452629,
0.04739497974514961,
0.08043137937784195,
0.03336402401328087,
0.0856536328792572,
-0.0792699083685875,
-0.04094230756163597,
-0.020212681964039803,
0.022368041798472404,
0.005920854397118092,
0.04974296689033508,
0.03245239704847336,
0.003088461235165596,
0.03306247293949127,
0.05133149027824402,
0.011468397453427315,
0.023340333253145218,
0.0023552943021059036,
0.030275778844952583,
-0.016135066747665405,
-0.038671039044857025,
-0.025271810591220856,
0.01701785810291767,
0.12412843853235245,
-0.07847391813993454,
-4.590268787296736e-8,
0.006302231457084417,
0.03282799571752548,
0.04575279355049133,
0.037973884493112564,
0.04069773852825165,
0.03633905202150345,
-0.05829685926437378,
0.012219768017530441,
0.06059762462973595,
-0.04288183152675629,
0.0801662802696228,
-0.006774232257157564,
-0.028581179678440094,
-0.07309678196907043,
-0.02350718155503273,
0.09302537888288498,
0.04893292486667633,
-0.00969157088547945,
0.017657622694969177,
-0.020181208848953247,
0.021906275302171707,
0.048193857073783875,
0.00456066383048892,
0.00570612633600831,
0.059172648936510086,
-0.07550514489412308,
0.0014751986600458622,
0.032511379569768906,
0.04584004357457161,
0.057557884603738785,
0.06342291831970215,
0.02612270787358284,
0.0256443340331316,
0.06096271798014641,
-0.008911441080272198,
0.055052515119314194,
0.018000489100813866,
-0.04841572418808937,
-0.01671564020216465,
-0.02345733717083931,
-0.024889614433050156,
0.03337465971708298,
-0.06884916126728058,
-0.02503027208149433,
-0.025107763707637787,
0.027756541967391968,
-0.007500075735151768,
-0.041023045778274536,
0.0666566863656044,
0.033699892461299896,
0.024242272600531578,
-0.05175100266933441,
0.022708941251039505,
0.02140769548714161,
0.087569959461689,
0.003767263377085328,
0.058151375502347946,
-0.02953924611210823,
0.03895019367337227,
0.00564403785392642,
0.004395634867250919,
0.004943227861076593,
-0.01226165983825922,
0.0633637011051178
] |
Learning to Adapt Frozen CLIP for Few-Shot Test-Time Domain Adaptation | https://openreview.net/forum?id=TD3SGJfBC7 | [
"Zhixiang Chi",
"Li Gu",
"Huan Liu",
"Ziqiang Wang",
"Yanan Wu",
"Yang Wang",
"Konstantinos N Plataniotis"
] | Poster | Few-shot Test-Time Domain Adaptation focuses on adapting a model at test time to a specific domain using only a few unlabeled examples, addressing domain shift. Prior methods leverage CLIP's strong out-of-distribution (OOD) abilities by generating domain-specific prompts to guide its generalized, frozen features. However, since downstream datasets are not explicitly seen by CLIP, solely depending on the feature space knowledge is constrained by CLIP's prior knowledge. Notably, when using a less robust backbone like ViT-B/16, performance significantly drops on challenging real-world benchmarks. Departing from the state-of-the-art of inheriting the intrinsic OOD capability of CLIP, this work introduces learning directly on the input space to complement the dataset-specific knowledge for frozen CLIP. Specifically, an independent side branch is attached in parallel with CLIP and enforced to learn exclusive knowledge via revert attention. To better capture the dataset-specific label semantics for downstream adaptation, we propose to enhance the inter-dispersion among text features via greedy text ensemble and refinement. The text and visual features are then progressively fused in a domain-aware manner by a generated domain prompt to adapt toward a specific domain. Extensive experiments show our method's superiority on 5 large-scale benchmarks (WILDS and DomainNet), notably improving over smaller networks like ViT-B/16 with gains of \textbf{+5.1} in F1 for iWildCam and \textbf{+3.1\%} in WC Acc for FMoW. \href{https://github.com/chi-chi-zx/L2C}{Our Code: L2C} | Distribution shifts, Visual prompt, Foundation model, Few-Shot Test-Time Domain Adaptation | We propose to learn data-specific knowledge from the input space to complement frozen CLIP, and adapt CLIP towards domain shift. | 452 | null | [
-0.02979949302971363,
-0.0681326761841774,
0.04976746067404747,
0.0543118491768837,
0.13385209441184998,
-0.029553605243563652,
-0.0205076951533556,
-0.014904852025210857,
-0.012595619075000286,
-0.04891805350780487,
0.0241798497736454,
-0.04721040278673172,
-0.008036993443965912,
0.015000092796981335,
0.01765482686460018,
-0.01125542912632227,
0.16423015296459198,
0.06908256560564041,
-0.11252914369106293,
-0.008574658073484898,
0.014887114986777306,
-0.03386387974023819,
0.07757484167814255,
0.03051356039941311,
0.04236229509115219,
-0.005459521431475878,
-0.009180507622659206,
0.02046431601047516,
0.06033814698457718,
-0.024141913279891014,
0.018261125311255455,
0.02855321578681469,
-0.05094544589519501,
-0.0018733078613877296,
-0.052899882197380066,
0.03286632150411606,
-0.02255687303841114,
0.06947480887174606,
-0.0022000966127961874,
-0.03543434292078018,
0.027478348463773727,
-0.02737564593553543,
0.046236369758844376,
0.005952998995780945,
0.036981236189603806,
0.02877882495522499,
-0.01668507605791092,
-0.04324943199753761,
0.028899339959025383,
0.03977563604712486,
-0.0633464977145195,
0.009304450824856758,
-0.09891779720783234,
0.12321902066469193,
-0.02689243108034134,
0.08782626688480377,
0.07088085263967514,
-0.009050040505826473,
0.01224481500685215,
-0.008348013274371624,
-0.04543980956077576,
-0.09284551441669464,
0.017872493714094162,
0.003966341260820627,
-0.008919669315218925,
-0.02770237997174263,
0.018063979223370552,
0.04962367191910744,
-0.03810925781726837,
-0.02699163183569908,
-0.009336932562291622,
0.11149343103170395,
-0.028956202790141106,
0.02249973453581333,
0.02214217372238636,
0.07755515724420547,
0.006700360681861639,
-0.018781090155243874,
0.04598786681890488,
-0.05351482704281807,
0.10018905997276306,
0.009524893015623093,
0.033799171447753906,
0.03685382008552551,
0.07574938237667084,
0.037148479372262955,
0.02217257022857666,
0.03219829499721527,
-0.028167199343442917,
0.030165348201990128,
-0.05671272799372673,
-0.08485754579305649,
0.1199292540550232,
-0.0018482981249690056,
0.012594938278198242,
0.013237249106168747,
-0.004801811184734106,
-0.0795212984085083,
-0.028838463127613068,
0.06150194630026817,
-0.02721494249999523,
0.05396400764584541,
0.022224552929401398,
-0.03901493549346924,
-0.03099941648542881,
-0.035175345838069916,
0.033820345997810364,
0.012020514346659184,
0.015134114772081375,
-0.09450723975896835,
0.09948433190584183,
0.07153872400522232,
0.04885782301425934,
-0.08062878251075745,
0.06320652365684509,
-0.03844613581895828,
-0.08182380348443985,
-0.08810915052890778,
-0.028061071410775185,
0.028102383017539978,
-0.04759505018591881,
0.03621821478009224,
-0.08076461404561996,
0.00792797189205885,
0.08228251338005066,
0.03202452138066292,
0.03750406950712204,
2.7179359773140528e-33,
0.054922524839639664,
0.03208063915371895,
0.05550746992230415,
0.0023472036700695753,
0.038821496069431305,
-0.02636134810745716,
0.02342693880200386,
0.04385148733854294,
-0.08016910403966904,
-0.027205627411603928,
0.017433175817131996,
-0.04738904535770416,
-0.06106887385249138,
0.014041908085346222,
0.012947465293109417,
-0.009263147599995136,
-0.037751153111457825,
0.041916072368621826,
0.013543779961764812,
0.09918151795864105,
0.000711983535438776,
-0.012072192505002022,
-0.04849215969443321,
-0.056616272777318954,
0.011325767263770103,
-0.02213525027036667,
-0.021018099039793015,
-0.016700543463230133,
-0.04355481639504433,
0.028594201430678368,
-0.14076273143291473,
0.031013285741209984,
0.011672591790556908,
0.04284466803073883,
0.03288467228412628,
-0.08941642940044403,
-0.0016691351775079966,
-0.10344062000513077,
-0.0017423523822799325,
-0.009659995324909687,
-0.017208606004714966,
0.04911601170897484,
-0.03943224251270294,
0.007275094743818045,
-0.0687016025185585,
-0.0888603925704956,
0.017419753596186638,
0.030481988564133644,
-0.09539031982421875,
0.017638105899095535,
0.08192981034517288,
0.039165887981653214,
-0.020562533289194107,
-0.12728199362754822,
-0.07248122990131378,
-0.01901724934577942,
0.008213872089982033,
-0.01876457966864109,
0.01851743273437023,
0.07501183450222015,
-0.023668494075536728,
0.06974770128726959,
-0.06575923413038254,
0.13860563933849335,
0.022599797695875168,
0.06810952723026276,
0.02072226256132126,
0.012716582976281643,
-0.0036928383633494377,
-0.01217908225953579,
-0.04983367398381233,
-0.003125756047666073,
-0.0722840279340744,
-0.03553834930062294,
0.05774899572134018,
-0.008674575947225094,
0.029980450868606567,
-0.08262348175048828,
0.041913121938705444,
0.052041925489902496,
0.024174103513360023,
-0.026409205049276352,
-0.03223760798573494,
-0.05352737754583359,
-0.026977596804499626,
-0.05397450178861618,
0.05780346691608429,
-0.10186171531677246,
-0.014078471809625626,
-0.10616819560527802,
-0.025781599804759026,
-0.02587006613612175,
-0.05215182155370712,
-0.03317403793334961,
0.09188516438007355,
-1.0850625602230712e-33,
0.03698458895087242,
-0.018266072496771812,
-0.047906480729579926,
0.05328097566962242,
0.00869305245578289,
0.03567206859588623,
0.03167780488729477,
0.04700281098484993,
-0.01301400363445282,
-0.11074628680944443,
-0.05314551293849945,
-0.07885897159576416,
-0.02971620485186577,
-0.04123530164361,
-0.033923547714948654,
0.019523289054632187,
-0.0664583072066307,
0.014121118001639843,
-0.03573371842503548,
0.04519294947385788,
0.032650984823703766,
0.014385250397026539,
-0.02964583784341812,
0.09322835505008698,
-0.05647461488842964,
0.06506257504224777,
-0.04263991490006447,
0.08348345756530762,
0.002083294792100787,
-0.05295484885573387,
-0.028079111129045486,
0.04734344780445099,
-0.010853631421923637,
-0.031681325286626816,
-0.09891387820243835,
0.08434280008077621,
0.010663214139640331,
-0.0369425006210804,
-0.026389529928565025,
0.12006621062755585,
0.07476676255464554,
-0.0075831632129848,
-0.05404365062713623,
0.02709583193063736,
-0.05811145156621933,
0.0311907809227705,
-0.0548742413520813,
-0.008719809353351593,
0.047870904207229614,
0.05386994779109955,
-0.07785200327634811,
0.07051899284124374,
-0.05889322608709335,
0.05059228837490082,
-0.026104027405381203,
-0.09679213166236877,
-0.01045515201985836,
0.006130930967628956,
0.05325433984398842,
0.009763482958078384,
0.0038015784230083227,
0.001576752052642405,
-0.015290276147425175,
-0.0278614591807127,
0.09377432614564896,
-0.02467290498316288,
-0.011445000767707825,
-0.002973447786644101,
-0.012925878167152405,
-0.023567013442516327,
0.035111114382743835,
-0.03130742907524109,
0.030726328492164612,
-0.09494640678167343,
0.03897155076265335,
-0.01196461170911789,
0.02227315865457058,
-0.004810838494449854,
-0.014854755252599716,
0.009328631684184074,
0.029281895607709885,
-0.03570855036377907,
0.02248716540634632,
0.10714485496282578,
0.05734523385763168,
0.17202746868133545,
0.03541436046361923,
0.04260082542896271,
0.04923618584871292,
-0.029168564826250076,
-0.07213485240936279,
0.025527220219373703,
-0.030490189790725708,
0.0632345974445343,
0.023312389850616455,
-4.1546655893398565e-8,
-0.07327751070261002,
0.024815669283270836,
-0.0553128756582737,
0.060864467173814774,
0.0033595976419746876,
-0.051448434591293335,
-0.06269502639770508,
0.06645381450653076,
-0.018513254821300507,
-0.00537867471575737,
0.07675646990537643,
0.035224005579948425,
0.01787271350622177,
-0.018569963052868843,
-0.02869069203734398,
0.0499921590089798,
0.03439553454518318,
0.04623083770275116,
-0.02484995126724243,
-0.0014503641286864877,
0.027701159939169884,
0.008755069226026535,
0.09090250730514526,
-0.0596478246152401,
-0.02217162400484085,
-0.10533981025218964,
0.014950335025787354,
0.04286174848675728,
0.15046299993991852,
-0.03198990225791931,
-0.017665646970272064,
-0.040752310305833817,
-0.03208437189459801,
-0.06572602689266205,
0.016702722758054733,
0.04401518031954765,
0.07597333192825317,
0.011145077645778656,
0.01934252493083477,
0.016402482986450195,
-0.02045905776321888,
0.058785777539014816,
-0.0711122453212738,
-0.010244706645607948,
-0.03547452762722969,
-0.01458654459565878,
0.04923610761761665,
-0.04301232099533081,
0.03515544906258583,
-0.015973055735230446,
-0.023932864889502525,
-0.03451045975089073,
-0.005185138434171677,
0.07085701078176498,
0.04669906571507454,
0.030722199007868767,
0.006175294052809477,
-0.013747621327638626,
0.016937632113695145,
0.036155980080366135,
0.01631789095699787,
-0.011910808272659779,
-0.0018345535499975085,
0.016754228621721268
] |
Tuning-Free Bilevel Optimization: New Algorithms and Convergence Analysis | https://openreview.net/forum?id=A4aG3XeIO7 | [
"Yifan Yang",
"Hao Ban",
"Minhui Huang",
"Shiqian Ma",
"Kaiyi Ji"
] | Poster | Bilevel optimization has recently attracted considerable attention due to its abundant applications in machine learning problems. However, existing methods rely on prior knowledge of problem parameters to determine stepsizes, resulting in significant effort in tuning stepsizes when these parameters are unknown. In this paper, we propose two novel tuning-free algorithms, D-TFBO and S-TFBO. D-TFBO employs a double-loop structure with stepsizes adaptively adjusted by the "inverse of cumulative gradient norms" strategy. S-TFBO features a simpler fully single-loop structure that updates three variables simultaneously with a theory-motivated joint design of adaptive stepsizes for all variables. We provide a comprehensive convergence analysis for both algorithms and show that D-TFBO and S-TFBO respectively require $\mathcal{O}(\frac{1}{\epsilon})$ and $\mathcal{O}(\frac{1}{\epsilon}\log^4(\frac{1}{\epsilon}))$ iterations to find an $\epsilon$-accurate stationary point, (nearly) matching their well-tuned counterparts using the information of problem parameters. Experiments on various problems show that our methods achieve performance comparable to existing well-tuned approaches, while being more robust to the selection of initial stepsizes.
To the best of our knowledge, our methods are the first to completely eliminate the need for stepsize tuning, while achieving theoretical guarantees. | Bilevel Optimization, Tuning-Free, Adaptive Optimization | null | 449 | 2410.05140 | [
-0.02438291907310486,
-0.07461460679769516,
-0.04533182084560394,
0.009285615757107735,
-0.04586018994450569,
-0.07804132997989655,
-0.029498925432562828,
-0.03500489518046379,
-0.05635197088122368,
-0.03580727055668831,
-0.021811477839946747,
0.07392539829015732,
0.0058747692964971066,
-0.017115183174610138,
0.005153038538992405,
0.03412884101271629,
0.029589476063847542,
0.05222322791814804,
0.01754344068467617,
-0.07145113497972488,
-0.05211775749921799,
-0.02880963310599327,
0.03986004367470741,
0.03629313036799431,
0.019384093582630157,
-0.023119289427995682,
-0.024855269119143486,
0.03914719074964523,
-0.028639113530516624,
-0.03888341411948204,
0.03227057680487633,
0.0043639144860208035,
-0.015043975785374641,
-0.08414191752672195,
-0.050398875027894974,
0.017982659861445427,
0.014275520108640194,
0.07380130141973495,
-0.014247799292206764,
0.018999386578798294,
-0.012770908884704113,
0.02598441205918789,
0.02551659196615219,
0.01102373469620943,
0.07271374762058258,
0.0075135924853384495,
-0.027239026501774788,
-0.051789481192827225,
0.009149059653282166,
-0.050493109971284866,
-0.031793635338544846,
-0.031453829258680344,
-0.04519191384315491,
0.021866608411073685,
0.0346059650182724,
0.03584316372871399,
0.02041466347873211,
0.012847197242081165,
0.06932529807090759,
-0.044974979013204575,
0.007249122019857168,
-0.12958799302577972,
0.04741964116692543,
-0.035809893161058426,
-0.019411595538258553,
-0.0254451185464859,
0.02310844324529171,
0.000951441703364253,
-0.014782768674194813,
0.04798392206430435,
-0.06077871099114418,
0.039659131318330765,
-0.03152714669704437,
0.054807744920253754,
0.05295741185545921,
0.0856972336769104,
0.05724165216088295,
0.05078732967376709,
0.016187161207199097,
-0.06335324048995972,
-0.05611804872751236,
0.02103319950401783,
0.0009000324062071741,
-0.050947658717632294,
0.0699097216129303,
-0.060146745294332504,
-0.042743295431137085,
-0.049460362643003464,
0.1623525470495224,
0.01727980561554432,
0.01879543624818325,
0.02677866816520691,
-0.09829604625701904,
-0.07618901133537292,
-0.004044566303491592,
-0.022526711225509644,
0.059152621775865555,
0.014905654825270176,
0.013040981255471706,
0.11391013115644455,
0.036234986037015915,
0.08908440917730331,
0.017785530537366867,
-0.020618204027414322,
0.02305196411907673,
-0.026704058051109314,
0.07705976814031601,
0.07474779337644577,
0.042843595147132874,
-0.1222613975405693,
0.016076279804110527,
0.015938101336359978,
0.07614178210496902,
-0.03640191629528999,
-0.02806413732469082,
-0.012734838761389256,
-0.051066525280475616,
-0.06610558927059174,
-0.0094350790604949,
0.09248465299606323,
-0.0510195717215538,
0.010572719387710094,
0.014644068665802479,
0.0844370424747467,
0.0015518831787630916,
-0.02002444490790367,
-0.08749619126319885,
2.435578560826693e-33,
-0.020041463896632195,
0.13587871193885803,
0.09556113928556442,
-0.06838087737560272,
-0.005589147098362446,
-0.0064259790815413,
0.02867649495601654,
0.003325417172163725,
-0.016665607690811157,
0.07667415589094162,
-0.01808461733162403,
0.011324546299874783,
-0.11299803853034973,
0.03691689297556877,
0.057167332619428635,
-0.060015954077243805,
0.020326385274529457,
0.052399516105651855,
-0.005473829340189695,
-0.05890624225139618,
0.08983862400054932,
-0.09597175568342209,
-0.045624710619449615,
-0.049848128110170364,
0.017019616439938545,
-0.030049482360482216,
0.09806715697050095,
-0.013871675357222557,
-0.1171131357550621,
0.024494532495737076,
-0.030186252668499947,
0.044637490063905716,
-0.019905364140868187,
0.04559169337153435,
0.035356588661670685,
-0.06889073550701141,
-0.02668396756052971,
-0.024030597880482674,
-0.02239554561674595,
0.09060502797365189,
-0.0032879936043173075,
-0.003008653875440359,
0.02289697527885437,
-0.028096193447709084,
-0.11054476350545883,
-0.04438946768641472,
-0.04585295915603638,
-0.01812894269824028,
0.03957038372755051,
-0.04764724522829056,
-0.05119195953011513,
0.038814544677734375,
-0.051065005362033844,
-0.02373487688601017,
0.022129565477371216,
0.002969641238451004,
0.057287950068712234,
0.02263457141816616,
-0.010707519017159939,
0.09384699910879135,
0.018116774037480354,
0.07230629771947861,
-0.11651047319173813,
0.018110375851392746,
-0.023192308843135834,
0.010420285165309906,
0.04803371801972389,
0.01927301660180092,
0.004991461522877216,
-0.06841734796762466,
0.001247038715519011,
-0.011617042124271393,
0.001095551298931241,
-0.022204279899597168,
0.024225831031799316,
-0.03566388413310051,
0.11805102229118347,
-0.09282135218381882,
-0.01585211418569088,
-0.14335811138153076,
-0.013004043139517307,
0.0951898917555809,
0.03607012331485748,
-0.05271191522479057,
-0.005633241962641478,
-0.10603169351816177,
0.03606294468045235,
-0.05654722452163696,
-0.13274195790290833,
-0.05279436707496643,
-0.14554539322853088,
0.08424517512321472,
-0.00349688739515841,
0.026238134130835533,
0.010529261082410812,
-2.8332784217526127e-33,
0.023487331345677376,
-0.09197018295526505,
0.03609226644039154,
0.027266696095466614,
0.010456787422299385,
0.046600691974163055,
-0.05021471530199051,
-0.08515430241823196,
0.03629491478204727,
-0.1063266322016716,
-0.04429766908288002,
-0.025264238938689232,
-0.012187854386866093,
0.044723015278577805,
0.0756581723690033,
0.0628308355808258,
-0.059183232486248016,
0.045340728014707565,
0.041133224964141846,
0.04996064677834511,
-0.05630337446928024,
0.04762118309736252,
-0.00738459313288331,
0.04179152846336365,
-0.05405723676085472,
-0.020524796098470688,
0.01634468510746956,
0.10514000058174133,
0.01154556404799223,
0.04332762584090233,
-0.07855087518692017,
0.04448790103197098,
-0.01920389011502266,
0.05266208201646805,
-0.027694713324308395,
0.08796293288469315,
-0.01360075082629919,
-0.0021724204998463392,
0.08376219868659973,
0.10861753672361374,
0.03655767813324928,
0.01185163389891386,
-0.025501461699604988,
-0.033999964594841,
-0.0035980537068098783,
0.027812648564577103,
-0.01681719347834587,
0.019477922469377518,
0.014701089821755886,
0.02986552193760872,
0.062241073697805405,
-0.034069109708070755,
-0.059835825115442276,
0.029366107657551765,
-0.010371960699558258,
0.028539931401610374,
-0.03235610947012901,
-0.0176263228058815,
0.050583187490701675,
-0.05512550473213196,
-0.01790483295917511,
0.014248842373490334,
-0.01503467932343483,
0.025482915341854095,
0.03869277238845825,
0.028136907145380974,
0.006534748710691929,
0.001397768035531044,
0.014154408127069473,
0.004770045168697834,
-0.05419379472732544,
0.002314100507646799,
0.0048906500451266766,
0.0018287405837327242,
-0.03210838884115219,
0.008749955333769321,
0.046795718371868134,
-0.03527407348155975,
-0.04004603624343872,
-0.03735172003507614,
0.042253267019987106,
0.05860588327050209,
0.0030949132051318884,
0.028754115104675293,
0.013729125261306763,
0.0781993642449379,
0.033733218908309937,
0.0995694026350975,
0.010430166497826576,
-0.07860584557056427,
0.04239717498421669,
0.022064868360757828,
0.003508113557472825,
-0.0029654030222445726,
-0.034666020423173904,
-4.893740879197139e-8,
-0.09061124920845032,
0.008615861646831036,
-0.02050110511481762,
0.05503687635064125,
0.1332133561372757,
-0.03385576605796814,
0.01188070047646761,
-0.018402332440018654,
0.0047214943915605545,
-0.1153736487030983,
0.03242415189743042,
0.011795062571763992,
0.026425320655107498,
0.04418487101793289,
0.005747513379901648,
0.032880302518606186,
-0.04450046271085739,
-0.0021943014580756426,
-0.03612246736884117,
0.0311551783233881,
-0.07974094897508621,
0.059290919452905655,
-0.014314080588519573,
-0.08321316540241241,
0.07350461184978485,
-0.1349320113658905,
-0.028861956670880318,
0.07798204571008682,
0.15077725052833557,
-0.0006520245224237442,
0.023855356499552727,
0.0018089207587763667,
-0.0396551750600338,
0.009396294131875038,
0.040008027106523514,
0.02805274911224842,
0.01889636740088463,
0.02283390425145626,
-0.041194524616003036,
-0.01740661822259426,
-0.009719133377075195,
0.08403211832046509,
-0.02404754050076008,
-0.021075917407870293,
-0.010470024310052395,
-0.006119673606008291,
-0.0027323660906404257,
-0.013956432230770588,
-0.027663815766572952,
-0.02408614009618759,
0.0013321825535967946,
0.059550147503614426,
0.0379619263112545,
0.006739472504705191,
0.046687085181474686,
-0.00922160129994154,
0.023869147524237633,
-0.01890913024544716,
0.007441693916916847,
0.07699139416217804,
-0.04117541387677193,
-0.0790805071592331,
0.000939427874982357,
-0.028260674327611923
] |
ParetoFlow: Guided Flows in Multi-Objective Optimization | https://openreview.net/forum?id=mLyyB4le5u | [
"Ye Yuan",
"Can Chen",
"Christopher Pal",
"Xue Liu"
] | Poster | In offline multi-objective optimization (MOO), we leverage an offline dataset of designs and their associated labels to simultaneously minimize multiple objectives. This setting more closely mirrors complex real-world problems compared to single-objective optimization. Recent works mainly employ evolutionary algorithms and Bayesian optimization, with limited attention given to the generative modeling capabilities inherent in such data. In this study, we explore generative modeling in offline MOO through flow matching, noted for its effectiveness and efficiency. We introduce \textit{ParetoFlow}, specifically designed to guide flow sampling to approximate the Pareto front. Traditional predictor~(classifier) guidance is inadequate for this purpose because it models only a single objective. In response, we propose a \textit{multi-objective predictor guidance} module that
assigns each sample a weight vector, representing a weighted distribution across multiple objective predictions. A local filtering scheme is introduced to address non-convex Pareto fronts. These weights uniformly cover the entire objective space, effectively directing sample generation towards the Pareto front. Since distributions with similar weights tend to generate similar samples, we introduce a \textit{neighboring evolution} module to foster knowledge sharing among neighboring distributions. This module generates offspring from these distributions, and selects the most promising one for the next iteration. Our method achieves state-of-the-art performance across various tasks. Our code is available. | Multi-objective optimization; flow matching; classifier guidance. | In offline multi-objective optimization, we introduce a ParetoFlow method, specifically designed to guide flow sampling to approximate the Pareto front. | 442 | 2412.03718 | [
-0.09290182590484619,
-0.08474478125572205,
-0.014655972830951214,
0.01680310256779194,
0.04306913912296295,
-0.045390576124191284,
-0.0026343935169279575,
-0.05424875393509865,
-0.05139530077576637,
-0.05774219334125519,
-0.09679634124040604,
-0.14312711358070374,
0.03460043668746948,
-0.019931349903345108,
-0.030576331540942192,
0.05984295532107353,
0.1731012463569641,
0.0942651554942131,
-0.059271227568387985,
-0.025317592546343803,
0.042127110064029694,
0.06898751854896545,
0.00929044745862484,
-0.005628658924251795,
0.01371862180531025,
0.05051257461309433,
0.04113002121448517,
-0.02200956456363201,
-0.018327949568629265,
0.008117049001157284,
0.056010257452726364,
0.019357629120349884,
-0.047907859086990356,
0.012893839739263058,
-0.03360443189740181,
0.07671485841274261,
-0.1316801905632019,
0.091514952480793,
0.02298394963145256,
-0.0174113679677248,
-0.034950077533721924,
0.022029204294085503,
0.011207115836441517,
-0.011380949057638645,
-0.004694926552474499,
0.006284922361373901,
-0.013625722378492355,
-0.020589837804436684,
-0.004164733458310366,
0.0712335854768753,
-0.03938288986682892,
-0.040539681911468506,
-0.09497351944446564,
0.09209434688091278,
0.013150927610695362,
0.01743117906153202,
0.05172212794423103,
-0.0426713190972805,
0.05451894924044609,
0.010239583440124989,
-0.07696285843849182,
-0.1314895898103714,
-0.02999415062367916,
-0.019991589710116386,
-0.031054247170686722,
-0.10919485986232758,
-0.016264591366052628,
0.060217324644327164,
0.0188616830855608,
-0.018912317231297493,
0.0523785762488842,
0.046546392142772675,
0.0058901687152683735,
-0.002868463983759284,
0.004980870056897402,
0.07822709530591965,
0.08475923538208008,
0.008574358187615871,
0.008191944099962711,
-0.08624005317687988,
0.017016533762216568,
0.04992242157459259,
0.02729712799191475,
0.05913874879479408,
0.0778467133641243,
-0.0170740969479084,
-0.029613755643367767,
0.020731421187520027,
0.10994277149438858,
0.07229278236627579,
-0.03538813441991806,
0.06473729014396667,
-0.006262310314923525,
-0.028644494712352753,
-0.01533757708966732,
-0.01803414709866047,
0.0059914072044193745,
-0.11562453210353851,
0.025311654433608055,
0.0795440822839737,
-0.025978202000260353,
0.06480350345373154,
0.08226406574249268,
-0.006843144539743662,
0.04701138660311699,
-0.0839308574795723,
0.06081918254494667,
0.03774789348244667,
0.0136702386662364,
-0.07353776693344116,
0.01911727525293827,
0.04294528067111969,
0.07006612420082092,
0.021516364067792892,
-0.017719246447086334,
-0.008781656622886658,
0.007303272373974323,
0.014335145242512226,
-0.047995708882808685,
0.1456300914287567,
-0.12451809644699097,
0.01234347652643919,
-0.010385925881564617,
-0.01440422236919403,
0.035084068775177,
0.00032422313233837485,
-0.049678679555654526,
8.075093340462244e-34,
-0.003800006117671728,
-0.0361454002559185,
0.04309222474694252,
0.05085140839219093,
0.04900841787457466,
-0.042065609246492386,
-0.010500754229724407,
0.01433857623487711,
-0.015428051352500916,
-0.028811056166887283,
0.04997200146317482,
-0.04886600375175476,
-0.08644630014896393,
0.07399482280015945,
0.04073120653629303,
-0.0296355951577425,
-0.04613112285733223,
0.08376516401767731,
-0.025967814028263092,
-0.006176048889756203,
0.024909302592277527,
-0.042238835245370865,
-0.02521316334605217,
-0.00706798629835248,
0.06356150656938553,
0.011899988166987896,
0.07612188160419464,
-0.0492556169629097,
-0.04989180341362953,
0.04778937250375748,
-0.06172408536076546,
-0.02506164088845253,
0.06924780458211899,
0.0033903506118804216,
-0.041813526302576065,
0.017598764970898628,
-0.011589858680963516,
-0.06609274446964264,
-0.055089935660362244,
0.024753302335739136,
-0.05830883979797363,
0.024385549128055573,
0.053285758942365646,
-0.007567605469375849,
-0.08763062953948975,
-0.12892408668994904,
0.016367202624678612,
0.022696351632475853,
0.04891114681959152,
-0.05600311607122421,
0.023841962218284607,
-0.030414920300245285,
0.06893139332532883,
-0.015445665456354618,
-0.0464395210146904,
-0.03905024006962776,
-0.039313722401857376,
0.04600328952074051,
0.048512525856494904,
-0.0025319538544863462,
-0.09110140800476074,
0.07775957882404327,
-0.06034942343831062,
-0.000253912789048627,
0.06662064045667648,
0.025259384885430336,
0.027341695502400398,
0.011391104198992252,
0.119794562458992,
-0.0049433633685112,
-0.03734621778130531,
0.004035951569676399,
0.005718146450817585,
-0.06472262740135193,
0.008367709815502167,
0.05596066638827324,
0.02670971304178238,
0.004331156611442566,
0.036459580063819885,
0.010844998992979527,
-0.019508838653564453,
0.10991127043962479,
0.028649363666772842,
0.004690590780228376,
0.10427607595920563,
0.0048459311947226524,
0.039653241634368896,
-0.04440489783883095,
-0.10600820928812027,
0.03301245719194412,
-0.0378391370177269,
-0.03263159468770027,
-0.1263984888792038,
-0.04043947532773018,
-0.0035797907039523125,
-4.047354242043315e-34,
-0.09504511952400208,
-0.03913270682096481,
0.011412600986659527,
0.004723111167550087,
0.055085889995098114,
-0.0306417066603899,
-0.10889866203069687,
-0.02149861305952072,
0.027511730790138245,
-0.03550437092781067,
-0.03399030864238739,
0.005005904007703066,
0.03807893767952919,
-0.02157844603061676,
0.0045286123640835285,
0.036326006054878235,
0.005555733572691679,
0.002104545943439007,
-0.03088420443236828,
-0.031177004799246788,
0.06192991882562637,
0.02055492252111435,
-0.04473455995321274,
0.015419022180140018,
-0.06519325822591782,
0.014275127090513706,
-0.04905061423778534,
0.09251927584409714,
0.025179142132401466,
0.03928673639893532,
0.006159091368317604,
-0.04691656306385994,
-0.008479504846036434,
-0.021409297361969948,
-0.018124829977750778,
0.07987292855978012,
-0.05106339603662491,
0.09341783076524734,
-0.04678816720843315,
0.12210488319396973,
-0.034777846187353134,
-0.023863181471824646,
-0.043361764401197433,
0.013125811703503132,
-0.040520768612623215,
0.013124029152095318,
0.018557270988821983,
-0.04136981815099716,
0.05196493864059448,
0.017759419977664948,
-0.038216397166252136,
0.10204317420721054,
-0.06437190622091293,
0.024711262434720993,
-0.07281333953142166,
-0.11138063669204712,
0.0034186416305601597,
-0.06054281070828438,
0.0465053990483284,
-0.058042947202920914,
0.07167776674032211,
0.002125238534063101,
0.017393186688423157,
-0.03832951560616493,
0.0074890246614813805,
-0.040254466235637665,
0.03625374659895897,
0.012607102282345295,
-0.0566776841878891,
0.02053748071193695,
-0.034026987850666046,
-0.046519625931978226,
0.02711579017341137,
-0.018362855538725853,
0.004910911899060011,
0.04400341585278511,
0.08226784318685532,
-0.009268664754927158,
0.031857699155807495,
-0.0036331939045339823,
0.03271933272480965,
0.08013366907835007,
0.018418947234749794,
0.042685337364673615,
0.1219235509634018,
-0.06436694413423538,
0.03386104106903076,
0.0875442624092102,
0.06961246579885483,
0.0346202589571476,
-0.0018401216948404908,
-0.02469080127775669,
0.03420523554086685,
-0.03652003034949303,
-0.06717368960380554,
-4.216406779278259e-8,
-0.05178084969520569,
0.01277719996869564,
0.04870883747935295,
0.04147838428616524,
0.045007627457380295,
0.07823623716831207,
-0.04520116001367569,
0.09468657523393631,
-0.008399171754717827,
-0.07642124593257904,
0.08368098735809326,
0.04855629429221153,
-0.0008972043870016932,
-0.01745549961924553,
0.03634350001811981,
0.009944909252226353,
-0.0010944624664261937,
-0.004803793970495462,
-0.01761581562459469,
-0.0776636004447937,
0.005770018324255943,
0.004320081323385239,
-0.005987814161926508,
-0.022071614861488342,
0.07631093263626099,
-0.025085415691137314,
0.0011329972185194492,
-0.04831117391586304,
0.06269412487745285,
0.007294420152902603,
-0.019509712234139442,
-0.02965814247727394,
-0.02764890529215336,
0.0029396822210401297,
-0.002539712470024824,
0.07123372703790665,
0.009884999133646488,
-0.02754446119070053,
-0.03192775696516037,
-0.03827561438083649,
0.021792154759168625,
0.025840289890766144,
-0.04508599266409874,
0.006947794929146767,
0.0510612353682518,
0.02919759787619114,
-0.02591073513031006,
-0.07467366755008698,
0.01861715130507946,
0.028274431824684143,
-0.018604815006256104,
-0.061313655227422714,
0.009568697772920132,
0.022715190425515175,
0.05866561830043793,
-0.04193725809454918,
-0.022974791005253792,
-0.09885957092046738,
0.07014927268028259,
0.0249325018376112,
0.005160783417522907,
0.05140124633908272,
-0.05080191791057587,
0.011844528838992119
] |
Quantized Spike-driven Transformer | https://openreview.net/forum?id=5J9B7Sb8rO | [
"Xuerui Qiu",
"Malu Zhang",
"Jieyuan Zhang",
"Wenjie Wei",
"Honglin Cao",
"Junsheng Guo",
"Rui-Jie Zhu",
"Yimeng Shan",
"Yang Yang",
"Haizhou Li"
] | Poster | Spiking neural networks (SNNs) are emerging as a promising energy-efficient alternative to traditional artificial neural networks (ANNs) due to their spike-driven paradigm.
However, recent research in the SNN domain has mainly focused on enhancing accuracy by designing large-scale Transformer structures, which typically rely on substantial computational resources, limiting their deployment on resource-constrained devices.
To overcome this challenge, we propose a quantized spike-driven Transformer baseline (QSD-Transformer), which achieves reduced resource demands by utilizing a low bit-width parameter.
Regrettably, the QSD-Transformer often suffers from severe performance degradation.
In this paper, we first conduct empirical analysis and find that the bimodal distribution of quantized spike-driven self-attention (Q-SDSA) leads to spike information distortion (SID) during quantization, causing significant performance degradation. To mitigate this issue, we take inspiration from mutual information entropy and propose a bi-level optimization strategy to rectify the information distribution in Q-SDSA.
Specifically, at the lower level, we introduce an information-enhanced LIF to rectify the information distribution in Q-SDSA.
At the upper level, we propose a fine-grained distillation scheme for the QSD-Transformer to align the distribution in Q-SDSA with that in the counterpart ANN.
By integrating the bi-level optimization strategy, the QSD-Transformer can attain enhanced energy efficiency without sacrificing its high-performance advantage.
We validate the QSD-Transformer on various visual tasks, and experimental results indicate that our method achieves state-of-the-art results in the SNN domain.
For instance, when compared to the prior SNN benchmark on ImageNet, the QSD-Transformer achieves 80.3\% top-1 accuracy, accompanied by significant reductions of 6.0$\times$ and 8.1$\times$ in power consumption and model size, respectively. Code is available at https://github.com/bollossom/QSD-Transformer. | Spiking Neural Network+Spike-driven+Quantized Spiking Transformer+ Neuromorphic Computing | We proposed a quantized spike-driven Transformer that achieves state-of-the-art results on various vision tasks and maintains a tiny model size. | 436 | 2501.13492 | [
-0.06091408059000969,
0.0027377698570489883,
0.051966343075037,
-0.0013493377482518554,
-0.037117063999176025,
-0.030011886730790138,
0.058016087859869,
-0.029480434954166412,
0.08479203283786774,
-0.01212149579077959,
-0.05788236856460571,
-0.03480861335992813,
-0.01636878214776516,
0.06240087002515793,
-0.008733208291232586,
0.056085485965013504,
0.04938601702451706,
0.05664488300681114,
-0.08209637552499771,
-0.02345617674291134,
0.13571138679981232,
-0.027707211673259735,
0.015690524131059647,
0.020319873467087746,
0.07069092988967896,
0.014572473242878914,
0.007352861110121012,
-0.090037040412426,
-0.006256699562072754,
-0.029933970421552658,
0.0010972851887345314,
-0.0618366114795208,
-0.025681132450699806,
0.006941413972526789,
-0.03469713032245636,
0.03058764897286892,
-0.062143344432115555,
0.00773412873968482,
0.027136001735925674,
-0.01967838779091835,
0.06103737652301788,
0.0057243830524384975,
0.03651927411556244,
0.030898988246917725,
0.02913982979953289,
0.06158634275197983,
0.04695836454629898,
-0.020780231803655624,
-0.018039308488368988,
-0.044376276433467865,
0.07338022440671921,
-0.02055678516626358,
0.0329236276447773,
0.06267945468425751,
0.00845260638743639,
0.12313048541545868,
0.01831158623099327,
0.07516689598560333,
-0.014922301284968853,
0.05237700790166855,
-0.024429207667708397,
-0.018601996824145317,
0.10060764849185944,
-0.04686445742845535,
-0.07435295730829239,
-0.014060791581869125,
0.05249854922294617,
0.021609511226415634,
0.002263671485707164,
-0.002568187890574336,
-0.032414667308330536,
0.020194727927446365,
-0.05259955674409866,
0.003516820026561618,
0.07509151101112366,
0.025344595313072205,
0.0594131276011467,
0.019698958843946457,
0.037361305207014084,
-0.07333365827798843,
0.02316412515938282,
0.027039743959903717,
0.03185011073946953,
-0.0021144049242138863,
0.17730242013931274,
0.011228756979107857,
0.01081632450222969,
0.08368153870105743,
-0.03977859020233154,
-0.00567167904227972,
-0.08748908340930939,
-0.06009146198630333,
0.022978749126195908,
-0.024308621883392334,
0.0517389141023159,
-0.015104081481695175,
-0.0094284163787961,
-0.03226083144545555,
-0.03842860087752342,
0.01985258236527443,
0.029107946902513504,
-0.016195831820368767,
0.043729573488235474,
0.011496729217469692,
0.03432052582502365,
-0.07278077304363251,
0.031355082988739014,
0.11332747340202332,
-0.014240315183997154,
-0.011173341423273087,
0.08385669440031052,
0.07082445919513702,
0.02489054948091507,
-0.05868321657180786,
0.05611637234687805,
0.005012763664126396,
-0.01793590746819973,
0.10412273555994034,
-0.02858346328139305,
-0.01922571286559105,
-0.0115223228931427,
-0.05837075412273407,
-0.08490893244743347,
-0.025593988597393036,
-0.000948633998632431,
0.02881905995309353,
-0.10557419061660767,
1.4591933691141453e-34,
-0.04454006627202034,
0.044075798243284225,
-0.007388036232441664,
-0.043010860681533813,
0.040528908371925354,
-0.06312642246484756,
0.015769682824611664,
0.028312688693404198,
0.01433187909424305,
0.02395760454237461,
-0.007874933071434498,
0.050201691687107086,
-0.054559800773859024,
0.06532822549343109,
0.03457549214363098,
-0.01946469023823738,
-0.09489647299051285,
-0.029738733544945717,
0.0041745430789887905,
-0.1850692331790924,
0.04092776030302048,
-0.05284329503774643,
-0.01674260012805462,
-0.007558383978903294,
-0.04190647974610329,
-0.0321892574429512,
0.06894524395465851,
-0.04527183622121811,
-0.026321211829781532,
-0.03748047351837158,
-0.11992368847131729,
0.03193572908639908,
-0.0074866157956421375,
-0.13652203977108002,
0.08554571866989136,
-0.06600725650787354,
0.07777807861566544,
0.0023689328227192163,
-0.038932085037231445,
0.029714815318584442,
0.018053874373435974,
0.04298513010144234,
0.016597969457507133,
-0.06679819524288177,
-0.05476310849189758,
-0.017092792317271233,
0.07447448372840881,
-0.01941029354929924,
-0.008006188087165356,
-0.1095026507973671,
0.020150018855929375,
-0.09839737415313721,
0.007655538152903318,
0.08619491010904312,
0.13544899225234985,
-0.03211270645260811,
0.1070738360285759,
0.021939611062407494,
0.02698853425681591,
0.10127659887075424,
-0.015031030401587486,
0.005944770760834217,
0.012799549847841263,
0.05264632776379585,
0.005084272474050522,
0.012497645802795887,
0.05381907895207405,
0.07269443571567535,
0.0029746631626039743,
-0.03269970789551735,
0.03454899042844772,
0.012386548332870007,
-0.02461162395775318,
-0.06780920177698135,
0.014224565587937832,
0.022479532286524773,
0.048852141946554184,
0.009441280737519264,
-0.08291872590780258,
-0.0073107467032969,
0.030167333781719208,
-0.06955544650554657,
0.001729886862449348,
0.05694032087922096,
-0.015256015583872795,
0.020131954923272133,
0.04808959737420082,
-0.0548858568072319,
-0.11157208681106567,
0.025520138442516327,
-0.047366321086883545,
-0.007450534962117672,
0.062060512602329254,
-0.05656734108924866,
-0.12375098466873169,
3.218970056871906e-34,
-0.09063591808080673,
0.04296378046274185,
-0.005075410474091768,
0.03325536102056503,
-0.007636623457074165,
-0.03400402516126633,
0.035611290484666824,
-0.003394561819732189,
0.014758295379579067,
-0.059671491384506226,
-0.11259040236473083,
-0.04942324385046959,
-0.013974811881780624,
0.009936301037669182,
0.053592123091220856,
-0.008626577444374561,
-0.06459550559520721,
0.03628738597035408,
-0.05256876349449158,
0.00849529355764389,
-0.0027828076854348183,
0.08282520622015,
-0.049670152366161346,
0.004122653976082802,
-0.019515592604875565,
0.07745125889778137,
-0.07131700962781906,
0.10920214653015137,
0.04232459515333176,
-0.07486863434314728,
-0.04207443818449974,
-0.048111435025930405,
-0.028472768142819405,
-0.10008812695741653,
-0.00913979671895504,
0.027564257383346558,
0.04803626239299774,
0.0026882574893534184,
-0.017462365329265594,
-0.006478970870375633,
0.06415146589279175,
-0.03054121322929859,
-0.0376291498541832,
0.07848980277776718,
0.02033107355237007,
-0.03705073520541191,
-0.08116964250802994,
-0.02287805639207363,
-0.032378047704696655,
0.12321571260690689,
0.011480488814413548,
0.0021065319888293743,
-0.057278670370578766,
0.026457644999027252,
-0.10721690952777863,
-0.11930273473262787,
-0.026964303106069565,
0.04243101552128792,
0.06019067391753197,
0.03396369889378548,
-0.03461340814828873,
-0.06149500980973244,
0.015369359403848648,
-0.06030477583408356,
-0.03883447125554085,
-0.04183780029416084,
0.0440676249563694,
-0.023741330951452255,
0.07252664864063263,
-0.035631321370601654,
0.027698921039700508,
-0.008358060382306576,
0.07266724854707718,
-0.021611973643302917,
-0.01882242225110531,
-0.033416636288166046,
0.006674416828900576,
-0.028519609943032265,
0.029279800131917,
-0.007410841062664986,
0.020116930827498436,
0.037570200860500336,
-0.009274689480662346,
-0.023765411227941513,
0.010659418068826199,
0.10182392597198486,
0.04729897901415825,
0.041520796716213226,
0.06646886467933655,
-0.026145240291953087,
-0.016573768109083176,
0.033138010650873184,
-0.01740112341940403,
0.06288015842437744,
-0.050437454134225845,
-4.2225476448720656e-8,
-0.014970434829592705,
0.02248353138566017,
0.006587150972336531,
-0.033751681447029114,
0.038369257003068924,
-0.01566108874976635,
0.0388181246817112,
0.09219304472208023,
-0.028937075287103653,
-0.036288630217313766,
0.09856658428907394,
-0.040146078914403915,
-0.009542540647089481,
0.001844460959546268,
-0.0055242059752345085,
0.031982991844415665,
0.03429477661848068,
0.011174916289746761,
0.03330645337700844,
-0.07769271731376648,
-0.055592168122529984,
0.06008053943514824,
-0.037223946303129196,
-0.06695649027824402,
0.0015645483508706093,
-0.04543082416057587,
0.019562307745218277,
0.027363594621419907,
0.061514098197221756,
-0.06553927809000015,
0.011566216126084328,
0.0038894189056009054,
0.08818991482257843,
0.0449860505759716,
-0.03189999237656593,
0.053715288639068604,
0.0042306529358029366,
0.021151460707187653,
0.02466140128672123,
-0.03632109612226486,
-0.027715221047401428,
-0.0125571358948946,
-0.008524361997842789,
0.06108282506465912,
-0.10337354242801666,
-0.06731132417917252,
-0.03695596382021904,
0.016739046201109886,
0.02299976721405983,
-0.0009148253011517227,
0.047463636845350266,
0.019779833033680916,
-0.0194641575217247,
-0.02177055925130844,
0.01756315678358078,
-0.04927913099527359,
0.036683328449726105,
0.01631682552397251,
0.00473433593288064,
0.05977002903819084,
0.05543603003025055,
-0.02257215976715088,
-0.11146658658981323,
0.021240931004285812
] |
Matérn Kernels for Tunable Implicit Surface Reconstruction | https://openreview.net/forum?id=Ox4AJ2Vurb | [
"Maximilian Weiherer",
"Bernhard Egger"
] | Poster | We propose to use the family of Matérn kernels for implicit surface reconstruction, building upon the recent success of kernel methods for 3D reconstruction of oriented point clouds. As we show from a theoretical and practical perspective, Matérn kernels have some appealing properties which make them particularly well suited for surface reconstruction---outperforming state-of-the-art methods based on the arc-cosine kernel while being significantly easier to implement, faster to compute, and scalable. Being stationary, we demonstrate that Matérn kernels allow for tunable surface reconstruction in the same way as Fourier feature mappings help coordinate-based MLPs overcome spectral bias. Moreover, we theoretically analyze Matérn kernels' connection to SIREN networks as well as their relation to previously employed arc-cosine kernels. Finally, based on recently introduced Neural Kernel Fields, we present data-dependent Matérn kernels and conclude that especially the Laplace kernel (being part of the Matérn family) is extremely competitive, performing almost on par with state-of-the-art methods in the noise-free case while having a more than five times shorter training time. | surface reconstruction, kernel methods, neural tangent kernel | We use and theoretically analyze the family of Matérn kernels for tunable implicit surface reconstruction, demonstrating that they outperform recently employed arc-cosine kernels. | 434 | null | [
-0.0796079933643341,
-0.06068629026412964,
0.10333839803934097,
-0.007381201256066561,
0.10167993605136871,
-0.00013114699686411768,
0.0015001638093963265,
-0.0753968209028244,
0.03221447020769119,
-0.05810617655515671,
-0.061809878796339035,
-0.04530061036348343,
-0.058455076068639755,
-0.007371300831437111,
0.022067613899707794,
-0.06658904999494553,
0.08279727399349213,
0.03812548890709877,
0.015279814600944519,
0.04567326605319977,
0.005615243222564459,
0.0637442097067833,
-0.07556851953268051,
0.009966613724827766,
0.054284073412418365,
0.028102464973926544,
0.12023679912090302,
-0.02711688168346882,
-0.025859622284770012,
-0.03066004253923893,
0.015255820006132126,
-0.015201130881905556,
-0.0620141439139843,
-0.006699390709400177,
-0.022015709429979324,
-0.0400543250143528,
0.05061196908354759,
0.037816908210515976,
-0.029728589579463005,
-0.004529427736997604,
0.049892883747816086,
0.04042470455169678,
-0.013991758227348328,
0.021826298907399178,
0.12630847096443176,
0.06406967341899872,
0.06841430068016052,
-0.11135508865118027,
0.09261550009250641,
0.0009476484847255051,
0.0016957363113760948,
0.012131193652749062,
0.015034305863082409,
0.0005561760044656694,
-0.0020124197471886873,
0.09942417591810226,
-0.004332319367676973,
0.006443789228796959,
0.09193948656320572,
0.03898877277970314,
0.02356388233602047,
-0.06486634910106659,
0.01648472063243389,
-0.02375347726047039,
-0.011768211610615253,
-0.05977291613817215,
0.04714914411306381,
-0.034632060676813126,
0.028277266770601273,
0.06055348739027977,
-0.028726160526275635,
0.04863099753856659,
-0.03454706817865372,
-0.0010967068374156952,
0.029372723773121834,
0.13523612916469574,
0.006976851262152195,
-0.03508009761571884,
-0.01102062314748764,
0.020572664216160774,
-0.0023672725073993206,
0.054595302790403366,
0.03868180885910988,
0.027304433286190033,
-0.017777498811483383,
-0.009140708483755589,
-0.09391339868307114,
0.09031552821397781,
-0.030425533652305603,
-0.004395711235702038,
0.000013578357538790442,
-0.05158795788884163,
-0.06716109067201614,
-0.07257647067308426,
-0.08274612575769424,
-0.04221184179186821,
0.030809583142399788,
-0.0023598195984959602,
0.03382554650306702,
0.06615100055932999,
-0.039613958448171616,
-0.003981167450547218,
-0.008793070912361145,
0.015477726235985756,
0.009317237883806229,
-0.05014079809188843,
0.03544261306524277,
0.03782321885228157,
0.06149515509605408,
0.025191647931933403,
0.025952523574233055,
-0.058440230786800385,
-0.050047215074300766,
-0.0026085651479661465,
0.030474869534373283,
-0.06088902801275253,
-0.05293911695480347,
-0.05726906284689903,
0.003551127854734659,
0.01980947144329548,
-0.009932460263371468,
-0.06433209776878357,
-0.024859309196472168,
0.07226688414812088,
0.015188335441052914,
-0.012606138363480568,
-0.04318968951702118,
1.2505769402646495e-33,
-0.023325808346271515,
0.17217326164245605,
-0.0165290217846632,
-0.05014794319868088,
0.05736318603157997,
-0.009305763989686966,
0.014050041325390339,
-0.012233765795826912,
0.12959198653697968,
0.04782893881201744,
-0.16417959332466125,
0.05510076507925987,
-0.043433256447315216,
0.015013974159955978,
0.006152369547635317,
-0.03351476415991783,
0.0024839937686920166,
-0.044873226433992386,
-0.10165809839963913,
0.022812150418758392,
0.027330562472343445,
-0.009375599212944508,
-0.0034933683928102255,
-0.016174448654055595,
0.00076211680425331,
0.0399051271378994,
-0.006068659946322441,
-0.0038451296277344227,
0.009003574028611183,
0.005278419703245163,
-0.025186752900481224,
-0.018064741045236588,
0.04861222207546234,
-0.0059725623577833176,
0.059948403388261795,
-0.03327992558479309,
0.0006978645687922835,
-0.05943818762898445,
-0.03785230964422226,
-0.03248654678463936,
-0.038151126354932785,
0.05130751058459282,
-0.022497529163956642,
-0.003759888932108879,
-0.0390203595161438,
-0.01761062629520893,
0.02714775688946247,
0.025076517835259438,
0.01173736248165369,
-0.07272368669509888,
0.05243350565433502,
-0.009700934402644634,
-0.05995117500424385,
-0.04844481498003006,
0.017919868230819702,
0.06620190292596817,
0.016048088669776917,
-0.053906966000795364,
0.06601493060588837,
0.030679570510983467,
-0.03204170614480972,
-0.013330246321856976,
0.07072228938341141,
0.003655361710116267,
-0.06646676361560822,
0.016308913007378578,
0.06598326563835144,
0.04701600968837738,
0.03536902368068695,
0.01969575509428978,
0.03608646243810654,
0.06730607151985168,
-0.01776237040758133,
-0.010337070561945438,
-0.0013263830915093422,
0.047408077865839005,
0.04880089312791824,
-0.032053444534540176,
0.06961716711521149,
0.05154506117105484,
-0.12860657274723053,
0.0022076107561588287,
-0.0017617447301745415,
-0.10187724977731705,
-0.028451384976506233,
-0.02941042184829712,
-0.0555228553712368,
-0.03973299637436867,
-0.02683616802096367,
-0.03279825672507286,
-0.08211497962474823,
-0.012814232148230076,
0.04932886362075806,
-0.05228745564818382,
-0.09331147372722626,
-2.4190039068091065e-33,
0.032057251781225204,
0.05600735545158386,
-0.06140999495983124,
-0.023020468652248383,
-0.079190194606781,
0.05044436454772949,
-0.01872561313211918,
0.054936010390520096,
-0.024965915828943253,
-0.042783528566360474,
-0.0691082626581192,
-0.003998495172709227,
0.06905559450387955,
0.022412288933992386,
0.03940650075674057,
0.053041838109493256,
-0.09300070255994797,
-0.025740766897797585,
-0.01442988682538271,
-0.06901616603136063,
-0.030973514541983604,
0.03564000502228737,
0.012301670387387276,
0.012183942832052708,
-0.07010235637426376,
0.059735655784606934,
-0.03782857581973076,
0.007896283641457558,
-0.0009016950498335063,
-0.01142711192369461,
-0.025179997086524963,
-0.08819698542356491,
0.07507524639368057,
-0.04552902281284332,
0.0010858314344659448,
0.0732802301645279,
0.09532912075519562,
0.016861505806446075,
-0.009545554406940937,
0.002503421390429139,
0.033532872796058655,
-0.05792764201760292,
-0.002384357387199998,
-0.0544111505150795,
-0.0512513667345047,
0.016917359083890915,
0.021446578204631805,
0.04353535920381546,
0.00679717818275094,
0.057314470410346985,
0.020161719992756844,
-0.012523365207016468,
-0.11419068276882172,
0.02349100634455681,
-0.05527123436331749,
-0.02731737121939659,
-0.027165133506059647,
-0.01945878192782402,
0.008618654683232307,
0.034943267703056335,
-0.07353705167770386,
-0.0852523148059845,
-0.031860288232564926,
0.07915125787258148,
0.002088767010718584,
0.09553143382072449,
0.03991322964429855,
-0.0512724369764328,
-0.006747606676071882,
0.0661286860704422,
0.0024933358654379845,
0.01743452250957489,
0.03074268065392971,
0.0555296316742897,
-0.015745004639029503,
-0.036360640078783035,
-0.005187372677028179,
-0.005597943440079689,
-0.0018007613252848387,
-0.06895856559276581,
0.024188805371522903,
0.008359162136912346,
0.015520680695772171,
0.08968329429626465,
0.11889730393886566,
0.04157082363963127,
0.016603058204054832,
-0.05008010193705559,
0.005891324486583471,
-0.02265676110982895,
0.025000497698783875,
0.023287218064069748,
-0.04214106872677803,
0.05266546085476875,
0.014251355081796646,
-4.7903046862529663e-8,
-0.04951634258031845,
0.07644905149936676,
0.06443279981613159,
-0.07368206977844238,
0.030955620110034943,
-0.07675466686487198,
0.08184000849723816,
0.004867736250162125,
-0.05137717351317406,
-0.06551352888345718,
-0.03611044958233833,
-0.17006462812423706,
-0.09485401958227158,
-0.038055628538131714,
0.11794032901525497,
0.03958217054605484,
0.03637116402387619,
-0.01764882542192936,
-0.022319383919239044,
0.0011191399535164237,
-0.03672455996274948,
0.06741374731063843,
0.00014318556350190192,
-0.004080786369740963,
0.0672702044248581,
0.004756371024996042,
-0.018911579623818398,
0.09332596510648727,
0.06754887849092484,
0.021534619852900505,
-0.0998118594288826,
-0.0038576284423470497,
-0.08041323721408844,
0.0577470026910305,
0.025013066828250885,
-0.035283129662275314,
0.007195455953478813,
-0.07909297943115234,
-0.03833600878715515,
-0.02195417881011963,
-0.04859313741326332,
0.13159802556037903,
-0.04397325962781906,
-0.012881673872470856,
0.0689476802945137,
0.0978584811091423,
0.10237348824739456,
-0.060027364641427994,
-0.013220100663602352,
0.05753430724143982,
-0.033740460872650146,
0.009788076393306255,
-0.005503232125192881,
0.04496984928846359,
-0.016869591549038887,
0.04248740151524544,
0.001860403106547892,
-0.08793551474809647,
-0.004580959212034941,
0.05556371435523033,
-0.030820626765489578,
-0.016127200797200203,
0.020553085952997208,
-0.02459271065890789
] |
Hierarchically Encapsulated Representation for Protocol Design in Self-Driving Labs | https://openreview.net/forum?id=9nUBh4V6SA | [
"Yu-Zhe Shi",
"Mingchen Liu",
"Fanxu Meng",
"Qiao Xu",
"Zhangqian Bi",
"Kun He",
"Lecheng Ruan",
"Qining Wang"
] | Poster | Self-driving laboratories have begun to replace human experimenters in performing single experimental skills or predetermined experimental protocols. However, as the pace of idea iteration in scientific research has been intensified by Artificial Intelligence, the demand for rapid design of new protocols for new discoveries become evident. Efforts to automate protocol design have been initiated, but the capabilities of knowledge-based machine designers, such as Large Language Models, have not been fully elicited, probably for the absence of a systematic representation of experimental knowledge, as opposed to isolated, flatten pieces of information. To tackle this issue, we propose a multi-faceted, multi-scale representation, where instance actions, generalized operations, and product flow models are hierarchically encapsulated using Domain-Specific Languages. We further develop a data-driven algorithm based on non-parametric modeling that autonomously customizes these representations for specific domains. The proposed representation is equipped with various machine designers to manage protocol design tasks, including planning, modification, and adjustment. The results demonstrate that the proposed method could effectively complement Large Language Models in the protocol design process, serving as an auxiliary module in the realm of machine-assisted scientific exploration. | Self-driving laboratories, protocol design, automated design, domain-specific language | null | 433 | null | [
0.004994104616343975,
0.0006513107800856233,
0.051563236862421036,
-0.0015204671071842313,
-0.03638988733291626,
-0.0615696981549263,
-0.024174602702260017,
0.04216489568352699,
0.028372226282954216,
-0.03424512594938278,
-0.07634880393743515,
-0.11145088076591492,
0.014051939360797405,
0.08083593845367432,
0.018884776160120964,
-0.022715024650096893,
0.09010504186153412,
0.0022518986370414495,
-0.03850582242012024,
-0.10066186636686325,
0.042911067605018616,
0.036883555352687836,
0.019537795335054398,
0.03783765807747841,
-0.03496720269322395,
0.024057721719145775,
0.008615035563707352,
-0.056648995727300644,
0.01282345037907362,
0.005905653350055218,
0.044237300753593445,
0.01647912710905075,
-0.01277634035795927,
0.04450256749987602,
0.0042540524154901505,
0.015129460953176022,
0.006546424236148596,
0.07400358468294144,
-0.08355100452899933,
-0.08817587047815323,
-0.0423450693488121,
-0.0936574935913086,
0.01677965559065342,
0.0712624341249466,
0.03368839994072914,
0.0000031992053663998377,
-0.0319707877933979,
-0.020441925153136253,
-0.051844965666532516,
0.031175309792160988,
-0.07967076450586319,
-0.05945926532149315,
-0.0011157938279211521,
-0.06267106533050537,
0.08047304302453995,
-0.015078667551279068,
0.05160015821456909,
-0.009079495444893837,
0.0058769164606928825,
0.030030930414795876,
0.0026495724450796843,
-0.057863421738147736,
-0.09124292433261871,
-0.006686327047646046,
-0.05225489288568497,
0.01893674209713936,
0.019646042957901955,
0.0615626722574234,
0.0359923280775547,
0.04820679500699043,
-0.03413251414895058,
0.0327242873609066,
-0.021245025098323822,
-0.016175853088498116,
0.04594004899263382,
0.023266224190592766,
-0.055681321769952774,
0.017737610265612602,
0.01912248134613037,
-0.10098086297512054,
0.015472201630473137,
-0.02195432037115097,
-0.014151844196021557,
0.1243102177977562,
0.041102297604084015,
-0.011671251617372036,
0.007635048590600491,
0.09248823672533035,
0.07129604369401932,
0.03205256164073944,
-0.08846833556890488,
-0.07382972538471222,
0.012930907309055328,
-0.044762834906578064,
0.03310277312994003,
0.044991184026002884,
0.0450747050344944,
-0.12577791512012482,
0.013660088181495667,
0.04765275493264198,
-0.0022425393108278513,
0.03526872768998146,
-0.005173207726329565,
-0.05279456824064255,
-0.06130215898156166,
0.02877359464764595,
-0.021288437768816948,
-0.019539890810847282,
0.0882820188999176,
-0.013279329985380173,
0.0021857134997844696,
-0.00010716982069425285,
-0.027935078367590904,
-0.04645383358001709,
-0.041074153035879135,
0.023387402296066284,
-0.034405678510665894,
0.001242950325831771,
-0.010736824944615364,
0.05881693214178085,
-0.03687667474150658,
-0.0016532590379938483,
0.05579445883631706,
-0.017653807997703552,
0.0631411224603653,
0.04647000506520271,
-0.04125204309821129,
2.123648323784285e-34,
0.048204775899648666,
0.017459852620959282,
-0.03318079188466072,
0.13001544773578644,
0.06188679486513138,
-0.014660459943115711,
0.06881381571292877,
0.03110555373132229,
-0.040798936039209366,
-0.037666697055101395,
0.02437232807278633,
0.04818142578005791,
-0.0563608817756176,
0.04849499464035034,
0.0441720075905323,
0.09350696951150894,
-0.0071032107807695866,
0.06284308433532715,
-0.03551062196493149,
-0.06368176639080048,
0.03506884351372719,
-0.011774077080190182,
-0.0055993967689573765,
-0.012379412539303303,
0.07791177183389664,
0.05568636581301689,
-0.0006714100600220263,
-0.005507874768227339,
-0.034845199435949326,
0.03257431089878082,
-0.08875161409378052,
0.03665047883987427,
0.024157453328371048,
0.048721592873334885,
0.04057224467396736,
-0.0073553468100726604,
0.0008858860819600523,
-0.07528303563594818,
0.0645679384469986,
0.04134410247206688,
-0.003400794230401516,
-0.006381964776664972,
-0.03247082605957985,
0.019077235832810402,
-0.08888661116361618,
-0.050535451620817184,
0.0004613962082657963,
0.027699721977114677,
0.018648652359843254,
-0.037024613469839096,
0.025499969720840454,
0.03716028109192848,
0.10812804847955704,
-0.1464456021785736,
0.10718061774969101,
-0.033112455159425735,
-0.009459284134209156,
0.06578102707862854,
0.017885174602270126,
0.09213372319936752,
-0.07280803471803665,
0.14123956859111786,
-0.05842963606119156,
0.11966010183095932,
0.025829637423157692,
-0.012001773342490196,
-0.04672367498278618,
0.023146823048591614,
0.10004688799381256,
0.013114778324961662,
-0.08818498998880386,
0.08783462643623352,
-0.06322703510522842,
-0.04921278357505798,
0.006134680472314358,
-0.004644499626010656,
0.07363487780094147,
-0.03318710997700691,
-0.011320310644805431,
0.01606426201760769,
-0.022314874455332756,
-0.03102724626660347,
-0.015768185257911682,
0.040881555527448654,
0.018355976790189743,
-0.03105776570737362,
-0.031069019809365273,
0.016911376267671585,
-0.04392886161804199,
-0.07308334112167358,
-0.05131734907627106,
-0.06642932444810867,
-0.0131979463621974,
0.02209288440644741,
0.019341111183166504,
-2.4166734892586013e-33,
-0.049659643322229385,
-0.03633417934179306,
-0.014766832813620567,
0.023370252922177315,
0.0594678558409214,
-0.05814117565751076,
0.027048736810684204,
-0.06173248216509819,
0.03617729991674423,
-0.02040974423289299,
-0.04726613312959671,
-0.07119198143482208,
0.020334050059318542,
-0.02080320008099079,
0.02131245844066143,
-0.029611334204673767,
-0.12911243736743927,
-0.07056945562362671,
0.027994902804493904,
0.004098964389413595,
-0.074740931391716,
0.07677996903657913,
-0.058992814272642136,
0.06152428686618805,
0.007122999522835016,
-0.03931339457631111,
-0.017415234819054604,
-0.04624120146036148,
0.047892648726701736,
0.0339050330221653,
-0.03829292580485344,
-0.017287278547883034,
0.000531016499735415,
-0.0059006791561841965,
-0.00022244687716010958,
0.10338489711284637,
0.0642383024096489,
-0.005873057525604963,
-0.009798244573175907,
-0.028832538053393364,
0.009049417451024055,
-0.06978273391723633,
-0.08708541840314865,
-0.001325238961726427,
0.044883884489536285,
0.02374713309109211,
-0.09147052466869354,
0.011467558331787586,
-0.01062715332955122,
-0.009763306006789207,
-0.025039661675691605,
0.019473077729344368,
-0.032722145318984985,
-0.1066235899925232,
-0.07171676307916641,
-0.02405543439090252,
0.03473704308271408,
-0.052209969609975815,
0.041438568383455276,
0.04221300035715103,
0.018679512664675713,
-0.08857360482215881,
0.06794730573892593,
0.01714565046131611,
-0.033137381076812744,
-0.05373244360089302,
-0.02546432800590992,
0.023900991305708885,
-0.059626709669828415,
-0.006273743696510792,
0.017633263021707535,
-0.02416817843914032,
-0.045485787093639374,
0.015439198352396488,
0.09791955351829529,
-0.08921808749437332,
-0.005931272637099028,
-0.07630316913127899,
0.0008910453179851174,
0.03960966318845749,
0.02711002342402935,
0.10589181631803513,
0.09021669626235962,
0.04632635414600372,
0.07413474470376968,
0.034363165497779846,
-0.042513344436883926,
0.15640874207019806,
0.032759733498096466,
0.021282266825437546,
-0.02227422036230564,
0.026618117466568947,
-0.0338757261633873,
0.13774272799491882,
-0.07309255748987198,
-4.7311885964518297e-8,
0.009838752448558807,
0.007463298738002777,
0.10435231029987335,
0.020093871280550957,
0.006198553368449211,
0.07780973613262177,
-0.06793159991502762,
-0.003619882045313716,
0.007786835078150034,
-0.0001654695370234549,
-0.01563601940870285,
0.01868068426847458,
-0.01908850483596325,
-0.015822889283299446,
-0.009876512922346592,
0.05420969799160957,
0.002939044963568449,
0.02326934225857258,
-0.0499158538877964,
-0.026500578969717026,
0.03754590451717377,
0.05944211035966873,
0.010556185618042946,
-0.006509249098598957,
0.028819510713219643,
-0.11837432533502579,
-0.002708966378122568,
0.060590703040361404,
0.032529521733522415,
-0.02794002555310726,
0.0000022796359644416953,
-0.05408693850040436,
0.003296325681731105,
0.06440798193216324,
0.031031981110572815,
0.00957091897726059,
0.02506195940077305,
0.0338534452021122,
0.020145533606410027,
-0.07967044413089752,
-0.04965335130691528,
0.062259335070848465,
-0.11894377321004868,
0.04863082244992256,
0.011698050424456596,
0.032400354743003845,
-0.07103561609983444,
-0.07317260652780533,
-0.030046403408050537,
0.044362328946590424,
-0.006442449055612087,
-0.03090067394077778,
-0.04200673848390579,
0.051934871822595596,
0.00833241455256939,
0.12998925149440765,
-0.0278706643730402,
-0.07385128736495972,
0.053308743983507156,
-0.03457644209265709,
0.002092640846967697,
0.061345022171735764,
0.022330159321427345,
0.010750999674201012
] |
Mechanism and emergence of stacked attention heads in multi-layer transformers | https://openreview.net/forum?id=rUC7tHecSQ | [
"Tiberiu Mușat"
] | Poster | In this paper, I introduce the retrieval problem, a simple yet common reasoning task that can be solved only by transformers with a minimum number of layers, which grows logarithmically with the input size. I empirically show that large language models can solve the task under different prompting formulations without any fine-tuning. To understand how transformers solve the retrieval problem, I train several transformers on a minimal formulation. Successful learning occurs only under the presence of an implicit curriculum. I uncover the learned mechanisms by studying the attention maps in the trained transformers. I also study the training process, uncovering that attention heads always emerge in a specific sequence guided by the implicit curriculum. | mechanistic interpretability, large language models, transformers, emergent abilities, curriculum learning, reasoning | null | 427 | 2411.12118 | [
-0.00846809707581997,
-0.02910606376826763,
0.037776920944452286,
0.01899336837232113,
0.02917669713497162,
0.06634782999753952,
0.011891704052686691,
0.05897902697324753,
0.0524936243891716,
-0.03465024754405022,
0.0338493250310421,
0.0013861383777111769,
0.0405881293118,
0.05843302235007286,
0.0330822616815567,
-0.005116195883601904,
0.017573216930031776,
0.05035843327641487,
-0.03696141764521599,
-0.07027190923690796,
0.026644041761755943,
0.018566111102700233,
0.07930799573659897,
-0.007910633459687233,
-0.037610288709402084,
0.02719959057867527,
-0.04114086180925369,
-0.09299276769161224,
0.03645215556025505,
-0.022324886173009872,
0.0740499496459961,
0.0035112483892589808,
-0.048467785120010376,
0.10264268517494202,
-0.04186317324638367,
0.09973316639661789,
-0.08803962916135788,
0.01829768903553486,
-0.038351960480213165,
-0.04749033972620964,
-0.05289577692747116,
0.007253955118358135,
-0.04068467766046524,
0.011879276484251022,
0.08201798051595688,
-0.0393436923623085,
-0.001723285298794508,
0.020225469022989273,
0.0024496300611644983,
-0.07697934657335281,
-0.05692664533853531,
-0.08504500985145569,
-0.005792027339339256,
0.07648750394582748,
-0.022352097555994987,
0.026151472702622414,
0.03287619724869728,
0.03846767544746399,
-0.03221624344587326,
-0.0025468464009463787,
-0.11667832732200623,
-0.07607710361480713,
-0.08940447866916656,
-0.06445267051458359,
0.026262735947966576,
0.007953433319926262,
0.03566941246390343,
-0.0013600356178358197,
-0.047596804797649384,
0.04390385001897812,
0.014249738305807114,
0.05065426975488663,
-0.01348155178129673,
-0.0468696691095829,
0.1441340446472168,
0.04397977888584137,
0.017425131052732468,
-0.0847221314907074,
0.021117709577083588,
-0.08604708313941956,
0.0035351526457816362,
-0.02100919559597969,
0.02388543076813221,
0.008229533210396767,
0.12056382745504379,
-0.007445871364325285,
-0.03470165282487869,
-0.03304535523056984,
-0.018307944759726524,
-0.0034037160221487284,
-0.05317889153957367,
-0.11144062876701355,
0.012477044016122818,
0.01817215047776699,
0.020773885771632195,
0.05134338140487671,
-0.02514890767633915,
-0.0004019273619633168,
-0.02304002456367016,
0.034058041870594025,
-0.007812130264937878,
0.0005852984031662345,
0.0336853452026844,
-0.005555290263146162,
-0.015041625127196312,
-0.05610079690814018,
0.016844969242811203,
0.05182396247982979,
0.015367342159152031,
-0.16021288931369781,
0.025671765208244324,
-0.044652119278907776,
0.02506916970014572,
-0.008012879639863968,
0.016441261395812035,
-0.016308100894093513,
0.047233469784259796,
-0.045832034200429916,
-0.0879104882478714,
0.020062562078237534,
0.02575613372027874,
0.03591194376349449,
0.0012190720299258828,
-0.01564703695476055,
0.05435545742511749,
-0.04875500127673149,
-0.012659868225455284,
1.772274281964539e-33,
0.08010118454694748,
0.06354289501905441,
0.029926104471087456,
0.08177580684423447,
-0.023819420486688614,
-0.018084969371557236,
0.09926123172044754,
0.02007863111793995,
-0.001655378844588995,
0.04608722776174545,
-0.028056582435965538,
0.06579495221376419,
-0.06093253195285797,
0.11065491288900375,
-0.0218619666993618,
-0.03265887126326561,
-0.0027842982672154903,
0.07291518151760101,
-0.06765951961278915,
-0.016665052622556686,
0.007860914804041386,
0.043090105056762695,
0.05865326523780823,
-0.10827963054180145,
-0.02025122568011284,
0.0023577874526381493,
0.012023095041513443,
-0.053543638437986374,
-0.031630150973796844,
0.009727495722472668,
-0.06662163883447647,
0.031213754788041115,
-0.0652560293674469,
0.03886386752128601,
0.045899540185928345,
-0.02632729336619377,
0.020448852330446243,
-0.010941626504063606,
0.08857404440641403,
0.011989034712314606,
0.05515103042125702,
0.030791835859417915,
0.05313563346862793,
0.027446966618299484,
-0.027697879821062088,
-0.0021096901036798954,
-0.03977200761437416,
0.027751948684453964,
-0.03756967931985855,
-0.02736041136085987,
0.026142779737710953,
0.01125395018607378,
-0.06907258182764053,
-0.06278543174266815,
0.1083400771021843,
0.03303234651684761,
0.0016860960749909282,
0.09774398058652878,
0.05073448643088341,
0.039149969816207886,
0.05728713423013687,
0.033498212695121765,
-0.046103715896606445,
0.08943317085504532,
0.06262975931167603,
-0.016560276970267296,
0.030478456988930702,
0.046002525836229324,
0.0959162786602974,
0.010298679582774639,
-0.02560705505311489,
-0.0018855093512684107,
-0.0361924022436142,
-0.05001349747180939,
0.06942889094352722,
-0.020134836435317993,
0.05625234171748161,
-0.09610380232334137,
0.06966762244701385,
-0.02279103547334671,
0.026280665770173073,
-0.05462728068232536,
-0.011797196231782436,
-0.0020313013810664415,
-0.0015328919980674982,
0.006580557208508253,
0.031651027500629425,
-0.09120877087116241,
-0.0470123291015625,
0.019521564245224,
-0.013121355324983597,
0.02667664922773838,
-0.05529794842004776,
-0.023573167622089386,
-0.01001387182623148,
-1.0139851922746477e-33,
0.05201968550682068,
0.029340360313653946,
-0.08911509811878204,
0.012209919281303883,
-0.07214009761810303,
-0.02492002584040165,
0.024393681436777115,
0.013873593881726265,
-0.11435778439044952,
0.0040065934881567955,
-0.03946535661816597,
-0.09332766383886337,
0.044098466634750366,
-0.02734663523733616,
0.02500046417117119,
-0.004418742377310991,
0.0043419538997113705,
0.004838150925934315,
0.026210658252239227,
0.022634821012616158,
-0.01831127516925335,
0.06620059162378311,
-0.10919644683599472,
0.009881473146378994,
-0.06737328320741653,
-0.00875898078083992,
-0.10945425927639008,
0.09632007777690887,
0.04350702092051506,
0.02284478396177292,
-0.07215329259634018,
-0.08953513205051422,
-0.021529411897063255,
0.007294538896530867,
-0.09453083574771881,
0.07530923187732697,
-0.008772836066782475,
-0.0009562612394802272,
-0.05575190484523773,
0.06699387729167938,
-0.025850750505924225,
-0.11314672976732254,
-0.061886195093393326,
-0.0073979562148451805,
-0.008031905628740788,
-0.06455632299184799,
-0.05205850303173065,
0.01351415179669857,
-0.056530632078647614,
0.048216722905635834,
-0.030901776626706123,
0.005859214346855879,
-0.026157615706324577,
-0.07930269092321396,
-0.0807410329580307,
0.017022674903273582,
0.02198188379406929,
-0.06269143521785736,
0.13081692159175873,
-0.060303036123514175,
-0.007642092648893595,
-0.01607063226401806,
0.036395132541656494,
-0.07199731469154358,
0.011071426793932915,
-0.012935556471347809,
0.014867307618260384,
-0.04027683287858963,
0.052899278700351715,
-0.023423271253705025,
0.09420037269592285,
0.013051825575530529,
0.04666268453001976,
-0.09380533546209335,
0.030179522931575775,
0.07588264346122742,
-0.04674743860960007,
-0.019583331421017647,
-0.013224482536315918,
-0.07077354937791824,
-0.04570981115102768,
0.009389334358274937,
0.013951818458735943,
0.028968704864382744,
0.03221048414707184,
0.07431836426258087,
0.10019486397504807,
0.05240796506404877,
0.014149809256196022,
-0.013053031638264656,
0.02442781999707222,
0.049034688621759415,
0.06003437936306,
0.049745965749025345,
-0.05190456286072731,
-4.2364060703903306e-8,
-0.1037432998418808,
0.037749283015728,
-0.004077406600117683,
0.028967753052711487,
0.07931841164827347,
-0.061616234481334686,
0.02544294483959675,
0.15294340252876282,
-0.08839169889688492,
0.026472976431250572,
0.01633160375058651,
0.03270430490374565,
0.0029059341177344322,
0.004927038215100765,
0.04301486164331436,
0.0680796355009079,
0.02933874912559986,
0.0639902725815773,
-0.06399047374725342,
-0.054252203553915024,
0.031425926834344864,
0.0732949897646904,
0.03327377140522003,
0.019963126629590988,
-0.019799156114459038,
0.009187296032905579,
-0.028666257858276367,
0.0856042206287384,
0.03725716844201088,
0.03507307544350624,
0.026098093017935753,
-0.005320720374584198,
-0.027236323803663254,
-0.07615014165639877,
0.012849087826907635,
0.10156827419996262,
0.046188462525606155,
-0.04798535630106926,
-0.03327651694417,
-0.06431202590465546,
0.04718461632728577,
0.03419632837176323,
-0.03564472869038582,
0.036736421287059784,
0.052154988050460815,
0.026895759627223015,
-0.017140893265604973,
-0.07818084955215454,
0.02334105409681797,
-0.003869815031066537,
-0.007253399584442377,
-0.04609083756804466,
-0.050083328038454056,
0.04306318238377571,
0.03318669646978378,
0.01551815029233694,
0.04110240936279297,
-0.0789099633693695,
-0.054363202303647995,
0.05655132979154587,
0.0021326036658138037,
0.10285303741693497,
-0.019235916435718536,
-0.0005099551635794342
] |
Progressive Parameter Efficient Transfer Learning for Semantic Segmentation | https://openreview.net/forum?id=YNbLUGDAX5 | [
"Nan Zhou",
"Huiqun Wang",
"Yaoyan Zheng",
"Di Huang"
] | Poster | Parameter Efficient Transfer Learning (PETL) excels in downstream classification fine-tuning with minimal computational overhead, demonstrating its potential within the pre-train and fine-tune paradigm. However, recent PETL methods consistently struggle when fine-tuning for semantic segmentation tasks, limiting their broader applicability. In this paper, we identify that fine-tuning for semantic segmentation requires larger parameter adjustments due to shifts in semantic perception granularity. Current PETL approaches are unable to effectively accommodate these shifts, leading to significant performance degradation. To address this, we introduce ProPETL, a novel approach that incorporates an additional midstream adaptation to progressively align pre-trained models for segmentation tasks. Through this process, ProPETL achieves state-of-the-art performance on most segmentation benchmarks and, for the first time, surpasses full fine-tuning on the challenging COCO-Stuff10k dataset. Furthermore, ProPETL demonstrates strong generalization across various pre-trained models and scenarios, highlighting its effectiveness and versatility for broader adoption in segmentation tasks. Code is available at: https://github.com/weeknan/ProPETL. | Parameter Efficient Transfer Learning, Semantic Segmentation | null | 421 | null | [
0.02726144716143608,
-0.08556041121482849,
0.06850200146436691,
0.06068621948361397,
0.051796745508909225,
0.011056480929255486,
0.012841812334954739,
-0.054857611656188965,
-0.02833595499396324,
-0.111600361764431,
-0.08394347131252289,
-0.09618555754423141,
-0.07730592042207718,
0.03375895693898201,
-0.032956160604953766,
-0.015010821633040905,
0.11376141756772995,
0.11193455010652542,
-0.12150426208972931,
-0.0010574081679806113,
0.003055302891880274,
0.021371955052018166,
0.0468580536544323,
0.019351014867424965,
0.027055999264121056,
-0.02672908641397953,
-0.041128870099782944,
-0.09706031531095505,
0.06532049924135208,
-0.041185613721609116,
-0.0006644402747042477,
0.08037354797124863,
-0.03454660624265671,
0.07134280353784561,
-0.017235275357961655,
0.04379444941878319,
0.009701386094093323,
-0.002343367785215378,
0.02036387287080288,
-0.05459028482437134,
0.06583011895418167,
-0.023031534627079964,
-0.03179476037621498,
0.024245193228125572,
-0.0196201354265213,
0.058002639561891556,
0.055457741022109985,
-0.04321267083287239,
0.06165303662419319,
-0.033356327563524246,
-0.06521112471818924,
-0.039706964045763016,
-0.07260355353355408,
0.09272889792919159,
-0.0548812597990036,
0.04841869696974754,
0.08373076468706131,
-0.01615990698337555,
0.03933975473046303,
-0.008152680471539497,
-0.053905557841062546,
-0.06358009576797485,
0.028135158121585846,
0.015297199599444866,
-0.05311272293329239,
-0.04977397993206978,
-0.00417611887678504,
0.01665901020169258,
0.024708595126867294,
-0.009813802316784859,
-0.03967922553420067,
0.07314079254865646,
0.015612157061696053,
0.005639181938022375,
0.018628763034939766,
0.03466736152768135,
0.05257752165198326,
0.011974173597991467,
0.04970515891909599,
-0.12590321898460388,
0.011705659329891205,
-0.00302705611102283,
0.06443964689970016,
0.016498150303959846,
0.11606353521347046,
0.03973035514354706,
-0.010053384117782116,
0.04651063308119774,
-0.03374746814370155,
-0.01673233136534691,
-0.05871400982141495,
-0.05096900090575218,
0.039352841675281525,
-0.004060112405568361,
-0.0016631984617561102,
0.00847876537591219,
-0.0355093851685524,
-0.12111254036426544,
-0.017369842156767845,
0.0930357575416565,
0.02085038833320141,
0.02231638692319393,
0.012219228781759739,
0.023521501570940018,
0.0005337087204679847,
-0.08183973282575607,
0.06513935327529907,
0.025260787457227707,
0.02041451819241047,
-0.006904230453073978,
0.07259901612997055,
0.08713407069444656,
0.044321853667497635,
-0.03287307173013687,
0.011668520979583263,
-0.0030386410653591156,
-0.0876656323671341,
-0.025623232126235962,
0.03574389964342117,
0.0020481101237237453,
-0.03938421607017517,
-0.005517087876796722,
-0.07305240631103516,
0.04651226848363876,
-0.011691100895404816,
-0.062180209904909134,
-0.057636283338069916,
2.2010132548949135e-33,
0.036576393991708755,
-0.021821383386850357,
0.07443693280220032,
-0.03158344328403473,
-0.02775636874139309,
-0.02944379486143589,
0.014178657904267311,
0.019279342144727707,
-0.03447970002889633,
0.0006156523013487458,
-0.03123706765472889,
-0.027990328148007393,
-0.06618793308734894,
0.08960726112127304,
-0.006928085349500179,
-0.07747963815927505,
-0.07379157096147537,
0.03501797094941139,
-0.029824987053871155,
-0.01591220684349537,
-0.02998284064233303,
0.0019544134847819805,
-0.06750020384788513,
0.010157879441976547,
0.041365187615156174,
0.018670804798603058,
-0.0011729768011718988,
-0.03095625340938568,
-0.00920127797871828,
0.034277770668268204,
-0.09581151604652405,
0.02639852464199066,
0.07483302801847458,
0.06392485648393631,
-0.04910123348236084,
-0.05623415485024452,
-0.03498426452279091,
-0.01831810735166073,
0.02245524153113365,
0.016861166805028915,
0.005814730655401945,
0.06665775179862976,
0.050683390349149704,
-0.013986624777317047,
-0.07207521051168442,
-0.11757232248783112,
-0.020243357867002487,
0.10335354506969452,
-0.09031598269939423,
0.09343243390321732,
0.027458202093839645,
0.0031426777131855488,
0.022532319650053978,
-0.024167941883206367,
-0.1040111631155014,
-0.030385898426175117,
0.07923617213964462,
0.027285603806376457,
-0.011078798212110996,
0.0493936650454998,
-0.0033429469913244247,
0.029522579163312912,
-0.04126211628317833,
0.0858459621667862,
0.010255196131765842,
0.03490074351429939,
-0.007274857722222805,
0.07506763190031052,
-0.02562207169830799,
0.07369651645421982,
-0.07892308384180069,
0.06985113024711609,
-0.02569064125418663,
-0.009497028775513172,
0.06874669343233109,
-0.026183640584349632,
-0.013263719156384468,
-0.03618212789297104,
-0.05337028205394745,
0.058361515402793884,
-0.0011535774683579803,
0.08776188641786575,
-0.022536147385835648,
-0.02634519711136818,
0.0162387453019619,
-0.03186943754553795,
0.08054184168577194,
-0.0880112573504448,
-0.009898725897073746,
0.049815934151411057,
-0.029413923621177673,
0.0017013950273394585,
0.004798828158527613,
0.05605277046561241,
0.0022368216887116432,
-7.04934181707785e-34,
0.05472103878855705,
0.021229013800621033,
-0.0413861982524395,
0.07584585249423981,
-0.07701709866523743,
-0.035430509597063065,
0.020995566621422768,
0.05687969550490379,
-0.05134410411119461,
-0.054138775914907455,
-0.0389435850083828,
-0.014619855210185051,
-0.013092830777168274,
-0.009901996701955795,
-0.010460440069437027,
-0.005894375499337912,
-0.0738639310002327,
0.021153327077627182,
-0.0012587076053023338,
0.03149202838540077,
0.027034131810069084,
0.08161435276269913,
-0.1055198535323143,
0.04432323947548866,
-0.14136295020580292,
0.02739102952182293,
-0.015145399607717991,
0.05301857739686966,
0.0004326945636421442,
-0.020170608535408974,
-0.0574958436191082,
0.0037175770848989487,
0.0026176352985203266,
-0.05994509905576706,
-0.07158593088388443,
0.034372661262750626,
0.0409458763897419,
-0.004792005755007267,
-0.03344815596938133,
0.04427187144756317,
-0.002094693947583437,
-0.0312986858189106,
-0.05885707959532738,
0.03548945114016533,
-0.0685388520359993,
-0.01112979743629694,
-0.012173919007182121,
0.014356052502989769,
-0.04868697002530098,
0.06879569590091705,
-0.060065437108278275,
0.02356983721256256,
-0.06981071084737778,
-0.02907712198793888,
-0.00828301440924406,
-0.028027117252349854,
-0.023848354816436768,
-0.04276392236351967,
0.012035281397402287,
-0.06050945073366165,
-0.038139138370752335,
0.016585230827331543,
0.035459090024232864,
-0.052738916128873825,
0.04885440692305565,
0.003908778075128794,
-0.05861658602952957,
-0.06856562942266464,
0.057598743587732315,
0.07869192957878113,
0.001179039478302002,
0.020451411604881287,
-0.019644351676106453,
-0.02051975205540657,
-0.05318453535437584,
-0.006468005944043398,
0.029906049370765686,
-0.03592197969555855,
0.02029302716255188,
0.011789838783442974,
0.014736304990947247,
-0.009566634893417358,
0.010368507355451584,
0.07252179086208344,
0.044846370816230774,
0.12079981714487076,
0.053675197064876556,
0.0949532762169838,
0.04842512309551239,
0.02385709062218666,
-0.042471013963222504,
0.016477450728416443,
0.06094048544764519,
0.024481015279889107,
-0.09463533759117126,
-3.6289730331873216e-8,
-0.11597630381584167,
0.04647155478596687,
-0.047792669385671616,
0.07135356962680817,
-0.02611573226749897,
-0.012384164147078991,
0.001845244667492807,
0.17752967774868011,
-0.03629278019070625,
0.039298348128795624,
0.006395472679287195,
0.008692868053913116,
0.016517285257577896,
-0.020364994183182716,
0.001969028264284134,
0.04419310390949249,
0.05621383339166641,
0.018724655732512474,
-0.020083243027329445,
-0.0078703248873353,
-0.005414051469415426,
0.07608629018068314,
0.05518839880824089,
-0.055142298340797424,
0.11212122440338135,
-0.12766166031360626,
0.007904255762696266,
0.13587014377117157,
0.04666978120803833,
-0.06560830771923065,
-0.035446248948574066,
0.010810058563947678,
-0.05412004142999649,
-0.009894322603940964,
0.057234421372413635,
0.10329604148864746,
-0.03904610872268677,
0.03930264338850975,
0.027971778064966202,
0.0025101839564740658,
0.06914234161376953,
0.052558477967977524,
-0.00785173662006855,
0.019156498834490776,
-0.012408647686243057,
-0.0027178344316780567,
0.012602142058312893,
0.013679102063179016,
0.03744400292634964,
-0.02203345112502575,
0.03589949756860733,
0.020677806809544563,
-0.011326120235025883,
0.04126041382551193,
0.018238414078950882,
0.007093632593750954,
-0.01930019073188305,
-0.0011962108546867967,
0.045561376959085464,
0.1277206540107727,
-0.0025768191553652287,
0.01265188492834568,
-0.10166137665510178,
-0.016365909948945045
] |
Dynamic Low-Rank Sparse Adaptation for Large Language Models | https://openreview.net/forum?id=oXh0939Zzq | [
"Weizhong Huang",
"Yuxin Zhang",
"Xiawu Zheng",
"Liuyang",
"Jing Lin",
"Yiwu Yao",
"Rongrong Ji"
] | Poster | Despite the efficacy of network sparsity in alleviating the deployment strain of Large Language Models (LLMs), it endures significant performance degradation. Applying Low-Rank Adaptation (LoRA) to fine-tune the sparse LLMs offers an intuitive approach to counter this predicament, while it holds shortcomings include: 1) The inability to integrate LoRA weights into sparse LLMs post-training, and 2) Insufficient performance recovery at high sparsity ratios. In this paper, we introduces dynamic $\textbf{Lo}$w-rank $\textbf{S}$parse $\textbf{A}$daptation $\textbf{(LoSA)}$, a novel method that seamlessly integrates low-rank adaptation into LLM sparsity within a unified framework, thereby enhancing the performance of sparse LLMs without increasing the inference latency. In particular, LoSA dynamically sparsifies the LoRA outcomes based on the corresponding sparse weights during fine-tuning, thus guaranteeing that the LoRA module can be integrated into the sparse LLMs post-training. Besides, to achieve the optimal sparse model architecture, LoSA leverages Representation Mutual Information (RMI) as an indicator to determine the importance of layers, thereby dynamically determining the optimal layer-wise sparsity rates during fine-tuning. Predicated on this, LoSA adjusts the rank of the LoRA module based on the variability in layer-wise reconstruction errors, allocating an appropriate fine-tuning for each layer to reduce the output discrepancies between dense and sparse LLMs. Extensive experiments tell that LoSA can efficiently boost the efficacy of sparse LLMs within a few hours, without introducing any additional inferential burden. For example, LoSA reduced the perplexity of sparse LLaMA-2-7B by $\textbf{68.73}$$\downarrow$ and increased zero-shot accuracy by $\textbf{16.32}$%$\uparrow$, achieving a $\textbf{2.60$\times$}$ speedup on CPU and $\textbf{2.23$\times$}$ speedup on GPU, requiring only $\textbf{45 minutes}$ of fine-tuning on $\textbf{a single}$ NVIDIA A100 80GB GPU. Code is available at https://github.com/wzhuang-xmu/LoSA. | Large Language Models; Network Sparsity; Low-Rank Adaptation | We present Dynamic Low-Rank Sparse Adaptation, an efficient fine-tuning method to enhance the performance of sparse Large Language Models. | 417 | 2502.14816 | [
-0.05949345976114273,
-0.11273130029439926,
0.006502480246126652,
0.054152119904756546,
0.06673890352249146,
0.03371550515294075,
-0.021610723808407784,
-0.009454835206270218,
0.011025767773389816,
-0.08470525592565536,
-0.036458998918533325,
0.00686027342453599,
0.02766571380198002,
-0.016031883656978607,
-0.016489695757627487,
0.038788795471191406,
0.0992148146033287,
0.09275101870298386,
-0.05218303203582764,
-0.0302504263818264,
-0.005958149209618568,
-0.0005463443812914193,
0.012018559500575066,
0.02822471596300602,
0.10681216418743134,
-0.040556348860263824,
-0.03272094205021858,
0.022519877180457115,
-0.004584962502121925,
-0.056974075734615326,
0.02974603697657585,
0.06523188948631287,
-0.014557153917849064,
0.01406603679060936,
-0.05851738527417183,
0.0098722530528903,
-0.01756184920668602,
0.017550179734826088,
0.01787581294775009,
0.0010078467894345522,
0.06692391633987427,
0.05138899013400078,
-0.0037490795366466045,
0.026862071827054024,
0.006623950321227312,
-0.04199374467134476,
-0.00003549479151843116,
-0.010727173648774624,
-0.015346342697739601,
0.0292966291308403,
-0.0016819036100059748,
-0.021428173407912254,
-0.05224777013063431,
0.054764922708272934,
-0.09796988219022751,
-0.03192265331745148,
0.0055909729562699795,
0.03792434185743332,
-0.029698053374886513,
-0.035284508019685745,
-0.042054206132888794,
-0.07135185599327087,
0.039461277425289154,
0.01792152225971222,
0.029226701706647873,
-0.053949788212776184,
0.06280363351106644,
0.046131156384944916,
-0.06567583978176117,
0.06291114538908005,
-0.05282922834157944,
0.030989104881882668,
-0.011877723969519138,
0.09942836314439774,
0.019519587978720665,
-0.0020098246168345213,
0.04291936382651329,
0.029509693384170532,
0.10105504840612411,
0.024762216955423355,
0.04501495510339737,
0.037192948162555695,
-0.015011698007583618,
-0.026013977825641632,
0.08495483547449112,
-0.017790138721466064,
0.010433259420096874,
-0.02052971161901951,
0.038985107094049454,
-0.00563431391492486,
-0.021591126918792725,
-0.06100266054272652,
0.005128262098878622,
0.06007937714457512,
-0.013299261219799519,
0.012272080406546593,
0.03465358540415764,
-0.04532769322395325,
-0.008318159729242325,
0.1313639134168625,
0.04047444090247154,
0.006631169468164444,
0.00559312803670764,
0.00046364040463231504,
-0.019457608461380005,
-0.0335020013153553,
0.04048236086964607,
0.0477946512401104,
0.01820100098848343,
-0.11509877443313599,
0.02903039939701557,
0.09135974943637848,
0.020703410729765892,
0.04310942813754082,
-0.0070675453171133995,
-0.057113438844680786,
0.05849163234233856,
-0.09888157993555069,
0.010289615951478481,
-0.00038589962059631944,
-0.05399887263774872,
-0.013216250576078892,
0.027709873393177986,
-0.027156250551342964,
0.01240032073110342,
0.055241409689188004,
-0.05262171849608421,
2.2846074336276865e-33,
0.11633840203285217,
0.0841553583741188,
0.019240520894527435,
-0.042913228273391724,
-0.012685982510447502,
-0.04303836077451706,
0.06935233622789383,
0.027814721688628197,
-0.10483493655920029,
-0.030013063922524452,
-0.09760762751102448,
-0.03426903486251831,
0.009631398133933544,
0.04093103110790253,
0.021441271528601646,
-0.01133103109896183,
0.029438039287924767,
0.07084258645772934,
0.029791321605443954,
-0.10206947475671768,
0.0737755075097084,
0.0092495521530509,
0.0073130470700562,
-0.04678419232368469,
0.05296608433127403,
0.008001082576811314,
0.09167132526636124,
-0.09155070036649704,
-0.0885559692978859,
0.03228999674320221,
-0.08579634130001068,
-0.08059512823820114,
0.07563399523496628,
-0.0023688217625021935,
0.032370664179325104,
-0.05682365223765373,
-0.02176435850560665,
-0.033572789281606674,
0.04224254563450813,
-0.11432866752147675,
-0.0005547963664866984,
0.005632699932903051,
-0.0031612194143235683,
-0.047228533774614334,
-0.08667344599962234,
-0.022605152800679207,
0.005391560960561037,
0.058410026133060455,
-0.08439771085977554,
-0.05315601825714111,
0.03690299764275551,
-0.04332023486495018,
-0.04974209889769554,
0.03858878090977669,
-0.028973253443837166,
-0.053651150315999985,
0.03657205030322075,
0.01883470080792904,
0.016630806028842926,
0.03411661460995674,
0.013031448237597942,
0.038002073764801025,
-0.02933991141617298,
0.0677512139081955,
0.11219611018896103,
-0.052211981266736984,
-0.020711058750748634,
0.03095148131251335,
0.04741927236318588,
-0.025276971980929375,
-0.028648104518651962,
-0.02882467396557331,
-0.04771827906370163,
-0.02354743704199791,
0.033411841839551926,
-0.06798791885375977,
0.08910432457923889,
-0.0945243164896965,
0.0457971915602684,
0.06166085600852966,
0.03863650932908058,
-0.01680050976574421,
0.0271356999874115,
-0.109835185110569,
-0.07710491865873337,
-0.003253784030675888,
0.07842793315649033,
-0.08487200736999512,
0.01747419498860836,
-0.02910885028541088,
-0.012527708895504475,
-0.04995356500148773,
-0.017020776867866516,
-0.10280104726552963,
0.04431242123246193,
-1.144879151468098e-33,
-0.04484708234667778,
-0.008871066384017467,
-0.023823589086532593,
0.0816982239484787,
-0.0006189057603478432,
-0.05811142921447754,
-0.01217092014849186,
0.08739646524190903,
-0.07587054371833801,
-0.06907206773757935,
-0.058620281517505646,
-0.07047858089208603,
0.0001740914594847709,
-0.036207765340805054,
0.07754508405923843,
0.011957666836678982,
-0.10517944395542145,
0.00793030858039856,
0.024488817900419235,
0.054291170090436935,
0.01787356473505497,
0.08990930765867233,
-0.004333463031798601,
0.010811051353812218,
-0.00792415626347065,
-0.031617727130651474,
-0.00998389720916748,
0.12107043713331223,
-0.027723174542188644,
-0.03512848541140556,
-0.015633292496204376,
0.008068599738180637,
-0.050575390458106995,
-0.029325410723686218,
-0.0821339339017868,
0.07936767488718033,
0.04293417930603027,
0.01584465615451336,
-0.04179418832063675,
0.09841644018888474,
0.010823138058185577,
-0.040591802448034286,
-0.06254711747169495,
0.005409677047282457,
0.019606243818998337,
-0.005739502143114805,
-0.13726255297660828,
-0.04829012230038643,
0.03283895179629326,
0.04759901389479637,
0.027606595307588577,
0.008465181104838848,
0.02839907817542553,
0.09688223898410797,
-0.03381045535206795,
-0.08725142478942871,
-0.021744994446635246,
-0.0356597937643528,
-0.012749001383781433,
0.002522372407838702,
-0.004460143856704235,
-0.04142909497022629,
-0.018607737496495247,
-0.08841846138238907,
0.04586804285645485,
-0.02551678568124771,
0.007299733813852072,
0.0030130951199680567,
-0.02564568817615509,
-0.0006153276772238314,
0.05580701678991318,
-0.00835895724594593,
-0.010964229702949524,
0.013825075700879097,
-0.014153974130749702,
-0.014833973720669746,
-0.01958135887980461,
-0.0010351744713261724,
-0.03259892016649246,
-0.03066493198275566,
-0.02223079279065132,
0.05435420572757721,
0.025331569835543633,
0.009615831077098846,
0.0475316047668457,
0.09930451959371567,
0.07757100462913513,
0.08722290396690369,
0.010913767851889133,
0.032770246267318726,
-0.009786184877157211,
-0.005525615066289902,
0.06860041618347168,
-0.008445767685770988,
-0.0673125684261322,
-4.1330121547389354e-8,
-0.12700572609901428,
0.06049026548862457,
-0.08121740818023682,
0.09885746240615845,
0.018526839092373848,
-0.015326086431741714,
-0.017755724489688873,
0.036175403743982315,
0.050858985632658005,
-0.04346631094813347,
0.014532225206494331,
-0.03322641924023628,
-0.011965418234467506,
-0.033701661974191666,
0.01837700605392456,
0.08419515937566757,
-0.018378494307398796,
0.0724836066365242,
-0.030878152698278427,
-0.006918739527463913,
0.03158821910619736,
0.110037662088871,
0.02227001264691353,
0.035334352403879166,
0.08322737365961075,
-0.027813350781798363,
-0.01627749763429165,
0.0955275148153305,
0.147277370095253,
0.020504813641309738,
-0.05831468105316162,
0.0051740542985498905,
0.0005691501428373158,
-0.05542721226811409,
0.0005409109871834517,
0.12112431228160858,
0.005466856993734837,
-0.027715446427464485,
-0.03100688010454178,
-0.0089688366279006,
-0.015452452003955841,
0.0531018041074276,
-0.07264493405818939,
0.01926788128912449,
0.028926022350788116,
-0.05026997625827789,
-0.08277806639671326,
-0.08001603186130524,
0.05715394392609596,
-0.07990261912345886,
0.060504622757434845,
-0.044436197727918625,
-0.04293131083250046,
0.0590386837720871,
0.05334137752652168,
0.016873706132173538,
0.035280417650938034,
0.0002620479208417237,
0.0659979060292244,
0.022240392863750458,
0.0108670424669981,
-0.01842532865703106,
-0.053220316767692566,
0.038311026990413666
] |
A Coefficient Makes SVRG Effective | https://openreview.net/forum?id=twtTLZnG0B | [
"Yida Yin",
"Zhiqiu Xu",
"Zhiyuan Li",
"Trevor Darrell",
"Zhuang Liu"
] | Poster | Stochastic Variance Reduced Gradient (SVRG), introduced by Johnson & Zhang (2013), is a theoretically compelling optimization method. However, as Defazio & Bottou (2019) highlight, its effectiveness in deep learning is yet to be proven. In this work, we demonstrate the potential of SVRG in optimizing real-world neural networks. Our empirical analysis finds that, for deeper neural networks, the strength of the variance reduction term in SVRG should be smaller and decrease as training progresses. Inspired by this, we introduce a multiplicative coefficient $\alpha$ to control the strength and adjust it through a linear decay schedule. We name our method $\alpha$-SVRG. Our results show $\alpha$-SVRG better optimizes models, consistently reducing training loss compared to the baseline and standard SVRG across various model architectures and multiple image classification datasets. We hope our findings encourage further exploration into variance reduction techniques in deep learning. Code is available at github.com/davidyyd/alpha-SVRG. | Optimization; Variance Reduction; SGD | Introducing a coefficient to control the variance reduction strength in SVRG makes it effective for deep networks. | 411 | 2311.05589 | [
-0.07086407393217087,
-0.06210353597998619,
0.05978382006287575,
0.022808820009231567,
0.012314444407820702,
0.04205719381570816,
-0.050629012286663055,
-0.012523251585662365,
-0.028745362535119057,
-0.09212221205234528,
-0.050819553434848785,
0.03190435469150543,
-0.052636392414569855,
0.003980822861194611,
-0.0742448940873146,
0.020601069554686546,
0.05019092559814453,
0.16531115770339966,
-0.14872269332408905,
-0.020116491243243217,
-0.013209117576479912,
-0.0135422982275486,
0.015491913072764874,
0.021069012582302094,
0.10020706802606583,
-0.056611452251672745,
-0.07577140629291534,
-0.09068137407302856,
0.03517422825098038,
-0.032657623291015625,
0.0978512167930603,
-0.10088295489549637,
0.04637982323765755,
0.0030666571110486984,
-0.01338223833590746,
0.0677449032664299,
-0.017666039988398552,
0.012486959807574749,
-0.0053766872733831406,
0.037638261914253235,
-0.024700583890080452,
0.01565251685678959,
0.024517446756362915,
0.026422392576932907,
0.06958471238613129,
0.060117147862911224,
0.036189667880535126,
-0.11572549492120743,
0.01701134629547596,
0.07433541119098663,
-0.03191572055220604,
-0.05160791799426079,
-0.037496261298656464,
0.07598615437746048,
-0.05182695388793945,
0.005424341652542353,
0.024857083335518837,
0.05328980088233948,
-0.02721368707716465,
0.05954442918300629,
0.04798591136932373,
-0.07312173396348953,
0.004495021887123585,
-0.023110944777727127,
-0.03788764774799347,
-0.010566312819719315,
0.09928743541240692,
-0.004451163578778505,
0.026615461334586143,
0.023977937176823616,
0.06121591851115227,
0.09402376413345337,
-0.06902192533016205,
0.03551262617111206,
0.029821116477251053,
0.008102284744381905,
0.10709922015666962,
-0.07754581421613693,
0.09681619703769684,
-0.036338143050670624,
0.0504235178232193,
0.012145236134529114,
0.01732311211526394,
-0.04246486350893974,
0.05255725234746933,
-0.004000501707196236,
-0.05862991139292717,
-0.046634040772914886,
0.014246940612792969,
-0.011529599316418171,
-0.01753232441842556,
-0.01735691912472248,
-0.0018923005554825068,
-0.051265835762023926,
-0.010557497851550579,
0.0059813885018229485,
0.005163370165973902,
-0.06995823234319687,
-0.06617649644613266,
0.09891793876886368,
-0.00729773286730051,
-0.0007866357918828726,
-0.02715228497982025,
-0.06893307715654373,
0.05688619613647461,
-0.014638470485806465,
0.09378158301115036,
-0.0072280107997357845,
-0.0334048792719841,
-0.057123102247714996,
0.03371673449873924,
0.02151273377239704,
-0.03873257339000702,
-0.006738397758454084,
0.028011919930577278,
0.03306357562541962,
0.04862166568636894,
0.0267278254032135,
-0.013422365300357342,
0.049557607620954514,
-0.09095089137554169,
0.02859717421233654,
-0.032135456800460815,
0.016493557021021843,
-0.004040098749101162,
-0.0008736913441680372,
-0.051342520862817764,
2.775052145794554e-33,
-0.034096043556928635,
0.01967461034655571,
0.02039678953588009,
-0.048945922404527664,
0.06387010216712952,
-0.04254736378788948,
-0.0055518303997814655,
-0.01765676960349083,
-0.06470603495836258,
0.02880767360329628,
-0.050419434905052185,
-0.02926049381494522,
-0.09115270525217056,
0.07261116802692413,
0.05314389616250992,
-0.048146747052669525,
-0.031240735203027725,
0.025848163291811943,
0.011385812424123287,
-0.06549518555402756,
0.0824417695403099,
-0.03357284143567085,
-0.014624892733991146,
-0.0024058884009718895,
0.017694571986794472,
-0.08036093413829803,
0.06654544174671173,
0.01842723973095417,
-0.03263314813375473,
0.01804717257618904,
-0.08484135568141937,
0.04522550851106644,
0.00025287552853114903,
0.05646083876490593,
0.033432960510253906,
-0.018023373559117317,
-0.024080565199255943,
0.03784829378128052,
0.03590591251850128,
-0.0027821906842291355,
-0.020934872329235077,
0.07002391666173935,
-0.018103783950209618,
-0.03629717603325844,
-0.043338045477867126,
-0.08412197977304459,
0.06353424489498138,
0.0031099070329219103,
-0.017093626782298088,
-0.09642565995454788,
-0.013604936189949512,
0.010055232793092728,
-0.043453801423311234,
0.03370187059044838,
0.039765097200870514,
-0.01979087106883526,
0.07560785114765167,
0.05479077249765396,
-0.01394020114094019,
0.0004208158061373979,
0.028114112094044685,
0.034387823194265366,
-0.005855742841959,
0.08151024580001831,
0.006695601157844067,
-0.043576136231422424,
0.061834998428821564,
0.06306850165128708,
-0.010593618266284466,
-0.037315741181373596,
-0.04973317310214043,
0.027076715603470802,
-0.015745457261800766,
-0.09712057560682297,
0.009164479561150074,
-0.001180025516077876,
0.0799439549446106,
-0.027840236201882362,
-0.040861327201128006,
0.042362626641988754,
-0.0714181661605835,
0.10685456544160843,
0.020846661180257797,
-0.017642710357904434,
-0.004370531532913446,
0.05075559392571449,
0.015130473300814629,
-0.08826320618391037,
-0.009385352022945881,
-0.00699874060228467,
-0.05802947282791138,
-0.05412023887038231,
-0.009092574939131737,
-0.03913106769323349,
-0.06340853124856949,
-1.6744812334997865e-33,
-0.08453663438558578,
0.0924932137131691,
-0.06262592226266861,
0.08859435468912125,
-0.044139694422483444,
-0.027892261743545532,
-0.02541402168571949,
0.01454619225114584,
-0.06511018425226212,
-0.03767574578523636,
0.02128203585743904,
0.017857033759355545,
-0.04068073630332947,
0.005768591538071632,
0.03882203251123428,
-0.05757981166243553,
0.026738232001662254,
-0.012208976782858372,
-0.04091864451766014,
-0.02907290682196617,
0.02378040924668312,
0.10297233611345291,
-0.03788621351122856,
0.027537349611520767,
-0.05505576729774475,
-0.0039982870221138,
-0.028213273733854294,
0.1252623349428177,
0.014778673648834229,
-0.0705629363656044,
-0.0565827451646328,
-0.0147513747215271,
-0.04298704117536545,
0.002371440874412656,
0.0546569861471653,
0.06870928406715393,
0.03190800920128822,
0.002693666610866785,
-0.01289050281047821,
0.06320556253194809,
0.09334169328212738,
-0.03111273981630802,
-0.052633173763751984,
-0.015768788754940033,
0.012752680107951164,
0.013551458716392517,
-0.060595352202653885,
-0.02699158899486065,
0.011312966234982014,
0.03767351806163788,
0.023342013359069824,
-0.05490310862660408,
-0.08505205810070038,
0.09716133028268814,
-0.012510020285844803,
-0.11071757972240448,
-0.032488077878952026,
-0.01043763104826212,
0.06591463088989258,
0.10159294307231903,
-0.08588636666536331,
-0.031220998615026474,
-0.060657959431409836,
-0.018635932356119156,
-0.052020519971847534,
-0.021037602797150612,
-0.05463448166847229,
0.06271593272686005,
0.07094857096672058,
0.02848510444164276,
0.03766249865293503,
0.01530543714761734,
0.013642879202961922,
0.06302222609519958,
-0.09443855285644531,
-0.033045604825019836,
0.05304688215255737,
-0.021179066970944405,
0.034902092069387436,
-0.029714742675423622,
0.0586569644510746,
0.02731950581073761,
0.022973278537392616,
-0.0005060916882939637,
0.1523650735616684,
0.07882191240787506,
-0.04462512955069542,
0.08394533395767212,
0.07310563325881958,
0.04513709619641304,
-0.036207593977451324,
0.008882577531039715,
0.011255170218646526,
0.020902052521705627,
-0.0273230392485857,
-4.612504511669613e-8,
0.03655557706952095,
0.11799666285514832,
0.06551910936832428,
-0.00769902765750885,
0.09281356632709503,
-0.05521029606461525,
0.02673606388270855,
0.07022587954998016,
-0.04077417030930519,
0.03820905461907387,
0.10273407399654388,
-0.019683845341205597,
0.004256243351846933,
-0.020867327228188515,
-0.026413923129439354,
0.07820207625627518,
-0.034761056303977966,
0.03689168021082878,
0.032284438610076904,
-0.0011672853725031018,
0.0056212907657027245,
0.007339823059737682,
0.024832403287291527,
-0.03194192796945572,
0.04506522789597511,
-0.10934358090162277,
-0.018959831446409225,
0.05627654865384102,
0.027510955929756165,
-0.021064845845103264,
0.013588967733085155,
0.04939641058444977,
0.02268405817449093,
-0.00837892759591341,
0.0026458476204425097,
0.11942295730113983,
0.029588062316179276,
0.01178422849625349,
-0.03516397997736931,
0.026017896831035614,
-0.03158898279070854,
-0.00010734376701293513,
-0.02298445627093315,
-0.00873508770018816,
-0.05187080428004265,
0.03365334868431091,
0.0013698606053367257,
-0.08754909038543701,
-0.021058641374111176,
-0.011844782158732414,
0.0225056279450655,
-0.01853995956480503,
0.04890885949134827,
0.12470150738954544,
0.023894138634204865,
-0.04423017427325249,
0.03152240440249443,
-0.063649483025074,
-0.017005253583192825,
0.08365314453840256,
-0.019225869327783585,
-0.05877985060214996,
-0.09555424749851227,
-0.0026157801039516926
] |
DiffPC: Diffusion-based High Perceptual Fidelity Image Compression with Semantic Refinement | https://openreview.net/forum?id=RL7PycCtAO | [
"Yichong Xia",
"Yimin Zhou",
"Jinpeng Wang",
"Baoyi An",
"Haoqian Wang",
"Yaowei Wang",
"Bin Chen"
] | Poster | Reconstructing high-quality images under low bitrates conditions presents a challenge, and previous methods have made this task feasible by leveraging the priors of diffusion models. However, the effective exploration of pre-trained latent diffusion models and semantic information integration in image compression tasks still needs further study. To address this issue, we introduce Diffusion-based High Perceptual Fidelity Image Compression with Semantic Refinement (DiffPC), a two-stage image compression framework based on stable diffusion. DiffPC efficiently encodes low-level image information, enabling the highly realistic reconstruction of the original image by leveraging high-level semantic features and the prior knowledge inherent in diffusion models. Specifically, DiffPC utilizes a multi-feature compressor to represent crucial low-level information with minimal bitrates and employs pre-embedding to acquire more robust hybrid semantics, thereby providing additional context for the decoding end. Furthermore, we have devised a control module tailored for image compression tasks, ensuring structural and textural consistency in reconstruction even at low bitrates and preventing decoding collapses induced by condition leakage. Extensive experiments demonstrate that our method achieves state-of-the-art perceptual fidelity and surpasses previous perceptual image compression methods by a significant margin in statistical fidelity. | lossy image compression, diffusion model | null | 409 | null | [
0.03536229953169823,
-0.0611288882791996,
0.033992063254117966,
0.06469462811946869,
0.08030138909816742,
-0.06024571880698204,
-0.06761482357978821,
-0.09499752521514893,
0.00967518612742424,
0.023607205599546432,
-0.01584126614034176,
-0.03422769159078598,
0.016716446727514267,
0.06685248762369156,
-0.0258452408015728,
0.0013401060132309794,
0.08199071884155273,
0.13066186010837555,
-0.13880528509616852,
-0.0458441860973835,
-0.0000602088839514181,
-0.06248031556606293,
0.016104890033602715,
-0.04511568695306778,
0.07670191675424576,
0.057222168892621994,
0.013900364749133587,
-0.017003633081912994,
0.07100044190883636,
-0.041653331369161606,
0.057120345532894135,
-0.040644802153110504,
-0.026750845834612846,
0.0192288588732481,
-0.023037217557430267,
0.10835923999547958,
0.032890211790800095,
0.01287643238902092,
-0.02271919697523117,
0.009084457531571388,
0.0018994008423760533,
0.08274752646684647,
-0.049111414700746536,
0.06464171409606934,
0.05670174956321716,
0.031226862221956253,
-0.025079011917114258,
0.02163415029644966,
0.02427304908633232,
-0.014115937054157257,
0.02588137611746788,
0.0437638983130455,
-0.029611771926283836,
0.056853361427783966,
-0.02892351523041725,
0.04907241091132164,
0.006902874447405338,
-0.01505048293620348,
-0.04488294944167137,
-0.00007342812023125589,
-0.06133109703660011,
-0.030738631263375282,
0.05130411311984062,
-0.005885804537683725,
-0.0006818276597186923,
0.005987408570945263,
0.027261247858405113,
-0.006494765169918537,
0.04160614311695099,
-0.0038956168573349714,
-0.06162125989794731,
-0.01585066318511963,
-0.04436226934194565,
-0.015406888909637928,
-0.0348452590405941,
-0.013653070665895939,
0.06687866896390915,
-0.01106271892786026,
0.02100292034447193,
-0.04203645884990692,
0.05582781508564949,
-0.01924602873623371,
0.06833387166261673,
0.015200315974652767,
0.042509596794843674,
-0.049772512167692184,
0.04487701877951622,
0.02794288657605648,
-0.03737161308526993,
-0.00045633743866346776,
-0.07078804820775986,
0.007550246547907591,
-0.013568052090704441,
0.0019005781505256891,
0.0449645072221756,
0.020226432010531425,
-0.031021365895867348,
-0.04527869448065758,
-0.031534343957901,
0.06727685779333115,
0.004753682296723127,
-0.047793373465538025,
-0.05863097682595253,
0.0009205065434798598,
0.028574815019965172,
-0.08638905733823776,
0.06883391737937927,
0.0706295371055603,
-0.02619221806526184,
0.06524764746427536,
0.06543391197919846,
0.07035864144563675,
0.0513567216694355,
-0.03818344697356224,
0.07576600462198257,
-0.05043192207813263,
-0.04347243160009384,
0.009590129368007183,
0.06634124368429184,
-0.03214184567332268,
-0.025319069623947144,
0.012672405689954758,
-0.08842147886753082,
-0.041510555893182755,
0.0453503355383873,
-0.08100749552249908,
-0.020325977355241776,
2.4984811002491168e-33,
-0.0674177035689354,
-0.014409220777451992,
0.023234792053699493,
0.06405622512102127,
0.04265141487121582,
0.0057245283387601376,
-0.04378671571612358,
-0.023032864555716515,
-0.04346754774451256,
-0.029906602576375008,
0.0010823929915204644,
-0.025548217818140984,
-0.08009479194879532,
0.08241543918848038,
-0.04833703860640526,
-0.047314681112766266,
-0.09195394068956375,
0.05364951863884926,
0.03018808923661709,
0.048400118947029114,
-0.013353263027966022,
0.02285291813313961,
-0.04272545501589775,
-0.03448963165283203,
-0.013584853149950504,
-0.038306280970573425,
0.06121658906340599,
-0.08477550745010376,
0.06470751017332077,
-0.005278064403682947,
-0.1730717569589615,
-0.05738505721092224,
0.03107442334294319,
-0.0006040747975930572,
-0.05301347002387047,
-0.013962716795504093,
-0.009783454239368439,
-0.03869355097413063,
-0.028251057490706444,
-0.0015754246851429343,
0.0055787828750908375,
0.06325752288103104,
-0.0643770769238472,
-0.047596532851457596,
-0.0980912372469902,
-0.07009240239858627,
0.062292274087667465,
0.022235868498682976,
-0.11266407370567322,
-0.00863636378198862,
0.06086314842104912,
-0.004541418515145779,
0.020659398287534714,
-0.05501018092036247,
-0.08188247680664062,
-0.07165294140577316,
0.11451601982116699,
-0.07158669084310532,
0.05649561062455177,
0.012415093369781971,
0.06848782300949097,
0.03685488551855087,
-0.0552690327167511,
0.09267935901880264,
0.008369211107492447,
-0.05008629709482193,
-0.006790098734200001,
0.08714944124221802,
-0.03096482902765274,
0.037151407450437546,
-0.09445450454950333,
0.03510429710149765,
0.025120262056589127,
-0.09790405631065369,
0.025482680648565292,
0.00782373920083046,
0.010200043208897114,
0.013160061091184616,
0.033670131117105484,
-0.012296569533646107,
-0.05208277329802513,
-0.03569325432181358,
-0.017620114609599113,
-0.06228053569793701,
-0.022874407470226288,
0.013171848841011524,
0.0055966670624911785,
-0.06223512068390846,
-0.016448521986603737,
-0.08194465190172195,
-0.0372358076274395,
-0.03741510957479477,
0.04456544667482376,
0.021718354895710945,
0.06478242576122284,
-7.485437823440937e-34,
0.012502570636570454,
0.07970181107521057,
-0.011679301038384438,
0.06410618871450424,
-0.05729891359806061,
-0.021014057099819183,
0.040555644780397415,
0.01278444193303585,
0.10935629904270172,
-0.07810617983341217,
-0.0005520504200831056,
-0.07217883318662643,
-0.036036521196365356,
-0.05526348948478699,
-0.02045990712940693,
-0.026533909142017365,
0.003485914319753647,
0.0029471113812178373,
-0.03492623195052147,
0.026270780712366104,
-0.018066948279738426,
0.06506534665822983,
-0.035383615642786026,
0.04779549315571785,
-0.0553773008286953,
0.0688762217760086,
-0.05025312304496765,
0.04634690284729004,
0.046102408319711685,
-0.02851221151649952,
-0.008016294799745083,
-0.022309355437755585,
0.056765224784612656,
-0.06834927946329117,
-0.06957616657018661,
0.0016167686553671956,
0.005833454895764589,
0.007222766522318125,
-0.056673917919397354,
0.05742892250418663,
0.03098481521010399,
-0.05339096859097481,
-0.08358971774578094,
0.07219960540533066,
0.014232010580599308,
-0.054840389639139175,
-0.04838361591100693,
-0.059158772230148315,
0.06366868317127228,
0.0713130459189415,
0.02359018661081791,
0.06296348571777344,
-0.10020066052675247,
0.11028705537319183,
-0.014673738740384579,
-0.05613623932003975,
-0.12049185484647751,
0.012689259834587574,
0.07996776700019836,
0.03200333192944527,
0.01755714975297451,
-0.09367281943559647,
0.016619892790913582,
-0.06902633607387543,
0.012607995420694351,
-0.012492699548602104,
-0.051443006843328476,
0.019195597618818283,
0.000029175571398809552,
0.0956805869936943,
-0.043182410299777985,
0.035968080163002014,
0.04078221693634987,
0.012148384004831314,
0.08798667788505554,
-0.05313710868358612,
0.03408415988087654,
-0.0012179302284494042,
0.04610909894108772,
0.06417875736951828,
0.03540818765759468,
-0.002250904683023691,
-0.002487128134816885,
0.05029526352882385,
0.05843203887343407,
0.03080892749130726,
-0.028156233951449394,
0.0003252726746723056,
0.038343314081430435,
0.05840916559100151,
-0.003042759606614709,
-0.016673676669597626,
0.011257453821599483,
0.02413013204932213,
-0.04730289429426193,
-3.823480909659338e-8,
-0.12014815211296082,
0.02049490064382553,
0.023284615948796272,
0.016980504617094994,
-0.0308309867978096,
-0.11189387738704681,
0.07056643813848495,
0.057262372225522995,
0.024538934230804443,
-0.06706321984529495,
0.1251903772354126,
-0.012232176028192043,
-0.06017661467194557,
-0.08619525283575058,
0.016349248588085175,
0.11850094795227051,
0.028144001960754395,
0.0017031593015417457,
0.07436186820268631,
-0.008262003771960735,
-0.04286419227719307,
0.013922534883022308,
0.04197307676076889,
-0.03085259720683098,
0.058364465832710266,
0.01528001856058836,
0.012605338357388973,
0.0477825365960598,
0.0759621262550354,
-0.008020535111427307,
0.0029772939160466194,
-0.032247696071863174,
0.03891560807824135,
-0.06261377781629562,
-0.05985506996512413,
0.09407512098550797,
0.052087798714637756,
0.018730273470282555,
0.004100262653082609,
-0.07712166756391525,
-0.015909606590867043,
0.04739892855286598,
-0.003772399388253689,
-0.0014527533203363419,
0.0315055213868618,
0.032122254371643066,
0.04812532290816307,
0.052106279879808426,
-0.010992459952831268,
0.024292204529047012,
0.022364869713783264,
0.028688514605164528,
0.02007768675684929,
0.07891032099723816,
0.07998215407133102,
-0.06243966519832611,
0.006756972521543503,
0.044612932950258255,
0.07254460453987122,
0.07154520601034164,
0.027116382494568825,
0.04888824746012688,
0.0789019912481308,
0.03809734061360359
] |
Unleashing the Power of Task-Specific Directions in Parameter Efficient Fine-tuning | https://openreview.net/forum?id=RYrJqz44p4 | [
"Chongjie Si",
"Zhiyi Shi",
"Shifan Zhang",
"Xiaokang Yang",
"Hanspeter Pfister",
"Wei Shen"
] | Poster | Large language models demonstrate impressive performance on downstream tasks, yet requiring extensive resource consumption when fully fine-tuning all parameters. To mitigate this, Parameter Efficient Fine-Tuning (PEFT) strategies, such as LoRA, have been developed.
In this paper, we delve into the concept of task-specific directions (TSDs)—critical for transitioning large models from pretrained states to task-specific enhancements in PEFT. We propose a framework to clearly define these directions and explore their properties, and practical utilization challenges. We then introduce a novel approach, LoRA-Dash, which aims to maximize the impact of TSDs during the fine-tuning process, thereby enhancing model performance on targeted tasks. Extensive experiments have conclusively demonstrated the effectiveness of LoRA-Dash, and in-depth analyses further reveal the underlying mechanisms of LoRA-Dash. | parameter efficient fine-tuning, low-rank adaptation, task-specific directions | It provides a clear definition of task-specific directions and propose a method unleashing the power of these directions. | 404 | 2409.01035 | [
-0.031512968242168427,
-0.07359682023525238,
0.05319303274154663,
0.025944789871573448,
0.04115778207778931,
-0.016362270340323448,
0.011025812476873398,
-0.034508440643548965,
0.0332290418446064,
-0.03917941078543663,
-0.09866947680711746,
-0.0146551001816988,
-0.023882390931248665,
0.0221028421074152,
0.07993762940168381,
-0.045979514718055725,
0.11836150288581848,
0.006896920967847109,
-0.09327757358551025,
-0.048192158341407776,
0.1118868961930275,
-0.006704296916723251,
0.06380617618560791,
0.015961628407239914,
-0.012372555211186409,
-0.0034386860206723213,
-0.04290951415896416,
-0.07933303713798523,
0.05044322460889816,
-0.05295799672603607,
-0.03916030377149582,
0.043477632105350494,
0.001118886168114841,
0.02654939703643322,
-0.06471771746873856,
0.05639435350894928,
-0.07154543697834015,
0.010808493010699749,
0.0008249857346527278,
-0.030541587620973587,
-0.00036203343188390136,
-0.010941936634480953,
0.000058286994317313656,
0.03468026965856552,
0.047017667442560196,
-0.059499286115169525,
0.033137064427137375,
-0.06236632913351059,
-0.0610865093767643,
0.012606016360223293,
-0.07494615018367767,
-0.030552063137292862,
0.022068966180086136,
0.03246579319238663,
0.009688358753919601,
0.05848177522420883,
0.036557797342538834,
0.05808471143245697,
-0.006128959823399782,
-0.019753051921725273,
-0.07932141423225403,
-0.09238579124212265,
-0.03684287518262863,
-0.03023003600537777,
0.023081526160240173,
-0.001632041996344924,
0.016721582040190697,
-0.004146016202867031,
-0.041519179940223694,
0.042747173458337784,
-0.033546604216098785,
0.032088372856378555,
-0.005262942519038916,
0.0061855860985815525,
0.04575159773230553,
0.05356902256608009,
0.011254637502133846,
-0.016749607399106026,
0.012845497578382492,
-0.04532164707779884,
0.012038547545671463,
-0.03159229829907417,
0.02003207430243492,
-0.004277951084077358,
0.17650669813156128,
0.03894677758216858,
-0.00802264828234911,
-0.04014814272522926,
0.08523266762495041,
-0.0018423853907734156,
-0.03974970057606697,
-0.11501678079366684,
0.047525618225336075,
-0.021022852510213852,
-0.03214012086391449,
0.09040210396051407,
0.015955651178956032,
-0.10779423266649246,
-0.015169711783528328,
0.05092795938253403,
0.06616492569446564,
0.03331976756453514,
0.017687134444713593,
0.014038649387657642,
-0.06684534251689911,
-0.039915163069963455,
0.06741942465305328,
0.08903251588344574,
-0.05052487179636955,
-0.052248675376176834,
0.019231664016842842,
0.055414944887161255,
0.08504576236009598,
0.044668588787317276,
-0.04212452098727226,
0.005659095942974091,
-0.07727444171905518,
-0.06177440285682678,
0.035075049847364426,
0.09450431913137436,
-0.04147005453705788,
-0.03351878747344017,
0.009838690049946308,
0.029555972665548325,
-0.007870730012655258,
0.022620949894189835,
-0.08402656018733978,
1.0859786192972472e-33,
0.05237255245447159,
0.024948304519057274,
0.0006643488886766136,
0.0005932794301770627,
-0.03566459193825722,
-0.013637591153383255,
0.03178669139742851,
0.024505119770765305,
-0.07609749585390091,
-0.0016998270293697715,
-0.007359841372817755,
0.02686259150505066,
-0.029864154756069183,
0.08197495341300964,
0.022392194718122482,
-0.04759954661130905,
-0.05134878307580948,
0.09224695712327957,
-0.05526266247034073,
-0.026722434908151627,
0.11089257895946503,
-0.02481999062001705,
0.014756702817976475,
-0.07945506274700165,
0.04741169512271881,
0.03149867802858353,
0.05551019310951233,
-0.06587301939725876,
-0.08876873552799225,
0.03139461576938629,
-0.07830195128917694,
-0.047811634838581085,
0.04318121820688248,
0.0372798778116703,
-0.009484504349529743,
-0.02947680838406086,
-0.029622288420796394,
0.003617106704041362,
0.062649667263031,
-0.0007850458496250212,
0.010379991494119167,
0.059895727783441544,
0.03249221667647362,
-0.033798858523368835,
-0.07931352406740189,
-0.06023290380835533,
-0.0035996197257190943,
0.024724405258893967,
-0.05385420843958855,
-0.014148959890007973,
-0.005842586979269981,
0.033983469009399414,
0.0013851277763023973,
-0.04070029407739639,
-0.014355902560055256,
-0.0035177357494831085,
0.036636192351579666,
0.050656720995903015,
-0.0027690567076206207,
0.07619436830282211,
0.04060178995132446,
0.0331607349216938,
-0.07094927877187729,
0.043535999953746796,
0.08542068302631378,
0.014535553753376007,
-0.032745953649282455,
0.1133619099855423,
0.05371933430433273,
-0.007673836313188076,
-0.023335883393883705,
-0.04509183391928673,
0.01730593666434288,
-0.010493782348930836,
0.0801815390586853,
-0.058621130883693695,
0.03802802041172981,
-0.08331973105669022,
-0.003499810816720128,
-0.004586812574416399,
0.03731239587068558,
0.03337220102548599,
0.007203401066362858,
-0.015872931107878685,
0.0000945793217397295,
-0.10821022093296051,
-0.03663519024848938,
-0.06369157135486603,
-0.07558215409517288,
-0.007211499381810427,
-0.0008202927419915795,
0.02249445952475071,
-0.006454084534198046,
-0.007403483614325523,
-0.04052786901593208,
-8.323744374794736e-34,
0.0005346836987882853,
-0.057987213134765625,
-0.030847052112221718,
0.08987705409526825,
0.007266872562468052,
-0.022653188556432724,
0.04684309661388397,
-0.03393710032105446,
-0.022403817623853683,
-0.018697388470172882,
-0.04436488077044487,
-0.06993036717176437,
-0.040757812559604645,
-0.023573018610477448,
0.050162870436906815,
-0.03189627453684807,
-0.035606980323791504,
0.023408299311995506,
0.050059929490089417,
0.029444394633173943,
0.003433960024267435,
0.03321878984570503,
-0.0443551205098629,
0.05418775975704193,
0.024553850293159485,
-0.00823612418025732,
-0.0642649382352829,
0.01576814241707325,
-0.03199291229248047,
0.006214559078216553,
-0.12541969120502472,
0.03774983435869217,
-0.01314080785959959,
-0.0048460885882377625,
-0.015248636715114117,
0.024991171434521675,
-0.002301358850672841,
0.010251311585307121,
-0.01728532649576664,
0.08086122572422028,
0.08393982797861099,
-0.03491959720849991,
-0.02666458487510681,
0.06873229146003723,
0.01055261678993702,
0.02720828540623188,
-0.15349209308624268,
-0.034201543778181076,
-0.030546408146619797,
-0.003610256826505065,
0.036688365042209625,
-0.010226638987660408,
0.004309951793402433,
-0.03891320899128914,
-0.031086953356862068,
-0.10365363210439682,
-0.02471793442964554,
-0.09545067697763443,
0.021597372367978096,
0.007922207936644554,
-0.03412800282239914,
-0.01362660899758339,
0.0378325879573822,
-0.10356584936380386,
0.05622077360749245,
0.001822569640353322,
-0.02888447605073452,
-0.01943116821348667,
0.07910751551389694,
-0.016681592911481857,
0.004994372371584177,
-0.03213515132665634,
-0.01221644226461649,
0.015743890777230263,
-0.04471362754702568,
-0.035977233201265335,
0.001421519322320819,
-0.06379540264606476,
0.004431934095919132,
-0.025956319645047188,
-0.03683985769748688,
0.05831335857510567,
-0.050651490688323975,
0.015864837914705276,
0.008071902208030224,
0.12271146476268768,
-0.051263175904750824,
0.14470210671424866,
0.004989343695342541,
0.016164839267730713,
-0.03032197244465351,
0.009795030578970909,
0.056359756737947464,
0.08039315044879913,
-0.08789906650781631,
-4.324178348724672e-8,
-0.06199894845485687,
0.02289542555809021,
-0.03195524960756302,
0.13444450497627258,
0.000056837903684936464,
-0.012319211848080158,
-0.008448651991784573,
0.10615943372249603,
0.04445517808198929,
-0.03715113550424576,
0.049164965748786926,
-0.012923765927553177,
-0.007279595825821161,
-0.0040769451297819614,
0.09443797171115875,
0.03408540412783623,
0.017765535041689873,
0.12647344172000885,
-0.044884294271469116,
-0.03504828363656998,
0.07211334258317947,
0.08211594074964523,
-0.04617688059806824,
0.04316087067127228,
0.06429141759872437,
-0.0561947226524353,
-0.047579582780599594,
0.07655732333660126,
0.12238287180662155,
-0.01167488843202591,
-0.008873973041772842,
0.003857721108943224,
-0.02824774570763111,
0.020385656505823135,
-0.002437083749100566,
0.06243357062339783,
-0.020081087946891785,
-0.006549414712935686,
0.009229432791471481,
0.03396105393767357,
0.07801821082830429,
0.13223987817764282,
-0.08675628155469894,
0.0255129374563694,
-0.009774832054972649,
-0.003915087785571814,
-0.07499929517507553,
-0.06629332900047302,
0.004379930440336466,
-0.11758768558502197,
-0.005567227490246296,
0.014584616757929325,
-0.04824240878224373,
0.07264101505279541,
-0.0006052234093658626,
0.07145310938358307,
0.04219507426023483,
-0.10442529618740082,
0.014333735220134258,
0.026039840653538704,
0.09487581998109818,
0.04302177578210831,
-0.058664437383413315,
0.005162758752703667
] |
Learning Shape-Independent Transformation via Spherical Representations for Category-Level Object Pose Estimation | https://openreview.net/forum?id=D4xztKoz0Y | [
"Huan Ren",
"Wenfei Yang",
"Xiang Liu",
"Shifeng Zhang",
"Tianzhu Zhang"
] | Poster | Category-level object pose estimation aims to determine the pose and size of novel objects in specific categories. Existing correspondence-based approaches typically adopt point-based representations to establish the correspondences between primitive observed points and normalized object coordinates. However, due to the inherent shape-dependence of canonical coordinates, these methods suffer from semantic incoherence across diverse object shapes. To resolve this issue, we innovatively leverage the sphere as a shared proxy shape of objects to learn shape-independent transformation via spherical representations. Based on this insight, we introduce a novel architecture called SpherePose, which yields precise correspondence prediction through three core designs. Firstly, We endow the point-wise feature extraction with SO(3)-invariance, which facilitates robust mapping between camera coordinate space and object coordinate space regardless of rotation transformation. Secondly, the spherical attention mechanism is designed to propagate and integrate features among spherical anchors from a comprehensive perspective, thus mitigating the interference of noise and incomplete point cloud. Lastly, a hyperbolic correspondence loss function is designed to distinguish subtle distinctions, which can promote the precision of correspondence prediction. Experimental results on CAMERA25, REAL275 and HouseCat6D benchmarks demonstrate the superior performance of our method, verifying the effectiveness of spherical representations and architectural innovations. | category-level object pose estimation, spherical representations, shape-independence, correspondence prediction | null | 403 | 2503.13926 | [
0.004141791258007288,
-0.060904815793037415,
0.00036413149791769683,
-0.02367238886654377,
0.12621696293354034,
0.006984780076891184,
-0.028244152665138245,
-0.07091512531042099,
0.06953515857458115,
-0.01548099797219038,
-0.00981177669018507,
-0.02762012556195259,
0.005528634414076805,
0.04656870663166046,
0.025034643709659576,
-0.06391679495573044,
0.09703876078128815,
0.11454334110021591,
-0.060986969619989395,
0.054271914064884186,
0.01905432902276516,
-0.010066474787890911,
0.05217204988002777,
0.0586840994656086,
0.06699391454458237,
0.031313683837652206,
0.06929422914981842,
-0.08501118421554565,
-0.08987728506326675,
-0.07975908368825912,
0.0037973651196807623,
-0.017759570851922035,
-0.08671463280916214,
-0.007859433069825172,
-0.06247827410697937,
0.06643325090408325,
-0.0028732167556881905,
0.008355951867997646,
0.0077902222983539104,
-0.08790760487318039,
0.00007796175486873835,
0.039658185094594955,
-0.005538725759834051,
0.028816189616918564,
0.07687005400657654,
0.1590786874294281,
0.0050257486291229725,
-0.040063098073005676,
0.020998213440179825,
0.0369800366461277,
-0.025064906105399132,
-0.030501313507556915,
-0.040703631937503815,
0.001932712853886187,
-0.011475174687802792,
0.04837404564023018,
0.005753827281296253,
-0.0915185734629631,
0.04511874541640282,
0.016369616612792015,
0.017242632806301117,
-0.01565220020711422,
0.03305772319436073,
0.01107358280569315,
-0.0490931011736393,
-0.09064610302448273,
0.046845946460962296,
-0.06266150623559952,
-0.03857097029685974,
-0.0053880056366324425,
-0.07452666759490967,
0.03166085481643677,
0.04052817448973656,
-0.009493445046246052,
0.09542249888181686,
0.07094676047563553,
0.02084469236433506,
-0.04505003243684769,
-0.008939852006733418,
-0.07107050716876984,
-0.002170862862840295,
0.09615517407655716,
0.04714607074856758,
0.06456878781318665,
0.08136869966983795,
0.012402912601828575,
-0.0691535621881485,
0.09883274137973785,
-0.08014851808547974,
-0.039873138070106506,
-0.01640600897371769,
-0.09366676211357117,
-0.005637473426759243,
-0.07801783829927444,
0.02425016090273857,
-0.036468178033828735,
0.032227836549282074,
-0.029560966417193413,
0.005180337466299534,
0.05428819730877876,
0.0031783573795109987,
0.04168000444769859,
0.06820257753133774,
-0.04608627408742905,
0.05023745447397232,
-0.01756119169294834,
-0.02564217522740364,
-0.04766015335917473,
0.03723551332950592,
-0.060417670756578445,
0.010963601991534233,
-0.12669043242931366,
-0.03505537286400795,
0.022908464074134827,
-0.0324070118367672,
0.030158577486872673,
-0.020349854603409767,
-0.0053089819848537445,
-0.00696814339607954,
-0.05857124924659729,
0.015513393096625805,
-0.03306250274181366,
0.0064810654148459435,
0.07756757736206055,
-0.0031600461807101965,
0.031343184411525726,
-0.07396817207336426,
-3.871944311784224e-34,
0.043831754475831985,
0.08125932514667511,
0.07287229597568512,
0.035537686198949814,
0.04275178909301758,
-0.004295743070542812,
-0.03808550536632538,
0.04592148959636688,
0.027755387127399445,
0.013418862596154213,
-0.050757817924022675,
-0.06294289976358414,
0.04813259840011597,
0.031744375824928284,
0.011527090333402157,
0.05608949437737465,
-0.029174959287047386,
0.021385259926319122,
-0.045179080218076706,
0.05363287031650543,
-0.02471929043531418,
-0.01578022539615631,
-0.026401041075587273,
-0.02366134524345398,
-0.00819482933729887,
0.04989844560623169,
-0.014100230298936367,
-0.007534218952059746,
-0.08155450969934464,
-0.006367950700223446,
0.006025152280926704,
0.03643999993801117,
-0.03062378615140915,
0.0012121836189180613,
0.03736910596489906,
-0.11010213941335678,
0.0033114582765847445,
-0.06311637908220291,
0.010478394106030464,
-0.0670635923743248,
0.08521945774555206,
0.06934183835983276,
-0.0941942110657692,
0.05281863361597061,
-0.05251915752887726,
-0.001866225735284388,
0.024278972297906876,
0.0633946880698204,
-0.03543965145945549,
0.04204903542995453,
0.06554552167654037,
0.007629720028489828,
-0.030554836615920067,
-0.0483611635863781,
-0.11795970797538757,
-0.014343521557748318,
-0.009500262327492237,
-0.04919208958745003,
-0.0014957848470658064,
-0.04106779769062996,
-0.01997470296919346,
0.033204805105924606,
-0.028666753321886063,
0.00573792215436697,
-0.00650360994040966,
-0.03352298587560654,
0.024461466819047928,
0.06507989019155502,
0.026598656550049782,
0.08654369413852692,
0.054566267877817154,
0.0001953620812855661,
-0.00982440635561943,
-0.008928600698709488,
-0.038389451801776886,
0.032624054700136185,
-0.03422802314162254,
-0.08468391001224518,
0.024379415437579155,
0.08383680135011673,
-0.07733573019504547,
0.08384345471858978,
0.007271069101989269,
-0.04986628517508507,
-0.08624691516160965,
-0.04831061139702797,
0.05247410759329796,
-0.09987739473581314,
0.03021814674139023,
-0.0034238609950989485,
-0.03786410763859749,
-0.07902713119983673,
-0.06397698819637299,
-0.0008031659526750445,
-0.02240641601383686,
-1.325417405421391e-34,
-0.001908698701299727,
-0.02697055973112583,
-0.021752791479229927,
-0.017740927636623383,
-0.025656580924987793,
-0.060619644820690155,
-0.028256922960281372,
-0.008085737936198711,
-0.016399366781115532,
-0.037698689848184586,
-0.04640030860900879,
-0.02247043512761593,
0.004632234573364258,
-0.021939929574728012,
0.08218642324209213,
0.060584522783756256,
-0.06083477661013603,
0.013952214270830154,
-0.07656224817037582,
0.008480090647935867,
0.08174221217632294,
0.03692363202571869,
-0.021079685539007187,
0.04624832421541214,
-0.0701269879937172,
0.09342097491025925,
0.004710136912763119,
0.021703707054257393,
0.006999780423939228,
-0.009923812933266163,
-0.03554175794124603,
-0.01729515939950943,
0.0294187031686306,
-0.018527166917920113,
-0.054684828966856,
0.040636658668518066,
-0.06779501587152481,
0.023126060143113136,
-0.012758252210915089,
0.04558517783880234,
0.010273637250065804,
-0.07621848583221436,
-0.05447673052549362,
0.04567285254597664,
0.008873463608324528,
-0.019060691818594933,
0.05641336366534233,
0.060933295637369156,
0.04659939557313919,
0.12016044557094574,
-0.09728914499282837,
0.015637526288628578,
-0.04148688539862633,
0.0033499267883598804,
-0.030194343999028206,
0.016639264300465584,
-0.11849494278430939,
0.001995048951357603,
0.10860302299261093,
-0.02740008942782879,
-0.01969706080853939,
0.03267355263233185,
-0.03438638150691986,
0.07642111927270889,
0.006804751697927713,
-0.03354713320732117,
-0.02207832597196102,
-0.05534449219703674,
0.06566902250051498,
0.046956367790699005,
0.02537199854850769,
-0.049517951905727386,
0.0049842363223433495,
-0.004157637245953083,
-0.02350003458559513,
-0.04812061786651611,
0.08457322418689728,
-0.0025735197123140097,
0.021496422588825226,
0.004370148293673992,
0.002771043684333563,
-0.04021555930376053,
0.015429491177201271,
0.16741372644901276,
0.09954196214675903,
0.09571707993745804,
0.01914847269654274,
0.0273186806589365,
-0.004542357288300991,
0.05283474922180176,
-0.003252766327932477,
0.10797129571437836,
-0.028915563598275185,
0.09034799784421921,
-0.009106130339205265,
-3.8449325501233034e-8,
-0.05854747071862221,
0.0599827878177166,
-0.0018348623998463154,
0.0350453145802021,
-0.09611830860376358,
-0.05279441177845001,
0.01147028524428606,
0.10088040679693222,
-0.07833363115787506,
0.011727451346814632,
-0.02546086721122265,
-0.08217687904834747,
-0.00834641233086586,
0.010074153542518616,
0.03125734254717827,
0.06266756355762482,
0.009954098612070084,
0.0652894452214241,
-0.047038331627845764,
0.041667889803647995,
-0.02888234704732895,
0.005469098687171936,
0.04742364585399628,
-0.029816124588251114,
-0.00631761085242033,
-0.09365931898355484,
-0.038987308740615845,
0.08756768703460693,
0.06290213018655777,
-0.03745793551206589,
-0.04526177793741226,
0.05773436278104782,
0.013138104230165482,
0.003059178125113249,
0.022296879440546036,
-0.0034747724421322346,
0.005032723769545555,
-0.01818527653813362,
0.04072568565607071,
-0.025984706357121468,
0.03936963528394699,
0.08342228084802628,
-0.006030331831425428,
0.029976025223731995,
0.01246377732604742,
0.09385410696268082,
0.08598247915506363,
-0.05790974199771881,
-0.0002873122866731137,
-0.01703072153031826,
-0.037359606474637985,
0.04850656911730766,
-0.02120283991098404,
0.07040334492921829,
0.01969713158905506,
0.03540048748254776,
0.04626746475696564,
-0.048264164477586746,
0.09256550669670105,
0.060566890984773636,
-0.013169209472835064,
-0.011138556525111198,
-0.1002107784152031,
-0.04011473432183266
] |
SuperCorrect: Advancing Small LLM Reasoning with Thought Template Distillation and Self-Correction | https://openreview.net/forum?id=PyjZO7oSw2 | [
"Ling Yang",
"Zhaochen Yu",
"Tianjun Zhang",
"Minkai Xu",
"Joseph E. Gonzalez",
"Bin CUI",
"Shuicheng YAN"
] | Poster | Large language models (LLMs) like GPT-4, DeepSeek-R1, and ReasonFlux have shown significant improvements in various reasoning tasks. However, smaller LLMs still struggle with complex mathematical reasoning because they fail to effectively identify and correct reasoning errors. Recent reflection-based methods aim to address these issues by enabling self-reflection and self-correction, but they still face challenges in independently detecting errors in their reasoning steps. To overcome these limitations, we propose SuperCorrect, a novel two-stage framework that uses a large teacher model to supervise and correct both the reasoning and reflection processes of a smaller student model. In the first stage, we extract hierarchical high-level and detailed thought templates from the teacher model to guide the student model in eliciting more fine-grained reasoning thoughts. In the second stage, we introduce cross-model collaborative direct preference optimization (DPO) to enhance the self-correction abilities of the student model by following the teacher's correction traces during training. This cross-model DPO approach teaches the student model to effectively locate and resolve erroneous thoughts with error-driven insights from the teacher model, breaking the bottleneck of its thoughts and acquiring new skills and knowledge to tackle challenging problems. Extensive experiments consistently demonstrate our superiority over previous methods. Notably, our SuperCorrect-7B model significantly surpasses powerful DeepSeekMath-7B by 7.8\%/5.3\% and Qwen2.5-Math-7B by 15.1\%/6.3\% on MATH/GSM8K benchmarks, achieving new SOTA performance among all 7B models. Code is available at: https://github.com/YangLing0818/SuperCorrect-llm | Large Language Models, LLM Reasoning | null | 397 | 2410.09008 | [
0.0064413296058773994,
-0.07257843017578125,
0.07904423773288727,
0.006545427720993757,
0.014750659465789795,
-0.021184904500842094,
-0.01863030157983303,
0.10876068472862244,
-0.044060491025447845,
0.0397314615547657,
-0.04374638572335243,
0.0306285098195076,
-0.018688969314098358,
0.09733524173498154,
0.023856202140450478,
0.027850870043039322,
0.11039246618747711,
0.03834686800837517,
-0.08942192792892456,
-0.04458405822515488,
-0.000601745443418622,
-0.031058670952916145,
-0.022881075739860535,
-0.015186171978712082,
-0.047054145485162735,
0.02393917180597782,
0.029031801968812943,
-0.033197738230228424,
-0.009194890037178993,
0.05275730788707733,
0.027571281418204308,
0.027468405663967133,
-0.05855345353484154,
-0.002998339245095849,
-0.03311353176832199,
0.08466342836618423,
-0.0586976520717144,
0.06891602277755737,
0.0059864334762096405,
-0.05626522749662399,
-0.09948039799928665,
-0.02054375596344471,
-0.008051734417676926,
0.00884074904024601,
0.016799280419945717,
-0.036584850400686264,
-0.02222350426018238,
0.018149275332689285,
0.02306530438363552,
0.02037438005208969,
-0.048888567835092545,
-0.0007152038160711527,
-0.05424853786826134,
-0.02453802153468132,
0.07384567707777023,
0.07936698198318481,
0.043074410408735275,
0.024199314415454865,
-0.037738386541604996,
-0.04816928133368492,
-0.08818793296813965,
-0.08993514627218246,
-0.05628843605518341,
-0.023393943905830383,
-0.03575846180319786,
0.04164658486843109,
-0.020398462191224098,
0.10047156363725662,
-0.03592803329229355,
0.08551578223705292,
0.015496980398893356,
0.11346200853586197,
-0.024542219936847687,
0.017818208783864975,
0.003789737122133374,
0.04786684736609459,
0.0006577651365660131,
-0.018602395430207253,
0.04436146095395088,
-0.02885029837489128,
-0.049831341952085495,
0.06220921501517296,
-0.008902071043848991,
0.004717949312180281,
0.1500317007303238,
-0.016932018101215363,
0.04493144527077675,
0.022538883611559868,
0.03957926109433174,
0.017508165910840034,
0.04008918255567551,
-0.04434312507510185,
0.00689192209392786,
0.047882355749607086,
0.015777627006173134,
0.09600286185741425,
-0.008075205609202385,
-0.022362539544701576,
-0.05702037364244461,
0.05603545531630516,
0.028428496792912483,
0.10225224494934082,
0.010921264067292213,
-0.05065060779452324,
-0.01906088925898075,
-0.03601039946079254,
0.06535828113555908,
0.09931128472089767,
0.0798342227935791,
-0.0969214141368866,
0.025002984330058098,
-0.007047710940241814,
-0.03520283102989197,
0.03943025320768356,
-0.05672001838684082,
-0.044727616012096405,
0.103724904358387,
-0.06909602880477905,
0.004325535614043474,
-0.022999228909611702,
0.001857968047261238,
0.04034150391817093,
0.07150290906429291,
0.023104721680283546,
0.05037667229771614,
-0.03422015905380249,
-0.05290167033672333,
9.3669864275999e-34,
0.05488613620400429,
0.12619663774967194,
0.013483700342476368,
-0.008405884727835655,
0.009297286160290241,
-0.025351636111736298,
0.002347060479223728,
0.059818025678396225,
-0.05099660903215408,
0.02037014625966549,
0.03065476194024086,
-0.014294041320681572,
-0.049808189272880554,
0.042463503777980804,
0.03785387799143791,
0.010932260192930698,
-0.0811714306473732,
-0.013820242136716843,
-0.08484957367181778,
-0.04343559965491295,
0.03649871423840523,
-0.0006421746220439672,
0.0049339886754751205,
-0.04273953661322594,
-0.005421831738203764,
-0.007349688559770584,
0.06749843060970306,
-0.013412346132099628,
0.028137896209955215,
0.007500836160033941,
-0.04719903692603111,
-0.08045807480812073,
-0.02288486622273922,
0.060005489736795425,
-0.042574722319841385,
-0.09708426892757416,
0.005013675428926945,
-0.06671717017889023,
0.04909365996718407,
-0.04995446652173996,
0.015941036865115166,
0.00020324958313722163,
0.043097157031297684,
0.0012927462812513113,
-0.025030603632330894,
-0.02864484116435051,
-0.004685757681727409,
0.08147650957107544,
-0.012396388687193394,
-0.04790549725294113,
0.039792001247406006,
0.03091074712574482,
-0.06413667649030685,
-0.023952407762408257,
-0.09668339788913727,
-0.0030379951931536198,
0.0007740341243334115,
0.037346839904785156,
0.08319174498319626,
-0.04571874067187309,
0.060355957597494125,
0.057123132050037384,
-0.004924791399389505,
0.06697814911603928,
0.02407338283956051,
0.017174210399389267,
-0.08063553273677826,
0.014396942220628262,
0.12954497337341309,
-0.03000013716518879,
-0.07724584639072418,
0.027588658034801483,
0.03331707417964935,
-0.0345691442489624,
0.010638786479830742,
-0.035162653774023056,
-0.012844939716160297,
-0.086509570479393,
0.04980824887752533,
-0.04291956499218941,
0.05749937891960144,
-0.007461118046194315,
-0.05320756882429123,
-0.03830810263752937,
-0.09286122024059296,
-0.04333527758717537,
0.020194729790091515,
-0.07545105367898941,
-0.018927736207842827,
-0.05507007986307144,
-0.011586270295083523,
-0.011219313368201256,
-0.01840679533779621,
-0.04926241189241409,
-0.028707049787044525,
-8.690351675457437e-34,
0.0071479869075119495,
-0.009943801909685135,
0.04527168720960617,
0.07663644105195999,
0.03147975727915764,
-0.07848141342401505,
0.03340497612953186,
0.008821874856948853,
0.013131318613886833,
-0.08189720660448074,
-0.0682353526353836,
-0.03925266116857529,
0.003967905417084694,
0.013914517126977444,
0.00712635787203908,
-0.005540454760193825,
-0.04944450780749321,
-0.026684653013944626,
0.007813182659447193,
0.0006122953491285443,
0.008651434443891048,
0.06463745981454849,
-0.08006313443183899,
0.0708024874329567,
-0.04345650598406792,
0.024653321132063866,
0.023913254961371422,
0.028990911319851875,
0.016584182158112526,
0.04923439770936966,
-0.014560825191438198,
-0.007709802128374577,
0.02903367206454277,
-0.04801788553595543,
-0.07614532113075256,
0.0631353035569191,
-0.050038356333971024,
-0.0031548093538731337,
-0.10072919726371765,
0.0866401270031929,
0.014245095662772655,
-0.09613312780857086,
-0.0367535762488842,
-0.006995349656790495,
-0.040833842009305954,
0.054876357316970825,
-0.07204407453536987,
0.037376150488853455,
0.020032601431012154,
0.009448463097214699,
-0.01211249828338623,
-0.0644603818655014,
-0.018173325806856155,
0.028882211074233055,
0.02550683170557022,
-0.04294312000274658,
0.058066461235284805,
-0.06904122233390808,
0.04071339964866638,
0.06934934109449387,
-0.0044730124063789845,
-0.019056955352425575,
0.10970783233642578,
-0.08021620661020279,
0.0030267229303717613,
-0.06801671534776688,
-0.05408288165926933,
-0.013915007933974266,
-0.02423004060983658,
0.012570131570100784,
-0.0066330027766525745,
0.004678410477936268,
-0.021038848906755447,
-0.07980670034885406,
0.07760341465473175,
0.00918686855584383,
-0.042053863406181335,
-0.012879925779998302,
-0.04667174071073532,
-0.059117745608091354,
0.0016950569115579128,
-0.02738780900835991,
0.025463363155722618,
0.12147597968578339,
-0.007884048856794834,
-0.0432344451546669,
0.05326361209154129,
0.15534411370754242,
-0.008067464455962181,
-0.004287415649741888,
-0.07392290979623795,
0.07577425241470337,
0.08685217797756195,
0.018591228872537613,
-0.10068632662296295,
-4.52728308175665e-8,
-0.05274135619401932,
0.033358436077833176,
0.01534715760499239,
0.1589980572462082,
-0.0029067613650113344,
-0.09744013100862503,
0.014534599147737026,
0.04005502164363861,
-0.0368855819106102,
0.01548411138355732,
-0.006859352812170982,
-0.029450975358486176,
-0.07790648192167282,
-0.06371097266674042,
0.027654558420181274,
0.046734411269426346,
0.017758013680577278,
-0.0023311034310609102,
-0.03867373615503311,
-0.04422670602798462,
0.04304782301187515,
-0.018945422023534775,
-0.03263523057103157,
0.024424515664577484,
0.06952156871557236,
-0.029879577457904816,
-0.02152782492339611,
0.15410451591014862,
0.07689561694860458,
-0.008684718050062656,
0.0014886180870234966,
0.005237181670963764,
0.040278781205415726,
-0.024453388527035713,
0.051552578806877136,
0.046941127628088,
0.030168652534484863,
0.05536084994673729,
0.06054750084877014,
-0.04454653337597847,
0.04840609058737755,
0.033963870257139206,
-0.022568603977560997,
0.07312510162591934,
0.035837944597005844,
-0.042237360030412674,
-0.08029603958129883,
-0.07021468877792358,
-0.0007199090905487537,
0.043584052473306656,
-0.011419190093874931,
-0.001644114381633699,
-0.056820038706064224,
0.04804471507668495,
0.008346349000930786,
0.03751254826784134,
-0.03400089964270592,
0.006804975681006908,
-0.04548270255327225,
-0.029562335461378098,
0.09240886569023132,
0.08708608895540237,
-0.0855196937918663,
0.08008873462677002
] |
Beyond Sequence: Impact of Geometric Context for RNA Property Prediction | https://openreview.net/forum?id=9htTvHkUhh | [
"Junjie Xu",
"Artem Moskalev",
"Tommaso Mansi",
"Mangal Prakash",
"Rui Liao"
] | Poster | Accurate prediction of RNA properties, such as stability and interactions, is crucial for advancing our understanding of biological processes and developing RNA-based therapeutics. RNA structures can be represented as 1D sequences, 2D topological graphs, or 3D all-atom models, each offering different insights into its function. Existing works predominantly focus on 1D sequence-based models, which overlook the geometric context provided by 2D and 3D geometries. This study presents the first systematic evaluation of incorporating explicit 2D and 3D geometric information into RNA property prediction, considering not only performance but also real-world challenges such as limited data availability, partial labeling, sequencing noise, and computational efficiency. To this end, we introduce a newly curated set of RNA datasets with enhanced 2D and 3D structural annotations, providing a resource for model evaluation on RNA data. Our findings reveal that models with explicit geometry encoding generally outperform sequence-based models, with an average prediction RMSE reduction of around 12% across all various RNA tasks and excelling in low-data and partial labeling regimes, underscoring the value of explicitly incorporating geometric context. On the other hand, geometry-unaware sequence-based models are more robust under sequencing noise but often require around 2-5x training data to match the performance of geometry-aware models. Our study offers further insights into the trade-offs between different RNA representations in practical applications and addresses a significant gap in evaluating deep learning models for RNA tasks. | Geometric deep learning, Graph Neural Networks, GNNs, RNA property prediction, Datasets and Benchmarks | null | 395 | 2410.11933 | [
-0.03993610665202141,
-0.07621803879737854,
-0.015275814570486546,
0.0030303001403808594,
-0.008771649561822414,
-0.01451341062784195,
-0.10704445838928223,
0.008547798730432987,
0.021338820457458496,
-0.05027240514755249,
-0.0575917549431324,
-0.03880741819739342,
-0.013426271267235279,
0.14601701498031616,
-0.00685845036059618,
0.06210091710090637,
-0.06101664900779724,
-0.0007326615741476417,
0.006543783936649561,
0.02611495740711689,
0.00828167051076889,
0.10952437669038773,
0.06757527589797974,
0.03053794801235199,
0.04856717959046364,
-0.004962901584804058,
0.10459466278553009,
-0.01462512370198965,
-0.012479209341108799,
-0.038294047117233276,
0.09429890662431717,
-0.01982845366001129,
0.0347885936498642,
-0.037518396973609924,
-0.018850458785891533,
-0.0645812600851059,
-0.03330458328127861,
0.057341404259204865,
0.012863942421972752,
0.019104953855276108,
0.09285076707601547,
0.07036557048559189,
0.043596554547548294,
0.056601446121931076,
0.004022729583084583,
0.010855968110263348,
-0.025812242180109024,
-0.06778144836425781,
0.02718399651348591,
0.006036571692675352,
-0.030366746708750725,
-0.056103501468896866,
-0.07103759050369263,
-0.009836582466959953,
0.047267407178878784,
0.0653214156627655,
0.03900017589330673,
-0.048123907297849655,
0.015612619929015636,
-0.06636201590299606,
0.1230040043592453,
-0.10625964403152466,
-0.010534094646573067,
-0.015014505945146084,
0.033671315759420395,
0.0010988005669787526,
0.03746731951832771,
0.028225574642419815,
-0.03400712087750435,
0.025819463655352592,
-0.02895977348089218,
0.07955619692802429,
-0.06335387378931046,
0.061144642531871796,
0.019075797870755196,
0.024905797094106674,
0.03257688507437706,
0.052165672183036804,
-0.00022911130508873612,
-0.04439665004611015,
-0.04937583580613136,
0.039933618158102036,
0.08630889654159546,
0.021170003339648247,
-0.0322071835398674,
0.10866578668355942,
0.023330440744757652,
0.08161848783493042,
-0.047455545514822006,
-0.008719280362129211,
0.030700895935297012,
0.018650582060217857,
-0.01435762643814087,
-0.027280734851956367,
-0.029557326808571815,
0.057729870080947876,
0.056552592664957047,
-0.07419371604919434,
-0.016626935452222824,
0.014970093034207821,
0.05847641080617905,
0.02918754518032074,
0.0036696859169751406,
-0.013975664973258972,
0.07638668268918991,
-0.02738988772034645,
0.07821065932512283,
-0.006404784508049488,
-0.013231529854238033,
0.027801888063549995,
0.04644130542874336,
0.022014938294887543,
0.014554671943187714,
0.027826333418488503,
-0.010592144913971424,
-0.04250049591064453,
0.0073170107789337635,
-0.02865193411707878,
0.1009158343076706,
0.09061022102832794,
0.021643782034516335,
-0.03188643231987953,
-0.00026651800726540387,
0.010704405605793,
-0.019139939919114113,
-0.006181926466524601,
-0.08074277639389038,
2.1014220674153876e-33,
0.024299154058098793,
0.03357163816690445,
0.042057741433382034,
0.04651160538196564,
0.07611344009637833,
-0.055019181221723557,
-0.050901323556900024,
-0.06180078163743019,
-0.026631474494934082,
0.040578216314315796,
-0.110951267182827,
-0.04560091719031334,
-0.010060308501124382,
0.03342784196138382,
-0.014766001142561436,
0.03394726291298866,
-0.00018207679386250675,
0.04824758693575859,
-0.10547976940870285,
0.02410154789686203,
0.013189100660383701,
-0.02540881745517254,
-0.025136802345514297,
0.008128211833536625,
-0.009776921011507511,
0.005133878439664841,
-0.023668834939599037,
0.009210029616951942,
-0.0060155210085213184,
-0.007065217476338148,
-0.05090051144361496,
0.06121810898184776,
-0.010660320520401001,
0.027511853724718094,
0.046809207648038864,
-0.010499485768377781,
-0.029749596491456032,
-0.05330389738082886,
-0.016743548214435577,
0.002229988342151046,
0.04364932328462601,
0.05222150683403015,
-0.008088833652436733,
0.004829023499041796,
-0.03503968194127083,
-0.028932493180036545,
-0.00890897773206234,
0.019016966223716736,
-0.09251990169286728,
-0.0301109217107296,
0.09926717728376389,
0.011344389989972115,
-0.005662344396114349,
-0.021430930122733116,
-0.0385165736079216,
-0.030444329604506493,
-0.05629970133304596,
-0.0499662309885025,
0.025115344673395157,
0.055454276502132416,
-0.025044506415724754,
0.003415709361433983,
0.034499477595090866,
-0.0033579031005501747,
-0.05524562671780586,
0.021805347874760628,
-0.0730106383562088,
-0.04205829277634621,
-0.0002753993321675807,
0.09305793792009354,
-0.009976858273148537,
0.038659706711769104,
0.01750818081200123,
-0.027233609929680824,
-0.0036174010019749403,
-0.02180282026529312,
0.022859975695610046,
-0.10871647298336029,
-0.0009130784310400486,
0.1382603645324707,
-0.032583437860012054,
-0.034381698817014694,
-0.030091602355241776,
-0.013517598621547222,
-0.06934751570224762,
-0.13479144871234894,
0.09946063160896301,
-0.0793057307600975,
-0.06032125651836395,
-0.0036473157815635204,
-0.032889097929000854,
-0.0453207828104496,
-0.04342096298933029,
-0.054188281297683716,
0.01585950329899788,
-3.3194825380488886e-33,
-0.009200082160532475,
0.017270904034376144,
0.021650632843375206,
-0.0411963015794754,
-0.06117886304855347,
0.008936724625527859,
-0.07895777374505997,
0.0012228254927322268,
0.02491561695933342,
-0.06121934950351715,
0.0431908518075943,
0.045103322714567184,
0.04801268130540848,
-0.0650918260216713,
0.07272915542125702,
0.014839736744761467,
-0.030567865818738937,
-0.0651366338133812,
-0.01896503195166588,
-0.003908733371645212,
0.06552386283874512,
-0.014648967422544956,
-0.08874741941690445,
0.010943870060145855,
-0.033934336155653,
0.013890962116420269,
-0.013757451437413692,
-0.002311118645593524,
-0.020454179495573044,
0.013271803967654705,
-0.0759853944182396,
-0.016385817900300026,
0.007025995757430792,
-0.01913844607770443,
-0.13399682939052582,
-0.007012384012341499,
0.050950922071933746,
-0.040021803230047226,
0.03526437282562256,
-0.04587234929203987,
-0.027469826862215996,
0.013286125846207142,
-0.09776962548494339,
0.0931992381811142,
0.014537816867232323,
-0.002651312155649066,
0.021289627999067307,
0.008521716110408306,
0.029644053429365158,
0.023118050768971443,
-0.05876799672842026,
0.08396247029304504,
0.019040392711758614,
0.03771002218127251,
0.020732160657644272,
-0.1112748235464096,
-0.11031799763441086,
-0.03553488850593567,
0.07740923762321472,
-0.02671266533434391,
-0.09360866248607635,
-0.0542699433863163,
-0.05373760685324669,
0.07428735494613647,
-0.0276994276791811,
-0.02845923788845539,
0.03830805420875549,
-0.043659232556819916,
-0.013936595991253853,
-0.011186446994543076,
-0.01126549020409584,
-0.02115379460155964,
0.024254700168967247,
0.015275673009455204,
0.015762558206915855,
-0.07266397029161453,
0.027636680752038956,
-0.06312780827283859,
-0.05451218783855438,
0.007683774456381798,
-0.012873688712716103,
0.033385004848241806,
0.0033297166228294373,
0.06824366748332977,
0.12992212176322937,
0.007252242881804705,
0.074849933385849,
0.10048343986272812,
-0.0017625211039558053,
0.03640948608517647,
-0.004751631524413824,
0.03901471197605133,
-0.0812387615442276,
0.07895370572805405,
0.02776525914669037,
-3.915955204547572e-8,
-0.03200593590736389,
-0.00561040174216032,
0.05064552649855614,
-0.045729026198387146,
-0.028021525591611862,
0.07224135100841522,
-0.006851258687674999,
0.15436294674873352,
-0.002844970440492034,
0.016053099185228348,
0.032485056668519974,
-0.01888934150338173,
-0.10720381140708923,
-0.039913371205329895,
0.060112159699201584,
0.13425573706626892,
-0.020648658275604248,
-0.025287756696343422,
-0.024110665544867516,
0.025012953206896782,
-0.11186112463474274,
0.08231571316719055,
0.0024749187286943197,
-0.0367494635283947,
-0.008218175731599331,
-0.007342880591750145,
-0.010649113915860653,
0.06343399733304977,
0.06781374663114548,
-0.08467649668455124,
0.039972588419914246,
-0.010771294124424458,
0.10541178286075592,
0.004520705435425043,
-0.1415916383266449,
0.01947125792503357,
0.09243151545524597,
0.028592785820364952,
-0.019431551918387413,
-0.007197549566626549,
-0.011589973233640194,
0.0849921777844429,
-0.03497444465756416,
-0.027717871591448784,
0.006954070646315813,
0.05222918838262558,
0.050681132823228836,
0.055326782166957855,
0.044051021337509155,
-0.02249429188668728,
-0.026467792689800262,
0.023921897634863853,
-0.008949311450123787,
0.015213951468467712,
0.006560290697962046,
0.10992728918790817,
-0.10799234360456467,
-0.01817162148654461,
0.05549449473619461,
0.023774314671754837,
-0.04959811642765999,
-0.09298989921808243,
-0.0293025691062212,
-0.05641280859708786
] |
FaceShot: Bring Any Character into Life | https://openreview.net/forum?id=oJA1GUqRww | [
"Junyao Gao",
"Yanan SUN",
"Fei Shen",
"Xin Jiang",
"Zhening Xing",
"Kai Chen",
"Cairong Zhao"
] | Poster | In this paper, we present ***FaceShot***, a novel training-free portrait animation framework designed to bring any character into life from any driven video without fine-tuning or retraining.
We achieve this by offering precise and robust reposed landmark sequences from an appearance-guided landmark matching module and a coordinate-based landmark retargeting module.
Together, these components harness the robust semantic correspondences of latent diffusion models to produce facial motion sequence across a wide range of character types.
After that, we input the landmark sequences into a pre-trained landmark-driven animation model to generate animated video.
With this powerful generalization capability, FaceShot can significantly extend the application of portrait animation by breaking the limitation of realistic portrait landmark detection for any stylized character and driven video.
Also, FaceShot is compatible with any landmark-driven animation model, significantly improving overall performance.
Extensive experiments on our newly constructed character benchmark CharacBench confirm that FaceShot consistently surpasses state-of-the-art (SOTA) approaches across any character domain.
More results are available at our project website https://faceshot2024.github.io/faceshot/. | portrait animation, diffusion model | null | 387 | 2503.00740 | [
-0.05463531240820885,
-0.05845758691430092,
0.01661226712167263,
-0.03797857463359833,
0.08300091326236725,
0.06922417879104614,
-0.08065561950206757,
-0.06955686956644058,
0.007277776952832937,
-0.07460562884807587,
0.02911136858165264,
-0.10374541580677032,
-0.0023768325336277485,
0.050185322761535645,
0.033691566437482834,
-0.09143712371587753,
0.05289459973573685,
0.14639149606227875,
-0.0016676776576787233,
0.027339056134223938,
-0.0416124165058136,
-0.06547179073095322,
0.04743781313300133,
-0.005731926299631596,
0.05377979576587677,
-0.001767137204296887,
0.044298410415649414,
-0.014901178888976574,
0.02900964766740799,
-0.07355178892612457,
0.09723790735006332,
-0.05314309522509575,
-0.01640092395246029,
-0.0115306805819273,
-0.08312579244375229,
0.08397521078586578,
-0.012730196118354797,
0.036034174263477325,
-0.07199940830469131,
-0.04030060023069382,
-0.03441270440816879,
0.08493787050247192,
0.01480125728994608,
0.04226265847682953,
0.1105211079120636,
0.0015395312802866101,
-0.003817350836470723,
-0.0613013356924057,
-0.006456080358475447,
-0.022307489067316055,
-0.08108583837747574,
-0.07238215208053589,
0.006789389066398144,
0.02248375490307808,
-0.04863189160823822,
0.05174217373132706,
0.06995885074138641,
-0.06124935299158096,
0.03038915805518627,
-0.01320220809429884,
-0.08087074011564255,
-0.02795874886214733,
0.019237324595451355,
-0.052827633917331696,
-0.019521454349160194,
-0.07540429383516312,
0.02848639525473118,
-0.05299123749136925,
0.015379302203655243,
0.033638231456279755,
-0.03736387938261032,
-0.033182479441165924,
-0.04773952066898346,
0.007785734720528126,
-0.012142033316195011,
0.00794452615082264,
0.04396045580506325,
-0.025063302367925644,
0.028579439967870712,
-0.02942490391433239,
0.07726108282804489,
-0.05452018231153488,
0.08273007720708847,
0.04241475462913513,
0.011834805831313133,
-0.007608314044773579,
-0.05642462149262428,
0.012317878194153309,
-0.07862231880426407,
0.05995551496744156,
-0.04114392772316933,
0.016018429771065712,
0.06684199720621109,
-0.09008152782917023,
0.057733695954084396,
-0.010627028532326221,
-0.011116931214928627,
-0.07633532583713531,
-0.008954526856541634,
0.07107964903116226,
-0.030445940792560577,
-0.054454557597637177,
-0.01560253370553255,
0.0318707600235939,
0.08689220249652863,
0.0001360975147690624,
0.052870117127895355,
-0.0087098004296422,
0.019194547086954117,
-0.07451452314853668,
0.04404037445783615,
-0.04240351915359497,
-0.03384747356176376,
0.004189903847873211,
0.09273974597454071,
0.016635295003652573,
-0.05339575558900833,
-0.06438684463500977,
-0.009235749021172523,
-0.028912117704749107,
0.031295295804739,
0.02889299765229225,
-0.03566313534975052,
-0.023977188393473625,
0.05320893973112106,
-0.02580973319709301,
-0.0157616026699543,
1.0139892330364786e-33,
0.056334927678108215,
0.09913540631532669,
0.0509699285030365,
0.08734769374132156,
0.022649269551038742,
-0.019205467775464058,
-0.03161733224987984,
0.0003280417004134506,
-0.039660852402448654,
-0.03768990561366081,
-0.021115833893418312,
-0.04249948635697365,
-0.03657127544283867,
0.0967126414179802,
-0.0184226855635643,
-0.007131209596991539,
-0.032211754471063614,
-0.005100567359477282,
0.031135596334934235,
0.12443918734788895,
0.054809197783470154,
0.008648686110973358,
-0.010453244671225548,
-0.017674613744020462,
-0.04982386529445648,
-0.0014645399060100317,
0.027264293283224106,
-0.06499578803777695,
-0.02642638608813286,
-0.005409237928688526,
-0.11815100908279419,
0.02801893651485443,
-0.01840398460626602,
-0.03086785599589348,
0.08219033479690552,
-0.07319267839193344,
-0.03885918855667114,
-0.08316726982593536,
0.038813382387161255,
0.018339892849326134,
-0.01662103831768036,
0.054403144866228104,
-0.08766001462936401,
-0.03416666015982628,
-0.08474219590425491,
0.007616421673446894,
0.041992805898189545,
0.09569315612316132,
-0.029181811958551407,
0.061756011098623276,
0.07325492799282074,
0.07066987454891205,
-0.03253936767578125,
-0.03197897970676422,
-0.12691408395767212,
0.007121841423213482,
0.045257553458213806,
-0.02062506414949894,
0.000004954652013111627,
0.03094896860420704,
0.03046441450715065,
0.029194006696343422,
-0.0394178107380867,
0.012989887036383152,
-0.005558835342526436,
-0.07834222912788391,
0.0056549240835011005,
0.03372309356927872,
0.005827269516885281,
0.04882614314556122,
0.029243338853120804,
0.011978513561189175,
0.023832961916923523,
-0.05464961752295494,
0.020016074180603027,
-0.00943447183817625,
0.07901902496814728,
0.04004702717065811,
-0.03180352598428726,
0.06524921953678131,
-0.019311076030135155,
-0.0012606248492375016,
-0.03215201944112778,
-0.09141810983419418,
0.03502565994858742,
-0.08177898079156876,
0.0038832975551486015,
-0.10020424425601959,
-0.019836656749248505,
-0.04477126896381378,
0.03757282719016075,
-0.08291605114936829,
0.0010828174417838454,
0.06712520867586136,
-0.03182530403137207,
-2.906027976332829e-34,
0.0689312219619751,
0.006834842264652252,
-0.028589507564902306,
0.003637459594756365,
0.0011752620339393616,
-0.001839140779338777,
0.06106602028012276,
0.08904584497213364,
0.035025130957365036,
-0.0339001826941967,
-0.03106573037803173,
-0.03586367890238762,
-0.008626630529761314,
-0.10356011986732483,
0.05484888330101967,
-0.004365101456642151,
0.033236581832170486,
-0.023473504930734634,
-0.08983572572469711,
0.007276717573404312,
0.02252131886780262,
0.03851921111345291,
-0.06395764648914337,
0.01803356595337391,
-0.03289801999926567,
0.11029718816280365,
0.0007173906778916717,
0.039574749767780304,
0.03266913816332817,
-0.020522411912679672,
-0.02169020101428032,
-0.012879098765552044,
0.024972891435027122,
0.020860865712165833,
-0.07731641829013824,
0.042908839881420135,
-0.01361062005162239,
-0.01773342862725258,
-0.036887720227241516,
0.10440909117460251,
0.050598904490470886,
-0.034967731684446335,
-0.033963996917009354,
0.07235666364431381,
0.04195870831608772,
0.013823824934661388,
-0.09585271775722504,
0.024990763515233994,
-0.016895757988095284,
0.07929624617099762,
-0.06821669638156891,
0.05411071330308914,
-0.06174791604280472,
0.05089332163333893,
-0.0406382791697979,
-0.06299924105405807,
0.011485784314572811,
0.019996393471956253,
0.06361570954322815,
0.03300825133919716,
-0.013581863604485989,
-0.04026925936341286,
-0.04913463443517685,
-0.024448415264487267,
-0.03239931911230087,
0.005511376075446606,
-0.03743124008178711,
-0.04241442307829857,
0.03708803281188011,
-0.014372006058692932,
0.04633408784866333,
-0.006389364134520292,
0.0008145993924699724,
0.024368928745388985,
-0.04152712598443031,
-0.012206853367388248,
0.06314714252948761,
0.017514999955892563,
0.014285220764577389,
-0.1053374707698822,
-0.04883607476949692,
-0.04219065234065056,
0.029968498274683952,
0.014980509877204895,
0.10226699709892273,
0.13963189721107483,
-0.004140249919146299,
-0.0059640780091285706,
0.024178888648748398,
0.05483751371502876,
-0.061328161507844925,
0.05816278234124184,
-0.0016073758015409112,
0.01467418484389782,
-0.020406518131494522,
-3.556256444881001e-8,
-0.16860799491405487,
0.04259589686989784,
-0.02801329642534256,
0.016917575150728226,
-0.12493740022182465,
-0.038754966109991074,
0.0959593802690506,
0.008096951991319656,
0.03817116096615791,
0.0033761202357709408,
0.09658382833003998,
0.0025272262282669544,
0.028206458315253258,
-0.034297626465559006,
-0.019343694671988487,
0.05944090336561203,
-0.03452999144792557,
0.09101998060941696,
-0.04468240216374397,
0.00015910335059743375,
-0.04241110011935234,
0.06931048631668091,
-0.01813601702451706,
-0.04710804671049118,
-0.029643865302205086,
-0.03247781842947006,
-0.018166733905673027,
0.11443083733320236,
0.03773946687579155,
-0.04964375123381615,
-0.048985544592142105,
0.05394449457526207,
0.03384943678975105,
-0.08719354122877121,
0.04319065809249878,
0.03969758749008179,
0.005459983833134174,
-0.028141478076577187,
0.001157586695626378,
-0.012854151427745819,
-0.002051210729405284,
0.06433473527431488,
-0.009223702363669872,
0.030056102201342583,
0.004413689952343702,
0.07105280458927155,
0.07047624886035919,
-0.007565218955278397,
0.09243322908878326,
0.015195559710264206,
-0.005919318646192551,
-0.07316344231367111,
-0.03196918964385986,
0.07425747811794281,
0.05911102145910263,
0.07352019846439362,
0.010923285968601704,
0.011677718721330166,
0.08205809444189072,
0.05464400351047516,
0.048228420317173004,
-0.029947401955723763,
0.024724991992115974,
-0.000053644940635422245
] |
IgGM: A Generative Model for Functional Antibody and Nanobody Design | https://openreview.net/forum?id=zmmfsJpYcq | [
"Rubo Wang",
"Fandi Wu",
"Xingyu Gao",
"Jiaxiang Wu",
"Peilin Zhao",
"Jianhua Yao"
] | Poster | Immunoglobulins are crucial proteins produced by the immune system to identify and bind to foreign substances, playing an essential role in shielding organisms from infections and diseases. Designing specific antibodies opens new pathways for disease treatment. With the rise of deep learning, AI-driven drug design has become possible, leading to several methods for antibody design. However, many of these approaches require additional conditions that differ from real-world scenarios, making it challenging to incorporate them into existing antibody design processes. Here, we introduce IgGM, a generative model for the de novo design of immunoglobulins with functional specificity. IgGM simultaneously generates antibody sequences and structures for a given antigen, consisting of three core components: a pre-trained language model for extracting sequence features, a feature learning module for identifying pertinent features, and a prediction module that outputs designed antibody sequences and the predicted complete antibody-antigen complex structure. IgGM effectively predicts structures and designs novel antibodies and nanobodies. This makes it highly applicable in a wide range of practical situations related to antibody and nanobody design. Code is available at: https://github.com/TencentAI4S/IgGM. | de novo antibody design, complex structure prediction, protein design | Efficient and accurate design methods for antibody and nanobody sequences and structures tailored for real-world design scenarios. | 385 | null | [
-0.015539688989520073,
-0.0948290079832077,
0.020308319479227066,
0.005470216739922762,
-0.002772914944216609,
-0.07880640774965286,
0.04637061059474945,
0.04885931313037872,
-0.0360257625579834,
-0.031124338507652283,
-0.07121559232473373,
-0.0575188547372818,
-0.018202513456344604,
0.0566846989095211,
-0.017363641411066055,
0.09299390763044357,
0.013961334712803364,
0.04669132083654404,
0.004594577010720968,
-0.041617877781391144,
0.043198827654123306,
0.015726998448371887,
0.017550241202116013,
-0.017602862790226936,
-0.053543251007795334,
0.0291457436978817,
0.019885702058672905,
0.010514426045119762,
-0.018844958394765854,
-0.09759297221899033,
0.05991504713892937,
-0.051795825362205505,
-0.035368911921978,
-0.020715603604912758,
-0.05906466394662857,
0.01253336388617754,
-0.0688508003950119,
0.03700713813304901,
-0.05718885734677315,
-0.10843633115291595,
0.003400404006242752,
0.0030092005617916584,
0.029323937371373177,
-0.02215999737381935,
0.0740034282207489,
-0.032898012548685074,
-0.04966742545366287,
0.049925196915864944,
0.07901396602392197,
0.06192389503121376,
-0.07014285027980804,
-0.01796889118850231,
-0.04232969507575035,
0.09667553007602692,
0.03767729178071022,
0.006175812799483538,
-0.06617312133312225,
-0.04674164205789566,
-0.028990091755986214,
0.034288518130779266,
-0.015908200293779373,
-0.08521173149347305,
0.06477794796228409,
-0.03207094594836235,
0.0002632500836625695,
-0.010769866406917572,
0.01479586772620678,
0.07912130653858185,
0.0030841596890240908,
-0.017878852784633636,
-0.014193670824170113,
0.02413204126060009,
-0.05061713233590126,
0.07298559695482254,
0.07359039038419724,
0.08822319656610489,
0.08833790570497513,
0.08693999797105789,
0.10193775594234467,
-0.04389319568872452,
-0.013516980223357677,
0.08265470713376999,
0.06647609919309616,
0.03661532700061798,
0.07903589308261871,
0.0711861252784729,
-0.03392814099788666,
0.06664677709341049,
-0.031314801424741745,
-0.007674024440348148,
0.04565262794494629,
-0.034001946449279785,
0.028796697035431862,
-0.02095927484333515,
0.019202765077352524,
-0.0031893595587462187,
0.04703245684504509,
-0.06325824558734894,
0.008871947415173054,
0.05914603918790817,
-0.05532343313097954,
0.001148509792983532,
0.07957537472248077,
-0.06346166878938675,
0.05711694061756134,
-0.062215253710746765,
0.031299494206905365,
0.004111551214009523,
0.03864873945713043,
0.025477668270468712,
0.01884441077709198,
0.0224214605987072,
-0.0027101824525743723,
0.0020639647264033556,
0.05884213000535965,
0.04170035570859909,
-0.03604797273874283,
-0.011168702505528927,
0.07101950794458389,
0.08808265626430511,
-0.00842688512057066,
-0.016579147428274155,
-0.0692896619439125,
0.013362825848162174,
-0.056410446763038635,
0.02674686908721924,
-0.13550099730491638,
-1.9749087991419177e-34,
-0.01152096688747406,
0.02483658492565155,
0.06639754772186279,
0.1211462914943695,
0.03437110036611557,
-0.10222860425710678,
0.023719817399978638,
0.05796557664871216,
-0.05224914476275444,
0.07936803251504898,
-0.0750325545668602,
-0.05933230370283127,
-0.08120747655630112,
0.19158798456192017,
0.035648930817842484,
-0.005148989614099264,
-0.018979880958795547,
-0.05746384337544441,
-0.005623740144073963,
-0.019886519759893417,
-0.013773171231150627,
-0.011169686913490295,
-0.04187017306685448,
-0.02243022620677948,
0.03411009535193443,
0.08634080737829208,
-0.00011487225128803402,
0.030570106580853462,
-0.10930336266756058,
0.006023510824888945,
-0.11934109777212143,
0.03510766848921776,
-0.0362614206969738,
-0.020119979977607727,
0.057107340544462204,
-0.05754544958472252,
0.01472328882664442,
-0.04180380329489708,
0.029539400711655617,
0.000003496765430099913,
0.02607598900794983,
-0.016085993498563766,
-0.0388483889400959,
0.018420910462737083,
-0.0034973258152604103,
0.02891307882964611,
0.0030503214802592993,
-0.058188438415527344,
-0.026418011635541916,
-0.041420966386795044,
0.058044370263814926,
0.014958837069571018,
0.012625781819224358,
-0.08644071221351624,
-0.005277041345834732,
0.04792686179280281,
-0.0582418292760849,
0.017900293692946434,
0.07858815789222717,
0.05788750573992729,
-0.046627070754766464,
0.03433992341160774,
-0.020454736426472664,
0.0353422574698925,
-0.014409596100449562,
0.01670791581273079,
0.0003308023442514241,
0.03164038434624672,
-0.00724484957754612,
0.01161094568669796,
0.02892741933465004,
-0.007940859533846378,
0.013086867518723011,
-0.05143139511346817,
-0.004393980372697115,
0.005213662516325712,
0.05846499279141426,
-0.05442831665277481,
0.005169679876416922,
0.02650570124387741,
0.053804751485586166,
0.118215411901474,
-0.08521027863025665,
0.007739492692053318,
-0.03741033002734184,
-0.07964371889829636,
0.034424085170030594,
-0.049699291586875916,
-0.08241726458072662,
-0.030485613271594048,
-0.04652908816933632,
0.009875686839222908,
-0.0013600066304206848,
-0.0008074040524661541,
-0.06901364028453827,
-2.465122774624688e-33,
-0.01653529331088066,
0.02220011129975319,
-0.05253971368074417,
-0.030370276421308517,
-0.019248347729444504,
-0.004549110773950815,
-0.0013026302913203835,
0.004034503363072872,
0.06500079482793808,
0.044580161571502686,
0.0486004464328289,
-0.029508426785469055,
0.009512043558061123,
-0.02570435404777527,
-0.00392138073220849,
0.08289917558431625,
-0.10630741715431213,
-0.02473863586783409,
-0.014263569377362728,
0.04071781784296036,
-0.028153585270047188,
0.1409469097852707,
0.02416815608739853,
-0.030906101688742638,
-0.09613489359617233,
-0.017221994698047638,
-0.013265702873468399,
0.06035713478922844,
0.05485542491078377,
-0.02336248941719532,
-0.061769723892211914,
-0.041191037744283676,
-0.017502957955002785,
-0.012400148436427116,
-0.061922840774059296,
-0.03245401754975319,
0.02362970821559429,
-0.06544898450374603,
0.07172715663909912,
-0.07489759474992752,
0.027393914759159088,
-0.03766269236803055,
-0.09512270987033844,
0.026676425710320473,
0.035156961530447006,
0.02651900425553322,
0.010672450065612793,
0.030470451340079308,
0.07230968773365021,
0.02377476915717125,
-0.0227917842566967,
-0.025597183033823967,
-0.04607796296477318,
-0.05087859183549881,
-0.13401080667972565,
-0.06317922472953796,
0.008737035095691681,
-0.02173314429819584,
0.08010795712471008,
0.020431609824299812,
-0.0823320597410202,
-0.0726369321346283,
0.07799363136291504,
-0.05505043640732765,
-0.041576966643333435,
0.0026831047143787146,
-0.015120997093617916,
-0.008219894021749496,
-0.009981528855860233,
-0.056714750826358795,
-0.030435390770435333,
0.005180126987397671,
0.04935012385249138,
0.032340165227651596,
0.031138569116592407,
-0.05808848515152931,
-0.03535503149032593,
-0.05282612144947052,
-0.04114016890525818,
-0.015562444925308228,
-0.01983269304037094,
-0.0024711908772587776,
-0.042302362620830536,
0.08073023706674576,
0.11508380621671677,
0.033000729978084564,
0.009198387153446674,
0.06993481516838074,
0.07416614890098572,
0.010630298405885696,
-0.018093859776854515,
0.09070220589637756,
-0.023093175143003464,
0.06050490215420723,
-0.08089015632867813,
-4.10626483926535e-8,
0.07842953503131866,
-0.04011494293808937,
0.05429963022470474,
-0.03665607422590256,
0.03290727362036705,
0.03232530504465103,
-0.07189756631851196,
0.013416357338428497,
-0.013879588805139065,
0.017672495916485786,
0.05546735227108002,
0.061316948384046555,
-0.061672575771808624,
-0.007910944521427155,
-0.005092752166092396,
0.04311036318540573,
0.017324173822999,
0.005875756498426199,
-0.0179152749478817,
-0.07132978737354279,
0.011443275026977062,
0.07670213282108307,
0.006590079516172409,
0.0094391955062747,
0.10309509932994843,
-0.12023486942052841,
-0.023486007004976273,
-0.06458225101232529,
-0.02485915645956993,
0.07914330065250397,
-0.07999303936958313,
0.050560299307107925,
-0.03990792855620384,
-0.007108448538929224,
-0.03634337708353996,
0.01660657301545143,
0.047805409878492355,
-0.04506721720099449,
-0.007551664486527443,
-0.06170888617634773,
0.05009313300251961,
0.0524466410279274,
-0.03106616996228695,
-0.06116288900375366,
-0.07455166429281235,
-0.027013810351490974,
-0.009552199393510818,
-0.08037324994802475,
0.058549921959638596,
0.033015988767147064,
-0.022374218329787254,
0.0027820342220366,
-0.011741048656404018,
-0.010767227970063686,
0.035326626151800156,
0.029897937551140785,
-0.03813678398728371,
-0.08567366003990173,
0.1201334297657013,
-0.03028922714293003,
-0.0012397117679938674,
-0.029528114944696426,
0.042699333280324936,
-0.0017169022466987371
] |
Refining CLIP's Spatial Awareness: A Visual-Centric Perspective | https://openreview.net/forum?id=38No4B8sx6 | [
"Congpei Qiu",
"Yanhao Wu",
"Wei Ke",
"Xiuxiu Bai",
"Tong Zhang"
] | Poster | Contrastive Language-Image Pre-training (CLIP) excels in global alignment with language but exhibits limited sensitivity to spatial information, leading to strong performance in zero-shot classification tasks but underperformance in tasks requiring precise spatial understanding. Recent approaches have introduced Region-Language Alignment (RLA) to enhance CLIP's performance in dense multimodal tasks by aligning regional visual representations with corresponding text inputs. However, we find that CLIP ViTs fine-tuned with RLA suffer from notable loss in spatial awareness, which is crucial for dense prediction tasks. To address this, we propose the Spatial Correlation Distillation (SCD) framework, which preserves CLIP's inherent spatial structure and mitigates above degradation. To further enhance spatial correlations, we introduce a lightweight Refiner that extracts refined correlations directly from CLIP before feeding them into SCD, based on an intriguring finding that CLIP naturally capture high-quality dense features. Together, these components form a robust distillation framework that enables CLIP ViTs to integrate both visual-language and visual-centric improvements, achieving state-of-the-art results across various open-vocabulary dense prediction benchmarks. | Self-distillation; CLIP; Open-vocabulary dense prediction | We refine CLIP’s region-language alignment by enhancing its spatial awareness, improving performance from both visual-centric and vision-language perspectives. | 382 | null | [
0.007042792160063982,
-0.1584174484014511,
0.08469546586275101,
-0.0012176425661891699,
0.0740208625793457,
0.011529196985065937,
-0.03861179202795029,
-0.08651095628738403,
0.04509546980261803,
-0.024214453995227814,
-0.05640602856874466,
-0.03933189436793327,
-0.034840431064367294,
0.11085444688796997,
-0.03582950308918953,
-0.018410969525575638,
0.06316159665584564,
0.10841245204210281,
-0.09147514402866364,
-0.0631609633564949,
0.06223931536078453,
-0.06286954879760742,
0.022940417751669884,
0.04038941115140915,
0.14551261067390442,
0.04033905267715454,
0.020394494757056236,
0.008879926055669785,
0.00911112129688263,
-0.03191014379262924,
0.031299889087677,
0.0012320408131927252,
0.032267387956380844,
0.0021914683748036623,
-0.09373310953378677,
0.08327822387218475,
-0.023571165278553963,
-0.019496720284223557,
0.020612750202417374,
-0.0447218157351017,
-0.02859075553715229,
0.053965143859386444,
0.03531263768672943,
0.023531602695584297,
0.03985443338751793,
0.02446373552083969,
0.022231562063097954,
-0.012951480224728584,
0.008689827285706997,
-0.0007709731580689549,
-0.12541250884532928,
-0.03135181963443756,
-0.03527718782424927,
0.028417762368917465,
-0.017774373292922974,
0.07953528314828873,
0.10192226618528366,
-0.010384705848991871,
0.02124478481709957,
-0.023283669725060463,
0.037557296454906464,
-0.08144428580999374,
0.01786443591117859,
0.03512493148446083,
-0.009072048589587212,
-0.011876720935106277,
-0.005287244450300932,
0.03884347900748253,
-0.007733565289527178,
-0.03147236630320549,
-0.05525059252977371,
0.08658228814601898,
0.001971966354176402,
0.05305412784218788,
0.001241384306922555,
0.09930852055549622,
-0.020273461937904358,
-0.0656685158610344,
0.006305455695837736,
-0.051704030483961105,
0.11781225353479385,
0.053647398948669434,
0.07177270948886871,
-0.0111159672960639,
0.10871099680662155,
0.016109619289636612,
0.021254153922200203,
-0.011608715169131756,
0.00636309664696455,
-0.031205344945192337,
-0.0258715208619833,
-0.06404542177915573,
0.03347444906830788,
-0.008358259685337543,
-0.0177370086312294,
0.030830183997750282,
-0.017935026437044144,
-0.0606658048927784,
-0.03955638408660889,
0.022715026512742043,
0.04839494451880455,
0.0073409732431173325,
-0.012423661537468433,
-0.028215335682034492,
0.005348988343030214,
-0.023271407932043076,
0.048819247633218765,
0.008957414887845516,
-0.02559363655745983,
-0.05165765807032585,
0.06679563969373703,
-0.018701672554016113,
0.009833065792918205,
0.05875451862812042,
-0.05101553723216057,
0.06792625039815903,
-0.024432098492980003,
-0.046539150178432465,
0.028304290026426315,
-0.03278641402721405,
-0.06141723692417145,
-0.0048040952533483505,
-0.13770200312137604,
0.016709916293621063,
0.04127001017332077,
0.021300911903381348,
0.0068417759612202644,
2.659960228589513e-33,
0.020867284387350082,
0.016197718679904938,
0.07130108773708344,
0.07292403280735016,
0.0229926947504282,
-0.05389637500047684,
-0.036149099469184875,
-0.02758549340069294,
-0.06932296603918076,
-0.013150003738701344,
0.03881438449025154,
0.021442200988531113,
-0.05562509223818779,
0.10366516560316086,
-0.044280920177698135,
-0.03438091650605202,
0.0041289422661066055,
0.05563586205244064,
-0.045936908572912216,
0.05620793625712395,
-0.015041286125779152,
0.0009214228484779596,
-0.06804722547531128,
-0.029383400455117226,
-0.03631427139043808,
-0.016366541385650635,
0.02174466848373413,
-0.05488811060786247,
-0.056877512484788895,
-0.006905409041792154,
-0.13702209293842316,
0.02671864442527294,
0.02013172209262848,
0.02965359203517437,
-0.015777157619595528,
-0.10263774544000626,
-0.05155309662222862,
-0.07521137595176697,
0.013148421421647072,
0.04977240040898323,
-0.016328206285834312,
0.08160101622343063,
-0.07828342914581299,
-0.02563643455505371,
-0.04171857237815857,
-0.008863803930580616,
-0.01414567418396473,
0.059734757989645004,
-0.06167985126376152,
0.03288314864039421,
0.05643533542752266,
0.04381164163351059,
0.015352525748312473,
-0.05628654360771179,
-0.07295816391706467,
-0.041635576635599136,
0.07372111082077026,
-0.06904295086860657,
0.029565779492259026,
-0.018055252730846405,
-0.029642727226018906,
0.05714713782072067,
-0.07066816836595535,
0.05687415972352028,
0.039444684982299805,
0.00687569472938776,
0.08416073769330978,
0.08792214095592499,
0.048835862427949905,
0.06400284171104431,
-0.07708720117807388,
0.008599324151873589,
0.005936812609434128,
-0.026347819715738297,
0.05201558768749237,
0.010823294520378113,
0.033206820487976074,
-0.047615811228752136,
0.04360809549689293,
0.04548905789852142,
-0.036879174411296844,
-0.017104484140872955,
-0.016668308526277542,
-0.05354480445384979,
-0.10421761125326157,
-0.06253991276025772,
0.04745886102318764,
-0.10485096275806427,
0.0017190157668665051,
-0.06003795936703682,
-0.016586462035775185,
-0.033143699169158936,
0.040586091578006744,
-0.01396886259317398,
0.014443287625908852,
-1.3102105952013296e-33,
0.06409762054681778,
-0.0034049253445118666,
0.041089605540037155,
0.05072782188653946,
-0.054628025740385056,
0.03726677969098091,
0.010898616164922714,
0.002365899970754981,
0.0013503708178177476,
-0.09097296744585037,
-0.020872022956609726,
-0.06722421944141388,
0.004321895074099302,
0.006630697753280401,
0.02901609055697918,
0.020219454541802406,
0.013230479322373867,
0.07233214378356934,
-0.064242422580719,
0.03978089615702629,
0.06730563193559647,
0.022203519940376282,
0.011381974443793297,
0.05046016350388527,
-0.06316884607076645,
0.05002330616116524,
0.04610816389322281,
0.0012528945226222277,
-0.05398694798350334,
-0.03455230966210365,
-0.006788732949644327,
0.032584238797426224,
0.03406026214361191,
-0.05281546711921692,
-0.1303032785654068,
0.008802876807749271,
0.03226301819086075,
-0.09363691508769989,
-0.019248872995376587,
0.0410337932407856,
0.059073418378829956,
-0.03247981145977974,
-0.08499854803085327,
0.023533470928668976,
-0.08408869802951813,
-0.04639986902475357,
-0.027586013078689575,
-0.0013702069409191608,
-0.009008673951029778,
0.11169654130935669,
-0.027451233938336372,
0.0766654759645462,
-0.11463005840778351,
0.008677889592945576,
0.024276722222566605,
-0.10152898728847504,
-0.07121539115905762,
-0.01171986386179924,
0.03059449791908264,
0.027478914707899094,
-0.022900421172380447,
0.005998747423291206,
-0.016711298376321793,
-0.04434552788734436,
0.0368524044752121,
0.002123742364346981,
-0.010968212969601154,
0.024973811581730843,
0.041401833295822144,
0.002820954192429781,
0.07369713485240936,
-0.018455903977155685,
0.016550201922655106,
-0.029678022488951683,
-0.004220002330839634,
0.01900709792971611,
0.038605157285928726,
0.004326821304857731,
-0.016978934407234192,
0.031347356736660004,
0.04291557893157005,
-0.037646807730197906,
-0.017668545246124268,
0.09120449423789978,
0.15009140968322754,
0.12940388917922974,
-0.016725318506360054,
0.03094625659286976,
0.08401839435100555,
0.06393074244260788,
-0.0736132562160492,
0.011471484787762165,
0.003524287836626172,
0.04098714143037796,
-0.011048678308725357,
-4.179392831815676e-8,
-0.10534703731536865,
0.037554625421762466,
-0.04802127927541733,
0.06247975677251816,
-0.10493072867393494,
-0.04863956198096275,
0.023662330582737923,
0.11281437426805496,
-0.011381484568119049,
0.01557940524071455,
0.07236761599779129,
-0.043265026062726974,
-0.025717690587043762,
-0.045221518725156784,
0.0313623882830143,
0.05072366073727608,
-0.030920887365937233,
0.07428009063005447,
-0.023336857557296753,
0.09796908497810364,
0.05704627186059952,
0.024556586518883705,
0.03525212034583092,
-0.00012051108933519572,
0.026354778558015823,
-0.06226939335465431,
-0.05881153792142868,
0.06015603989362717,
0.12111233174800873,
-0.06517782807350159,
0.02528095804154873,
-0.022312533110380173,
0.05040673539042473,
-0.050406672060489655,
-0.014043953269720078,
0.026943502947688103,
0.010934107005596161,
0.003658831585198641,
-0.03781986981630325,
-0.00016504219092894346,
-0.021863525733351707,
0.01847340725362301,
-0.009507471695542336,
0.016878491267561913,
-0.06698340177536011,
0.008249820210039616,
0.037788085639476776,
-0.037241846323013306,
0.008649644441902637,
-0.011260274797677994,
-0.05313654989004135,
-0.009564138017594814,
0.011058035306632519,
0.06764984130859375,
0.05323268845677376,
-0.0022866479121148586,
-0.02508716657757759,
0.04966995120048523,
0.021646201610565186,
0.07855468988418579,
0.025461873039603233,
0.010598482564091682,
-0.030565356835722923,
-0.04342605546116829
] |
Do Deep Neural Network Solutions Form a Star Domain? | https://openreview.net/forum?id=QjO0fUlVYK | [
"Ankit Sonthalia",
"Alexander Rubinstein",
"Ehsan Abbasnejad",
"Seong Joon Oh"
] | Poster | It has recently been conjectured that neural network solution sets reachable via stochastic gradient descent (SGD) are convex, considering permutation invariances. This means that a linear path can connect two independent solutions with low loss, given the weights of one of the models are appropriately permuted. However, current methods to test this theory often require very wide networks to succeed. In this work, we conjecture that more generally, the SGD solution set is a star domain that contains a star model that is linearly connected to all the other solutions via paths with low loss values, modulo permutations. We propose the Starlight algorithm that finds a star model of a given learning task. We validate our claim by showing that this star model is linearly connected with other independently found solutions. As an additional benefit of our study, we demonstrate better uncertainty estimates on Bayesian Model Averaging over the obtained star domain. Further, we demonstrate star models as potential substitutes for model ensembles. | mode connectivity, loss landscapes, neural networks, parameter space, star domain | We conjecture that SGD solution sets for DNNs are star-domains modulo permutation invariances. | 380 | 2403.07968 | [
-0.10040074586868286,
-0.1632397174835205,
0.08798065036535263,
0.011674339883029461,
-0.0022144427057355642,
-0.017642026767134666,
0.001820878591388464,
-0.022638125345110893,
0.05694793909788132,
-0.03656305372714996,
-0.05444609373807907,
0.051177978515625,
0.061287906020879745,
-0.0027970431838184595,
-0.10951204597949982,
0.020356768742203712,
-0.005195667501538992,
0.07841144502162933,
-0.06287841498851776,
0.029367437586188316,
0.007053461391478777,
-0.011717921122908592,
-0.0602286271750927,
0.008846716955304146,
0.07273693382740021,
0.00423419987782836,
0.0185680091381073,
-0.028182154521346092,
-0.039588458836078644,
-0.0014235908165574074,
0.07688112556934357,
-0.01415039598941803,
-0.11036435514688492,
0.005692905746400356,
-0.012000979855656624,
0.045545510947704315,
-0.05971692502498627,
-0.03166777640581131,
-0.03366009145975113,
-0.04691494256258011,
0.056829389184713364,
-0.007582934573292732,
0.09135080873966217,
0.025364112108945847,
0.05423664301633835,
0.021017543971538544,
0.0036269419360905886,
-0.1472199708223343,
0.0022246893495321274,
-0.03221433609724045,
-0.024789266288280487,
-0.041022077202796936,
-0.09431593865156174,
-0.0021884797606617212,
0.059227168560028076,
0.016177112236618996,
0.05054471269249916,
-0.026390740647912025,
-0.023763177916407585,
0.0375819057226181,
0.06944674998521805,
-0.09676391631364822,
-0.025310669094324112,
-0.04289313778281212,
0.034275609999895096,
-0.014658733271062374,
-0.027574293315410614,
0.04188758507370949,
-0.07043434679508209,
0.014587195590138435,
0.00984826311469078,
0.13628661632537842,
-0.03469792380928993,
0.030216975137591362,
0.03792904317378998,
0.013425443321466446,
0.17948012053966522,
-0.0676860511302948,
0.02469848282635212,
-0.043917253613471985,
-0.0016436749137938023,
0.03614027425646782,
0.05538307502865791,
0.06574385613203049,
0.12366286665201187,
-0.05572975054383278,
-0.020731335505843163,
0.009597470983862877,
-0.0172573309391737,
-0.01785694807767868,
-0.09573659300804138,
-0.07541704177856445,
-0.05093427002429962,
-0.05200731381773949,
-0.010175604373216629,
0.0016000651521608233,
0.014699546620249748,
-0.14160074293613434,
0.039641864597797394,
0.07933971285820007,
-0.06034738942980766,
0.03401754051446915,
0.08190988004207611,
-0.07722779363393784,
0.055686838924884796,
0.024811794981360435,
0.04096942022442818,
-0.007920331321656704,
0.034734416753053665,
-0.10936858505010605,
0.014257638715207577,
-0.0038714080583304167,
0.0018803426064550877,
0.08967997133731842,
-0.024546723812818527,
-0.05584462732076645,
0.015368450433015823,
0.017893223091959953,
-0.008192694745957851,
0.044247739017009735,
-0.1384986788034439,
0.027976995334029198,
0.06487596035003662,
-0.0073429592885077,
-0.02417006902396679,
0.029306303709745407,
-0.029361985623836517,
2.485845086990754e-33,
0.042595114558935165,
0.037274979054927826,
0.09153784066438675,
0.017381051555275917,
0.07486999034881592,
0.017063068225979805,
-0.015439484268426895,
-0.009221096523106098,
0.011921949684619904,
-0.021826406940817833,
-0.04908207058906555,
-0.003463205648586154,
-0.07449186593294144,
0.0357617549598217,
0.09288059920072556,
-0.024004515260457993,
0.028895191848278046,
-0.03642663359642029,
0.028854474425315857,
-0.09247307479381561,
0.0685420036315918,
-0.061312176287174225,
-0.010073613375425339,
-0.05830332636833191,
0.02840513177216053,
-0.05139348655939102,
0.04983093589544296,
0.001741122337989509,
-0.02775299735367298,
0.0026433966122567654,
-0.08068934828042984,
0.054821114987134933,
0.01309958565980196,
0.086839459836483,
0.0666380226612091,
0.06253425031900406,
-0.0017116657691076398,
0.053941596299409866,
0.01616235263645649,
-0.09637754410505295,
0.05874292925000191,
-0.023545237258076668,
0.020474525168538094,
0.011221732012927532,
-0.050325218588113785,
-0.07733780145645142,
-0.016981322318315506,
0.009170354343950748,
-0.021988719701766968,
0.007473998237401247,
-0.026963911950588226,
0.0006916651618666947,
-0.04329106956720352,
-0.07456024736166,
-0.014278016053140163,
0.016705041751265526,
-0.011901015415787697,
0.021672911942005157,
0.03302278369665146,
0.005338437855243683,
0.03666059672832489,
0.008699744939804077,
-0.03966226801276207,
0.034093812108039856,
0.05971147492527962,
0.1244121566414833,
-0.009946053847670555,
0.015416168607771397,
0.03533108904957771,
-0.019735099747776985,
-0.03866932913661003,
0.016027960926294327,
-0.013951905071735382,
-0.08354836702346802,
0.0775739774107933,
0.015260888263583183,
0.05902404710650444,
-0.05640091747045517,
0.012489343993365765,
-0.018847599625587463,
-0.03949365019798279,
0.022259147837758064,
-0.024372078478336334,
-0.05197526887059212,
-0.003016450209543109,
-0.0028383126482367516,
0.0037819461431354284,
-0.08655355125665665,
-0.020047562196850777,
-0.05546911433339119,
0.0016722528962418437,
-0.04940587654709816,
0.019292393699288368,
0.01413232646882534,
-0.04916589334607124,
-3.030382559433609e-33,
-0.05815623328089714,
-0.0005881924880668521,
0.01263189222663641,
0.07801911234855652,
0.07871155440807343,
-0.048509981483221054,
0.010136434808373451,
-0.04298270493745804,
-0.06304749846458435,
-0.010758131742477417,
-0.0803137868642807,
0.03636982664465904,
-0.05306842178106308,
-0.06385910511016846,
0.07417793571949005,
-0.03970550745725632,
-0.016779854893684387,
0.09879013150930405,
-0.003368327161297202,
-0.017376475036144257,
-0.011896264739334583,
0.1368575245141983,
-0.04610763117671013,
0.054617252200841904,
-0.010692800395190716,
-0.004257772583514452,
-0.026035597547888756,
0.0832766592502594,
-0.0810680165886879,
0.017297670245170593,
0.035614077001810074,
-0.048728372901678085,
-0.07372190803289413,
0.03421569988131523,
-0.04562946781516075,
0.04632927477359772,
-0.018174219876527786,
0.028778478503227234,
-0.047750312834978104,
0.09184038639068604,
0.017061496153473854,
-0.06663588434457779,
-0.0676642656326294,
0.010557089000940323,
0.04518289491534233,
0.013601509854197502,
0.01691359654068947,
-0.04959697276353836,
0.030370930209755898,
0.0002808817953336984,
0.023999132215976715,
-0.024134784936904907,
0.006630235817283392,
0.0566428005695343,
-0.04200456291437149,
0.019274501129984856,
-0.03709958866238594,
0.009713578037917614,
0.16892887651920319,
0.043263990432024,
-0.06428442895412445,
-0.03517519310116768,
0.018459497019648552,
0.06412991136312485,
-0.019130418077111244,
-0.043620672076940536,
-0.0047263880260288715,
0.035152778029441833,
0.01753915660083294,
0.013361698016524315,
0.025654198601841927,
-0.04909852147102356,
0.002397168893367052,
0.028973115608096123,
0.013889297842979431,
0.025578582659363747,
0.03458354249596596,
-0.014717601239681244,
-0.007818647660315037,
-0.06464250385761261,
0.03922618553042412,
-0.03146795555949211,
0.07498834282159805,
0.03232286125421524,
0.0763825923204422,
0.020841164514422417,
0.10448949038982391,
0.0018440946005284786,
0.04274117201566696,
-0.03948313370347023,
-0.024572359398007393,
0.0815819501876831,
0.02828306145966053,
-0.0659395158290863,
-0.009636281989514828,
-4.588678592654105e-8,
-0.05486087128520012,
0.08593595772981644,
-0.0019373218528926373,
0.02328835427761078,
0.08154159039258957,
-0.04916189983487129,
0.026112273335456848,
0.05771242082118988,
0.00959018711000681,
0.03107236512005329,
0.04084070771932602,
0.010852593928575516,
-0.013716904446482658,
-0.05168590322136879,
-0.10040280222892761,
0.05502832308411598,
0.019443629309535027,
0.036423228681087494,
0.02268146350979805,
0.09060145914554596,
-0.00829453393816948,
0.0358852855861187,
0.04623759910464287,
-0.015583357773721218,
0.014758015982806683,
-0.09847714006900787,
-0.06139691546559334,
0.061374541372060776,
0.021030746400356293,
0.048276886343955994,
-0.015677757561206818,
0.04364504665136337,
-0.0007311274530366063,
-0.03005039505660534,
-0.02800748497247696,
0.12371086329221725,
0.08795291930437088,
-0.030542364344000816,
0.005831455811858177,
0.0008881120011210442,
-0.011097838170826435,
0.06398086994886398,
-0.008234130218625069,
-0.004815632477402687,
-0.022448958829045296,
0.04533563181757927,
0.021189872175455093,
-0.07179185748100281,
-0.019481176510453224,
-0.016163654625415802,
-0.008273441344499588,
0.048520199954509735,
-0.059157226234674454,
0.00834230612963438,
-0.017784912139177322,
0.01451411284506321,
-0.007874415256083012,
-0.07825534045696259,
-0.017191806808114052,
0.007184951566159725,
-0.0006493062828667462,
-0.029380595311522484,
-0.06436501443386078,
0.011204785667359829
] |
Information Theoretic Text-to-Image Alignment | https://openreview.net/forum?id=Ugs2W5XFFo | [
"CHAO WANG",
"Giulio Franzese",
"Alessandro Finamore",
"Massimo Gallo",
"Pietro Michiardi"
] | Poster | Diffusion models for Text-to-Image (T2I) conditional generation have recently achieved
tremendous success. Yet, aligning these models with user’s intentions still involves a
laborious trial-and-error process, and this challenging alignment problem has attracted
considerable attention from the research community. In this work, instead of relying on
fine-grained linguistic analyses of prompts, human annotation, or auxiliary vision-language
models, we use Mutual Information (MI) to guide model alignment. In brief, our method
uses self-supervised fine-tuning and relies on a point-wise MI estimation between prompts
and images to create a synthetic fine-tuning set for improving model alignment. Our
analysis indicates that our method is superior to the state-of-the-art, yet it only requires
the pre-trained denoising network of the T2I model itself to estimate MI, and a simple
fine-tuning strategy that improves alignment while maintaining image quality. Code available at https://github.com/Chao0511/mitune. | Diffusion model, Text-image alignment, Mutual information | A novel fine-tuning method for text-to-image generative diffusion models, that uses mutual information to align generated images to user intentions through natural prompts. | 378 | 2405.20759 | [
-0.015608404763042927,
-0.047838244587183,
0.03796239197254181,
-0.02826097421348095,
0.07798406481742859,
-0.014219547621905804,
-0.0014545604353770614,
0.02058178000152111,
0.08057898283004761,
-0.06244217976927757,
0.041198331862688065,
-0.030548136681318283,
0.053954996168613434,
0.07079891115427017,
-0.01415808405727148,
-0.051967766135931015,
0.06537704914808273,
-0.006671569310128689,
-0.10872746258974075,
-0.05901851877570152,
-0.001807427266612649,
0.022618619725108147,
0.055044110864400864,
0.00547318859025836,
0.010507604107260704,
0.05577649176120758,
0.03718358278274536,
-0.045360371470451355,
-0.030484896153211594,
-0.08608077466487885,
-0.01550423912703991,
-0.0010845176875591278,
0.015927022323012352,
0.060207534581422806,
-0.017527427524328232,
0.05683092400431633,
-0.03608826920390129,
0.015311895869672298,
0.03153753653168678,
-0.055794112384319305,
-0.027684029191732407,
0.0020562561694532633,
0.03009326197206974,
0.02444997802376747,
0.12681454420089722,
-0.0041441055946052074,
-0.03498692438006401,
0.0074468194507062435,
0.005354556255042553,
0.03477490320801735,
-0.14932288229465485,
-0.003713784273713827,
0.015575223602354527,
0.04718560352921486,
-0.022525858134031296,
0.07149916887283325,
0.06625033915042877,
0.012646135874092579,
0.054278817027807236,
0.0016073203878477216,
-0.0010824375785887241,
-0.059776365756988525,
-0.014530802145600319,
0.0795811340212822,
0.03480001911520958,
0.007198763079941273,
0.03277844563126564,
0.0055448333732783794,
-0.07415150105953217,
0.010236751288175583,
-0.03757334500551224,
0.07108444720506668,
0.0048710862174630165,
0.041448432952165604,
-0.004552168771624565,
0.07939979434013367,
0.045404404401779175,
-0.013894384726881981,
0.047580793499946594,
-0.11407025158405304,
0.0027336387429386377,
0.006259514484554529,
0.14672787487506866,
0.011189579963684082,
0.05844530463218689,
-0.030791258439421654,
-0.011264567263424397,
0.05347020551562309,
0.061206452548503876,
-0.030993571504950523,
-0.12400554120540619,
-0.10200781375169754,
-0.024275975301861763,
0.03294014930725098,
0.02501548081636429,
0.09542376548051834,
-0.03270706534385681,
-0.08565255254507065,
-0.08096767961978912,
0.074742391705513,
-0.014340639114379883,
0.021738331764936447,
0.011860984377563,
-0.07265474647283554,
0.04808248206973076,
-0.03357262536883354,
0.0647401362657547,
-0.0050125038251280785,
0.05082133412361145,
-0.1025216355919838,
0.052415937185287476,
-0.013586966320872307,
-0.04617185518145561,
0.02008969895541668,
0.05889346823096275,
-0.060347650200128555,
0.032542672008275986,
0.009311070665717125,
0.04715578630566597,
-0.06451088935136795,
-0.03646445274353027,
-0.00799950584769249,
-0.09492399543523788,
-0.0028834070544689894,
0.0026742585469037294,
-0.0510837621986866,
-0.021397925913333893,
1.382206295152459e-34,
0.017848892137408257,
0.0681208148598671,
0.07898728549480438,
0.10311519354581833,
0.031687263399362564,
-0.016244446858763695,
-0.08643938601016998,
-0.014227964915335178,
-0.05608963221311569,
-0.01808442361652851,
-0.040637873113155365,
-0.05834498256444931,
-0.07806798070669174,
0.06328697502613068,
-0.015430550090968609,
-0.04954987019300461,
-0.05891665816307068,
0.07032867521047592,
-0.044667135924100876,
0.04878785461187363,
0.02064761146903038,
0.059999704360961914,
-0.05809197202324867,
-0.051306359469890594,
0.03376884013414383,
-0.005658046342432499,
0.05782202631235123,
-0.1158052384853363,
-0.01875152625143528,
-0.015224349685013294,
-0.09471991658210754,
-0.037202559411525726,
0.05256801098585129,
-0.021050240844488144,
0.05873890221118927,
-0.04041234031319618,
-0.06618600338697433,
-0.02650446444749832,
-0.004835775587707758,
0.03021143563091755,
0.048267632722854614,
0.11028579622507095,
0.005799473728984594,
0.004059114493429661,
-0.04866192489862442,
-0.02527073211967945,
-0.0198584645986557,
0.03804289549589157,
0.0050062621012330055,
-0.04929128289222717,
0.04478437080979347,
0.009263371117413044,
-0.05425712838768959,
-0.04398560896515846,
-0.06667058914899826,
-0.024006837978959084,
0.06187012791633606,
-0.03511866182088852,
0.011622585356235504,
-0.007871212437748909,
0.043168678879737854,
-0.0016488261753693223,
-0.03373558446764946,
0.03344862535595894,
0.08417513966560364,
-0.03625768795609474,
-0.027664434164762497,
0.03475143387913704,
0.04631883651018143,
0.00913193915039301,
-0.042603444308042526,
-0.006347568240016699,
-0.03081957995891571,
-0.01216708030551672,
0.029768047854304314,
0.025361571460962296,
0.05443335697054863,
-0.07855718582868576,
0.04838506132364273,
0.03238671272993088,
-0.030383458361029625,
-0.04222375527024269,
0.0257860254496336,
-0.06860452145338058,
-0.07075821608304977,
-0.020081259310245514,
0.07062223553657532,
-0.04331255331635475,
-0.03872260823845863,
0.04330366104841232,
0.053228750824928284,
0.05253848806023598,
-0.003913660999387503,
0.033357325941324234,
0.041400618851184845,
2.368634892231333e-34,
0.006833469495177269,
0.035479091107845306,
0.055274851620197296,
0.03907328099012375,
-0.04907572269439697,
-0.0077436515130102634,
0.04960539937019348,
0.0661572590470314,
0.05745682492852211,
-0.07464314997196198,
-0.019172850996255875,
-0.012922428548336029,
-0.020082857459783554,
-0.04249678924679756,
-0.046875957399606705,
0.0024361738469451666,
0.063569076359272,
0.029855703935027122,
0.010771343484520912,
0.012717283330857754,
0.0013415197608992457,
0.0730029046535492,
-0.07896637171506882,
0.0341506227850914,
-0.04947773739695549,
0.09130129963159561,
0.03927963972091675,
0.04021492600440979,
-0.030458921566605568,
-0.006732708774507046,
-0.05031101405620575,
-0.01137332059442997,
-0.029767949134111404,
-0.04835151880979538,
-0.06348571181297302,
0.04975776746869087,
-0.010256713256239891,
-0.052312448620796204,
-0.005595637951046228,
0.07833075523376465,
0.048400960862636566,
0.0173062514513731,
-0.1369912475347519,
-0.017156224697828293,
-0.08225832879543304,
-0.053490545600652695,
-0.0409918874502182,
-0.04849210008978844,
-0.015726035460829735,
0.09330441057682037,
-0.04665040969848633,
0.057044245302677155,
-0.09691934287548065,
-0.017972245812416077,
-0.08724518865346909,
-0.10703032463788986,
-0.09592120349407196,
-0.008116984739899635,
0.05748573690652847,
-0.02478392980992794,
-0.013371877372264862,
0.006429365836083889,
0.06434433162212372,
-0.049384478479623795,
-0.018262865021824837,
-0.058183733373880386,
-0.022211214527487755,
0.011962140910327435,
0.008136630989611149,
0.01722991093993187,
0.0761817991733551,
-0.06899267435073853,
-0.011064182035624981,
0.05868034437298775,
0.021171780303120613,
-0.09342063218355179,
0.09507258236408234,
-0.007924133911728859,
0.05158171430230141,
-0.11033153533935547,
0.06557277590036392,
0.029799161478877068,
-0.00020782100909855217,
0.053909458220005035,
0.028594424948096275,
0.09112900495529175,
0.07436592876911163,
0.047075364738702774,
0.05773484706878662,
0.07079276442527771,
-0.05058804154396057,
-0.006106369663029909,
0.03649437800049782,
0.03675512969493866,
-0.10618045181035995,
-4.461946190303934e-8,
-0.06280715763568878,
0.013023203238844872,
0.00157286343164742,
0.0824090838432312,
-0.10510467737913132,
-0.0606573186814785,
0.029856370761990547,
0.052939675748348236,
0.0323462150990963,
-0.068613201379776,
0.0037430054508149624,
0.06694754958152771,
-0.06918144971132278,
-0.042673271149396896,
-0.028486821800470352,
0.07684322446584702,
-0.011705854907631874,
0.06201043725013733,
0.0034167366102337837,
0.04431357607245445,
-0.021837374195456505,
0.016536308452486992,
-0.029160529375076294,
0.004288478288799524,
0.06983169913291931,
-0.007839346304535866,
-0.0406215563416481,
0.05275245010852814,
0.06358232349157333,
-0.06529983133077621,
0.0007815650897100568,
0.01967553421854973,
0.010262228548526764,
-0.02378498576581478,
0.007734217215329409,
0.06500144302845001,
0.03331151232123375,
0.005719686858355999,
0.006052494514733553,
-0.04015829414129257,
-0.0005145411123521626,
0.08915412425994873,
-0.03248150274157524,
-0.01060477178543806,
0.0653700977563858,
-0.021619010716676712,
0.09472765028476715,
-0.05868273973464966,
0.009488318115472794,
-0.06553413718938828,
0.0010237395763397217,
-0.038995061069726944,
0.030090736225247383,
0.015219851396977901,
-0.007117774803191423,
-0.029695063829421997,
0.06991516798734665,
0.03650406748056412,
0.10092715919017792,
0.06642845273017883,
0.014023867435753345,
0.04511977732181549,
-0.017164595425128937,
-0.041389644145965576
] |
Semantix: An Energy-guided Sampler for Semantic Style Transfer | https://openreview.net/forum?id=si37wk8U5D | [
"Huiang He",
"Minghui Hu",
"Chuanxia Zheng",
"Chaoyue Wang",
"Tat-Jen Cham"
] | Poster | Recent advances in style and appearance transfer are impressive, but most methods isolate global style and local appearance transfer, neglecting semantic correspondence. Additionally, image and video tasks are typically handled in isolation, with little focus on integrating them for video transfer. To address these limitations, we introduce a novel task, *Semantic Style Transfer*, which involves transferring style and appearance features from a reference image to a target visual content based on semantic correspondence. We subsequently propose a training-free method, *Semantix*, an energy-guided sampler designed for Semantic Style Transfer that simultaneously guides both style and appearance transfer based on semantic understanding capacity of pre-trained diffusion models. Additionally, as a sampler, *Semantix* can be seamlessly applied to both image and video models, enabling semantic style transfer to be generic across various visual media. Specifically, once inverting both reference and context images or videos to noise space by SDEs, *Semantix* utilizes a meticulously crafted energy function to guide the sampling process, including three key components: *Style Feature Guidance*, *Spatial Feature Guidance* and *Semantic Distance* as a regularisation term. Experimental results demonstrate that *Semantix* not only effectively accomplishes the task of semantic style transfer across images and videos, but also surpasses existing state-of-the-art solutions in both fields. | style transfer, diffusion model, energy guidance | An Energy-Guided Sampler for Semantic Style Transfer across Images and Videos | 374 | null | [
-0.0022679686080664396,
-0.06088320538401604,
0.0792483314871788,
0.031710825860500336,
0.1352234184741974,
-0.03171054646372795,
-0.046243973076343536,
-0.07830997556447983,
0.0004169755557086319,
-0.11227262020111084,
-0.07885679602622986,
-0.08832161873579025,
0.030225737020373344,
-0.013337502256035805,
0.07011843472719193,
0.04416022449731827,
0.12310189008712769,
0.08144420385360718,
-0.06262453645467758,
0.0338287279009819,
0.01864025741815567,
-0.038510460406541824,
0.03304368630051613,
-0.058805014938116074,
0.05470125749707222,
0.03399750962853432,
0.07303418964147568,
0.02036852203309536,
0.0678495317697525,
-0.04361538216471672,
0.05208628252148628,
-0.0022654009517282248,
-0.06861361116170883,
0.005404599476605654,
-0.054372288286685944,
0.046525679528713226,
-0.02049247734248638,
0.04612818360328674,
-0.048800352960824966,
0.013296092860400677,
-0.028840897604823112,
0.013278252445161343,
-0.05111406370997429,
0.0006231197621673346,
-0.009084676392376423,
0.02903418429195881,
-0.0009111469844356179,
-0.006265821401029825,
0.010415622033178806,
0.08695212751626968,
-0.02748691849410534,
-0.05324302241206169,
-0.10290107876062393,
0.04187420755624771,
-0.024690089747309685,
0.06435654312372208,
0.09817011654376984,
-0.024943815544247627,
0.03012297861278057,
-0.003909274470061064,
-0.045042745769023895,
-0.07586672157049179,
0.012139825150370598,
-0.025060487911105156,
-0.01685166172683239,
-0.05544384568929672,
0.03061634674668312,
-0.009015073999762535,
0.03286448121070862,
-0.0938090905547142,
-0.030669765546917915,
0.01046108826994896,
-0.007523145992308855,
0.07409146428108215,
-0.011939410120248795,
0.02902945689857006,
0.0217185840010643,
-0.05461814999580383,
-0.04444824904203415,
-0.12548214197158813,
0.05253520607948303,
-0.0319957435131073,
0.09375733137130737,
0.044253911823034286,
0.06621602177619934,
0.039832424372434616,
-0.02185654640197754,
0.0066221049055457115,
-0.026961777359247208,
0.03485267981886864,
-0.08614005148410797,
-0.022562842816114426,
-0.027157342061400414,
-0.08546432107686996,
0.061055418103933334,
-0.021834321320056915,
-0.0348508283495903,
-0.13659349083900452,
0.025343000888824463,
0.05618506297469139,
0.02383323200047016,
0.026461396366357803,
-0.0017841800581663847,
-0.0028698774985969067,
0.01794278807938099,
-0.05872355401515961,
0.05344858765602112,
0.041675180196762085,
0.018612639978528023,
-0.002623317763209343,
0.12752139568328857,
0.0592544786632061,
-0.02889249660074711,
-0.059804197400808334,
0.05453094467520714,
0.005328480154275894,
-0.046514373272657394,
-0.02305961400270462,
0.03755542263388634,
-0.028035394847393036,
0.012644406408071518,
-0.008488541468977928,
-0.07631547749042511,
-0.004861960653215647,
0.05160190910100937,
-0.0727199986577034,
-0.009138372726738453,
1.9725718153534372e-34,
0.04623202607035637,
-0.012246904894709587,
0.06931012868881226,
0.05535564199090004,
0.043988071382045746,
-0.03311449661850929,
-0.029799116775393486,
-0.018933331593871117,
-0.08518710732460022,
-0.052225492894649506,
0.06044912338256836,
-0.07609900832176208,
0.0037425293121486902,
0.12610262632369995,
-0.009162780828773975,
0.0005751657881774008,
-0.025538280606269836,
0.01457882672548294,
0.050603847950696945,
0.0383516401052475,
0.033253565430641174,
0.004513521213084459,
0.022711537778377533,
0.04472590237855911,
-0.03085182048380375,
-0.034676291048526764,
0.02366751618683338,
-0.030946139246225357,
0.054326314479112625,
-0.03717795014381409,
-0.07974902540445328,
-0.0052180164493620396,
0.02177068404853344,
0.004839516710489988,
0.024486800655722618,
-0.04363187402486801,
0.030506756156682968,
-0.05426120012998581,
0.06914524734020233,
-0.036664046347141266,
-0.040512360632419586,
0.06655984371900558,
-0.005084087140858173,
-0.02251650206744671,
-0.08803358674049377,
0.01292602438479662,
0.04123907908797264,
-0.0011538671096786857,
-0.012954155914485455,
0.016747405752539635,
0.05123881995677948,
-0.059289056807756424,
0.03844762220978737,
-0.029055584222078323,
-0.059482790529727936,
-0.030109860002994537,
0.05823104456067085,
-0.030273308977484703,
0.061103276908397675,
-0.03931372985243797,
0.07815081626176834,
0.0623575896024704,
-0.0012952289544045925,
0.04762522503733635,
0.06914014369249344,
0.022836267948150635,
0.055787552148103714,
-0.014199644327163696,
-0.03593411669135094,
0.004648017697036266,
-0.0468338206410408,
0.059835419058799744,
-0.03453825041651726,
-0.03764906898140907,
0.0644952580332756,
-0.06347990781068802,
-0.0009988037636503577,
-0.03897837921977043,
0.045386526733636856,
0.12295874208211899,
-0.07168903946876526,
0.035719480365514755,
-0.02832590974867344,
-0.08741738647222519,
-0.05236180126667023,
-0.004718806128948927,
-0.015967583283782005,
-0.09555774182081223,
0.0263536237180233,
-0.05810871347784996,
0.0038027360569685698,
-0.05042557045817375,
-0.07166173309087753,
0.014812727458775043,
-0.026585916057229042,
1.033360897801675e-34,
0.012965244241058826,
0.036857008934020996,
0.042113929986953735,
0.06769092381000519,
-0.015404634177684784,
-0.0016843288904055953,
0.025245407596230507,
0.06032488867640495,
0.07630102336406708,
-0.04243234544992447,
0.02245386876165867,
0.011253333650529385,
-0.07747364789247513,
-0.101232148706913,
-0.026035388931632042,
-0.007940806448459625,
-0.019255787134170532,
0.036411937326192856,
-0.1192888617515564,
0.03569173440337181,
0.012983010150492191,
0.06818147748708725,
-0.011544645763933659,
0.009915943257510662,
-0.0976572334766388,
0.08312027156352997,
-0.01775434985756874,
0.034208033233881,
-0.001955845858901739,
-0.029174180701375008,
-0.00845406949520111,
-0.058459725230932236,
0.03752404451370239,
-0.015429430641233921,
-0.07859576493501663,
0.03550459444522858,
0.04419254511594772,
0.05280659720301628,
-0.04570996016263962,
0.11782515794038773,
-0.026098335161805153,
-0.03858115151524544,
-0.009041687473654747,
0.053536556661129,
-0.0674574226140976,
-0.08323946595191956,
-0.08778258413076401,
-0.042191676795482635,
-0.02418866939842701,
0.017323315143585205,
-0.006603567861020565,
0.010691066272556782,
-0.10124232620000839,
0.01960669457912445,
-0.013040644116699696,
-0.07117801159620285,
-0.03320563957095146,
0.04727718234062195,
0.015905946493148804,
0.030090345069766045,
-0.005351214203983545,
-0.03511079400777817,
-0.06320129334926605,
-0.08091919869184494,
0.0013106102123856544,
0.0026244947221130133,
-0.05777556076645851,
-0.057804081588983536,
0.04935631901025772,
0.05347907543182373,
-0.0007781447493471205,
0.02194799855351448,
-0.02106093242764473,
0.032414600253105164,
0.005926342215389013,
-0.07803475856781006,
0.07549004256725311,
-0.003828397486358881,
-0.018154706805944443,
-0.056361764669418335,
-0.04666297510266304,
-0.07831788063049316,
0.0010333172976970673,
0.04411067068576813,
0.09839141368865967,
0.12580323219299316,
0.0014412605669349432,
0.004546992015093565,
0.02553492970764637,
-0.01094841305166483,
-0.02361430414021015,
0.0041094995103776455,
0.13612444698810577,
0.055623359978199005,
-0.09201659262180328,
-3.633979517303487e-8,
-0.11308659613132477,
0.027117298915982246,
0.02696797251701355,
0.06812916696071625,
-0.044527020305395126,
-0.00934529583901167,
0.09560995548963547,
0.026702918112277985,
0.005242977757006884,
-0.029434988275170326,
0.021728649735450745,
0.05639946833252907,
0.025133594870567322,
0.015563671477138996,
-0.024557773023843765,
0.016559695824980736,
0.05734066665172577,
0.06150209158658981,
-0.018731340765953064,
-0.041167471557855606,
-0.03457487374544144,
0.00208593113347888,
0.005453584250062704,
-0.0038957917131483555,
0.12794730067253113,
0.01252161618322134,
0.05378112196922302,
0.09284358471632004,
0.1033230796456337,
-0.02387375943362713,
-0.06527070701122284,
0.008147861808538437,
-0.014449105598032475,
-0.06605478376150131,
0.016323011368513107,
0.02674144133925438,
0.0059941294603049755,
0.0009811259806156158,
-0.018605008721351624,
-0.0060807871632277966,
-0.029278311878442764,
0.013632649555802345,
-0.014801169745624065,
0.04265006259083748,
0.005586407147347927,
0.01786230131983757,
0.02269642800092697,
-0.04003757983446121,
0.033988434821367264,
0.03724593296647072,
0.025010457262396812,
-0.05270583555102348,
-0.0010190880857408047,
0.09617449343204498,
0.0380227230489254,
-0.03556695207953453,
-0.049127284437417984,
0.08451861143112183,
0.0921311303973198,
0.09758102148771286,
0.03423848748207092,
-0.002688745269551873,
-0.008308080956339836,
-0.03982466086745262
] |
GaussianAnything: Interactive Point Cloud Flow Matching for 3D Generation | https://openreview.net/forum?id=P4DbTSDQFu | [
"Yushi LAN",
"Shangchen Zhou",
"Zhaoyang Lyu",
"Fangzhou Hong",
"Shuai Yang",
"Bo Dai",
"Xingang Pan",
"Chen Change Loy"
] | Poster | Recent advancements in diffusion models and large-scale datasets have revolutionized image and video generation, with increasing focus on 3D content generation. While existing methods show promise, they face challenges in input formats, latent space structures, and output representations. This paper introduces a novel 3D generation framework that addresses these issues, enabling scalable and high-quality 3D generation with an interactive Point Cloud-structured Latent space. Our approach utilizes a VAE with multi-view posed RGB-D-N renderings as input, features a unique latent space design that preserves 3D shape information, and incorporates a cascaded latent flow-based model for improved shape-texture disentanglement. The proposed method, GaussianAnything, supports multi-modal conditional 3D generation, allowing for point cloud, caption, and single-view image inputs. Experimental results demonstrate superior performance on various datasets, advancing the state-of-the-art in 3D content generation. | 3D Object Generation, Gaussian Splatting, Flow-based Generative Models | A novel interactive 3D generation framework for 3D Gaussians. | 371 | null | [
-0.0392901748418808,
-0.042739372700452805,
0.04938315227627754,
-0.05129169672727585,
0.06783708184957504,
0.043279848992824554,
-0.056145597249269485,
-0.11089775711297989,
0.0432950034737587,
-0.04151049628853798,
-0.0325658842921257,
-0.04889586567878723,
0.022597746923565865,
0.04674690216779709,
0.01420478243380785,
0.0108024375513196,
0.07555342465639114,
0.060717903077602386,
-0.06025363877415657,
0.008253275416791439,
0.006607773248106241,
0.030666392296552658,
-0.05416882038116455,
-0.02450147643685341,
0.07589757442474365,
0.06862728297710419,
0.11364810913801193,
-0.04802926257252693,
0.04333171620965004,
-0.02808026596903801,
0.07092306762933731,
-0.002418285934254527,
-0.07860307395458221,
0.06437502056360245,
-0.05562707036733627,
0.03195435181260109,
-0.02798367850482464,
0.01661485619843006,
-0.06629078090190887,
0.007157441694289446,
0.0005683375056833029,
0.04045795276761055,
-0.025329262018203735,
0.04085085913538933,
0.08474449068307877,
0.049550339579582214,
-0.06799083948135376,
-0.06038288399577141,
0.055004581809043884,
-0.002292418386787176,
0.0013530811993405223,
-0.05587133765220642,
0.015505295246839523,
-0.01815749704837799,
-0.028464006260037422,
0.028605297207832336,
0.03229110315442085,
-0.06158158555626869,
0.03349998593330383,
-0.011781871318817139,
0.03064713440835476,
-0.03393850103020668,
0.07180864363908768,
-0.055646613240242004,
0.0034187694545835257,
-0.08048728853464127,
0.0035595714580267668,
-0.04803056642413139,
-0.025480477139353752,
-0.056723881512880325,
-0.09046120196580887,
0.02982337400317192,
-0.01691790670156479,
-0.0424417145550251,
-0.02779073268175125,
0.0560181699693203,
-0.00028718338580802083,
-0.09291539341211319,
-0.012411299161612988,
0.0070073045790195465,
0.03734615072607994,
0.06004921346902847,
0.053833115845918655,
0.07818756997585297,
-0.0895240381360054,
0.034312017261981964,
-0.00692882901057601,
0.08501169085502625,
-0.031201526522636414,
-0.028903350234031677,
-0.04141542315483093,
0.033587366342544556,
-0.036068789660930634,
-0.017306534573435783,
-0.0007457343745045364,
-0.03560814633965492,
0.0008791831787675619,
-0.1207951009273529,
0.035588543862104416,
0.009481140412390232,
0.021009040996432304,
0.00047826365334913135,
0.07658170908689499,
0.0012658535270020366,
0.06304240226745605,
-0.05236929655075073,
0.05227678641676903,
0.038020405918359756,
0.013315110467374325,
-0.04390362650156021,
0.03141999989748001,
-0.03849397972226143,
0.0549510158598423,
0.02441161684691906,
-0.017088761553168297,
-0.005212653428316116,
-0.012482902966439724,
-0.04082534462213516,
-0.0019478372996672988,
-0.006322558503597975,
0.04069170355796814,
0.015886865556240082,
-0.04159785807132721,
-0.007837895303964615,
-0.008417151868343353,
-0.00888781063258648,
0.0004905648529529572,
1.0559355550841219e-33,
0.0294332392513752,
0.03142775967717171,
0.05657591298222542,
0.05978677421808243,
0.09844650328159332,
0.06497971713542938,
-0.02260492742061615,
-0.049094196408987045,
-0.0023244954645633698,
-0.035569336265325546,
-0.08518967777490616,
-0.10771314054727554,
-0.065425805747509,
0.10002963244915009,
0.01873004622757435,
-0.018018625676631927,
-0.06756645441055298,
-0.015357457101345062,
-0.004398460499942303,
0.13013237714767456,
0.05975471809506416,
-0.024375295266509056,
-0.023170433938503265,
0.007315917406231165,
-0.08503137528896332,
0.021403973922133446,
0.02430165559053421,
0.023488743230700493,
-0.0028996479231864214,
-0.04805878549814224,
-0.08645810931921005,
-0.02831619419157505,
0.017727604135870934,
0.011106799356639385,
-0.06774837523698807,
-0.02463747002184391,
-0.00732124038040638,
-0.05810539424419403,
0.013181544840335846,
-0.050349801778793335,
0.030019285157322884,
0.054716404527425766,
-0.024590136483311653,
-0.041692234575748444,
-0.1132281944155693,
0.03752235695719719,
0.07429774850606918,
0.0340563990175724,
0.01464516669511795,
0.00897970236837864,
0.1172274500131607,
0.07756589353084564,
-0.026170313358306885,
-0.0705944374203682,
-0.06643009930849075,
-0.07335354387760162,
-0.0017731105908751488,
-0.0189492367208004,
0.11701261252164841,
0.04081202670931816,
0.007242538966238499,
0.025402124971151352,
-0.03521448001265526,
-0.004687345586717129,
-0.0742453783750534,
-0.02951183170080185,
0.054859425872564316,
0.06075325980782509,
0.07524507492780685,
0.04563067480921745,
-0.042248185724020004,
0.014593463391065598,
-0.022690774872899055,
-0.09647493064403534,
0.018028253689408302,
-0.05937451869249344,
-0.01541423425078392,
-0.015151562169194221,
0.07723015546798706,
0.04699033498764038,
-0.10300611704587936,
0.05121351405978203,
0.0010586957214400172,
-0.059901002794504166,
-0.024951964616775513,
-0.060835711658000946,
-0.013721263036131859,
-0.03804323077201843,
-0.03160616755485535,
-0.07260086387395859,
-0.010443379171192646,
-0.06901697814464569,
0.0473083034157753,
0.007930800318717957,
0.035804640501737595,
-2.550331650968659e-34,
0.02741847187280655,
0.0041948952712118626,
-0.05272749438881874,
0.05840456485748291,
0.07750002294778824,
0.017959650605916977,
0.020169444382190704,
0.006299533415585756,
0.06013978272676468,
-0.05376860871911049,
0.006409777794033289,
0.06015715003013611,
0.012062931433320045,
0.011084247380495071,
0.0008742354111745954,
0.02821759507060051,
0.022114800289273262,
-0.05867144837975502,
-0.06681497395038605,
-0.030674943700432777,
0.009159279055893421,
-0.005328649654984474,
-0.03189634531736374,
-0.03628572076559067,
-0.09103288501501083,
0.11264576762914658,
-0.00957563053816557,
-0.049867842346429825,
0.0022987285628914833,
0.061192821711301804,
0.028428170830011368,
0.016531746834516525,
0.05853670835494995,
-0.007527082692831755,
-0.03720671311020851,
-0.01657402142882347,
0.05648699775338173,
0.06428451091051102,
-0.0767478421330452,
0.00022945615637581795,
-0.007466718554496765,
-0.09501762688159943,
-0.025381702929735184,
0.11045513302087784,
-0.040692541748285294,
0.08238182961940765,
0.017152437940239906,
-0.048785291612148285,
-0.00756032345816493,
0.10142461955547333,
-0.07040980458259583,
0.09226903319358826,
-0.07394111156463623,
0.0009471537196077406,
0.00860558170825243,
-0.05658501759171486,
-0.10032320767641068,
0.026854630559682846,
0.07721149921417236,
0.019805967807769775,
-0.02455388568341732,
-0.036957018077373505,
-0.04525281488895416,
-0.0015289679868146777,
-0.001243631704710424,
0.00260368874296546,
-0.03582038730382919,
0.0024835653603076935,
0.0501798577606678,
0.10060662031173706,
0.046596430242061615,
0.03985244780778885,
0.05700923874974251,
0.016854435205459595,
0.05561482906341553,
-0.07049237936735153,
0.06195373088121414,
0.02566358633339405,
0.08788971602916718,
-0.0029533461201936007,
0.033815279603004456,
0.03711770847439766,
0.03582054376602173,
0.083748959004879,
0.06718812137842178,
0.028041819110512733,
-0.008538106456398964,
-0.007090995088219643,
-0.009870261885225773,
0.08843287080526352,
-0.055352628231048584,
-0.007264233194291592,
-0.043664801865816116,
0.05521713197231293,
-0.02868528664112091,
-3.661714131908411e-8,
-0.029583323746919632,
0.01684836484491825,
0.045055363327264786,
-0.052858803421258926,
-0.07816147059202194,
-0.06804700940847397,
0.08858489245176315,
0.11278243362903595,
0.049390677362680435,
-0.02082378789782524,
0.0455896370112896,
-0.08898801356554031,
-0.031161854043602943,
-0.03253449499607086,
0.06835387647151947,
0.07959695160388947,
0.006409497931599617,
0.0546419583261013,
0.01568041369318962,
0.017481127753853798,
-0.0966876968741417,
0.04418192803859711,
-0.030307697132229805,
0.013825606554746628,
0.006698943208903074,
0.029873192310333252,
-0.020683323964476585,
0.062115397304296494,
0.09515313804149628,
-0.08118202537298203,
0.005323664750903845,
0.0016136595513671637,
0.005331790074706078,
0.015351393260061741,
-0.05520413815975189,
-0.011435551568865776,
-0.04607298597693443,
0.00038835566374473274,
0.0010582534596323967,
-0.0833001509308815,
-0.05187452584505081,
0.05726397782564163,
0.04137891158461571,
-0.05241133272647858,
0.030518213286995888,
0.11712568998336792,
0.09041214734315872,
-0.07733891159296036,
0.010331988334655762,
0.03362797945737839,
-0.04778272658586502,
-0.03347386419773102,
-0.0021390721667557955,
0.09416079521179199,
0.10232483595609665,
0.026115482673048973,
-0.028838060796260834,
-0.0321098268032074,
0.09468107670545578,
-0.019817110151052475,
-0.07417094707489014,
-0.010988625697791576,
0.021999230608344078,
0.010553495027124882
] |
TraceVLA: Visual Trace Prompting Enhances Spatial-Temporal Awareness for Generalist Robotic Policies | https://openreview.net/forum?id=b1CVu9l5GO | [
"Ruijie Zheng",
"Yongyuan Liang",
"Shuaiyi Huang",
"Jianfeng Gao",
"Hal Daumé III",
"Andrey Kolobov",
"Furong Huang",
"Jianwei Yang"
] | Poster | Although large vision-language-action (VLA) models pretrained on extensive robot datasets offer promising generalist policies for robotic learning, they still struggle with spatial-temporal dynamics in interactive robotics, making them less effective in handling complex tasks, such as manipulation. In this work, we introduce visual trace prompting, a simple yet effective approach to facilitate VLA models’ spatial-temporal awareness for action prediction by encoding state-action trajectories visually. We develop a new TraceVLA model by finetuning
OpenVLA on our own collected dataset of 150K robot manipulation trajectories using visual trace prompting. Evaluations of TraceVLA across 137 configurations in SimplerEnv and 4 tasks on a physical WidowX robot demonstrate state-of-the-art performance, outperforming OpenVLA by 10% on SimplerEnv and 3.5x on real-robot tasks and exhibiting robust generalization across diverse embodiments and scenarios. To further validate the effectiveness and generality of our method, we present a compact VLA model based on 4B Phi-3-Vision, pretrained on the Open-X-Embodiment and finetuned on our dataset, rivals the 7B OpenVLA baseline while significantly improving inference efficiency. | Vision Language Model, Robot Learning | Visual trace prompting enhances VLA models' spatial-temporal understanding, boosting robotic manipulation performance. | 365 | 2412.10345 | [
0.02507699467241764,
-0.1527027040719986,
0.02769138477742672,
0.015065395273268223,
0.03308003395795822,
0.014650735072791576,
-0.001932453946210444,
-0.0476851686835289,
0.009982739575207233,
-0.0036326758563518524,
0.01277540996670723,
-0.05339634418487549,
-0.056161992251873016,
0.07711610943078995,
0.0626392588019371,
-0.0009548617526888847,
0.08934705704450607,
0.04423382878303528,
-0.10116461664438248,
0.01181851327419281,
0.07493375986814499,
-0.020200561732053757,
0.10206107050180435,
0.017492791637778282,
-0.03147660568356514,
-0.0036846341099590063,
0.04204178974032402,
-0.0746871829032898,
0.0036880041006952524,
-0.07553905248641968,
0.012104945257306099,
-0.03907391056418419,
0.04420322924852371,
-0.038181204348802567,
-0.055005691945552826,
0.06954508274793625,
-0.07145523279905319,
0.031132841482758522,
-0.06959576904773712,
-0.09223242104053497,
-0.036336760967969894,
0.02228713035583496,
0.07981405407190323,
0.028938205912709236,
-0.00323098199442029,
0.05398482084274292,
0.0402589775621891,
-0.09788020700216293,
0.04038301482796669,
-0.0010803217301145196,
-0.08201707899570465,
-0.06637346744537354,
0.0677163228392601,
-0.007554478943347931,
-0.08328850567340851,
0.024191847071051598,
0.029595764353871346,
-0.061439163982868195,
0.06935632228851318,
-0.0671912282705307,
0.003303400706499815,
-0.058336012065410614,
0.020172694697976112,
-0.0005379525246098638,
-0.08339101076126099,
-0.041452836245298386,
0.033012185245752335,
-0.05902610346674919,
0.00027575864805839956,
0.0024194917641580105,
-0.028511440381407738,
0.04814884811639786,
0.02809007838368416,
0.003910903353244066,
0.002103095408529043,
0.03188741207122803,
0.004119428806006908,
-0.047323986887931824,
0.012765410356223583,
-0.044479046016931534,
0.02804393321275711,
-0.00875550415366888,
-0.03567114472389221,
0.11565942317247391,
0.09837044775485992,
0.023769276216626167,
-0.04144121706485748,
0.04377632215619087,
0.05794064700603485,
-0.013001718558371067,
0.010568033903837204,
-0.05190807580947876,
0.03708985820412636,
-0.07611384242773056,
-0.005653201602399349,
0.022505294531583786,
-0.03543151542544365,
-0.09344352036714554,
0.0029352661222219467,
0.006679462268948555,
0.00660795159637928,
-0.027744032442569733,
-0.0008753552101552486,
0.01535473857074976,
0.008413377217948437,
-0.016324885189533234,
0.06046928092837334,
-0.007984458468854427,
-0.011093474924564362,
-0.004888331517577171,
0.04780416190624237,
-0.08749041706323624,
0.06080381199717522,
0.0363025926053524,
0.01886504329741001,
0.07389513403177261,
-0.010560660623013973,
-0.0014516643714159727,
0.02717418223619461,
0.019663004204630852,
0.03888148069381714,
-0.05277566984295845,
0.055821940302848816,
-0.01786341890692711,
0.06652896851301193,
-0.02943674847483635,
-0.022535275667905807,
-8.717675490629324e-34,
0.07281382381916046,
0.027987627312541008,
0.0493040606379509,
0.0012014041421934962,
0.048969317227602005,
-0.02068823017179966,
0.013803109526634216,
-0.028964310884475708,
-0.051734402775764465,
0.01971859484910965,
-0.013144942000508308,
-0.021654771640896797,
-0.06359133124351501,
0.12669959664344788,
0.02653282694518566,
0.03241151198744774,
-0.00587462168186903,
0.04544072970747948,
-0.07909101247787476,
0.07755335420370102,
0.07980968803167343,
-0.01759687438607216,
-0.022351045161485672,
-0.06181126460433006,
0.021726476028561592,
0.023237263783812523,
0.008759400807321072,
0.03132031857967377,
-0.0049515520222485065,
-0.01346590369939804,
-0.05025981739163399,
0.06958817690610886,
-0.001331597683019936,
0.10529307276010513,
0.037240173667669296,
0.005316579714417458,
-0.07105740159749985,
-0.06283343583345413,
0.00216602417640388,
-0.005842522718012333,
0.02012569084763527,
0.019002173095941544,
-0.03646606579422951,
-0.003062398871406913,
-0.054782040417194366,
-0.02427934855222702,
0.040361080318689346,
0.08527476340532303,
-0.07885090261697769,
0.016539573669433594,
0.04207181558012962,
0.053804103285074234,
-0.04584071412682533,
-0.07077544927597046,
0.006398032885044813,
-0.002864493988454342,
0.05839818716049194,
0.0363239124417305,
0.0027567276265472174,
-0.02774827927350998,
0.019175352528691292,
0.04362671077251434,
0.012762357480823994,
0.07535343617200851,
-0.008923537097871304,
-0.00990021787583828,
0.027799470350146294,
0.14203070104122162,
0.07964275777339935,
0.002982818754389882,
-0.04729996994137764,
0.046951282769441605,
0.008064896799623966,
-0.08782008290290833,
0.09779505431652069,
0.01025133766233921,
0.03318432345986366,
-0.11494837701320648,
0.03753405809402466,
0.024709947407245636,
-0.035945385694503784,
-0.07521820068359375,
-0.015501382760703564,
0.05541767179965973,
0.01545779686421156,
-0.038649581372737885,
0.03679421916604042,
-0.020082993432879448,
-0.04699926823377609,
-0.0077881221659481525,
-0.048348113894462585,
-0.07522348314523697,
-0.08130844682455063,
0.04934773966670036,
-0.037292901426553726,
-4.022464067519614e-34,
0.032320477068424225,
0.035447943955659866,
0.03278261423110962,
0.053185418248176575,
-0.04164041206240654,
-0.09574571996927261,
0.0022782732266932726,
-0.000953728100284934,
0.006624176632612944,
-0.04268979653716087,
-0.042492933571338654,
-0.04576166346669197,
-0.005263849627226591,
0.06580210477113724,
0.0694260448217392,
-0.00363251194357872,
0.03362676128745079,
0.0372258685529232,
-0.023823492228984833,
-0.049052897840738297,
-0.02305440790951252,
0.05865432694554329,
-0.05935320258140564,
-0.007758680265396833,
-0.03595070168375969,
0.002040259074419737,
0.05232152342796326,
0.045707400888204575,
-0.07180610299110413,
0.045801565051078796,
-0.0564543679356575,
-0.010397380217909813,
0.0443476103246212,
0.02675393968820572,
-0.04621744155883789,
0.020588498562574387,
0.03359522670507431,
-0.04856106638908386,
0.02157139964401722,
0.04628234729170799,
0.056914277374744415,
-0.08748724311590195,
-0.053253285586833954,
0.014308006502687931,
-0.047654230147600174,
-0.04335179179906845,
-0.03478052094578743,
0.0445563904941082,
-0.04784301668405533,
0.02351258136332035,
-0.04386588931083679,
0.046391334384679794,
-0.07024460285902023,
-0.04022318497300148,
0.0005351926665753126,
-0.008012119680643082,
-0.03981463983654976,
-0.04149770736694336,
0.04828226566314697,
-0.0733855590224266,
-0.08508220314979553,
0.024647902697324753,
-0.01813591830432415,
0.04778381809592247,
0.02073078416287899,
0.040468938648700714,
-0.020444417372345924,
-0.00914966780692339,
0.009061693213880062,
-0.021198466420173645,
-0.004964346997439861,
-0.035206716507673264,
-0.01603405550122261,
0.002915077842772007,
0.06845834106206894,
-0.017561549320816994,
-0.024348223581910133,
-0.0783185362815857,
0.03896995633840561,
-0.013681848533451557,
-0.019615907222032547,
-0.08379955589771271,
0.08413929492235184,
0.06163762882351875,
0.051008421927690506,
0.1731250286102295,
-0.060455963015556335,
0.06761635094881058,
0.014563465490937233,
0.06490988284349442,
-0.05992523208260536,
0.05914267525076866,
-0.006762769538909197,
0.07770828902721405,
-0.08750128746032715,
-3.500468892525532e-8,
-0.10613005608320236,
0.07433312386274338,
0.038936205208301544,
-0.009212377481162548,
0.040305621922016144,
0.011136582121253014,
-0.019341548904776573,
0.04098657891154289,
-0.015093964524567127,
0.03827530890703201,
0.06504957377910614,
0.032388534396886826,
-0.018719037994742393,
-0.05326031148433685,
0.02066871151328087,
0.056003160774707794,
-0.01188222598284483,
-0.010287968441843987,
-0.014838770031929016,
-0.0023369903210550547,
0.027529912069439888,
-0.03777731955051422,
0.02124330960214138,
-0.021945610642433167,
-0.03277643024921417,
-0.04510198533535004,
-0.08847686648368835,
0.11214463412761688,
0.05124346539378166,
-0.06421104073524475,
0.07544171810150146,
0.001953020691871643,
0.037371620535850525,
-0.00484051275998354,
-0.038627754896879196,
0.012003140524029732,
-0.007105265744030476,
-0.0826663076877594,
0.032035041600465775,
0.03198138624429703,
0.05330008268356323,
0.07247915863990784,
-0.005421665962785482,
0.015375616028904915,
-0.055494509637355804,
0.03857249394059181,
-0.06238802894949913,
-0.21243083477020264,
-0.029316214844584465,
-0.03619314357638359,
-0.0837586373090744,
-0.0021701818332076073,
-0.02538193389773369,
0.16377922892570496,
0.08638410270214081,
0.022003447636961937,
0.035842157900333405,
-0.035260241478681564,
-0.045736074447631836,
0.12156245857477188,
0.025561366230249405,
0.024598928168416023,
-0.036183129996061325,
-0.010753658600151539
] |
Gnothi Seauton: Empowering Faithful Self-Interpretability in Black-Box Transformers | https://openreview.net/forum?id=UvMSKonce8 | [
"Shaobo Wang",
"Hongxuan Tang",
"Mingyang Wang",
"Hongrui Zhang",
"Xuyang Liu",
"Weiya Li",
"Xuming Hu",
"Linfeng Zhang"
] | Poster | The debate between self-interpretable models and post-hoc explanations for black-box models is central to Explainable AI (XAI). Self-interpretable models, such as concept-based networks, offer insights by connecting decisions to human-understandable concepts but often struggle with performance and scalability. Conversely, post-hoc methods like Shapley values, while theoretically robust, are computationally expensive and resource-intensive. To bridge the gap between these two lines of research, we propose a novel method that combines their strengths, providing theoretically guaranteed self-interpretability for black-box models without compromising prediction accuracy. Specifically, we introduce a parameter-efficient pipeline, AutoGnothi, which integrates a small side network into the black-box model, allowing it to generate Shapley value explanations without changing the original network parameters. This side-tuning approach significantly reduces memory, training, and inference costs, outperforming traditional parameter-efficient methods, where full fine-tuning serves as the optimal baseline. AutoGnothi enables the black-box model to predict and explain its predictions with minimal overhead. Extensive experiments show that AutoGnothi offers accurate explanations for both vision and language tasks, delivering superior computational efficiency with comparable interpretability. | Transformers, Interpretability, Efficient-AI, Shapley Value | We introduce a method that makes a black-box transformer model self-interpretable. By using a side network to generate Shapley values, it reduces memory, training, and inference costs while maintaining prediction accuracy. | 357 | 2410.21815 | [
-0.004883816000074148,
-0.0461299866437912,
-0.03584517538547516,
0.08868113905191422,
0.03617711365222931,
-0.015001922845840454,
-0.009669831022620201,
0.020549019798636436,
0.0022498448379337788,
-0.061617884784936905,
-0.12007138878107071,
-0.102955162525177,
0.00866919569671154,
-0.004886784590780735,
-0.05593642219901085,
-0.01908116601407528,
0.14556193351745605,
0.009928178042173386,
-0.1345558613538742,
-0.0853101834654808,
0.03679749369621277,
0.045309629291296005,
0.004918794613331556,
0.03797168657183647,
0.035938628017902374,
-0.04357771947979927,
-0.022503100335597992,
-0.012954680249094963,
-0.035869475454092026,
-0.02550380863249302,
0.0012551931431517005,
-0.05613179877400398,
-0.030496496707201004,
0.037512827664613724,
-0.04257771372795105,
0.014489036053419113,
-0.07667390257120132,
0.07016415148973465,
-0.04119262844324112,
-0.0773509219288826,
0.004161105491220951,
-0.021776676177978516,
-0.015374070964753628,
0.006316353101283312,
0.045455820858478546,
-0.007515670731663704,
0.015756573528051376,
-0.05030176788568497,
-0.009417089633643627,
0.005320188123732805,
-0.05846184119582176,
-0.012572581879794598,
-0.057140883058309555,
0.062054943293333054,
0.006050636526197195,
0.052377086132764816,
0.0033850078471004963,
0.05442666634917259,
-0.024505365639925003,
-0.03305364400148392,
-0.02803575061261654,
-0.07060220092535019,
0.028727639466524124,
0.01773836836218834,
-0.0069459290243685246,
0.01043584942817688,
0.05238482356071472,
0.043545935302972794,
-0.06683764606714249,
0.032434701919555664,
0.07181672006845474,
0.0768771693110466,
-0.06463559716939926,
-0.0015106278005987406,
0.061141110956668854,
0.11742673069238663,
0.019973257556557655,
-0.005565708037465811,
0.046229373663663864,
-0.006927900016307831,
-0.05778266489505768,
0.003770320676267147,
0.005743490066379309,
0.11377834528684616,
0.01997847482562065,
0.029676614329218864,
0.005803556181490421,
0.0459391251206398,
0.01603074185550213,
0.054379235953092575,
-0.04521101340651512,
-0.06914247572422028,
-0.015560375526547432,
-0.02243621088564396,
0.09017201513051987,
0.060354724526405334,
0.01400281023234129,
-0.11994951963424683,
-0.07634416967630386,
0.0610305517911911,
-0.07892873883247375,
0.08443278074264526,
0.06080292910337448,
-0.08678635954856873,
0.0127875292673707,
-0.01844729669392109,
0.06309212744235992,
0.028316820040345192,
0.049606963992118835,
-0.02556406706571579,
-0.022149592638015747,
0.0173504576086998,
-0.007979653775691986,
-0.09091169387102127,
0.06554808467626572,
0.07984112948179245,
0.04740479215979576,
-0.0024473390076309443,
0.014159939251840115,
0.07067185640335083,
0.011429128237068653,
-0.017934437841176987,
0.0482795387506485,
0.07775963842868805,
0.057152993977069855,
0.005328734405338764,
-0.10873876512050629,
8.833785863486757e-34,
-0.01821136847138405,
0.06733036786317825,
0.09530775994062424,
-0.00848451815545559,
0.07832910865545273,
-0.030380290001630783,
-0.0646531879901886,
0.038418885320425034,
0.006867869291454554,
0.018429504707455635,
-0.08017827570438385,
0.09938404709100723,
-0.07939742505550385,
0.01508274581283331,
0.050801362842321396,
0.03580061346292496,
-0.06693410873413086,
0.002479582792147994,
-0.030255725607275963,
-0.05458122119307518,
0.08356016129255295,
-0.06758186966180801,
-0.04719565063714981,
-0.04050798714160919,
0.011980745010077953,
0.00992038194090128,
0.11283571273088455,
0.029054608196020126,
-0.021762007847428322,
0.0024523697793483734,
-0.10215616971254349,
0.027758367359638214,
0.06203506886959076,
-0.03955445811152458,
0.04861816018819809,
-0.08689168095588684,
-0.043604958802461624,
-0.04293452203273773,
0.014473602175712585,
0.005597088020294905,
0.03257173299789429,
-0.00784287042915821,
-0.04074247181415558,
-0.05531228706240654,
-0.03787069022655487,
-0.0735144391655922,
-0.019798031076788902,
-0.016173256561160088,
-0.09609201550483704,
0.0013050958514213562,
0.05139631778001785,
0.05543798580765724,
-0.034623440355062485,
0.0013596562203019857,
-0.0280888881534338,
-0.01455928198993206,
0.019262585788965225,
0.040013447403907776,
0.03332487866282463,
0.06569098681211472,
0.03457559645175934,
0.032792557030916214,
0.008315951563417912,
0.061730872839689255,
0.007180726155638695,
0.07047168910503387,
-0.01919974386692047,
0.1284005343914032,
0.02218380942940712,
-0.027258440852165222,
-0.06494487076997757,
-0.005787465255707502,
-0.05343101546168327,
-0.03984670341014862,
0.02012360282242298,
0.01346539705991745,
0.013010154478251934,
-0.09684253484010696,
0.028926320374011993,
-0.014305205084383488,
0.012494361028075218,
-0.0008502827840857208,
0.0019372798269614577,
0.04560883715748787,
-0.05848611891269684,
-0.013308141380548477,
0.05551676079630852,
0.006205942947417498,
-0.046322669833898544,
-0.018473053351044655,
-0.048739414662122726,
0.004032496828585863,
-0.0010407838271930814,
-0.011298524215817451,
-0.08321118354797363,
-1.0300339962412336e-33,
-0.03695850819349289,
-0.018783917650580406,
-0.07898817211389542,
0.0697692334651947,
0.024263976141810417,
-0.08657617121934891,
-0.027978703379631042,
0.04102771356701851,
0.0465642511844635,
-0.017209986224770546,
-0.06884098798036575,
-0.027901541441679,
0.0616670697927475,
0.04684711620211601,
0.012188011780381203,
-0.04076108708977699,
-0.03990650922060013,
-0.01318959891796112,
0.011691934429109097,
0.04850418120622635,
-0.020806526765227318,
0.07975325733423233,
-0.13501283526420593,
-0.01605779305100441,
-0.09028378129005432,
0.040954820811748505,
-0.03300318866968155,
0.05285470560193062,
0.009465595707297325,
0.012271466664969921,
-0.06269782036542892,
0.07056384533643723,
-0.0123196542263031,
-0.024015121161937714,
-0.03360705077648163,
0.08466148376464844,
0.015845241025090218,
0.01765359193086624,
-0.034409187734127045,
0.04931161180138588,
0.07847080379724503,
-0.055441614240407944,
-0.09684697538614273,
-0.046821702271699905,
-0.02902238443493843,
-0.036300331354141235,
-0.03273395076394081,
0.07023222744464874,
0.04119297116994858,
0.052147649228572845,
-0.03292496129870415,
0.01490955799818039,
-0.02395932748913765,
0.02442176640033722,
-0.046990636736154556,
-0.04420797526836395,
0.008708171546459198,
0.013794584199786186,
0.06622390449047089,
0.03294568136334419,
-0.0619543120265007,
-0.020103702321648598,
0.021347418427467346,
-0.014632199890911579,
-0.05031596124172211,
-0.022352825850248337,
0.01598801463842392,
-0.016303718090057373,
0.058423545211553574,
-0.03085886687040329,
-0.002329090144485235,
-0.06282951682806015,
-0.031155329197645187,
0.02136804163455963,
-0.013102193363010883,
-0.017407407984137535,
-0.017920754849910736,
-0.00924608577042818,
0.02803574502468109,
-0.04376661404967308,
0.05170827358961105,
0.03850456699728966,
0.03895861655473709,
0.020879341289401054,
0.06961192935705185,
-0.055200282484292984,
-0.0035713366232812405,
0.1144535094499588,
0.04146605357527733,
0.05165112763643265,
-0.08017285168170929,
0.05578271672129631,
-0.09990815073251724,
0.12995195388793945,
-0.10068002343177795,
-4.137491416145167e-8,
-0.02860965020954609,
0.07160305976867676,
0.062218502163887024,
0.09127293527126312,
0.031245574355125427,
-0.036364614963531494,
-0.019275544211268425,
0.043353915214538574,
0.0011741064954549074,
0.02996007539331913,
0.017180200666189194,
-0.06094985082745552,
-0.05202711746096611,
-0.017186716198921204,
-0.027120446786284447,
0.10533905774354935,
0.023722510784864426,
0.013477368280291557,
0.021288005635142326,
-0.04563022404909134,
-0.02341640740633011,
-0.028738709166646004,
-0.010945837944746017,
-0.025810716673731804,
0.08536335825920105,
-0.06620142608880997,
-0.04495561495423317,
0.12741070985794067,
-0.013661514036357403,
0.11073070019483566,
-0.023823490366339684,
0.03629240766167641,
0.05444595217704773,
-0.07070660591125488,
-0.005943047348409891,
0.11370768398046494,
-0.0029366223607212305,
0.024897973984479904,
0.05452774092555046,
-0.054081179201602936,
0.012292531318962574,
0.08894951641559601,
-0.040507134050130844,
-0.020783325657248497,
-0.00007855916919652373,
-0.027468306943774223,
-0.012470312416553497,
-0.0916627049446106,
-0.015386102721095085,
0.013407548889517784,
0.028417304158210754,
-0.014349937438964844,
0.017560552805662155,
0.06568882614374161,
0.05589950457215309,
-0.020850831642746925,
0.01364103239029646,
-0.015592451207339764,
-0.01599961519241333,
0.1041560098528862,
-0.0264019425958395,
0.03400750830769539,
-0.03083903342485428,
0.005628443788737059
] |
DenseGrounding: Improving Dense Language-Vision Semantics for Ego-centric 3D Visual Grounding | https://openreview.net/forum?id=iGafR0hSln | [
"Henry Zheng",
"Hao Shi",
"Qihang Peng",
"Yong Xien Chng",
"Rui Huang",
"Yepeng Weng",
"zhongchao shi",
"Gao Huang"
] | Poster | Enabling intelligent agents to comprehend and interact with 3D environments through natural language is crucial for advancing robotics and human-computer interaction. A fundamental task in this field is ego-centric 3D visual grounding, where agents locate target objects in real-world 3D spaces based on verbal descriptions. However, this task faces two significant challenges: (1) loss of fine-grained visual semantics due to sparse fusion of point clouds with ego-centric multi-view images, (2) limited textual semantic context due to arbitrary language descriptions. We propose DenseGrounding, a novel approach designed to address these issues by enhancing both visual and textual semantics. For visual features, we introduce the Hierarchical Scene Semantic Enhancer, which retains dense semantics by capturing fine-grained global scene features and facilitating cross-modal alignment. For text descriptions, we propose a Language Semantic Enhancer that leverage large language models to provide rich context and diverse language descriptions with additional context during model training. Extensive experiments show that DenseGrounding significantly outperforms existing methods in overall accuracy, achieving improvements of **5.81%** and **7.56%** when trained on the comprehensive full training dataset and smaller mini subset, respectively, further advancing the SOTA in ego-centric 3D visual grounding. Our method also achieves **1st place** and receives **Innovation Award** in the 2024 Autonomous Grand Challenge Multi-view 3D Visual Grounding Track, validating its effectiveness and robustness. | 3D Visual Grounding, Embodied AI, Egocentric Vision | DenseGrounding enhances visual and textual semantics to significantly improve ego-centric 3D visual grounding, outperforming existing methods. | 355 | null | [
0.015251570381224155,
-0.12087463587522507,
0.046014297753572464,
-0.018784068524837494,
0.13218218088150024,
-0.029735414311289787,
-0.012001469731330872,
-0.010828533209860325,
0.023427817970514297,
-0.06828394532203674,
-0.01481324341148138,
-0.08744869381189346,
0.025071024894714355,
0.08114167302846909,
0.07033312320709229,
0.04449085518717766,
0.14187510311603546,
0.052060745656490326,
-0.11521530896425247,
0.016380539163947105,
0.03985859826207161,
0.020346583798527718,
0.06898658722639084,
0.026951156556606293,
0.05157240480184555,
0.06171764060854912,
0.014895778149366379,
-0.05821123346686363,
0.0020440593361854553,
0.005772430915385485,
0.024685274809598923,
0.008113161660730839,
0.02840607613325119,
0.05371106043457985,
0.0034768241457641125,
0.06888124346733093,
-0.030488165095448494,
0.017363766208291054,
-0.04440893605351448,
-0.059935104101896286,
-0.002641710452735424,
-0.01163521595299244,
0.024362778291106224,
0.02349831350147724,
0.09223763644695282,
0.058791618794202805,
-0.0042155710980296135,
-0.053921107202768326,
0.11011496186256409,
-0.003742360509932041,
-0.07548855990171432,
-0.0474790558218956,
-0.008899969980120659,
0.014649194665253162,
-0.030477197840809822,
0.04422267898917198,
0.04012061282992363,
-0.04203590005636215,
0.058778274804353714,
-0.0007211153279058635,
0.013013185001909733,
-0.006673600059002638,
0.07375656813383102,
-0.0020294450223445892,
0.003027592319995165,
-0.06759010255336761,
0.018174035474658012,
-0.019362343475222588,
-0.05638173222541809,
-0.007698004599660635,
0.0035928585566580296,
0.01856147311627865,
0.05155451223254204,
-0.046773605048656464,
0.011503683403134346,
0.02380916103720665,
-0.024032140150666237,
-0.10907801240682602,
-0.019202105700969696,
-0.07279503345489502,
0.02249065972864628,
0.016154108569025993,
0.06680034846067429,
0.03637311980128288,
0.025904666632413864,
0.025499491021037102,
0.008425885811448097,
0.012113379314541817,
-0.07416504621505737,
-0.01766173355281353,
-0.06046631559729576,
-0.16554297506809235,
-0.011076792143285275,
-0.028165848925709724,
0.028428804129362106,
0.018885575234889984,
-0.07216422259807587,
-0.1779804527759552,
-0.03757137805223465,
0.04988318309187889,
0.04342874512076378,
-0.029817577451467514,
0.001843805774115026,
-0.0419280007481575,
-0.019796110689640045,
0.007353255990892649,
0.03645053133368492,
0.010534659028053284,
-0.03621305152773857,
-0.03620195761322975,
0.07323044538497925,
-0.03800562769174576,
0.005707739852368832,
-0.004578134045004845,
-0.00965370237827301,
0.037177614867687225,
0.041214246302843094,
-0.03248484805226326,
0.07381168007850647,
-0.03373456373810768,
-0.047618716955184937,
-0.0002079637924907729,
0.039841726422309875,
0.012705597095191479,
0.05752858147025108,
-0.051553115248680115,
0.0350160077214241,
-1.059889991548685e-33,
0.04383568838238716,
0.04408645257353783,
0.07188525795936584,
0.03494218736886978,
0.08117784559726715,
-0.02177051082253456,
-0.03364058956503868,
0.03671368584036827,
-0.054407503455877304,
0.0032845225650817156,
-0.018260056152939796,
-0.029340064153075218,
0.00007625664147781208,
0.05595431104302406,
-0.04909452423453331,
-0.04617936909198761,
-0.06419617682695389,
0.044158633798360825,
-0.019953303039073944,
0.07946062088012695,
0.01742938905954361,
-0.01385871134698391,
-0.03351221978664398,
0.026850098744034767,
0.0305415578186512,
-0.045792821794748306,
0.08415687084197998,
-0.05190577358007431,
0.0062033929862082005,
-0.0383613184094429,
-0.0980052500963211,
0.04876380041241646,
0.07034605741500854,
0.048904117196798325,
-0.02446543611586094,
-0.018460696563124657,
-0.09555348753929138,
-0.051359985023736954,
0.0005110892234370112,
0.03390437364578247,
-0.02065098099410534,
0.03339662402868271,
-0.03335565701127052,
-0.06322477012872696,
-0.04287483170628548,
-0.000818139873445034,
-0.0013742485316470265,
0.08462674170732498,
-0.0677884891629219,
0.000480043760035187,
0.07040603458881378,
0.024892281740903854,
-0.007618582341820002,
-0.03503001481294632,
-0.0369182825088501,
-0.06126386299729347,
0.030100151896476746,
0.003136163344606757,
0.07350346446037292,
-0.04115996137261391,
0.037947073578834534,
0.0023758509196341038,
-0.03617033734917641,
0.07985436916351318,
0.06941433250904083,
-0.01169787347316742,
0.0805017501115799,
0.11627128720283508,
0.0747375562787056,
0.03606891632080078,
-0.07308029383420944,
-0.005887042265385389,
0.03793983906507492,
-0.01601478084921837,
0.06732343137264252,
-0.050634726881980896,
0.02887457050383091,
-0.09158559143543243,
0.0515047162771225,
0.0685589611530304,
-0.03501063212752342,
-0.059555262327194214,
0.01813836209475994,
-0.0399400070309639,
-0.106417216360569,
-0.0000025725551040522987,
0.012302742339670658,
-0.09352913498878479,
0.021592389792203903,
-0.04343800246715546,
-0.03614017739892006,
-0.1364925056695938,
-0.028753843158483505,
-0.031946878880262375,
-0.03653167188167572,
7.242834156530449e-34,
0.040148358792066574,
-0.006343497894704342,
-0.05237550660967827,
0.023178650066256523,
0.0027637360617518425,
-0.07100880891084671,
0.04305753484368324,
0.0011551097268238664,
0.023506639525294304,
-0.03609481081366539,
-0.03616784140467644,
0.07995254546403885,
0.04085312783718109,
0.035277679562568665,
0.04777088388800621,
0.0017469192389398813,
-0.02726869285106659,
0.05577344074845314,
-0.0297674722969532,
0.03075505420565605,
0.04925363138318062,
0.029981600120663643,
-0.07089533656835556,
0.0151591831818223,
-0.04861186444759369,
0.018076784908771515,
-0.044028230011463165,
0.06033780798316002,
-0.000024171980840037577,
-0.0016514654271304607,
0.019709324464201927,
0.007760972715914249,
-0.006378741003572941,
-0.037912338972091675,
-0.08199040591716766,
0.016788039356470108,
0.01565621979534626,
-0.04611083120107651,
-0.08455348759889603,
0.04919568821787834,
0.07727943360805511,
-0.06233316287398338,
-0.10235174000263214,
-0.0007682076538912952,
-0.06923557072877884,
-0.017001718282699585,
-0.06429428607225418,
0.00622613076120615,
-0.00015140966570470482,
0.09684359282255173,
-0.08244209736585617,
0.02304147183895111,
-0.0515177883207798,
0.005467208102345467,
-0.02858222834765911,
-0.08685173839330673,
-0.050391774624586105,
-0.003697796957567334,
0.050504788756370544,
-0.036957670003175735,
-0.07370953261852264,
-0.010087525472044945,
-0.005169394891709089,
0.029145892709493637,
0.038792066276073456,
-0.015274854376912117,
-0.07728944718837738,
-0.017227478325366974,
0.046949952840805054,
0.03494630381464958,
0.02452622726559639,
-0.04273172840476036,
-0.0035585167352110147,
0.05711554363369942,
0.0144018130376935,
0.05288776382803917,
0.07553613185882568,
-0.0065727801993489265,
0.018585164099931717,
-0.046727053821086884,
0.044308919459581375,
-0.05433851107954979,
0.02816459722816944,
0.14332906901836395,
0.05267824977636337,
0.1032099649310112,
-0.022258669137954712,
0.06215502321720123,
-0.006641541142016649,
0.02362694963812828,
-0.04900546371936798,
-0.008873826824128628,
-0.010029217228293419,
0.050212785601615906,
-0.09880375117063522,
-3.2655329107456055e-8,
-0.08830151706933975,
0.10099083930253983,
-0.03101373463869095,
0.009150322526693344,
-0.03367515653371811,
-0.04404352232813835,
0.01567980833351612,
0.050739701837301254,
-0.010449030436575413,
-0.04825213551521301,
-0.03110526315867901,
-0.06028882414102554,
-0.03609313815832138,
-0.08907616138458252,
0.04034454748034477,
0.08361416310071945,
-0.020540660247206688,
0.020903756842017174,
0.01974303089082241,
0.050242774188518524,
0.0091921491548419,
0.053839001804590225,
0.032566141337156296,
0.029001576825976372,
-0.026061655953526497,
0.05189057067036629,
-0.020614517852663994,
0.12392361462116241,
0.05994424968957901,
-0.029726725071668625,
0.04735582321882248,
-0.008788611739873886,
0.00005093864820082672,
-0.026524802669882774,
-0.009302522987127304,
0.06874687224626541,
0.015546937473118305,
-0.018046239390969276,
0.011133082211017609,
-0.022357042878866196,
0.01712692156434059,
0.06938125193119049,
0.028148388490080833,
-0.009844728745520115,
0.027965931221842766,
0.04006749764084816,
-0.04154086858034134,
-0.12202826887369156,
-0.004538785200566053,
-0.03802036494016647,
-0.08416271954774857,
0.02783706597983837,
-0.03104713000357151,
0.1451408416032791,
0.05220334604382515,
0.01108524203300476,
-0.0074557350017130375,
-0.061058253049850464,
0.037915658205747604,
0.09857942909002304,
-0.059356238692998886,
0.030599987134337425,
-0.013449207879602909,
-0.0014078621752560139
] |
Bayesian Analysis of Combinatorial Gaussian Process Bandits | https://openreview.net/forum?id=50cmx4SrkM | [
"Jack Sandberg",
"Niklas Åkerblom",
"Morteza Haghir Chehreghani"
] | Poster | We consider the combinatorial volatile Gaussian process (GP) semi-bandit problem. Each round, an agent is provided a set of available base arms and must select a subset of them to maximize the long-term cumulative reward. We study the Bayesian setting and provide novel Bayesian cumulative regret bounds for three GP-based algorithms: GP-UCB, GP-BayesUCB and GP-TS. Our bounds extend previous results for GP-UCB and GP-TS to the \emph{infinite}, \emph{volatile} and \emph{combinatorial} setting, and to the best of our knowledge, we provide the first regret bound for GP-BayesUCB. Volatile arms encompass other widely considered bandit problems such as contextual bandits.
Furthermore, we employ our framework to address the challenging real-world problem of online energy-efficient navigation, where we demonstrate its effectiveness compared to the alternatives. | Multi-armed bandits, Combinatorial bandits, Contextual bandits, Gaussian processes, Energy-efficient navigation | We present novel Bayesian regret bounds for GP-UCB, GP-BayesUCB and GP-TS for the combinatorial volatile Gaussian process semi-bandit problem and study the application of online energy-efficient navigation. | 354 | 2312.12676 | [
-0.11064748466014862,
-0.028775786980986595,
0.02392098680138588,
-0.008479557931423187,
0.0216203723102808,
-0.002780656097456813,
0.08122797310352325,
0.023259885609149933,
0.046614497900009155,
0.00022674634237773716,
-0.05441887304186821,
0.0151776522397995,
0.10695073753595352,
-0.02013731747865677,
0.003133980091661215,
-0.014165438711643219,
0.11103853583335876,
0.03519556671380997,
0.02242124453186989,
0.007686242461204529,
-0.08993914723396301,
-0.004669663030654192,
0.05476674437522888,
-0.07167152315378189,
0.034577589482069016,
-0.02988383173942566,
0.04442508891224861,
-0.03901224955916405,
0.03246806189417839,
0.026740869507193565,
0.014706126414239407,
0.007502482272684574,
-0.008453060872852802,
-0.06279265135526657,
0.025017956271767616,
0.039069052785634995,
-0.1193394586443901,
-0.05261532589793205,
0.023668058216571808,
0.06724011898040771,
0.015046767890453339,
0.020748894661664963,
0.004823641385883093,
0.09276483952999115,
-0.029558401554822922,
-0.028972208499908447,
-0.030177541077136993,
-0.004413402173668146,
0.05916949734091759,
-0.06460446119308472,
-0.015650108456611633,
-0.007074381224811077,
-0.0051641762256622314,
0.008717852644622326,
-0.04357583075761795,
-0.001653605024330318,
-0.007661439478397369,
-0.04652147740125656,
0.014033861458301544,
0.0442701056599617,
0.01658431626856327,
-0.15255744755268097,
-0.030792057514190674,
-0.048545558005571365,
-0.04622591286897659,
0.019643789157271385,
0.009567705914378166,
0.05472409725189209,
0.011412407271564007,
0.10772722214460373,
0.030516518279910088,
0.06424636393785477,
-0.08988569676876068,
0.02643713727593422,
-0.0019318785052746534,
0.007794865407049656,
-0.015876982361078262,
-0.03969253599643707,
-0.03103494644165039,
-0.017874104902148247,
-0.04767431691288948,
-0.04689905047416687,
0.023429883643984795,
-0.040782611817121506,
0.07391035556793213,
-0.023319724947214127,
0.06631659716367722,
0.03156096488237381,
0.15662004053592682,
-0.016850950196385384,
-0.03490670025348663,
-0.011781408451497555,
0.027540935203433037,
-0.025032654404640198,
-0.03967462107539177,
-0.012506119906902313,
-0.025413068011403084,
-0.06004991754889488,
-0.016524605453014374,
0.04425384849309921,
0.02560950070619583,
0.07754355669021606,
0.003046890255063772,
-0.038625869899988174,
-0.04859187826514244,
-0.04126838222146034,
-0.052011679857969284,
0.09954234212636948,
0.09279332309961319,
-0.058004554361104965,
0.05278927832841873,
0.02807583473622799,
0.13048116862773895,
0.06039270758628845,
-0.06435526907444,
0.03418142348527908,
-0.015841810032725334,
0.01641068607568741,
-0.024564141407608986,
0.09716461598873138,
0.013946831226348877,
0.02101590298116207,
0.04792589321732521,
0.017219124361872673,
-0.00838086660951376,
-0.03770005702972412,
0.025156328454613686,
-1.3549845321756774e-33,
0.04849904403090477,
-0.012554595246911049,
0.06221119314432144,
-0.012210994027554989,
0.02439829893410206,
0.08481087535619736,
0.02912021055817604,
-0.06409252434968948,
-0.08148116618394852,
0.02527204342186451,
0.017648868262767792,
0.08673875778913498,
-0.10119336098432541,
0.07790432870388031,
0.11084141582250595,
-0.015726642683148384,
-0.045973196625709534,
0.01189902238547802,
0.06404703110456467,
-0.07553305476903915,
0.07909383624792099,
-0.004733304493129253,
-0.0013296363176777959,
-0.010149985551834106,
-0.02083628624677658,
-0.03596172481775284,
-0.06876240670681,
-0.033301059156656265,
0.010811196640133858,
0.016890501603484154,
-0.09951336681842804,
0.03456880524754524,
-0.04891688749194145,
0.09989872574806213,
0.0033196841832250357,
-0.018774161115288734,
-0.03914998099207878,
-0.012491066940128803,
-0.02582891657948494,
-0.02179602161049843,
-0.06398169696331024,
0.020176541060209274,
0.09981079399585724,
-0.039427500218153,
-0.11322423070669174,
-0.11637800186872482,
0.008309132419526577,
-0.01913612335920334,
-0.024010945111513138,
-0.033546023070812225,
-0.0062908148393034935,
-0.020105483010411263,
0.022863440215587616,
0.031065626069903374,
-0.0818830206990242,
-0.0015261339722201228,
-0.0005858623189851642,
0.02513992413878441,
-0.024073757231235504,
0.023973964154720306,
0.11717917025089264,
0.02477601356804371,
0.04148975759744644,
-0.0006734602502547204,
-0.04466557130217552,
0.06804406642913818,
-0.03958422690629959,
0.008201943710446358,
0.04379081726074219,
0.02240736223757267,
0.011987311765551567,
-0.02366454340517521,
0.09921418875455856,
-0.060287874191999435,
0.08122255653142929,
0.0005796122131869197,
0.02199501544237137,
-0.011356201954185963,
-0.011786749586462975,
-0.042782045900821686,
-0.02117600105702877,
-0.0608057975769043,
-0.026274099946022034,
-0.0313740111887455,
0.006667389068752527,
-0.04271543771028519,
-0.03808844834566116,
-0.10081301629543304,
-0.050656139850616455,
0.02022665925323963,
-0.1148238331079483,
-0.002322931308299303,
-0.02292720042169094,
0.04203249514102936,
-0.061464205384254456,
-3.8289659883485057e-34,
0.011790141463279724,
-0.08156778663396835,
0.08452581614255905,
0.027595622465014458,
0.042619407176971436,
-0.06619099527597427,
0.007966091856360435,
-0.07358808815479279,
-0.008944086730480194,
-0.03961169719696045,
-0.08800750225782394,
0.10556329041719437,
0.025691011920571327,
0.004831445869058371,
0.049820952117443085,
-0.08488187193870544,
0.09072721749544144,
0.01559264026582241,
-0.06725472956895828,
-0.10841475427150726,
0.014857332222163677,
0.0004645395092666149,
-0.049481868743896484,
-0.017258627340197563,
0.034661952406167984,
0.04047182574868202,
0.06319740414619446,
0.0407230406999588,
-0.029013175517320633,
0.05223069712519646,
-0.04824022576212883,
0.010039825923740864,
-0.06782513856887817,
0.015431995503604412,
-0.027006806805729866,
0.01980874128639698,
0.126663476228714,
0.12029951065778732,
-0.06596685945987701,
0.09648407250642776,
0.08149448037147522,
-0.04866918548941612,
0.052659597247838974,
0.001031228108331561,
0.050452787429094315,
0.03531275689601898,
-0.046611446887254715,
0.02625587210059166,
0.010603928938508034,
-0.03962060064077377,
0.04727175086736679,
0.0909234955906868,
-0.06582390516996384,
0.0944444090127945,
-0.08021596819162369,
-0.04780546575784683,
-0.04468199238181114,
-0.06468116492033005,
0.10109665989875793,
0.029823968186974525,
-0.0049477312713861465,
0.0376005582511425,
-0.04124424234032631,
0.04638087376952171,
-0.02891029603779316,
-0.013554036617279053,
-0.062007613480091095,
0.021930059418082237,
-0.010639544576406479,
-0.023558637127280235,
-0.02324954979121685,
0.038071196526288986,
0.016408683732151985,
-0.00849159061908722,
-0.008199251256883144,
0.09788931161165237,
0.020175877958536148,
0.011241949163377285,
0.02501409873366356,
-0.023314587771892548,
-0.03388240188360214,
0.046895675361156464,
0.02581147663295269,
-0.0033032826613634825,
0.026090623810887337,
0.032326940447092056,
0.004525937605649233,
-0.03076159954071045,
0.006726942490786314,
-0.08598928898572922,
-0.0017076429212465882,
-0.07579529285430908,
0.05622147396206856,
0.044041622430086136,
-0.06547126919031143,
-3.930544778540934e-8,
-0.028029201552271843,
0.05363960936665535,
0.07514262944459915,
0.02111845090985298,
0.11191003024578094,
0.018377821892499924,
-0.009662996977567673,
0.05460461229085922,
-0.04445486143231392,
-0.012870381586253643,
0.08738060295581818,
-0.00401971023529768,
0.008303465321660042,
0.004849825985729694,
-0.03378162905573845,
0.005146705079823732,
-0.02904180996119976,
-0.07019704580307007,
0.011460908688604832,
-0.04256390035152435,
0.015453672967851162,
-0.011568335816264153,
0.04784923046827316,
-0.04196427762508392,
-0.023612959310412407,
-0.008732309564948082,
-0.01790880598127842,
0.04936351999640465,
0.053264085203409195,
0.10164905339479446,
-0.02635522559285164,
-0.010921594686806202,
0.0004825541691388935,
0.016252698376774788,
-0.003651473904028535,
-0.03784209489822388,
-0.11946314573287964,
-0.024023007601499557,
0.01501118391752243,
-0.05631354823708534,
-0.01966036483645439,
-0.01611252687871456,
-0.02550475113093853,
0.012981841340661049,
-0.00706253107637167,
0.011702793650329113,
-0.0345129631459713,
-0.012568306177854538,
0.05265772342681885,
-0.02553836815059185,
0.0016466075321659446,
-0.024861734360456467,
-0.014613406732678413,
0.06805044412612915,
0.0731462687253952,
0.007596744690090418,
-0.0017749045509845018,
-0.12326713651418686,
-0.005303066689521074,
0.009543758817017078,
-0.028401849791407585,
-0.00786727573722601,
-0.11583392322063446,
0.03926048055291176
] |
Test-time Adaptation for Image Compression with Distribution Regularization | https://openreview.net/forum?id=bsnRUkVn63 | [
"Kecheng Chen",
"Pingping Zhang",
"Tiexin Qin",
"Shiqi Wang",
"Hong Yan",
"Haoliang Li"
] | Poster | Current test- or compression-time adaptation image compression (TTA-IC) approaches, which leverage both latent and decoder refinements as a two-step adaptation scheme, have potentially enhanced the rate-distortion (R-D) performance of learned image compression models on cross-domain compression tasks, \textit{e.g.,} from natural to screen content images. However, compared with the emergence of various decoder refinement variants, the latent refinement, as an inseparable ingredient, is barely
tailored to cross-domain scenarios. To this end, we are interested in developing an advanced latent refinement method by extending the effective hybrid latent refinement (HLR) method, which is designed for \textit{in-domain} inference improvement but shows noticeable degradation of the rate cost in \textit{cross-domain} tasks. Specifically, we first provide theoretical analyses, in a cue of marginalization approximation from in- to cross-domain scenarios, to uncover that the vanilla HLR suffers from an underlying mismatch between refined Gaussian conditional and hyperprior distributions, leading to deteriorated joint probability approximation of marginal distribution with increased rate consumption. To remedy this issue, we introduce a simple Bayesian approximation-endowed \textit{distribution regularization} to encourage learning a better joint probability approximation in a plug-and-play manner. Extensive experiments on six in- and cross-domain datasets demonstrate that our proposed method not only improves the R-D performance compared with other latent refinement counterparts, but also can be flexibly integrated into existing TTA-IC methods with incremental benefits. | test-time adaptation, image compression, entropy coding | null | 347 | 2410.12191 | [
-0.02566417120397091,
-0.039712343364953995,
0.0463750921189785,
0.015136667527258396,
0.16411294043064117,
0.0023649497888982296,
-0.05807340145111084,
-0.06350275129079819,
0.016951080411672592,
-0.0012761959806084633,
-0.012271805666387081,
-0.03317037597298622,
0.006476039066910744,
0.03506823256611824,
-0.01475889515131712,
-0.0956580862402916,
0.04387965798377991,
0.1375952661037445,
-0.09617605060338974,
-0.04756626859307289,
0.04302074387669563,
-0.008361052721738815,
0.05893769487738609,
-0.05012346804141998,
0.12037507444620132,
-0.016989359632134438,
0.03583432361483574,
-0.026211518794298172,
0.01433106604963541,
-0.030703533440828323,
0.023623567074537277,
0.03175521641969681,
-0.06244468688964844,
-0.0024400390684604645,
-0.07449404150247574,
0.08305391669273376,
0.002939585829153657,
-0.007732215337455273,
0.021523797884583473,
0.06415550410747528,
-0.0170164592564106,
0.06725150346755981,
-0.012573989108204842,
0.014946750365197659,
0.02203000709414482,
-0.04785295948386192,
0.03357191011309624,
-0.06913931667804718,
-0.013238179497420788,
0.02896989695727825,
0.03129157796502113,
0.08259843289852142,
-0.08222988247871399,
0.0569973923265934,
-0.012616354040801525,
0.050691016018390656,
0.04424586147069931,
0.0021157506853342056,
0.03133954852819443,
0.04904176667332649,
-0.04733162745833397,
-0.080201156437397,
0.02963370829820633,
0.01949922740459442,
0.027348347008228302,
-0.03251657262444496,
0.033875420689582825,
-0.007648601196706295,
0.0021553365513682365,
-0.06028810143470764,
-0.14247675240039825,
0.046341922134160995,
-0.046849869191646576,
-0.0016456925077363849,
0.033049311488866806,
0.06979429721832275,
0.0683913379907608,
0.03628402203321457,
0.0657808929681778,
-0.09407087415456772,
0.036450352519750595,
0.003946795593947172,
0.020563632249832153,
-0.023954827338457108,
0.1317700892686844,
-0.0025833281688392162,
-0.001713273231871426,
0.07540689408779144,
-0.01254185289144516,
-0.009528237394988537,
0.011624951846897602,
-0.018126215785741806,
0.013159723952412605,
0.03644310683012009,
0.015003220178186893,
-0.019448304548859596,
-0.012445660308003426,
-0.06585068255662918,
0.04360859468579292,
0.07904144376516342,
0.005918440409004688,
0.025541674345731735,
0.02352972887456417,
0.02186831831932068,
-0.048734162002801895,
-0.12111146003007889,
0.1002516821026802,
0.11335944384336472,
0.001105248462408781,
0.0013762809103354812,
0.0031131471041589975,
0.03031078167259693,
0.05208739638328552,
-0.02717740647494793,
0.016602138057351112,
-0.09334823489189148,
-0.033092595636844635,
-0.05744101107120514,
0.04226187989115715,
-0.021177884191274643,
-0.052872948348522186,
0.020310837775468826,
-0.05096666142344475,
-0.0487266406416893,
0.09833413362503052,
-0.011244709603488445,
-0.008415528573095798,
2.5675781275583894e-33,
0.00251377047970891,
0.08943987637758255,
0.048612043261528015,
0.022127337753772736,
-0.007751008961349726,
0.018923064693808556,
-0.016666226089000702,
0.01259557530283928,
0.056956175714731216,
-0.004491646774113178,
-0.030257008969783783,
-0.014622382819652557,
-0.10841917991638184,
0.05584103614091873,
-0.0014133104123175144,
0.05498332902789116,
-0.12002475559711456,
0.06765420734882355,
0.007172057870775461,
0.035299152135849,
-0.016043497249484062,
-0.0029676933772861958,
-0.023295946419239044,
-0.08473687618970871,
-0.021231861785054207,
-0.08518350124359131,
0.04060498625040054,
-0.07401233166456223,
-0.01550512108951807,
-0.0179466400295496,
-0.14239166676998138,
0.01583833619952202,
0.016051655635237694,
0.03428689017891884,
-0.0164792463183403,
-0.03720688819885254,
0.07383088767528534,
-0.04681658372282982,
-0.08622009307146072,
0.009900450706481934,
-0.024116674438118935,
0.07174757122993469,
0.024556569755077362,
-0.05229523405432701,
-0.04387238994240761,
-0.11356057971715927,
0.06309336423873901,
0.0019164681434631348,
-0.01821863278746605,
-0.01886448822915554,
0.04852752387523651,
0.04339256510138512,
0.030625661835074425,
-0.04390879347920418,
-0.05092386528849602,
0.005239041522145271,
0.047163061797618866,
-0.03518283739686012,
0.08427181094884872,
0.04857870936393738,
-0.015987679362297058,
0.03182295337319374,
-0.03376723825931549,
0.1262868493795395,
0.007989090867340565,
0.006824736017733812,
0.009961862117052078,
0.0015629170229658484,
0.002311726799234748,
0.048725780099630356,
-0.018936196342110634,
0.031243130564689636,
-0.05284091457724571,
-0.0408756360411644,
0.04060531035065651,
-0.05713346600532532,
0.017564885318279266,
0.007250390946865082,
-0.012107293121516705,
-0.010768787935376167,
-0.03496961295604706,
0.00599999725818634,
-0.04322526231408119,
-0.02043652907013893,
-0.006567345000803471,
-0.02283092588186264,
0.02118554525077343,
-0.11083795875310898,
0.005825777538120747,
-0.07217776030302048,
0.0003315034555271268,
-0.03462592139840126,
-0.030565708875656128,
-0.030808627605438232,
0.07446733862161636,
-1.3772967924867268e-33,
-0.030001342296600342,
0.05074813961982727,
-0.026818858459591866,
0.04051124304533005,
0.019941600039601326,
-0.06084928289055824,
-0.0032617838587611914,
0.029074257239699364,
0.047642335295677185,
-0.1355632245540619,
0.005791554227471352,
-0.0932331532239914,
-0.025235850363969803,
-0.05123360455036163,
-0.04635007679462433,
0.01897905394434929,
-0.035660095512866974,
-0.0038023011293262243,
-0.030505970120429993,
-0.03310913220047951,
0.101848304271698,
0.0522378534078598,
0.028355427086353302,
0.011854352429509163,
0.016263553872704506,
0.033499374985694885,
-0.010014879517257214,
0.09920389950275421,
0.027127793058753014,
-0.06084458529949188,
0.021434884518384933,
-0.032184530049562454,
-0.03733713552355766,
0.022476762533187866,
-0.008061081171035767,
0.06737145036458969,
-0.00044431991409510374,
0.023182855919003487,
-0.052220407873392105,
0.12128061801195145,
0.08901996910572052,
-0.0060104141011834145,
-0.08634533733129501,
0.012345312163233757,
0.03901495411992073,
0.024778923019766808,
-0.044495612382888794,
-0.09727589786052704,
0.1359281688928604,
0.07177010178565979,
-0.012691483832895756,
0.021648406982421875,
-0.07098095118999481,
0.03307739272713661,
-0.028061380609869957,
-0.07719878852367401,
-0.045083124190568924,
-0.0300301443785429,
0.08543232083320618,
0.0237283892929554,
-0.019395392388105392,
-0.022199956700205803,
-0.005283537320792675,
-0.07883691787719727,
-0.028372840955853462,
-0.08463077247142792,
-0.019527066498994827,
0.024716002866625786,
0.06630242615938187,
0.07925610989332199,
0.0448480099439621,
-0.039495933800935745,
0.07534924149513245,
0.03598304092884064,
0.08051721006631851,
-0.024763938039541245,
0.06772703677415848,
0.0379975400865078,
-0.002887096954509616,
-0.0265426654368639,
0.03345554694533348,
0.0014506883453577757,
-0.017990456894040108,
0.0619097463786602,
0.018662618473172188,
0.14377442002296448,
-0.0002262427587993443,
-0.008721784688532352,
0.05031639337539673,
-0.053747594356536865,
-0.0446539930999279,
0.042386699467897415,
0.022460883483290672,
0.0438205860555172,
-0.07538396865129471,
-4.221033833573529e-8,
-0.018578460440039635,
0.03597123920917511,
-0.024447251111268997,
-0.00440917769446969,
-0.035138096660375595,
-0.05554712191224098,
-0.01792140118777752,
0.017995523288846016,
-0.013021018356084824,
-0.08324291557073593,
0.10335088521242142,
-0.05427961051464081,
0.03135876730084419,
-0.06694184243679047,
-0.011777645908296108,
0.07142581790685654,
0.025921763852238655,
0.006484657991677523,
-0.001105890958569944,
-0.04741060361266136,
-0.07056820392608643,
0.04617389291524887,
0.05489214509725571,
-0.020504234358668327,
0.00887544546276331,
-0.0164352897554636,
0.038330789655447006,
0.006898873019963503,
0.06551503390073776,
-0.014551541768014431,
0.004315125290304422,
0.027276063337922096,
-0.02376917190849781,
-0.038149382919073105,
-0.03965560719370842,
0.03901489078998566,
0.08278213441371918,
0.005531775299459696,
0.008908913470804691,
0.07453196495771408,
-0.05386385694146156,
-0.008325139991939068,
-0.06823323667049408,
-0.03202982246875763,
0.07190394401550293,
-0.03893459960818291,
0.01703069917857647,
-0.04080815985798836,
-0.03605400025844574,
-0.07954858988523483,
0.04810426011681557,
-0.009471341967582703,
-0.02072717621922493,
-0.0014591952785849571,
0.12698793411254883,
-0.04850853607058525,
0.07467576116323471,
-0.0020121117122471333,
0.04040955379605293,
0.03262867033481598,
0.0035638075787574053,
0.02770715206861496,
-0.02191511169075966,
0.010144323110580444
] |
Controllable Blur Data Augmentation Using 3D-Aware Motion Estimation | https://openreview.net/forum?id=Wvi8c0tgvt | [
"Insoo Kim",
"Hana Lee",
"Hyong-Euk Lee",
"Jinwoo Shin"
] | Poster | Existing realistic blur datasets provide insufficient variety in scenes and blur patterns to be trained, while expanding data diversity demands considerable time and effort due to complex dual-camera systems. To address the challenge, data augmentation can be an effective way to artificially increase data diversity. However, existing methods on this line are typically designed to estimate motions from a 2D perspective, e.g., estimating 2D non-uniform kernels disregarding 3D aspects of blur modeling, which leads to unrealistic motion patterns due to the fact that camera and object motions inherently arise in 3D space. In this paper, we propose a 3D-aware blur synthesizer capable of generating diverse and realistic blur images for blur data augmentation. Specifically, we estimate 3D camera positions within the motion blur interval, generate the corresponding scene images, and aggregate them to synthesize a realistic blur image. Since the 3D camera positions projected onto the 2D image plane inherently lie in 2D space, we can represent the 3D transformation as a combination of 2D transformation and projected 3D residual component. This allows for 3D transformation without requiring explicit depth measurements, as the 3D residual component is directly estimated via a neural network. Furthermore, our blur synthesizer allows for controllable blur data augmentation by modifying blur magnitude, direction, and scenes, resulting in diverse blur images. As a result, our method significantly improves deblurring performance, making it more practical for real-world scenarios. | Blur synthesis, Data augmentation, Blind motion deblurring, 3D motion modeling | We propose a controllable 3D-aware blur synthesizer to generate diverse blur images, improving deblurring performance. | 339 | null | [
-0.08068858087062836,
-0.1472312957048416,
0.018737917765975,
0.035422999411821365,
0.006989750545471907,
-0.013492667116224766,
-0.013982455246150494,
-0.07177845388650894,
-0.00866256095468998,
-0.03647992014884949,
-0.03074241802096367,
-0.08022072166204453,
-0.008973881602287292,
-0.04008391499519348,
-0.02127859555184841,
0.011471770703792572,
0.07333588600158691,
0.12668931484222412,
-0.08920129388570786,
-0.03630067780613899,
0.0943041518330574,
-0.017828520387411118,
0.0055228229612112045,
0.06087746098637581,
0.07950638979673386,
-0.013647370971739292,
0.05717803165316582,
-0.04995844140648842,
-0.03542497754096985,
0.0008233590633608401,
0.014006311073899269,
0.004745472688227892,
-0.02481788769364357,
-0.042154710739851,
-0.01001594215631485,
-0.0015171865234151483,
0.01981731317937374,
-0.0004915035679005086,
-0.019403299316763878,
-0.04467837139964104,
-0.001931138918735087,
0.04868776723742485,
-0.02494939975440502,
0.0457259826362133,
0.056608110666275024,
0.09793050587177277,
0.08289038389921188,
-0.09041395038366318,
0.0693991556763649,
0.06186400726437569,
-0.02753431536257267,
-0.030299656093120575,
0.07769341766834259,
0.056739501655101776,
0.014872937463223934,
0.02621413767337799,
-0.004913909826427698,
0.010223074816167355,
0.06896179169416428,
0.050618913024663925,
0.022889427840709686,
-0.003938811365514994,
0.07464908063411713,
-0.014705255627632141,
-0.03699488937854767,
-0.04203334450721741,
0.07332538068294525,
0.010114607401192188,
-0.06157011538743973,
-0.035213541239500046,
-0.08687462657690048,
0.020246975123882294,
0.04124637693166733,
0.021046457812190056,
0.02226938121020794,
-0.03340721130371094,
0.021007820963859558,
-0.10688047856092453,
-0.005852072965353727,
-0.055289529263973236,
0.07122468203306198,
-0.046851545572280884,
-0.009505939669907093,
0.033689435571432114,
-0.055040620267391205,
0.01973472163081169,
0.012243062257766724,
0.0846063494682312,
-0.025938855484128,
0.05483308434486389,
-0.0858326107263565,
-0.006973987445235252,
-0.02307492308318615,
-0.09733038395643234,
0.04728114977478981,
-0.015993451699614525,
0.018481263890862465,
-0.04697735980153084,
0.0749596580862999,
0.009427846409380436,
0.006483638193458319,
-0.04047869145870209,
0.021006057038903236,
-0.021524662151932716,
0.09928932040929794,
0.03628013655543327,
0.04382289573550224,
0.00026826848625205457,
0.06833209842443466,
0.012598047964274883,
0.02996165305376053,
-0.023411497473716736,
0.004793860483914614,
-0.037072475999593735,
0.07951406389474869,
0.05875042453408241,
0.00023725026403553784,
-0.02735971286892891,
-0.03193573281168938,
-0.016919581219553947,
-0.015385297127068043,
-0.006086862646043301,
-0.08769901841878891,
0.05851821228861809,
0.006054903846234083,
-0.016184410080313683,
-0.05657001584768295,
1.2955241708202474e-33,
0.05752120167016983,
0.11992621421813965,
0.047198936343193054,
0.0613323412835598,
-0.0004078514175489545,
-0.03526351600885391,
-0.001909303362481296,
0.009504764340817928,
0.019678181037306786,
-0.025966176763176918,
-0.02288816124200821,
-0.06756842136383057,
-0.015119054354727268,
0.0576242171227932,
0.016495877876877785,
-0.0045357877388596535,
-0.020063400268554688,
0.04176017642021179,
0.014219176955521107,
0.11483187228441238,
-0.0009167582029476762,
-0.0018230123678222299,
-0.034317925572395325,
0.032870013266801834,
-0.02063535712659359,
-0.013033339753746986,
0.03134249150753021,
-0.017669830471277237,
0.050011564046144485,
-0.05220472812652588,
-0.023168807849287987,
0.025596728548407555,
0.03933390602469444,
-0.03843842074275017,
0.04630713909864426,
-0.05545898899435997,
-0.013765949755907059,
0.002181680640205741,
-0.03432460501790047,
-0.002644406398758292,
-0.0013011581031605601,
0.08527268469333649,
-0.08003173768520355,
-0.013286520726978779,
-0.015608754009008408,
-0.02628989703953266,
0.04170869290828705,
0.12207194417715073,
-0.048659760504961014,
0.02344510518014431,
0.0744558796286583,
0.0019812604878097773,
0.018015282228589058,
-0.11183513700962067,
0.004748002626001835,
-0.06705577671527863,
0.01860698312520981,
-0.03227922320365906,
0.09821194410324097,
0.020231446251273155,
0.02499069646000862,
-0.012838653288781643,
-0.027979135513305664,
0.043341655284166336,
-0.014880123548209667,
0.0336005762219429,
0.00675506517291069,
0.07740981131792068,
0.011021485552191734,
-0.005207158625125885,
-0.040590304881334305,
-0.01452118530869484,
-0.030652372166514397,
-0.03555293008685112,
0.05005394667387009,
0.023524373769760132,
0.00024836123338900506,
-0.06144842505455017,
0.0303511805832386,
0.02816794067621231,
-0.04946919530630112,
-0.027734383940696716,
0.009413250721991062,
-0.012518013827502728,
-0.029848653823137283,
-0.02385965920984745,
-0.045779407024383545,
-0.019327186048030853,
-0.08436691761016846,
-0.044163208454847336,
-0.05355418846011162,
-0.0957789197564125,
-0.042859096080064774,
-0.04897312819957733,
-0.009774432517588139,
-4.808716422941141e-34,
0.09926383197307587,
0.06660666316747665,
-0.065968357026577,
0.038690537214279175,
0.0025601107627153397,
-0.0039812601171433926,
0.0032749921083450317,
-0.013032292947173119,
0.09138111025094986,
-0.013663189485669136,
-0.032940663397312164,
-0.004911027383059263,
0.08838695287704468,
-0.009581479243934155,
-0.03178678825497627,
-0.008872694335877895,
-0.0012157357996329665,
-0.06326227635145187,
-0.12351026386022568,
-0.01861688122153282,
0.005856378469616175,
0.033436957746744156,
0.058302368968725204,
0.013291295617818832,
-0.07774361222982407,
0.07785927504301071,
-0.0880136489868164,
0.0978889986872673,
0.030455131083726883,
-0.001059395493939519,
0.021826069802045822,
-0.07028353214263916,
0.05849817395210266,
-0.036944009363651276,
-0.08546169847249985,
0.021923275664448738,
0.07629362493753433,
-0.015796132385730743,
-0.09215927869081497,
-0.027518106624484062,
-0.013840030878782272,
-0.08208142220973969,
-0.023530159145593643,
0.06405702233314514,
-0.04331643506884575,
0.038986049592494965,
0.028669292107224464,
0.0015066618798300624,
0.01637699455022812,
0.06131347641348839,
-0.0743679478764534,
0.0880221351981163,
-0.08644698560237885,
-0.0008620643056929111,
0.03745116665959358,
-0.09545949846506119,
-0.13981066644191742,
-0.0035044753458350897,
0.11910492181777954,
-0.025994552299380302,
-0.03900687023997307,
-0.06189247965812683,
-0.06918061524629593,
0.014154259115457535,
-0.02887660637497902,
0.03582478687167168,
-0.04576561972498894,
-0.022273769602179527,
0.021087875589728355,
0.028219930827617645,
0.06278563290834427,
-0.011149115860462189,
0.05162197723984718,
-0.006996763404458761,
-0.009260863997042179,
-0.07407902926206589,
0.04434499517083168,
-0.0394013486802578,
0.019394526258111,
0.0047663068398833275,
0.05958189442753792,
-0.07876581698656082,
0.08299016952514648,
0.10622274875640869,
0.06067390739917755,
0.040229663252830505,
-0.040864214301109314,
-0.003927656449377537,
0.014602185226976871,
0.04647734388709068,
-0.05771920830011368,
0.024107230827212334,
-0.07646278291940689,
-0.008518338203430176,
-0.03305961191654205,
-3.827073058459973e-8,
-0.006960293278098106,
0.07819349318742752,
0.06424159556627274,
-0.015048455446958542,
-0.010798001661896706,
-0.07075940817594528,
0.059101421386003494,
0.12424509972333908,
0.014129444025456905,
-0.10216809809207916,
0.01966380886733532,
0.02225347049534321,
0.04624349623918533,
-0.04941611737012863,
0.07753356546163559,
0.03895079717040062,
-0.032181937247514725,
0.0069379121996462345,
0.0278937965631485,
0.08929748088121414,
-0.07160181552171707,
-0.014174070209264755,
-0.040551818907260895,
-0.011340826749801636,
0.0358344241976738,
-0.005507826805114746,
-0.02284551039338112,
-0.0015043155290186405,
0.1050836592912674,
0.007185548543930054,
0.008512449450790882,
-0.03911356255412102,
0.05171231925487518,
-0.0308651365339756,
-0.07921858131885529,
-0.00010517266491660848,
-0.028174040839076042,
0.0873742401599884,
-0.03108154982328415,
-0.07551193237304688,
-0.08015985041856766,
0.11708255112171173,
0.018190642818808556,
0.04542185366153717,
-0.00989212840795517,
-0.01333707571029663,
0.07435315102338791,
-0.0749259665608406,
0.029691699892282486,
-0.09078574925661087,
-0.042336322367191315,
0.002478501060977578,
0.062614805996418,
0.08497574180364609,
0.10179704427719116,
0.0068956464529037476,
-0.010720448568463326,
-0.01448075007647276,
0.02174418419599533,
0.08603502810001373,
-0.04201449453830719,
-0.04347263649106026,
-0.01053621992468834,
-0.009093105792999268
] |
MaxCutPool: differentiable feature-aware Maxcut for pooling in graph neural networks | https://openreview.net/forum?id=xlbXRJ2XCP | [
"Carlo Abate",
"Filippo Maria Bianchi"
] | Poster | We propose a novel approach to compute the MAXCUT in attributed graphs, i.e., graphs with features associated with nodes and edges. Our approach works well on any kind of graph topology and can find solutions that jointly optimize the MAXCUT along with other objectives. Based on the obtained MAXCUT partition, we implement a hierarchical graph pooling layer for Graph Neural Networks, which is sparse, trainable end-to-end, and particularly suitable for downstream tasks on heterophilic graphs. | Graph neural networks, graph pooling, graph coarsening, maxcut | A GNN-based approach for computing MAXCUT on attributed graphs, used to implement graph pooling. | 337 | 2409.05100 | [
-0.023644614964723587,
-0.0790662169456482,
0.05282701551914215,
0.015837904065847397,
0.02471008524298668,
0.013909822329878807,
-0.042932700365781784,
-0.030054058879613876,
-0.11095577478408813,
-0.022946637123823166,
0.0012434370582923293,
-0.07437422126531601,
-0.006660982966423035,
0.09506189823150635,
-0.04276411980390549,
0.07611653953790665,
0.11127787083387375,
0.05859021469950676,
-0.04117446392774582,
-0.0032697657588869333,
0.007298027630895376,
-0.0146543700248003,
-0.03944388031959534,
-0.008893430233001709,
0.09177695959806442,
0.04732554033398628,
-0.0601465106010437,
-0.03749191761016846,
0.07074487209320068,
-0.023902539163827896,
0.057288214564323425,
-0.03724629059433937,
-0.04086323454976082,
0.04818883538246155,
-0.004248581826686859,
0.03775840625166893,
-0.06273398548364639,
-0.03474131226539612,
0.059825170785188675,
0.005559437442570925,
0.01432162057608366,
0.02703329361975193,
-0.0042414735071361065,
0.03305567800998688,
0.07647933810949326,
0.08765248954296112,
-0.05311410874128342,
-0.003816756885498762,
-0.012594696134328842,
0.05720863491296768,
-0.023816080763936043,
-0.0061324117705225945,
-0.04220922663807869,
0.026190949603915215,
0.05161210894584656,
-0.03041018359363079,
0.03228078410029411,
0.019137363880872726,
0.028722360730171204,
0.06859796494245529,
0.04661448672413826,
-0.10202153772115707,
0.006744758225977421,
-0.03551766648888588,
-0.019289249554276466,
-0.017200151458382607,
-0.005940343253314495,
0.008372414857149124,
0.02084866352379322,
0.0703955665230751,
0.06327778100967407,
0.10431499034166336,
-0.08428100496530533,
-0.025938749313354492,
0.04071985185146332,
0.0682070329785347,
0.05677071213722229,
0.053113147616386414,
-0.03882201761007309,
-0.10099659860134125,
-0.03864205256104469,
0.0029081213288009167,
0.04633255675435066,
-0.030490167438983917,
0.057332344353199005,
0.01681061089038849,
-0.12283802032470703,
0.03763262927532196,
-0.04195016995072365,
0.03986600786447525,
-0.09760699421167374,
0.03200850263237953,
-0.021029654890298843,
-0.056963928043842316,
-0.03248324990272522,
-0.0010501854121685028,
-0.03444284945726395,
-0.006217717658728361,
-0.08539066463708878,
0.06053600087761879,
-0.05859244987368584,
-0.005643936339765787,
0.04551774263381958,
-0.032877273857593536,
0.09284038096666336,
0.04214669391512871,
-0.016460414975881577,
0.019764533266425133,
-0.024901363998651505,
-0.10499073565006256,
0.001782899838872254,
0.09222766011953354,
0.05973797291517258,
-0.031661879271268845,
0.014545822516083717,
0.010697088204324245,
0.0030628349632024765,
-0.0618487112224102,
0.06155696511268616,
0.13703246414661407,
-0.06869161874055862,
0.024903660640120506,
-0.0662686675786972,
0.04258362576365471,
-0.03209923580288887,
0.038621433079242706,
-0.011356629431247711,
1.1701150866348332e-34,
0.10868911445140839,
0.03449789434671402,
0.003565300954505801,
-0.07429282367229462,
0.04907698929309845,
0.015923550352454185,
0.02063009887933731,
0.029223905876278877,
0.015095225535333157,
-0.04784834757447243,
-0.08937575668096542,
-0.03426966071128845,
-0.04916742071509361,
0.10328961163759232,
0.020339662209153175,
-0.03404608741402626,
0.021013807505369186,
0.01996031403541565,
0.0072472491301596165,
-0.10507947951555252,
0.03437511995434761,
-0.04393867030739784,
0.0008922598790377378,
0.029811672866344452,
0.03256945312023163,
-0.013998875394463539,
0.019545936957001686,
0.019299359992146492,
-0.010802438482642174,
0.028525574132800102,
-0.07701479643583298,
0.026679757982492447,
0.005980269052088261,
0.05525670573115349,
-0.01401435025036335,
0.0007910634158179164,
0.006191158667206764,
-0.015900632366538048,
0.003788664937019348,
-0.019312042742967606,
0.03126199170947075,
0.0006438023410737514,
-0.0118188988417387,
-0.043668825179338455,
-0.09612853080034256,
-0.024533938616514206,
-0.0014151943614706397,
-0.007240117061883211,
-0.07633937895298004,
-0.004694332834333181,
-0.03329119458794594,
-0.048436231911182404,
0.016772979870438576,
0.028928060084581375,
-0.01662031002342701,
-0.07879967987537384,
-0.014149431139230728,
0.07850093394517899,
0.04623086377978325,
0.03822599723935127,
0.02950422279536724,
0.01589152403175831,
-0.06040851026773453,
0.0894567221403122,
-0.04460042342543602,
0.06184237450361252,
0.004104825668036938,
0.04540831223130226,
0.08227582275867462,
-0.04576348140835762,
-0.07732973247766495,
0.0073632244020700455,
-0.04374176263809204,
-0.05129784345626831,
-0.009962573647499084,
-0.045296698808670044,
-0.05192640423774719,
-0.10601667314767838,
0.014379938133060932,
0.00029934279154986143,
-0.06144409999251366,
0.06579696387052536,
0.03803889825940132,
-0.005935485940426588,
0.006751036271452904,
-0.006849132478237152,
0.051339875906705856,
-0.026665860787034035,
-0.002787315286695957,
-0.01959266886115074,
-0.06508469581604004,
-0.016502326354384422,
0.06502744555473328,
0.018597889691591263,
0.07315119355916977,
-2.364706169705694e-34,
0.026463571935892105,
0.09260519593954086,
0.03224734589457512,
0.03710506483912468,
-0.0017767074750736356,
-0.03300938010215759,
-0.05884263291954994,
-0.06002255156636238,
-0.08259560167789459,
0.09334069490432739,
0.016287565231323242,
-0.04407611861824989,
0.030571158975362778,
-0.06168494373559952,
-0.023826071992516518,
-0.04720478132367134,
-0.09537780284881592,
-0.005707497708499432,
0.044124994426965714,
0.01970978081226349,
-0.04498577117919922,
0.02683272585272789,
-0.06844769418239594,
0.05718177556991577,
0.05268603190779686,
0.03032410331070423,
-0.09300704300403595,
-0.04242665320634842,
0.038738638162612915,
-0.006467732600867748,
-0.07727254927158356,
0.025233279913663864,
-0.11892416328191757,
-0.04866798222064972,
0.05449371784925461,
0.05339362844824791,
-0.021209556609392166,
0.04099782183766365,
0.05068531259894371,
0.04176243767142296,
0.0422825962305069,
-0.06333852559328079,
-0.05042864382266998,
0.0754789486527443,
0.04417496174573898,
0.039303284138441086,
-0.03668668866157532,
-0.003330555045977235,
-0.12945252656936646,
0.01417918223887682,
-0.07478037476539612,
-0.00030472161597572267,
0.03260057047009468,
0.05132288485765457,
-0.0735975131392479,
-0.049968015402555466,
0.06696698069572449,
0.06196197494864464,
0.06346974521875381,
-0.04798705130815506,
0.0098459143191576,
-0.06709246337413788,
-0.080101877450943,
0.008887037634849548,
0.017170358449220657,
-0.04480176791548729,
-0.0017057836521416903,
-0.003961654379963875,
-0.020353103056550026,
0.02677936851978302,
0.054316382855176926,
-0.0071403831243515015,
0.022322585806250572,
-0.018899213522672653,
-0.007599313743412495,
-0.04389569163322449,
0.016034763306379318,
-0.023967592045664787,
-0.01005993876606226,
0.005539874546229839,
-0.012122261337935925,
0.04091509059071541,
0.05711423605680466,
0.05291014909744263,
0.11179006844758987,
0.03385204076766968,
-0.02358471229672432,
0.043116286396980286,
-0.0009007677435874939,
-0.02262260764837265,
-0.07477476447820663,
-0.03662113845348358,
-0.09842518717050552,
0.03773081302642822,
-0.032690174877643585,
-3.499888379110416e-8,
-0.10660310089588165,
0.08789976686239243,
-0.04836456477642059,
-0.01712232641875744,
0.0909712016582489,
-0.04181006923317909,
0.08193057030439377,
0.15651747584342957,
0.009632717818021774,
0.10398560017347336,
0.09953227639198303,
-0.01785089820623398,
0.027406232431530952,
-0.06828385591506958,
-0.031696122139692307,
-0.04443317651748657,
0.0004090286965947598,
-0.0003549023240339011,
0.04961390793323517,
0.017977813258767128,
0.031100977212190628,
-0.029136642813682556,
0.04685892537236214,
0.039636749774217606,
0.014853035099804401,
-0.14413736760616302,
-0.004805602133274078,
0.04419071599841118,
0.10430882126092911,
0.04501430690288544,
-0.05183828994631767,
-0.0034990848507732153,
0.05240026116371155,
0.006299150642007589,
0.029829703271389008,
0.12890557944774628,
0.01912079192698002,
0.04716044291853905,
-0.06162923201918602,
0.06668014824390411,
-0.03000720776617527,
0.01825466752052307,
-0.027776271104812622,
-0.04830525815486908,
-0.05234287679195404,
0.02777380496263504,
0.0450083464384079,
0.026424864307045937,
0.010542360134422779,
-0.002680992940440774,
-0.02983475662767887,
-0.0037556677125394344,
0.003381614340469241,
0.003356355242431164,
-0.018154975026845932,
0.0121380016207695,
-0.03597568720579147,
-0.08451318740844727,
0.0064806560985744,
0.016436396166682243,
0.03568820282816887,
0.034680239856243134,
-0.041157323867082596,
-0.04532245919108391
] |
FakeShield: Explainable Image Forgery Detection and Localization via Multi-modal Large Language Models | https://openreview.net/forum?id=pAQzEY7M03 | [
"Zhipei Xu",
"Xuanyu Zhang",
"Runyi Li",
"Zecheng Tang",
"Qing Huang",
"Jian Zhang"
] | Poster | The rapid development of generative AI is a double-edged sword, which not only facilitates content creation but also makes image manipulation easier and more difficult to detect. Although current image forgery detection and localization (IFDL) methods are generally effective, they tend to face two challenges: \textbf{1)} black-box nature with unknown detection principle, \textbf{2)} limited generalization across diverse tampering methods (e.g., Photoshop, DeepFake, AIGC-Editing). To address these issues, we propose the explainable IFDL task and design FakeShield, a multi-modal framework capable of evaluating image authenticity, generating tampered region masks, and providing a judgment basis based on pixel-level and image-level tampering clues. Additionally, we leverage GPT-4o to enhance existing IFDL datasets, creating the Multi-Modal Tamper Description dataSet (MMTD-Set) for training FakeShield's tampering analysis capabilities. Meanwhile, we incorporate a Domain Tag-guided Explainable Forgery Detection Module (DTE-FDM) and a Multi-modal Forgery Localization Module (MFLM) to address various types of tamper detection interpretation and achieve forgery localization guided by detailed textual descriptions. Extensive experiments demonstrate that FakeShield effectively detects and localizes various tampering techniques, offering an explainable and superior solution compared to previous IFDL methods. The code is available at https://github.com/zhipeixu/FakeShield. | Image Forgery Detection and Localization, Multi-modal Large Language Model, Tamper Detection | null | 328 | 2410.02761 | [
-0.07387315481901169,
0.01890587992966175,
0.012847185134887695,
0.020077873021364212,
0.14870940148830414,
-0.07311813533306122,
0.025372179225087166,
-0.09210656583309174,
0.04389650374650955,
-0.05617716535925865,
0.029627811163663864,
-0.07309827953577042,
0.06417977064847946,
0.04667702689766884,
-0.024336617439985275,
-0.04348035529255867,
0.04173482209444046,
0.08772719651460648,
-0.11404477804899216,
-0.02135641872882843,
0.12726126611232758,
-0.021524470299482346,
0.09802742302417755,
-0.04317200556397438,
0.014504271559417248,
0.03217999264597893,
-0.012495022267103195,
-0.02193027175962925,
0.006217097397893667,
-0.07608014345169067,
0.031995534896850586,
0.018275702372193336,
-0.05960160121321678,
0.05975717678666115,
-0.031544171273708344,
0.10350575298070908,
0.0015449277125298977,
0.059269510209560394,
0.010511773638427258,
-0.0792812630534172,
-0.059436194598674774,
-0.003580710617825389,
-0.01564108394086361,
0.04856909438967705,
0.02985271066427231,
0.005839021410793066,
0.017779134213924408,
0.004641399253159761,
0.04949548467993736,
0.014456071890890598,
-0.056743476539850235,
-0.012855430133640766,
-0.024147294461727142,
-0.0004457788891158998,
0.003660377115011215,
-0.04004635661840439,
0.08493988960981369,
-0.02418428100645542,
0.04286116361618042,
0.028023716062307358,
-0.03875000402331352,
0.04549156501889229,
0.022258905693888664,
0.02291346713900566,
0.00011261369218118489,
0.025629622861742973,
-0.012791701592504978,
-0.020510606467723846,
-0.033840421587228775,
-0.08666019886732101,
0.05575106665492058,
0.020069817081093788,
-0.006848842371255159,
0.04019983857870102,
-0.05692083388566971,
0.0553073026239872,
-0.06579212844371796,
-0.005525772925466299,
0.016278931871056557,
-0.0926734209060669,
0.05484285205602646,
-0.01820393092930317,
0.09172697365283966,
0.018780546262860298,
0.05055273696780205,
0.023699946701526642,
-0.027176424860954285,
0.03429209068417549,
0.023015504702925682,
0.07561995834112167,
-0.02246992290019989,
0.001665893360041082,
0.02413049340248108,
0.04915592074394226,
0.07513996958732605,
-0.03079754114151001,
-0.045918382704257965,
-0.0667952448129654,
0.02416994795203209,
0.06713192164897919,
-0.06667497754096985,
-0.036833591759204865,
-0.04246356338262558,
-0.06197255849838257,
0.08054789900779724,
-0.01700020022690296,
0.0303113404661417,
0.011275812052190304,
0.024705827236175537,
-0.037543825805187225,
0.041765592992305756,
0.07623490691184998,
-0.010022781789302826,
-0.06505895406007767,
0.06475388258695602,
-0.0015779805835336447,
0.008231990039348602,
0.023720206692814827,
-0.04776299372315407,
-0.0568934865295887,
-0.052057262510061264,
-0.009573780000209808,
-0.013553641736507416,
0.01065809465944767,
0.020687663927674294,
-0.07542847096920013,
-0.08390749245882034,
3.997587209142857e-34,
0.09684254229068756,
0.060391202569007874,
0.006681640166789293,
0.09126514941453934,
0.01817373000085354,
0.000947917636949569,
-0.07774445414543152,
0.03601939231157303,
-0.04938380792737007,
0.016630256548523903,
0.02149766869843006,
0.01704365573823452,
-0.09231489151716232,
0.10261192172765732,
0.0771033763885498,
0.04531622305512428,
-0.022358236834406853,
-0.052384693175554276,
-0.029262209311127663,
-0.01628350466489792,
-0.025816461071372032,
-0.04959413409233093,
-0.011785270646214485,
-0.012731348164379597,
0.004873996134847403,
0.019615059718489647,
0.06641032546758652,
-0.008149136789143085,
0.039151839911937714,
-0.0009003687300719321,
-0.14758433401584625,
-0.05238042771816254,
0.0514528825879097,
0.015336327254772186,
-0.0052681113593280315,
-0.017819572240114212,
0.006413071881979704,
-0.07519516348838806,
-0.02374502830207348,
0.06097223982214928,
0.0061368877068161964,
-0.04999125748872757,
-0.05181915685534477,
-0.07494261860847473,
-0.03770889714360237,
0.029129359871149063,
-0.024374784901738167,
0.04786768555641174,
-0.04764880612492561,
0.0401252843439579,
0.03531491383910179,
0.07891504466533661,
0.053718797862529755,
-0.07713985443115234,
-0.0723385214805603,
-0.026364527642726898,
0.014154191128909588,
-0.02400604449212551,
0.09860610961914062,
0.03194088488817215,
0.04503018036484718,
0.04972109943628311,
-0.04293879121541977,
0.11147822439670563,
-0.03636360168457031,
-0.04194696247577667,
0.04015360772609711,
0.015010640025138855,
0.024127231910824776,
-0.0038348583038896322,
-0.07690299302339554,
0.055869922041893005,
0.000251226854743436,
0.022509537637233734,
-0.04883953556418419,
-0.0590672567486763,
0.03453763946890831,
-0.007678022142499685,
0.04472983255982399,
0.008357139304280281,
-0.016977526247501373,
-0.001183851738460362,
0.028077607974410057,
-0.0010120385559275746,
-0.07572346925735474,
-0.04600616171956062,
0.03788459300994873,
-0.08398788422346115,
-0.017608722671866417,
-0.0050383200868964195,
-0.024624798446893692,
-0.08708006888628006,
-0.033133238554000854,
0.039016906172037125,
-0.02524055913090706,
-5.2421009862801635e-34,
0.010331555269658566,
-0.004400862846523523,
-0.036175437271595,
0.04255913197994232,
-0.01798812486231327,
-0.06446361541748047,
0.061361342668533325,
0.08553127944469452,
0.07344842702150345,
-0.0810151919722557,
-0.03326070308685303,
0.015028407797217369,
0.015220092609524727,
-0.040706533938646317,
-0.0115013113245368,
-0.05181252956390381,
-0.0018994861748069525,
0.04595565050840378,
-0.05792906507849693,
0.04260757192969322,
0.04628131538629532,
0.029509400948882103,
-0.061213403940200806,
0.04497089609503746,
-0.00720955478027463,
0.07840363681316376,
-0.04899268224835396,
0.007738352287560701,
0.008457288146018982,
0.012953843921422958,
0.05742013454437256,
0.07400139421224594,
0.06181533262133598,
0.04253270477056503,
-0.03554007038474083,
-0.010437767952680588,
0.10556898266077042,
-0.03621536120772362,
0.014112786389887333,
0.0803753063082695,
0.06257021427154541,
0.011652067303657532,
-0.1274666041135788,
0.011645710095763206,
-0.0819145143032074,
-0.039501041173934937,
-0.008424319326877594,
0.01279047504067421,
0.07573974132537842,
0.0791957899928093,
0.043553125113248825,
-0.057208023965358734,
-0.08328642696142197,
-0.01499443780630827,
-0.042974647134542465,
-0.07516911625862122,
-0.06407058984041214,
-0.017610039561986923,
0.09419001638889313,
0.02761373110115528,
-0.0037160892970860004,
-0.039516642689704895,
-0.037765976041555405,
-0.0011822321685031056,
-0.013514379039406776,
-0.03605591878294945,
-0.03300924226641655,
0.02356623485684395,
-0.022353358566761017,
0.015897652134299278,
0.0639631375670433,
-0.07151666283607483,
-0.02147550694644451,
0.036034733057022095,
0.014274470508098602,
-0.01667637750506401,
0.03258764371275902,
-0.02509615384042263,
-0.020495839416980743,
0.004775001667439938,
0.12598805129528046,
-0.07462626695632935,
0.04586578160524368,
0.07828839123249054,
0.061933547258377075,
0.06440107524394989,
-0.05071655288338661,
0.029263868927955627,
0.02497526817023754,
0.024208640679717064,
-0.1046234667301178,
0.05213800445199013,
-0.0037754178047180176,
0.0828806608915329,
0.00041706199408508837,
-3.954420435547945e-8,
-0.007481901440769434,
0.006905431393533945,
-0.03621647134423256,
-0.02370447665452957,
-0.0275475662201643,
0.008580449037253857,
-0.03916524350643158,
-0.03788132965564728,
0.02750045247375965,
-0.1130538284778595,
0.08599679917097092,
-0.0568566620349884,
-0.09736794978380203,
-0.15192024409770966,
-0.0580102801322937,
0.02181881107389927,
-0.04627961665391922,
0.02910715527832508,
0.037770409137010574,
-0.00787604320794344,
-0.03339804708957672,
0.047859735786914825,
0.019648009911179543,
-0.05355532094836235,
0.0028171369340270758,
-0.05033593252301216,
-0.07781226187944412,
0.027188675478100777,
0.019084591418504715,
0.09488660842180252,
0.016690293326973915,
-0.034085098654031754,
0.0618409737944603,
0.0654723048210144,
-0.03807428106665611,
0.10138636827468872,
-0.00829240120947361,
-0.011389841325581074,
0.008129850961267948,
-0.007527460809797049,
0.05931103602051735,
-0.00800242554396391,
-0.05176376551389694,
-0.026932472363114357,
-0.014613564126193523,
-0.037605173885822296,
0.04340747371315956,
-0.13944068551063538,
0.005325218662619591,
-0.006227856036275625,
0.00015620424528606236,
0.006572043057531118,
-0.03527872636914253,
0.12671753764152527,
0.013235128484666348,
-0.0028214447665959597,
0.05180150642991066,
0.03511451929807663,
0.06933100521564484,
0.09982788562774658,
0.022597774863243103,
-0.009791388176381588,
0.021426323801279068,
-0.02561471424996853
] |
SecureGS: Boosting the Security and Fidelity of 3D Gaussian Splatting Steganography | https://openreview.net/forum?id=H4FSx06FCZ | [
"Xuanyu Zhang",
"Jiarui Meng",
"Zhipei Xu",
"Shuzhou Yang",
"Yanmin Wu",
"Ronggang Wang",
"Jian Zhang"
] | Poster | 3D Gaussian Splatting (3DGS) has emerged as a premier method for 3D representation due to its real-time rendering and high-quality outputs, underscoring the critical need to protect the privacy of 3D assets. Traditional NeRF steganography methods fail to address the explicit nature of 3DGS since its point cloud files are publicly accessible. Existing GS steganography solutions mitigate some issues but still struggle with reduced rendering fidelity, increased computational demands, and security flaws, especially in the security of the geometric structure of the visualized point cloud. To address these demands, we propose a \textbf{SecureGS}, a secure and efficient 3DGS steganography framework inspired by Scaffold-GS's anchor point design and neural decoding. SecureGS uses a hybrid decoupled Gaussian encryption mechanism to embed offsets, scales, rotations, and RGB attributes of the hidden 3D Gaussian points in anchor point features, retrievable only by authorized users through privacy-preserving neural networks. To further enhance security, we propose a density region-aware anchor growing and pruning strategy that adaptively locates optimal hiding regions without exposing hidden information. Extensive experiments show that SecureGS significantly surpasses existing GS steganography methods in rendering fidelity, speed, and security. | 3DGS steganography, copyright protection, watermarking | null | 326 | 2503.06118 | [
-0.06523866206407547,
-0.011552966199815273,
-0.03435009717941284,
0.03515009954571724,
0.07527421414852142,
0.042884521186351776,
-0.00679754326120019,
-0.13561402261257172,
0.04224758595228195,
-0.043631792068481445,
0.006543766241520643,
0.06132090091705322,
0.06327079981565475,
-0.009642979130148888,
0.0020407161209732294,
-0.018083618953824043,
0.021144241094589233,
0.09647590667009354,
-0.035487934947013855,
0.03266295790672302,
0.004931213334202766,
-0.10828118771314621,
-0.015808669850230217,
0.01179986260831356,
0.049096621572971344,
-0.009084273129701614,
0.0866938978433609,
-0.00037342190626077354,
-0.020652949810028076,
0.006951808463782072,
0.08247785270214081,
-0.09280461072921753,
-0.02035122737288475,
0.013007859699428082,
-0.0027483191806823015,
0.030741121619939804,
-0.024200215935707092,
-0.01947350800037384,
-0.008177165873348713,
-0.003065014723688364,
0.026534385979175568,
-0.005199133884161711,
-0.021355705335736275,
0.03928317129611969,
0.04063650965690613,
0.08111245185136795,
0.015929345041513443,
-0.02163967303931713,
-0.011847437359392643,
-0.06310345232486725,
0.030088555067777634,
-0.043717652559280396,
-0.0680563673377037,
0.013941588811576366,
-0.03805402293801308,
-0.0441121906042099,
-0.028039492666721344,
-0.019239526242017746,
0.0038575318176299334,
0.047876689583063126,
0.014348166063427925,
0.00028062728233635426,
0.06848526746034622,
0.018780317157506943,
-0.020444106310606003,
0.012478534132242203,
0.011530522257089615,
-0.02234892174601555,
-0.03829474747180939,
0.050391990691423416,
-0.0346294529736042,
0.03215982764959335,
-0.0650423914194107,
-0.019679797813296318,
-0.015178529545664787,
0.1072063073515892,
-0.04527938738465309,
-0.0730920359492302,
-0.011125735007226467,
-0.04683752730488777,
0.07318059355020523,
0.002671327907592058,
0.050942886620759964,
0.1070825606584549,
0.007513678632676601,
0.009574993513524532,
-0.011672683991491795,
0.0398077517747879,
-0.016074510291218758,
0.015569104813039303,
0.02124851569533348,
0.05609465762972832,
-0.018159406259655952,
0.0038606764283031225,
-0.0009211943834088743,
-0.08941739797592163,
-0.017195142805576324,
-0.022067183628678322,
-0.049832042306661606,
0.0038259802386164665,
0.025978947058320045,
-0.03898603469133377,
0.09068186581134796,
-0.03653407469391823,
0.12587669491767883,
0.07770466804504395,
0.04150664806365967,
0.05007164552807808,
0.03738372027873993,
0.045901037752628326,
-0.002045079367235303,
0.049900420010089874,
-0.02572466991841793,
-0.022602727636694908,
-0.025957299396395683,
0.05604665353894234,
-0.09243281930685043,
-0.007557103876024485,
0.008571327663958073,
0.024199651554226875,
0.05589820072054863,
0.01832406409084797,
-0.025972917675971985,
-0.021807124838232994,
-0.10688002407550812,
-0.01228247582912445,
-0.07804150879383087,
1.6247473532134418e-33,
-0.009565837681293488,
0.0380793958902359,
0.006290333811193705,
-0.017963184043765068,
0.0628824532032013,
0.047905758023262024,
-0.027015507221221924,
0.020371316000819206,
-0.04787847399711609,
0.0283928532153368,
-0.0757608413696289,
-0.018583739176392555,
0.00783711951225996,
0.0858752653002739,
0.02316645160317421,
0.035787131637334824,
-0.021614020690321922,
0.0072617302648723125,
0.0067503745667636395,
-0.03597291558980942,
0.03700824826955795,
-0.045610133558511734,
0.02047162875533104,
-0.03465184569358826,
-0.023108113557100296,
0.0027279506903141737,
0.0003125630028080195,
-0.03355877101421356,
-0.03490764647722244,
0.00670221634209156,
-0.040861837565898895,
0.013731666840612888,
-0.0004735675756819546,
-0.03988351672887802,
0.017749466001987457,
-0.06602831184864044,
0.03100317344069481,
-0.1355031430721283,
-0.03327600657939911,
0.03363361954689026,
0.06356412172317505,
-0.02636418677866459,
-0.0020648955833166838,
-0.033182572573423386,
-0.057957809418439865,
0.09491551667451859,
0.06886889040470123,
0.04262593016028404,
-0.024065077304840088,
0.014355073682963848,
0.07610230147838593,
0.04050938040018082,
-0.09626495838165283,
-0.08363285660743713,
-0.03732028231024742,
-0.08473491668701172,
0.025277258828282356,
-0.02805308811366558,
0.09866099804639816,
0.0024970306549221277,
-0.06502485275268555,
0.06246878579258919,
-0.0052441926673054695,
0.0105684082955122,
-0.13238172233104706,
-0.023306751623749733,
0.047009240835905075,
-0.012976188212633133,
0.005175945349037647,
0.039704468101263046,
-0.02290927805006504,
0.031048646196722984,
0.04516347497701645,
0.0337313637137413,
-0.0876825824379921,
-0.035355065017938614,
0.03140420466661453,
-0.00015459751011803746,
0.11255928128957748,
-0.004471806809306145,
-0.08926105499267578,
-0.044491104781627655,
0.023023061454296112,
-0.03285372257232666,
-0.09579464793205261,
-0.12809479236602783,
0.003841697471216321,
0.01663723587989807,
-0.025737695395946503,
-0.0014480495592579246,
-0.028521055355668068,
-0.12297389656305313,
0.036583345383405685,
-0.05662012845277786,
-0.0704972892999649,
-1.5735163706687294e-33,
-0.012367363087832928,
0.007724709343165159,
-0.09049717336893082,
0.031057195737957954,
0.016291256994009018,
-0.017574017867445946,
-0.0077404119074344635,
0.003642449388280511,
0.02081349305808544,
-0.034388910979032516,
-0.04052512347698212,
0.04723871499300003,
0.06428535282611847,
0.023412145674228668,
0.05687910318374634,
-0.013947131112217903,
-0.038216520100831985,
-0.012887077406048775,
-0.10708164423704147,
-0.06541389971971512,
0.0021406833548098803,
-0.038892962038517,
-0.03931934013962746,
0.12262433022260666,
0.07273166626691818,
0.09214417636394501,
0.021257689222693443,
0.0009897141717374325,
0.07362271100282669,
0.08603827655315399,
-0.052666883915662766,
0.04923663288354874,
0.027631137520074844,
-0.050577037036418915,
-0.03184027969837189,
-0.02883840911090374,
0.10068240016698837,
0.03537030890583992,
-0.007999911904335022,
-0.019207602366805077,
0.01190083846449852,
-0.017093010246753693,
-0.03304186835885048,
0.056500278413295746,
0.02132081612944603,
0.08346527814865112,
0.02859395369887352,
0.0498095266520977,
0.019284985959529877,
0.06786185503005981,
-0.04328649863600731,
0.03935472294688225,
-0.060253143310546875,
0.01713864877820015,
0.024480657652020454,
-0.06961123645305634,
-0.14487650990486145,
0.035031262785196304,
0.041688840836286545,
0.01146666333079338,
0.05588653311133385,
-0.04275300353765488,
-0.01773720420897007,
0.0040801819413900375,
-0.010461282916367054,
0.03761634975671768,
-0.008826027624309063,
0.004847305361181498,
-0.05370572581887245,
0.04340949282050133,
0.002465080702677369,
-0.03382810205221176,
0.010291555896401405,
0.016382332891225815,
-0.028860678896307945,
-0.008248030208051205,
-0.01750384457409382,
-0.011182083748281002,
0.005398890934884548,
0.06215932220220566,
0.11857859790325165,
-0.03231459856033325,
0.047498270869255066,
0.08905297517776489,
0.10957515984773636,
0.009723217226564884,
-0.020899169147014618,
0.011615047231316566,
-0.0066085499711334705,
0.0006057152641005814,
-0.07066743075847626,
0.04771766811609268,
-0.0857648104429245,
0.08005321770906448,
0.04116254672408104,
-3.632122513863578e-8,
-0.009129523299634457,
0.0616513229906559,
0.040758874267339706,
-0.07552938908338547,
-0.05694827809929848,
0.0034987488761544228,
-0.008586810901761055,
0.004649460781365633,
0.013690419495105743,
-0.08229027688503265,
0.06322897970676422,
-0.1507900059223175,
-0.053238626569509506,
-0.07255929708480835,
0.04771101102232933,
0.08983300626277924,
-0.06278464943170547,
-0.07185252010822296,
0.011941284872591496,
-0.014100301079452038,
-0.0930963084101677,
0.09171439707279205,
-0.04409962147474289,
0.005154937040060759,
0.003431553952395916,
0.033781953155994415,
-0.05039950832724571,
0.09906290471553802,
0.090504489839077,
0.03303373232483864,
0.018202897161245346,
-0.05105748027563095,
0.05217026546597481,
0.08238618820905685,
-0.07568222284317017,
0.051307499408721924,
-0.014920104295015335,
0.007424422539770603,
0.03532234579324722,
0.04496687278151512,
-0.000025601928427931853,
0.006011687219142914,
0.0257726963609457,
-0.03507621958851814,
-0.09105495363473892,
0.0121860820800066,
0.12776847183704376,
-0.026006437838077545,
-0.0565716028213501,
0.042801328003406525,
-0.006570343393832445,
-0.05329133942723274,
-0.060095515102148056,
0.14239248633384705,
0.042973924428224564,
-0.019243964925408363,
-0.019300157204270363,
-0.052100878208875656,
0.030990727245807648,
0.05030953884124756,
-0.022747404873371124,
-0.02039911597967148,
0.014253027737140656,
0.04998721182346344
] |
GenDataAgent: On-the-fly Dataset Augmentation with Synthetic Data | https://openreview.net/forum?id=WoGnnggVCZ | [
"Zhiteng Li",
"Lele Chen",
"Jerone Andrews",
"Yunhao Ba",
"Yulun Zhang",
"Alice Xiang"
] | Poster | We propose a generative agent that augments training datasets with synthetic data for model fine-tuning. Unlike prior work, which uniformly samples synthetic data, our agent iteratively generates relevant samples on-the-fly, aligning with the target distribution. It prioritizes synthetic data that complements difficult training samples, focusing on those with high variance in gradient updates. Experiments across several image classification tasks demonstrate the effectiveness of our approach. | supervised learning, classification, computer vision, synthetic data, generative AI, responsible AI, fairness | GenDataAgent is an on-the-fly generative agent that augments training datasets with task-relevant synthetic data, improving model fine-tuning and classification performance | 312 | null | [
-0.06149045005440712,
-0.12462455034255981,
0.0225685965269804,
0.017471354454755783,
0.032670773565769196,
-0.004443982150405645,
-0.016574056819081306,
-0.11079376935958862,
-0.0919235348701477,
-0.07563568651676178,
0.013514714315533638,
-0.04742775484919548,
-0.0072362725622951984,
-0.017899712547659874,
0.018460402265191078,
-0.004197295289486647,
0.11198101937770844,
0.11429421603679657,
-0.06744418293237686,
0.004411360248923302,
-0.04808446019887924,
0.007481583394110203,
0.07766808569431305,
-0.047338180243968964,
0.07083772867918015,
-0.010062404908239841,
0.027679571881890297,
-0.045265082269907,
0.04388497397303581,
-0.08837129175662994,
0.05795220285654068,
0.061113882809877396,
-0.10628642141819,
0.008057869039475918,
-0.0022131979931145906,
0.04987400397658348,
-0.05298770219087601,
0.10430964082479477,
0.05872175469994545,
0.06426773220300674,
0.02697122097015381,
0.010873710736632347,
-0.03711667284369469,
0.007927493192255497,
0.018909690901637077,
0.006870773155242205,
0.014823905192315578,
-0.0787116214632988,
0.06926286965608597,
0.03825525566935539,
-0.05798695236444473,
-0.04705552011728287,
-0.04814606159925461,
0.09618192166090012,
-0.006631261669099331,
-0.012922887690365314,
0.11269529908895493,
-0.002398597076535225,
0.04775309935212135,
-0.03584161773324013,
0.04303852096199989,
-0.03739868104457855,
0.02337540313601494,
0.0030660463962703943,
-0.03943150117993355,
-0.062177207320928574,
-0.030299697071313858,
-0.0019236751832067966,
0.05111447349190712,
-0.08345471322536469,
-0.01211200188845396,
0.1100454032421112,
0.0024645552039146423,
-0.009070227853953838,
0.03295770660042763,
0.001982208341360092,
0.025318413972854614,
-0.0009583273204043508,
0.1407543569803238,
-0.11727896332740784,
0.019063448533415794,
-0.04640229418873787,
0.014325027354061604,
0.005598521325737238,
0.10637255012989044,
0.025176916271448135,
0.005459809210151434,
0.060193564742803574,
0.000214392741327174,
0.042415812611579895,
-0.04888356849551201,
0.0015972042456269264,
0.03929268196225166,
0.00042390558519400656,
-0.05900285765528679,
0.0786246806383133,
0.03569403290748596,
-0.1431802213191986,
-0.04500598460435867,
0.05930952727794647,
-0.05166663974523544,
0.02242090180516243,
-0.03479783236980438,
0.05097226798534393,
0.015071781352162361,
-0.08263885974884033,
0.07419300824403763,
0.003805942600592971,
-0.00244978372938931,
-0.09851335734128952,
0.058551013469696045,
0.007470117416232824,
0.017092421650886536,
-0.03378230705857277,
-0.019931523129343987,
0.00612768204882741,
-0.024046240374445915,
0.025852147489786148,
-0.08932053297758102,
-0.04484899714589119,
-0.0369601845741272,
0.0186177808791399,
-0.041878148913383484,
0.027644000947475433,
0.027330569922924042,
0.014705201610922813,
-0.07808627933263779,
1.735279454363246e-33,
0.025139588862657547,
0.03536449000239372,
0.07947271317243576,
0.012154906988143921,
0.06390883773565292,
-0.03130678832530975,
-0.012577385641634464,
-0.00713242357596755,
-0.05763804167509079,
-0.0019481390481814742,
-0.06552965193986893,
-0.025645192712545395,
-0.14603497087955475,
0.18444137275218964,
0.02100779488682747,
-0.024679051712155342,
-0.11564072221517563,
0.031568530946969986,
-0.011925574392080307,
0.0650448203086853,
-0.007192266173660755,
-0.01578832045197487,
0.005482485983520746,
-0.025943677872419357,
-0.005123886279761791,
0.024624083191156387,
0.019039161503314972,
-0.0015244113747030497,
-0.005356351844966412,
0.007909093983471394,
-0.06963671743869781,
0.06068224832415581,
0.03843385726213455,
0.09198325127363205,
-0.04504067078232765,
-0.05106253921985626,
0.008044018410146236,
-0.04616353288292885,
0.03324525058269501,
0.10510286688804626,
0.024187970906496048,
0.052967775613069534,
0.052342288196086884,
0.010897704400122166,
-0.04502592608332634,
-0.008527709171175957,
0.04619643837213516,
0.03805183619260788,
0.019952110946178436,
0.004832631442695856,
0.00957724079489708,
0.06277643889188766,
-0.024178482592105865,
-0.06715968996286392,
-0.07262080907821655,
0.04824136942625046,
0.006135047413408756,
0.011982292868196964,
0.004524983465671539,
-0.017826808616518974,
0.01562836952507496,
0.017532067373394966,
-0.014164607971906662,
0.11052241176366806,
-0.010527659207582474,
-0.029889428988099098,
0.054681446403265,
0.016286712139844894,
0.0008036372601054609,
0.06179121136665344,
-0.04279971867799759,
0.03124287538230419,
-0.09435001760721207,
-0.02332996018230915,
0.039664026349782944,
-0.020847653970122337,
0.017024995759129524,
-0.005433715879917145,
-0.002261972287669778,
-0.00046180616482160985,
0.026080632582306862,
0.05191138759255409,
-0.011043678037822247,
-0.03878113999962807,
-0.024864694103598595,
-0.0527883917093277,
-0.031044693663716316,
-0.038388825953006744,
0.022968266159296036,
0.029039984568953514,
-0.10233074426651001,
-0.014047144912183285,
-0.06603577733039856,
0.013801093213260174,
0.01468557771295309,
-1.7727801118773772e-33,
-0.00454927422106266,
0.029353316873311996,
-0.04271715506911278,
0.09625953435897827,
-0.02088022045791149,
-0.014812679961323738,
-0.02417684905230999,
0.034721244126558304,
-0.008004370145499706,
0.015993481501936913,
0.0017043715342879295,
0.01232913974672556,
-0.0334932804107666,
-0.08672907203435898,
-0.05117969587445259,
-0.056269172579050064,
-0.01628543995320797,
0.007523204665631056,
-0.06791525334119797,
-0.06446048617362976,
0.03832671418786049,
0.06503511220216751,
-0.012874062173068523,
0.035496633499860764,
-0.030811036005616188,
0.0499030239880085,
-0.035463448613882065,
0.1224125549197197,
0.03156593069434166,
0.02423333376646042,
0.004487029742449522,
-0.014341183006763458,
0.02945788763463497,
-0.049685440957546234,
-0.061481088399887085,
0.11267443746328354,
0.10692386329174042,
-0.009058746509253979,
-0.027257567271590233,
0.11411674320697784,
0.020390836521983147,
-0.017147034406661987,
-0.1415214240550995,
0.07962352782487869,
-0.015475240536034107,
-0.06700874865055084,
-0.04474205896258354,
0.021485161036252975,
0.06224023923277855,
0.055526066571474075,
-0.019320007413625717,
0.009658913128077984,
-0.037716854363679886,
0.008763249032199383,
0.0037368659395724535,
-0.07044541835784912,
0.032753974199295044,
-0.04743108153343201,
0.05957280471920967,
0.01790875941514969,
-0.08944159001111984,
-0.06119377911090851,
0.010386977344751358,
-0.07275547087192535,
-0.06676791608333588,
-0.006272254046052694,
-0.04692555218935013,
-0.022970058023929596,
0.06988222897052765,
0.025256982073187828,
0.07853711396455765,
-0.04443947225809097,
0.005470788571983576,
0.03318382054567337,
0.010509992018342018,
-0.07625015079975128,
0.06188873201608658,
0.008742932230234146,
0.08726099878549576,
-0.04530550539493561,
0.038802266120910645,
-0.04769019037485123,
0.033430036157369614,
0.0708082914352417,
0.0601339153945446,
0.11951706558465958,
0.0077756550163030624,
0.036099597811698914,
0.038141366094350815,
-0.03136398270726204,
-0.046654172241687775,
0.0015805255388841033,
-0.04993060976266861,
0.02669229917228222,
-0.08482663333415985,
-3.635070200402879e-8,
-0.024789266288280487,
0.01418975368142128,
0.07774120569229126,
-0.004488520324230194,
0.005805238615721464,
-0.05819407105445862,
-0.026485968381166458,
0.08160796761512756,
-0.012057986110448837,
-0.08994486182928085,
0.07109219580888748,
0.031882818788290024,
0.05030253529548645,
-0.011520658619701862,
0.068288654088974,
0.035474542528390884,
0.002091720700263977,
0.07234922796487808,
-0.04419732093811035,
0.01190316490828991,
-0.049948800355196,
0.06035340204834938,
-0.003121102461591363,
-0.04482432082295418,
0.09058630466461182,
-0.07213174551725388,
-0.012345163151621819,
-0.005237744189798832,
0.020407479256391525,
0.04021240025758743,
0.00019455599249340594,
-0.05157993733882904,
-0.04281793162226677,
-0.019245484843850136,
-0.027109185233712196,
0.04122588038444519,
0.004372106399387121,
-0.03817560151219368,
-0.032237276434898376,
-0.08239481598138809,
-0.010478571988642216,
0.05192375183105469,
-0.018099570646882057,
-0.026598332449793816,
-0.01953970640897751,
-0.003265008330345154,
-0.0032695576082915068,
-0.06151031702756882,
0.03241410851478577,
0.0015071457019075751,
0.017338216304779053,
-0.050662700086832047,
0.03113468363881111,
0.02277112938463688,
0.019143132492899895,
-0.028776485472917557,
-0.013877279125154018,
-0.060994457453489304,
0.046681277453899384,
0.04478616267442703,
-0.013372947461903095,
-0.020707260817289352,
-0.007145779673010111,
-0.04647044092416763
] |
VideoGrain: Modulating Space-Time Attention for Multi-Grained Video Editing | https://openreview.net/forum?id=SSslAtcPB6 | [
"Xiangpeng Yang",
"Linchao Zhu",
"Hehe Fan",
"Yi Yang"
] | Poster | Recent advancements in diffusion models have significantly improved video generation and editing capabilities. However, multi-grained video editing, which encompasses class-level, instance-level, and part-level modifications, remains a formidable challenge. The major difficulties in multi-grained editing include semantic misalignment of text-to-region control and feature coupling within the diffusion model. To address these difficulties, we present VideoGrain, a zero-shot approach that modulates space-time (cross- and self-) attention mechanisms to achieve fine-grained control over video content. We enhance text-to-region control by amplifying each local prompt's attention to its corresponding spatial-disentangled region while minimizing interactions with irrelevant areas in cross-attention. Additionally, we improve feature separation by increasing intra-region awareness and reducing inter-region interference in self-attention. Extensive experiments demonstrate our method achieves state-of-the-art performance in real-world scenarios. Our code, data, and demos are available on the [project page](https://knightyxp.github.io/VideoGrain_project_page/). | diffusion model, video editing | zero-shot method for class-level, instance-level and part-level video editing | 308 | 2502.17258 | [
0.04210646077990532,
-0.07943479716777802,
0.06310637295246124,
0.029903469607234,
0.07953566312789917,
-0.00973333790898323,
-0.07466324418783188,
-0.05999822914600372,
0.07199664413928986,
-0.07212743163108826,
-0.03709183633327484,
-0.02179139107465744,
-0.06157243251800537,
0.008306662552058697,
-0.021281544119119644,
-0.007865170016884804,
0.10527878999710083,
0.007143213879317045,
-0.1573985517024994,
-0.04055684432387352,
0.037309616804122925,
-0.0614623948931694,
0.1271701604127884,
0.025690581649541855,
0.05949915200471878,
0.017543239519000053,
0.004615615587681532,
-0.04835885763168335,
0.027533981949090958,
-0.020685922354459763,
0.031796302646398544,
-0.013375918380916119,
-0.019669059664011,
0.04829173907637596,
-0.10610821098089218,
0.053406402468681335,
-0.023808211088180542,
0.034600015729665756,
-0.06912676244974136,
-0.08995288610458374,
0.029613537713885307,
0.011789835058152676,
0.03149712458252907,
0.007419796194881201,
0.011296370066702366,
0.007363796234130859,
0.02703331969678402,
-0.0301689263433218,
0.05989581719040871,
-0.020027441903948784,
-0.05824326351284981,
-0.09133689850568771,
-0.07824074476957321,
0.04263518005609512,
-0.043000977486371994,
0.05858594924211502,
0.09984218329191208,
0.008747766725718975,
0.03358008340001106,
-0.011036518961191177,
-0.04083817079663277,
-0.015015876851975918,
0.036373063921928406,
-0.00893496349453926,
0.03702099993824959,
-0.028713718056678772,
0.05432920902967453,
0.00714012049138546,
-0.04985879361629486,
-0.027618758380413055,
0.01981280744075775,
0.028425166383385658,
-0.052814316004514694,
0.018782177940011024,
0.02567245066165924,
0.01203909795731306,
-0.030530957505106926,
-0.04879770800471306,
0.023007133975625038,
-0.04999382793903351,
0.10457818955183029,
-0.04379231855273247,
0.04934000223875046,
0.020803222432732582,
0.03591218590736389,
-0.03350117802619934,
0.01538627315312624,
0.04791790246963501,
0.010508585721254349,
0.03306501731276512,
-0.1059492826461792,
-0.013060611672699451,
0.013759308494627476,
-0.07577676326036453,
0.007543392479419708,
-0.039261121302843094,
-0.0190289206802845,
-0.07445526868104935,
-0.028654523193836212,
0.03264794126152992,
0.031171292066574097,
-0.059649910777807236,
-0.010668058879673481,
-0.021192271262407303,
0.07266803830862045,
-0.020068245008587837,
0.05964360013604164,
0.018772687762975693,
-0.029452959075570107,
-0.025090448558330536,
0.060165852308273315,
-0.012401227839291096,
-0.029296666383743286,
-0.007700867485255003,
0.02742638997733593,
0.0053786830976605415,
0.036513760685920715,
0.01935485750436783,
0.03578558191657066,
0.020939907059073448,
0.05337072163820267,
-0.08178479224443436,
-0.09100355952978134,
0.005853473674505949,
0.04372266307473183,
-0.06499570608139038,
-0.037159211933612823,
-1.8292234935130248e-33,
0.0919589027762413,
-0.017896166071295738,
0.03158485144376755,
0.07034830003976822,
0.04431276395916939,
0.028910063207149506,
0.005040197633206844,
-0.033089108765125275,
-0.07492076605558395,
-0.0680723637342453,
0.05877445265650749,
-0.08555184304714203,
-0.028845299035310745,
0.07236206531524658,
0.10719140619039536,
-0.05863793566823006,
-0.03428344056010246,
0.018005533143877983,
-0.009820893406867981,
0.059635650366544724,
-0.04473567008972168,
0.039569687098264694,
-0.0771004781126976,
-0.032005663961172104,
-0.041424572467803955,
-0.007785373367369175,
0.05827239528298378,
-0.10893820971250534,
0.012499925680458546,
0.006131903268396854,
-0.12631990015506744,
0.07298629730939865,
-0.008218495175242424,
0.030584122985601425,
-0.010064800269901752,
-0.015055311843752861,
-0.0571783185005188,
-0.09549951553344727,
0.06183813512325287,
0.017482679337263107,
0.0019626689609140158,
0.059301476925611496,
-0.019437281414866447,
-0.06537952274084091,
-0.07600555568933487,
-0.017928654327988625,
0.03795739635825157,
0.14896589517593384,
-0.05439256876707077,
0.024629324674606323,
0.07314937561750412,
0.07429095357656479,
0.008168157190084457,
-0.08799435943365097,
-0.03976203501224518,
0.003412614343687892,
0.10204575955867767,
-0.03744686394929886,
0.008210143074393272,
0.022621745243668556,
0.06705278903245926,
0.048346277326345444,
-0.013187428936362267,
0.09212800115346909,
0.027037175372242928,
0.030353063717484474,
0.03497549146413803,
0.0887189656496048,
0.03601600229740143,
0.01749257929623127,
-0.09821704030036926,
0.02845393493771553,
0.017436306923627853,
-0.024962764233350754,
0.027271166443824768,
-0.00521138496696949,
0.04049159958958626,
-0.06102631241083145,
0.04606432095170021,
0.018773291260004044,
-0.0006014008540660143,
-0.07407957315444946,
-0.03320090472698212,
0.001992839854210615,
-0.029935335740447044,
-0.02110198326408863,
-0.00958308670669794,
-0.01905989833176136,
-0.056046873331069946,
-0.06901285797357559,
0.020559262484312057,
-0.012590721249580383,
-0.004340391140431166,
0.030249368399381638,
-0.0077079348266124725,
2.0247183059593486e-33,
0.049112796783447266,
-0.005391660612076521,
-0.016085008159279823,
0.03814353421330452,
0.011749586090445518,
0.023041000589728355,
0.04576782509684563,
-0.009540854021906853,
0.10138910263776779,
-0.011115931905806065,
-0.08250802010297775,
-0.023675180971622467,
-0.025883350521326065,
-0.049081649631261826,
-0.033868081867694855,
-0.01896359585225582,
0.051027148962020874,
0.001620282419025898,
-0.07344484329223633,
-0.010158361867070198,
0.041775915771722794,
0.007468759082257748,
0.05014659836888313,
0.01568497344851494,
-0.07743600010871887,
-0.01978270336985588,
-0.07244108617305756,
0.06965915858745575,
0.021459445357322693,
-0.05293576791882515,
-0.045094892382621765,
0.040252167731523514,
0.02767019160091877,
-0.05237741023302078,
-0.12270975112915039,
0.039112675935029984,
0.05124989151954651,
-0.01114636193960905,
-0.03687875717878342,
0.05292409285902977,
0.07759534567594528,
-0.028809944167733192,
-0.05170971900224686,
0.038594357669353485,
-0.04840516671538353,
0.0303299892693758,
-0.1013282909989357,
0.023698771372437477,
-0.09001828730106354,
0.09143993258476257,
-0.11494330316781998,
0.0582147054374218,
-0.07666296511888504,
-0.03092186525464058,
-0.06354007869958878,
-0.057968270033597946,
-0.037400662899017334,
-0.007146541960537434,
0.08511049300432205,
-0.013988257385790348,
-0.03208385407924652,
0.028309017419815063,
0.006780822761356831,
-0.06391303241252899,
0.02984585426747799,
0.01381753757596016,
-0.0023224856704473495,
0.001714973128400743,
0.06514488160610199,
0.023994872346520424,
0.12639831006526947,
-0.012938241474330425,
0.033018726855516434,
-0.02057119831442833,
0.03310750424861908,
0.03755725547671318,
0.06668657809495926,
-0.041681382805109024,
0.010159132070839405,
-0.06794396042823792,
-0.0023072664625942707,
-0.0009614487062208354,
0.025016050785779953,
0.01700131967663765,
0.04690666124224663,
0.13117246329784393,
0.04677712544798851,
0.09626584500074387,
0.020468590781092644,
0.016777757555246353,
-0.047163695096969604,
0.07108281552791595,
0.040267784148454666,
0.01013211626559496,
-0.0359467975795269,
-3.567999584674908e-8,
-0.07007501274347305,
0.01767791248857975,
-0.07555655390024185,
0.08588556200265884,
-0.06272076070308685,
-0.06629505753517151,
0.07632207870483398,
0.09041514992713928,
0.05547056719660759,
-0.013099268078804016,
0.004722962621599436,
0.001639478257857263,
-0.05102842301130295,
-0.058777425438165665,
-0.06892894208431244,
0.023500515148043633,
-0.020540276542305946,
0.08414940536022186,
-0.025128738954663277,
0.04415363073348999,
0.008208539336919785,
0.014018227346241474,
0.04298871383070946,
-0.07249618321657181,
-0.018032627180218697,
-0.0318937785923481,
0.006030217744410038,
0.09107619524002075,
0.07511972635984421,
-0.05744439363479614,
0.01940295845270157,
-0.004594055004417896,
-0.017911117523908615,
0.010272929444909096,
-0.04737067595124245,
0.029395051300525665,
0.03654944896697998,
0.06988521665334702,
-0.0031310885678976774,
-0.007253733929246664,
0.023897860199213028,
0.013419609516859055,
0.039659712463617325,
0.04195639118552208,
-0.013884799554944038,
0.04151979833841324,
0.05019677057862282,
-0.11670456826686859,
0.10817816853523254,
-0.03437962383031845,
0.016635717824101448,
0.04555096477270126,
-0.03511671721935272,
0.05294232815504074,
0.02171449176967144,
0.04063423350453377,
0.056584764271974564,
0.03504278138279915,
0.021732782945036888,
0.02546548657119274,
0.022823119536042213,
0.029562359675765038,
-0.06915120780467987,
0.0024085219483822584
] |
Let the Code LLM Edit Itself When You Edit the Code | https://openreview.net/forum?id=zqzsZ5cXbB | [
"Zhenyu He",
"Jun Zhang",
"Shengjie Luo",
"Jingjing Xu",
"Zhi Zhang",
"Di He"
] | Poster | In this work, we investigate a typical scenario in code generation where a developer edits existing code in real time and requests a code assistant, e.g., a large language model, to re-predict the next token or next line on the fly. Naively, the LLM needs to re-encode the entire KV cache to provide an accurate prediction. However, this process is computationally expensive, especially when the sequence length is long. Simply encoding the edited subsequence and integrating it to the original KV cache meets the temporal confusion problem, leading to significantly worse performance. We address this efficiency and accuracy trade-off by introducing $\underline{\textbf{P}\text{ositional}\ \textbf{I}\text{ntegrity}\ \textbf{E}\text{ncoding}}$ (PIE). Building upon the rotary positional encoding, PIE first removes the rotary matrices in the Key cache that introduce temporal confusion and then reapplies the correct rotary matrices. This process ensures that positional relationships between tokens are correct and requires only a single round of matrix multiplication. We validate the effectiveness of PIE through extensive experiments on the RepoBench-C-8k dataset, utilizing DeepSeek-Coder models with 1.3B, 6.7B, and 33B parameters. Our evaluation includes three real-world coding tasks: code insertion, code deletion, and multi-place code editing. Results demonstrate that PIE reduces computational overhead by over 85% compared to the standard full recomputation approach across all model sizes and tasks while well approximating the model performance. | code generation, efficiency, large language model, code assistant | null | 307 | 2407.03157 | [
-0.07708349078893661,
0.00076724385144189,
0.013157304376363754,
-0.0004379733873065561,
-0.05111716687679291,
-0.02213093265891075,
-0.10108578205108643,
0.00828147679567337,
0.06113777682185173,
-0.012360230088233948,
-0.026293540373444557,
-0.023933980613946915,
-0.015799816697835922,
-0.0544322207570076,
-0.015950430184602737,
-0.04225309565663338,
0.0594249852001667,
0.11914791166782379,
-0.050281163305044174,
-0.08702725172042847,
0.0628620907664299,
0.003782594343647361,
0.0043126544915139675,
0.027417892590165138,
0.05114763602614403,
0.03787332400679588,
0.006590587552636862,
0.05644636228680611,
0.07617833465337753,
-0.02642168290913105,
0.01626623421907425,
0.10599518567323685,
0.036986760795116425,
0.01827867701649666,
-0.021185291931033134,
0.04613650590181351,
-0.08198211342096329,
-0.028011074289679527,
-0.028346456587314606,
-0.04538106918334961,
-0.025168707594275475,
-0.00011994573287665844,
-0.0014839492505416274,
0.045972466468811035,
0.11115701496601105,
0.030046282336115837,
0.04950379952788353,
-0.08508334308862686,
-0.13875333964824677,
-0.04094693809747696,
-0.023723991587758064,
-0.000012332285223237704,
-0.005204828921705484,
0.017594028264284134,
-0.06018957495689392,
-0.04883705824613571,
0.06953302025794983,
0.04177168756723404,
0.024209024384617805,
-0.04118918627500534,
0.009277778677642345,
-0.06652174890041351,
-0.009734977968037128,
-0.005405425559729338,
0.03889763727784157,
-0.023211605846881866,
0.07246527820825577,
0.016788816079497337,
0.003294026479125023,
0.006674817763268948,
-0.020669179037213326,
0.041122473776340485,
-0.0613909475505352,
0.07220458984375,
-0.03683169186115265,
0.11177942901849747,
0.05569784715771675,
0.02870643138885498,
0.0523253008723259,
-0.10448580235242844,
0.03308972716331482,
0.011564509011805058,
-0.02063511125743389,
-0.024256443604826927,
0.095652274787426,
-0.01533648930490017,
0.07684481889009476,
-0.023205885663628578,
0.07397405058145523,
0.03150539845228195,
0.05758850276470184,
-0.07604993879795074,
0.0733703151345253,
-0.044513680040836334,
0.03911645710468292,
0.003744548885151744,
0.06325773894786835,
0.016297679394483566,
0.04936639964580536,
0.09246654808521271,
0.023405781015753746,
0.08662944287061691,
-0.06772302091121674,
-0.0179935023188591,
0.026437321677803993,
0.007831484079360962,
0.0631847232580185,
0.047848086804151535,
-0.0042909677140414715,
-0.06577375531196594,
0.05094491317868233,
0.01836489327251911,
-0.003852772992104292,
-0.009321395307779312,
0.03935486078262329,
0.027775796130299568,
0.037771645933389664,
0.03229851275682449,
0.06387016177177429,
0.08622764050960541,
-0.02709856815636158,
-0.0397377610206604,
-0.005845594685524702,
0.07931581139564514,
-0.030876178294420242,
-0.05295785516500473,
-0.04084383323788643,
2.8575147275436682e-33,
-0.011511215940117836,
0.02737489342689514,
-0.016612565144896507,
-0.0019128104904666543,
0.026628024876117706,
-0.01789320632815361,
0.0198038499802351,
0.0047112624160945415,
-0.033378716558218,
-0.015338221564888954,
0.02324720285832882,
-0.01726345345377922,
0.026132427155971527,
0.04769628494977951,
0.04324961081147194,
0.028323695063591003,
-0.03240262717008591,
0.017636071890592575,
-0.06940395385026932,
-0.029832016676664352,
0.05822031944990158,
-0.045882489532232285,
0.006613511126488447,
-0.09980388730764389,
-0.03977683559060097,
0.06791207939386368,
0.07083749771118164,
-0.05483315885066986,
-0.021523112431168556,
0.020326271653175354,
-0.06518586724996567,
-0.024348320439457893,
-0.008222363889217377,
0.01626926101744175,
0.019430845975875854,
-0.02331230603158474,
0.053141701966524124,
-0.08960621058940887,
0.044559694826602936,
-0.06303940713405609,
0.04132194072008133,
0.01833449862897396,
0.00722189387306571,
-0.05515839904546738,
-0.057022515684366226,
-0.07769240438938141,
-0.0350700318813324,
0.06117232143878937,
-0.02030663564801216,
-0.003973463084548712,
0.09608405828475952,
0.061856891959905624,
-0.060936346650123596,
-0.023208891972899437,
-0.002113927621394396,
-0.027991892769932747,
0.035797227174043655,
-0.061554573476314545,
0.039965249598026276,
0.15227392315864563,
-0.02182909846305847,
0.04416366294026375,
0.022366425022482872,
0.08474936336278915,
0.0019634675700217485,
0.057526472955942154,
-0.01522225234657526,
0.06071526184678078,
0.03365908935666084,
-0.05994032323360443,
-0.06300253421068192,
-0.01765023171901703,
-0.018611811101436615,
-0.06820956617593765,
-0.050964634865522385,
-0.06467051804065704,
0.07521776854991913,
-0.03301779180765152,
-0.05444186553359032,
-0.024017656221985817,
0.042045217007398605,
0.07923019677400589,
-0.02545267716050148,
0.015546701848506927,
-0.001557262148708105,
-0.15888755023479462,
0.03389931842684746,
-0.033125605434179306,
-0.05012281984090805,
0.04592065513134003,
0.023025652393698692,
-0.07879890501499176,
0.14055737853050232,
0.032953668385744095,
-0.021109657362103462,
-1.0870827574675547e-33,
-0.02822742983698845,
-0.02455252967774868,
-0.023596545681357384,
0.09181395173072815,
-0.07488757371902466,
-0.12100273370742798,
-0.05374757945537567,
-0.023728100582957268,
-0.059368737041950226,
-0.05704348161816597,
-0.0722021535038948,
-0.07718277722597122,
0.01436327863484621,
0.014165176078677177,
0.05758887529373169,
-0.010976746678352356,
-0.029397930949926376,
-0.03548537567257881,
0.03210453689098358,
0.05937432870268822,
0.04002301022410393,
0.05032392963767052,
-0.03522854670882225,
0.02267862670123577,
-0.031132077798247337,
0.08338665962219238,
-0.04113147780299187,
0.0814124196767807,
-0.028091173619031906,
-0.020100446417927742,
-0.05717088654637337,
-0.028193030506372452,
-0.06150197237730026,
-0.01853221468627453,
-0.006767802871763706,
0.014489378780126572,
0.059766341000795364,
0.018549969419836998,
0.016805848106741905,
0.02391739934682846,
0.03618122264742851,
-0.008623464964330196,
-0.0785093680024147,
0.009991146624088287,
-0.004666099324822426,
0.03567027673125267,
-0.059215445071458817,
0.028389379382133484,
0.06543093919754028,
0.008109270595014095,
0.07696624100208282,
0.013405798003077507,
-0.04572802409529686,
0.029752690345048904,
0.001822628197260201,
0.008927734568715096,
-0.05388737842440605,
0.0040819537825882435,
0.02398412674665451,
-0.032553549855947495,
-0.03176579624414444,
-0.05443048104643822,
0.08513934910297394,
-0.07359562814235687,
0.05568411573767662,
-0.009191889315843582,
-0.015526402741670609,
-0.008813399821519852,
-0.014581789262592793,
-0.06723848730325699,
0.07570571452379227,
-0.041218843311071396,
-0.05524038150906563,
-0.029139196500182152,
0.018953921273350716,
-0.08334730565547943,
0.08110734075307846,
-0.05100588873028755,
-0.016460709273815155,
-0.05038873106241226,
-0.13688813149929047,
0.00682278536260128,
-0.011648055166006088,
0.016539882868528366,
-0.08254702389240265,
0.05822576582431793,
0.034849442541599274,
0.03283201903104782,
-0.006414773408323526,
-0.030715033411979675,
-0.025508031249046326,
0.052740346640348434,
0.04533218592405319,
0.029476821422576904,
0.001063282135874033,
-5.6986181107276934e-8,
0.03331562131643295,
-0.01939646527171135,
-0.03072989545762539,
0.09379294514656067,
0.10956323146820068,
-0.09941386431455612,
-0.037843480706214905,
0.06501781940460205,
0.02105383574962616,
-0.029826173558831215,
0.1300176978111267,
-0.03496451675891876,
-0.02777956798672676,
-0.04405922070145607,
0.026492387056350708,
0.10105828940868378,
-0.050973642617464066,
-0.009053446352481842,
-0.06380032002925873,
-0.015372821129858494,
-0.01572393625974655,
0.08572936803102493,
0.042967237532138824,
-0.010967499576508999,
-0.02384422905743122,
-0.04467490315437317,
0.050897303968667984,
0.12171570211648941,
0.07106926292181015,
-0.006690685171633959,
0.010303105227649212,
0.005217462778091431,
0.029667718335986137,
0.0442480705678463,
-0.005351212341338396,
-0.01603706181049347,
0.013812804594635963,
0.024101819843053818,
0.07138340920209885,
0.01910032331943512,
0.01743347942829132,
0.013269624672830105,
-0.0586676262319088,
0.07299474626779556,
-0.02688981592655182,
-0.06281763315200806,
-0.03882062807679176,
-0.05975620076060295,
-0.029964694753289223,
0.00010125769767910242,
-0.06222470849752426,
0.052249595522880554,
-0.0560150146484375,
0.03289297595620155,
0.02237127721309662,
0.06285566091537476,
-0.008451258763670921,
-0.05121597275137901,
0.054337210953235626,
0.09517695754766464,
0.024343635886907578,
-0.010679828934371471,
-0.06941618770360947,
-0.07246026396751404
] |
CityGaussianV2: Efficient and Geometrically Accurate Reconstruction for Large-Scale Scenes | https://openreview.net/forum?id=a3ptUbuzbW | [
"Yang Liu",
"Chuanchen Luo",
"Zhongkai Mao",
"Junran Peng",
"Zhaoxiang Zhang"
] | Poster | Recently, 3D Gaussian Splatting (3DGS) has revolutionized radiance field reconstruction, manifesting efficient and high-fidelity novel view synthesis. However, accurately representing surfaces, especially in large and complex scenarios, remains a significant challenge due to the unstructured nature of 3DGS. In this paper, we present CityGaussianV2, a novel approach for large-scale scene reconstruction that addresses critical challenges related to geometric accuracy and efficiency. Building on the favorable generalization capabilities of 2D Gaussian Splatting (2DGS), we address its convergence and scalability issues. Specifically, we implement a decomposed-gradient-based densification and depth regression technique to eliminate blurry artifacts and accelerate convergence. To scale up, we introduce an elongation filter that mitigates Gaussian count explosion caused by 2DGS degeneration. Furthermore, we optimize the CityGaussian pipeline for parallel training, achieving up to 10$\times$ compression, at least 25\% savings in training time, and a 50\% decrease in memory usage. We also established standard geometry benchmarks under large-scale scenes. Experimental results demonstrate that our method strikes a promising balance between visual quality, geometric accuracy, as well as storage and training costs. | neural rendering, novel view synthesis, large-scale scene, radiance field, surfel splatting, surfel reconstruction | CityGaussianV2 addresses scalability and convergence problem of surface reconstruction algorithms, achieving SOTA geometric accuracy and high training efficiency under large scale scenes. | 305 | 2411.00771 | [
-0.03866659477353096,
-0.07885979115962982,
0.055598460137844086,
0.004430317785590887,
0.08359900116920471,
-0.0508391410112381,
-0.05178754776716232,
-0.07878958433866501,
-0.062484417110681534,
-0.06464327871799469,
-0.016155241057276726,
-0.038539864122867584,
-0.023614222183823586,
0.07174623012542725,
-0.028171595185995102,
0.020046578720211983,
0.0635189414024353,
0.14325404167175293,
-0.03108961693942547,
0.003612237749621272,
0.013266024179756641,
-0.01901593990623951,
-0.0775936096906662,
0.015268576331436634,
0.1176149994134903,
0.025593403726816177,
0.06611040234565735,
-0.04074348509311676,
-0.04249925911426544,
-0.004166417755186558,
0.03265202045440674,
-0.008920079097151756,
-0.023209813982248306,
-0.04247802123427391,
-0.027773737907409668,
0.021528512239456177,
-0.007010749075561762,
0.026049157604575157,
0.0019598151557147503,
-0.015724819153547287,
0.00285120727494359,
0.0583638958632946,
-0.04209264740347862,
0.022202664986252785,
0.10471156984567642,
0.05788000300526619,
0.059774208813905716,
-0.08867349475622177,
0.08095959573984146,
-0.03477514535188675,
-0.025546930730342865,
0.00532476045191288,
-0.000013130323168297764,
-0.008672002702951431,
-0.010693608783185482,
0.04469845071434975,
-0.03538896143436432,
-0.04241696745157242,
0.045010678470134735,
0.005711437202990055,
0.04847761243581772,
-0.01036764495074749,
-0.0024656581226736307,
-0.0050087785348296165,
-0.01867414265871048,
-0.02258342131972313,
0.08543059974908829,
-0.05602394416928291,
0.009348208084702492,
0.07749830931425095,
-0.0801202729344368,
0.04559587687253952,
-0.013658162206411362,
-0.012587886303663254,
-0.01531780231744051,
0.016569189727306366,
0.005546310916543007,
-0.08359336107969284,
0.023486334830522537,
-0.03489422798156738,
0.07902166247367859,
-0.001282267621718347,
0.05539701133966446,
0.006021731533110142,
-0.018819188699126244,
-0.0005939552211202681,
-0.013129142113029957,
0.12246722728013992,
-0.0075371768325567245,
0.008490784093737602,
-0.04279150813817978,
0.02245722897350788,
-0.0539783239364624,
-0.03676554933190346,
-0.08812323957681656,
-0.03221144527196884,
-0.018913250416517258,
-0.054938919842243195,
0.03329940512776375,
0.05048075690865517,
0.041180990636348724,
-0.05503493919968605,
0.027118830010294914,
0.03237276151776314,
0.05328670144081116,
-0.02757534384727478,
0.09569786489009857,
0.06474568694829941,
-0.0002179862349294126,
0.039279572665691376,
0.07883266359567642,
0.026329997926950455,
-0.006607656367123127,
-0.015265998430550098,
0.018455952405929565,
0.11674109101295471,
0.0002801910450216383,
-0.007351536303758621,
-0.033959466964006424,
0.02505398727953434,
-0.011881995014846325,
-0.01890268363058567,
-0.08409200608730316,
0.07410455495119095,
0.012437173165380955,
-0.02209855243563652,
-0.0010275662643834949,
2.2710999015107846e-33,
-0.0160266924649477,
0.08263547718524933,
0.010791166685521603,
0.06469393521547318,
0.023022137582302094,
0.02789636142551899,
0.03633773699402809,
0.03135022148489952,
-0.024978214874863625,
0.050576262176036835,
-0.07206668704748154,
-0.07566796988248825,
-0.000012227829756739084,
0.07537113130092621,
0.031115463003516197,
-0.005109054502099752,
-0.02891843393445015,
0.02398662641644478,
-0.03153827041387558,
0.07181965559720993,
-0.0013716642279177904,
-0.06307659298181534,
0.0030523191671818495,
-0.02573457732796669,
-0.054572030901908875,
-0.024682005867362022,
0.026694338768720627,
-0.007012910209596157,
0.001644063857384026,
-0.011183555237948895,
-0.014344974420964718,
0.035641901195049286,
0.005936054512858391,
0.03592956066131592,
0.01551796030253172,
-0.04985012486577034,
0.04298936948180199,
-0.09622059762477875,
-0.027488572522997856,
-0.009349307045340538,
-0.08227505534887314,
0.1364033818244934,
-0.0012883128365501761,
-0.021620430052280426,
-0.048671308904886246,
0.023760497570037842,
0.06273973733186722,
0.06409723311662674,
0.00003525565261952579,
0.00038865397800691426,
0.13228683173656464,
0.0018408360192552209,
-0.10281313955783844,
0.020213164389133453,
0.005462305154651403,
-0.007862513884902,
0.05614660680294037,
-0.06365202367305756,
0.08272965252399445,
0.021102797240018845,
-0.004204512108117342,
0.0019301183056086302,
-0.052675120532512665,
0.05084891989827156,
-0.044741809368133545,
-0.027436979115009308,
0.05824114754796028,
0.0772973895072937,
-0.018838780000805855,
-0.016206612810492516,
-0.03346133604645729,
-0.027374107390642166,
0.04992122948169708,
-0.0234285369515419,
0.051363881677389145,
0.009307165630161762,
0.048412714153528214,
-0.05734223872423172,
0.030246544629335403,
0.06977032124996185,
-0.09435070306062698,
0.03533785790205002,
0.012192419730126858,
-0.08859720081090927,
0.010309852659702301,
-0.056199654936790466,
-0.005106639117002487,
-0.0384015254676342,
-0.026250245049595833,
-0.051816172897815704,
-0.07410784810781479,
-0.06062138453125954,
-0.027266496792435646,
-0.07334113121032715,
-0.06498002260923386,
-9.461700966562445e-34,
0.028177816420793533,
0.005877595394849777,
-0.05485653132200241,
0.03415359929203987,
-0.04942109435796738,
-0.04092485457658768,
-0.030416976660490036,
-0.00140247936360538,
0.04342719912528992,
-0.08913055062294006,
-0.058393385261297226,
0.017066512256860733,
0.09243922680616379,
0.02633923664689064,
0.03503856062889099,
0.020872309803962708,
0.009777473285794258,
-0.04142812266945839,
-0.14280709624290466,
-0.07710287719964981,
-0.006162246223539114,
0.033588871359825134,
-0.0001353169936919585,
0.0004487086262088269,
-0.09402693063020706,
0.10218043625354767,
-0.04751262813806534,
0.029336197301745415,
0.014743896201252937,
0.003891645697876811,
-0.041430339217185974,
0.001028458820655942,
0.01653008535504341,
-0.02091996930539608,
-0.040686219930648804,
0.06148003786802292,
0.07416695356369019,
0.01237130630761385,
-0.04650559276342392,
-0.034811802208423615,
-0.019199544563889503,
0.009918286465108395,
-0.025659900158643723,
0.041244987398386,
-0.0035586259327828884,
0.0387999452650547,
0.025607233867049217,
-0.0174685250967741,
0.008063019253313541,
0.07578496634960175,
-0.06053386628627777,
0.07568491250276566,
-0.10041482001543045,
0.07493606209754944,
0.004141849000006914,
-0.10340207815170288,
-0.14533378183841705,
-0.0017503579147160053,
0.05760665237903595,
-0.04190313443541527,
0.0032407748512923717,
-0.055141184478998184,
-0.09255863726139069,
0.0455135703086853,
0.011612256988883018,
0.05517110601067543,
0.01688593626022339,
0.017970064654946327,
0.041179537773132324,
0.0836491584777832,
-0.025702891871333122,
-0.02908542938530445,
-0.044684022665023804,
0.037919797003269196,
-0.05101415514945984,
-0.041747257113456726,
0.03193718194961548,
0.02312730997800827,
0.05895495042204857,
0.07354588061571121,
0.09596480429172516,
0.008155020885169506,
0.031081680208444595,
0.09700854867696762,
0.14369334280490875,
0.04601427912712097,
-0.04188085347414017,
0.02386498637497425,
0.08506422489881516,
-0.03533049300312996,
-0.02626476250588894,
-0.0013598455116152763,
-0.005912291817367077,
-0.03460278734564781,
0.015313398092985153,
-4.2718504289496195e-8,
-0.03267285227775574,
0.09709351509809494,
-0.012174312025308609,
-0.06511372327804565,
-0.03967546671628952,
-0.059611976146698,
0.004166869446635246,
0.1190546452999115,
-0.008591064251959324,
-0.07294843345880508,
0.010858154855668545,
-0.0753789022564888,
-0.04337375983595848,
0.006930339150130749,
0.0504474937915802,
0.07076678425073624,
0.0027583036571741104,
-0.019607560709118843,
-0.013431866653263569,
0.08363368362188339,
-0.07648546993732452,
0.024926787242293358,
-0.05505459010601044,
0.026776328682899475,
-0.0013094551395624876,
0.02072327770292759,
-0.049181465059518814,
0.0895012617111206,
0.17257489264011383,
-0.061407264322042465,
0.011238988488912582,
-0.01970032975077629,
0.04130217432975769,
0.02460739016532898,
-0.04244639351963997,
-0.035290371626615524,
0.005925674922764301,
0.046716444194316864,
0.05078507959842682,
-0.0978759378194809,
-0.06329336017370224,
0.009317238815128803,
0.017670871689915657,
0.047361016273498535,
-0.009144235402345657,
0.06461131572723389,
0.06290075182914734,
-0.08946269750595093,
-0.010205492377281189,
-0.02229996584355831,
-0.004601113498210907,
0.0007055175374262035,
-0.006447960156947374,
0.06176483631134033,
0.046776216477155685,
-0.018842196092009544,
0.004514692816883326,
-0.019861580803990364,
0.0798911452293396,
-0.02312455326318741,
-0.06002090871334076,
-0.007323941681534052,
-0.10092733055353165,
-0.019156156107783318
] |
Debiasing Mini-Batch Quadratics for Applications in Deep Learning | https://openreview.net/forum?id=Q0TEVKV2cp | [
"Lukas Tatzel",
"Bálint Mucsányi",
"Osane Hackel",
"Philipp Hennig"
] | Poster | Quadratic approximations form a fundamental building block of machine learning methods. E.g., second-order optimizers try to find the Newton step into the minimum of a local quadratic proxy to the objective function; and the second-order approximation of a network's loss function can be used to quantify the uncertainty of its outputs via the Laplace approximation. When computations on the entire training set are intractable - typical for deep learning - the relevant quantities are computed on mini-batches. This, however, distorts and biases the shape of the associated *stochastic* quadratic approximations in an intricate way with detrimental effects on applications. In this paper, we (i) show that this bias introduces a systematic error, (ii) provide a theoretical explanation for it, (iii) explain its relevance for second-order optimization and uncertainty quantification via the Laplace approximation in deep learning, and (iv) develop and evaluate debiasing strategies. | quadratic Taylor approximation, mini-batching, second-order optimizers, conjugate gradients, uncertainty quantification, Laplace approximation, stochastic curvature, GGN, KFAC | This paper shows that mini-batching introduces biases in quadratic approximations to deep learning loss functions, discusses their impact on second-order optimization and uncertainty quantification, and proposes debiasing strategies. | 303 | 2410.14325 | [
-0.083170585334301,
0.010572361759841442,
0.05603225156664848,
0.0024145119823515415,
-0.008614853955805302,
0.026139305904507637,
-0.05285504832863808,
0.0033144918270409107,
0.006828570272773504,
-0.019191166386008263,
0.019233034923672676,
0.02282167784869671,
0.04049794375896454,
-0.029712196439504623,
-0.09679577499628067,
0.01386052742600441,
0.09130049496889114,
0.09792714565992355,
-0.1079162210226059,
-0.04710964858531952,
0.0235377736389637,
0.023558015003800392,
-0.02981235459446907,
-0.023352326825261116,
0.12014060467481613,
-0.018533512949943542,
0.0024669861886650324,
0.03246905282139778,
-0.009986421093344688,
-0.020986482501029968,
0.05999983847141266,
-0.05891215801239014,
0.01819450967013836,
-0.06432235240936279,
0.04312121123075485,
0.02596336230635643,
0.007071590982377529,
0.04230845347046852,
0.05103333294391632,
-0.025280553847551346,
0.04768158495426178,
0.0137252202257514,
0.02365083061158657,
0.05456790328025818,
0.08763684332370758,
-0.08354251831769943,
0.04260754585266113,
-0.09274832159280777,
0.03464497625827789,
0.0017391470028087497,
0.022008558735251427,
-0.006759528536349535,
-0.015303805470466614,
0.07306167483329773,
0.009039157070219517,
0.03673604875802994,
0.025792550295591354,
0.07749714702367783,
-0.04462578520178795,
0.09242621064186096,
0.06259682774543762,
-0.10031462460756302,
-0.02486071176826954,
-0.0374167338013649,
-0.0705605298280716,
-0.014007586054503918,
0.0451836958527565,
-0.025547757744789124,
0.00445950124412775,
0.07086000591516495,
0.031038528308272362,
0.06536612659692764,
-0.07721215486526489,
0.031699445098638535,
0.032436735928058624,
-0.0432954877614975,
0.11242164671421051,
-0.040239762514829636,
0.03830043599009514,
-0.05922741815447807,
0.03854285553097725,
-0.03495502471923828,
0.0347602404654026,
-0.06573808193206787,
0.09510710090398788,
-0.06002964824438095,
-0.027219021692872047,
-0.01758769527077675,
0.029638301581144333,
-0.03267292678356171,
-0.06279116123914719,
-0.042237788438797,
-0.10448256880044937,
-0.04445160552859306,
0.015355831012129784,
-0.031735360622406006,
0.023981420323252678,
-0.025841230526566505,
-0.03490760922431946,
0.12246107310056686,
0.007247408851981163,
-0.014819263480603695,
-0.02649785205721855,
-0.03597554564476013,
0.06058178469538689,
-0.07158003002405167,
0.042281463742256165,
0.025296306237578392,
0.06589160114526749,
-0.08267100900411606,
-0.028884191066026688,
0.018720215186476707,
0.0065248701721429825,
0.06006716191768646,
0.031213046982884407,
-0.05535205081105232,
0.09232696890830994,
-0.045351363718509674,
-0.02902890555560589,
0.04694577679038048,
-0.10431286692619324,
0.0028304907027632,
0.007563740946352482,
0.034411028027534485,
0.0012573731364682317,
0.005733545403927565,
-0.049478884786367416,
9.444018960132185e-34,
-0.0027095461264252663,
0.004656791687011719,
0.060314930975437164,
-0.07796099036931992,
0.028803225606679916,
-0.004600226413458586,
0.04628393426537514,
-0.03147192671895027,
0.017657876014709473,
0.02453407645225525,
-0.04951934888958931,
-0.04434662684798241,
-0.06098150834441185,
0.06831224262714386,
0.06064928323030472,
0.010309955105185509,
-0.022483952343463898,
-0.0004458192561287433,
0.08553121984004974,
-0.05080162361264229,
0.03382816165685654,
-0.05880875140428543,
-0.03875008225440979,
-0.01125916838645935,
-0.027231547981500626,
0.028615353628993034,
0.08068954199552536,
0.056016966700553894,
0.04002847522497177,
0.05757468193769455,
-0.1107117310166359,
-0.027681762352585793,
0.0076737902127206326,
-0.06998372822999954,
0.08414502441883087,
-0.009545370936393738,
0.0033058892004191875,
0.028045911341905594,
0.044109996408224106,
-0.03974518924951553,
-0.019620707258582115,
0.04067523032426834,
0.07467322796583176,
-0.035503171384334564,
-0.11427458375692368,
-0.08432146906852722,
0.0888402909040451,
-0.0452532134950161,
0.03845049440860748,
-0.09445788711309433,
-0.02671751379966736,
0.03641866147518158,
0.020213192328810692,
0.07087605446577072,
0.05376273766160011,
-0.011658906005322933,
0.019674822688102722,
0.032354358583688736,
0.01826503314077854,
0.005198726896196604,
0.036748502403497696,
0.00811141636222601,
-0.03817090392112732,
0.06192491203546524,
0.039506442844867706,
0.11044789850711823,
0.0010459405602887273,
0.029608195647597313,
0.0328369177877903,
-0.024686120450496674,
-0.03247622400522232,
-0.028513262048363686,
0.030637120828032494,
-0.11249776929616928,
0.0785522311925888,
0.02622152492403984,
0.03879706934094429,
-0.05498229339718819,
-0.004189332015812397,
-0.011843369342386723,
-0.035733792930841446,
0.08023903518915176,
-0.0637657418847084,
-0.0307615976780653,
-0.08991067111492157,
-0.0455639511346817,
-0.013652774505317211,
-0.060191698372364044,
0.028942931443452835,
-0.00840859767049551,
-0.12634962797164917,
0.009220468811690807,
-0.0443563275039196,
0.022266965359449387,
-0.019842496141791344,
-1.5894158501185624e-33,
-0.04821903631091118,
0.04858023300766945,
-0.003408203599974513,
0.0007486671092920005,
-0.013970003463327885,
-0.030269622802734375,
-0.09736049920320511,
-0.051923707127571106,
-0.031191101297736168,
-0.03818722441792488,
-0.0506841205060482,
0.015989387407898903,
-0.04707968980073929,
-0.022068671882152557,
0.01846129074692726,
0.03036111779510975,
-0.018874524161219597,
0.008318610489368439,
0.004736177623271942,
-0.08106209337711334,
0.04139012098312378,
0.07166371494531631,
-0.01965874433517456,
0.06520943343639374,
0.026463128626346588,
0.009547668509185314,
-0.0069425106048583984,
0.042071860283613205,
0.020539434626698494,
-0.09871438890695572,
-0.10597880929708481,
-0.006478995084762573,
-0.05809259042143822,
-0.02455723285675049,
-0.03156404569745064,
0.09163908660411835,
0.031562209129333496,
0.025592010468244553,
-0.03640984743833542,
0.008835547603666782,
0.10352303087711334,
-0.03359464183449745,
-0.05371798202395439,
0.015558142215013504,
0.06010027602314949,
-0.059068530797958374,
0.0026344768702983856,
-0.025986839085817337,
0.012551107443869114,
0.07502539455890656,
0.006518503651022911,
-0.002490855287760496,
-0.10277099907398224,
0.07454205304384232,
-0.0398421473801136,
-0.056063976138830185,
-0.03240090608596802,
0.021872488781809807,
0.0962805300951004,
0.07705305516719818,
-0.10980224609375,
-0.046450499445199966,
-0.10033470392227173,
-0.0012254536850377917,
-0.06224473565816879,
0.05669417232275009,
-0.08740317076444626,
0.06636010110378265,
0.11584638804197311,
0.02276691421866417,
-0.001801864942535758,
-0.03268551453948021,
-0.0026094713248312473,
0.06188679486513138,
-0.08680244535207748,
-0.017273135483264923,
0.04008857533335686,
0.0003440525324549526,
-0.022221341729164124,
0.020343197509646416,
0.09800320118665695,
0.022009313106536865,
0.08248040825128555,
-0.004385500214993954,
0.059052724391222,
-0.0027629919350147247,
-0.002829980105161667,
0.06895508617162704,
0.009009912610054016,
-0.0523468442261219,
-0.04046608507633209,
0.03680877014994621,
-0.007901007309556007,
0.024211561307311058,
-0.08244817703962326,
-4.3110922831601783e-8,
-0.050728265196084976,
0.062265586107969284,
0.07004114985466003,
0.012850455939769745,
0.030469760298728943,
-0.021279288455843925,
-0.021463504061102867,
0.036348648369312286,
0.0021352150943130255,
-0.05047452822327614,
0.05549604818224907,
0.020935554057359695,
-0.018249472603201866,
-0.07827270776033401,
-0.07902657240629196,
0.14093269407749176,
0.020718205720186234,
-0.046927280724048615,
0.0381222702562809,
-0.027847768738865852,
0.002509188139811158,
0.021752845495939255,
-0.014225011691451073,
-0.007744069676846266,
0.015262650325894356,
-0.086713045835495,
-0.044482965022325516,
0.04324822872877121,
0.02586224488914013,
0.04748216271400452,
-0.05975983291864395,
0.07305101305246353,
0.01855018362402916,
0.0028892098926007748,
0.012427969835698605,
0.08536984771490097,
0.06293416023254395,
-0.023246243596076965,
0.0015412586508318782,
-0.025370996445417404,
-0.022246213629841805,
0.06172866001725197,
-0.004675896838307381,
-0.008277595043182373,
0.004790535196661949,
0.029314115643501282,
0.01233835518360138,
0.013230334967374802,
0.07053723931312561,
-0.04147527366876602,
0.015324907377362251,
0.00588188786059618,
-0.02372751571238041,
0.08280538022518158,
-0.009521329775452614,
-0.041521329432725906,
0.009769105352461338,
-0.12588584423065186,
-0.019756413996219635,
0.05060625076293945,
0.03855007141828537,
0.055231235921382904,
-0.06281372904777527,
0.012511790730059147
] |
Denoising as Adaptation: Noise-Space Domain Adaptation for Image Restoration | https://openreview.net/forum?id=jsBhmOCKYs | [
"Kang Liao",
"Zongsheng Yue",
"Zhouxia Wang",
"Chen Change Loy"
] | Poster | Although learning-based image restoration methods have made significant progress, they still struggle with limited generalization to real-world scenarios due to the substantial domain gap caused by training on synthetic data. Existing methods address this issue by improving data synthesis pipelines, estimating degradation kernels, employing deep internal learning, and performing domain adaptation and regularization. Previous domain adaptation methods have sought to bridge the domain gap by learning domain-invariant knowledge in either feature or pixel space. However, these techniques often struggle to extend to low-level vision tasks within a stable and compact framework. In this paper, we show that it is possible to perform domain adaptation via the noise space using diffusion models. In particular, by leveraging the unique property of how auxiliary conditional inputs influence the multi-step denoising process, we derive a meaningful *diffusion loss* that guides the restoration model in progressively aligning both restored synthetic and real-world outputs with a target clean distribution. We refer to this method as *denoising as adaptation*. To prevent shortcuts during joint training, we present crucial strategies such as channel-shuffling layer and residual-swapping contrastive learning in the diffusion model. They implicitly blur the boundaries between conditioned synthetic and real data and prevent the reliance of the model on easily distinguishable features. Experimental results on three classical image restoration tasks, namely denoising, deblurring, and deraining, demonstrate the effectiveness of the proposed method. | Image Restoration, Domain Adaptation, Diffusion Loss | We propose a novel domain adaptation method in the noise space for image restoration, guided by a diffusion loss that leverages auxiliary conditional inputs. | 302 | 2406.18516 | [
-0.04172983020544052,
-0.1259901374578476,
0.057993438094854355,
0.04114416986703873,
0.05114193260669708,
-0.07295309752225876,
-0.040380243211984634,
-0.1193152591586113,
-0.021485621109604836,
-0.009087832644581795,
-0.038020290434360504,
0.016505815088748932,
0.02359466440975666,
0.0028253616765141487,
-0.04753118008375168,
0.011586147360503674,
0.022581208497285843,
0.14073026180267334,
-0.1267554610967636,
-0.036875806748867035,
-0.0016767901834100485,
-0.027811391279101372,
-0.01904313638806343,
0.0296306274831295,
0.07056081295013428,
0.03980493173003197,
0.09449460357427597,
-0.022275861352682114,
-0.0025378535501658916,
-0.0688873827457428,
-0.009688097983598709,
-0.00956409890204668,
-0.047109074890613556,
-0.0618579164147377,
0.0031982562504708767,
0.058949731290340424,
-0.03899227827787399,
0.03953159600496292,
0.04904135689139366,
-0.046126674860715866,
-0.011646364815533161,
0.00024640042101964355,
-0.013695145025849342,
-0.039287496358156204,
0.027169210836291313,
0.04752639681100845,
0.003940875642001629,
-0.06668810546398163,
0.093633733689785,
-0.001222957973368466,
0.05604470148682594,
0.05144306272268295,
-0.06967316567897797,
0.12696726620197296,
0.0016657249070703983,
0.05862707272171974,
0.09404852241277695,
0.05253645032644272,
-0.021538088098168373,
0.02552490122616291,
0.018994146957993507,
-0.06292355805635452,
0.005936539266258478,
0.01836496777832508,
-0.04741920903325081,
-0.011405530385673046,
0.0351562425494194,
0.030042918398976326,
-0.039456721395254135,
0.04011712968349457,
-0.09489956498146057,
0.0701490044593811,
-0.017874639481306076,
0.011738422326743603,
0.050095412880182266,
0.050031304359436035,
0.07716725766658783,
-0.041457291692495346,
0.02858791872859001,
-0.06852930039167404,
0.12270757555961609,
0.05448344349861145,
0.023643959313631058,
-0.0035140207037329674,
0.11465154588222504,
0.03164253383874893,
-0.008936074562370777,
-0.0023970315232872963,
-0.00010573936742730439,
-0.03400756046175957,
-0.06808815151453018,
-0.001704439870081842,
-0.036813583225011826,
0.04304629936814308,
-0.004003599286079407,
0.01334652490913868,
0.03306349739432335,
-0.011475990526378155,
-0.03426998853683472,
0.08531370759010315,
-0.029150577262043953,
0.0022033913992345333,
-0.00015257135964930058,
-0.057233016937971115,
0.06848160177469254,
-0.03914802893996239,
0.09291226416826248,
0.06902419775724411,
0.0038084450643509626,
-0.046362586319446564,
0.035001520067453384,
0.045757800340652466,
0.044072773307561874,
0.023309003561735153,
0.06106056272983551,
-0.06558643281459808,
0.027751069515943527,
-0.03668740764260292,
-0.05732916668057442,
-0.02561338245868683,
-0.012424827553331852,
0.0091924536973238,
-0.0793209969997406,
-0.02390015684068203,
0.035120513290166855,
0.0026252788957208395,
0.00879396591335535,
4.682712612082607e-34,
0.029000939801335335,
0.02530129812657833,
0.045142896473407745,
-0.019202951341867447,
0.033808205276727676,
-0.03663382679224014,
0.016326984390616417,
-0.05783449113368988,
0.012552815489470959,
-0.012332488782703876,
0.012091903947293758,
-0.008025375194847584,
-0.10081426799297333,
0.0846029669046402,
0.015543628484010696,
0.02722746692597866,
-0.0030655746813863516,
0.05216771364212036,
0.06023024022579193,
0.009802840650081635,
-0.04070260003209114,
0.028516249731183052,
-0.09860730916261673,
0.006635066121816635,
0.01922982558608055,
0.013309205882251263,
0.05037982761859894,
-0.03537502512335777,
0.014418537728488445,
-0.001545337843708694,
-0.15362761914730072,
0.04509429261088371,
-0.024718591943383217,
0.02778247930109501,
0.030162706971168518,
-0.0322357602417469,
0.07269089668989182,
0.024970101192593575,
-0.048480045050382614,
-0.006991805974394083,
-0.00021748905419372022,
0.08090751618146896,
-0.016689348965883255,
-0.0316917821764946,
-0.06312749534845352,
-0.05670013278722763,
0.04818463325500488,
0.005785796325653791,
-0.0694953054189682,
-0.04355766251683235,
0.055751949548721313,
-0.033484481275081635,
-0.013776544481515884,
-0.10328107327222824,
-0.04358945041894913,
-0.023870371282100677,
0.09684569388628006,
-0.06856086850166321,
0.050904110074043274,
-0.010840319097042084,
0.051535140722990036,
0.008840047754347324,
-0.04924813657999039,
0.1362719088792801,
0.03305508941411972,
0.008816301822662354,
0.02132228948175907,
0.019349008798599243,
-0.06479083746671677,
-0.08462077379226685,
-0.04413791745901108,
0.05336955189704895,
-0.0520993247628212,
-0.008795227855443954,
0.07377027720212936,
0.008005158044397831,
0.01968766748905182,
-0.04275995492935181,
0.02816171385347843,
0.009204545989632607,
-0.008323279209434986,
-0.010945814661681652,
-0.08188959956169128,
-0.07896146178245544,
0.004742621909826994,
0.047501090914011,
-0.01156984455883503,
-0.12026309221982956,
-0.047204919159412384,
-0.09430994838476181,
-0.06857787072658539,
-0.06578364223241806,
-0.02452390268445015,
-0.0005590183427557349,
0.04419024661183357,
-1.2036822461775252e-33,
0.028183910995721817,
0.10820916295051575,
0.0023057449143379927,
0.04946751520037651,
-0.046299856156110764,
-0.0070346505381166935,
0.023371145129203796,
0.04548237845301628,
0.05239785462617874,
-0.0893436074256897,
-0.05233466625213623,
-0.04604518041014671,
-0.047781795263290405,
-0.004374413751065731,
-0.05024172365665436,
-0.0359586738049984,
-0.04705740511417389,
-0.0568372942507267,
-0.0924721211194992,
0.014328994788229465,
0.05087941512465477,
0.09199067950248718,
-0.031320057809352875,
0.08457682281732559,
-0.0897870659828186,
0.0824756920337677,
-0.03296886011958122,
0.14853248000144958,
0.022700129076838493,
-0.058683786541223526,
-0.005556368734687567,
-0.03877989947795868,
0.01424183789640665,
-0.028764955699443817,
-0.07304846495389938,
0.07659132778644562,
0.01373005285859108,
-0.03811357915401459,
-0.051442887634038925,
0.07107212394475937,
0.06199636310338974,
-0.039796918630599976,
-0.05304502695798874,
0.02440112829208374,
0.053464941680431366,
-0.04216153919696808,
-0.02104964479804039,
-0.05191430076956749,
0.03341053053736687,
0.04440019279718399,
0.01797773875296116,
0.07846303284168243,
-0.10433036834001541,
0.041891034692525864,
-0.006210084538906813,
-0.057717565447092056,
-0.02280735783278942,
0.012106585316359997,
0.12389158457517624,
0.034703802317380905,
-0.03477246314287186,
-0.043471865355968475,
-0.01443176157772541,
-0.0180019773542881,
-0.0698089450597763,
0.04208850488066673,
-0.060981154441833496,
0.05302979797124863,
0.03367775306105614,
0.07033645361661911,
0.025195840746164322,
-0.0022799880243837833,
-0.017459357157349586,
-0.03880694508552551,
0.03266328200697899,
-0.10581997781991959,
0.04699108377099037,
-0.02469579502940178,
0.023968903347849846,
0.020373469218611717,
0.029233872890472412,
-0.06891532242298126,
0.05461212992668152,
0.04084398224949837,
0.0490553043782711,
0.10640457272529602,
-0.006170647218823433,
-0.011383412405848503,
0.02438606321811676,
-0.05745982751250267,
-0.030190886929631233,
-0.013637445867061615,
-0.04210096597671509,
-0.021451447159051895,
-0.06605963408946991,
-4.1533212424837984e-8,
-0.0344417579472065,
0.08580031245946884,
0.02231680229306221,
-0.001620025490410626,
0.004047267138957977,
-0.12767037749290466,
0.030278585851192474,
0.058744627982378006,
0.01586259715259075,
-0.04109416902065277,
0.0396965928375721,
0.02113754115998745,
-0.03387683629989624,
-0.04065299779176712,
-0.040172029286623,
0.09108823537826538,
0.03160254657268524,
0.013778596185147762,
-0.0014280567411333323,
-0.009948018938302994,
-0.03396478295326233,
0.01867235265672207,
0.0000015703209328421508,
-0.020578661933541298,
0.058792002499103546,
-0.047422830015420914,
0.023314932361245155,
-0.01308477483689785,
0.08329011499881744,
-0.00984759721904993,
0.04236762225627899,
0.03298993036150932,
0.006056670099496841,
-0.0426970049738884,
-0.03453601896762848,
0.01296980306506157,
0.1084134429693222,
0.015836747363209724,
0.0031167413108050823,
-0.05610188841819763,
-0.08507239818572998,
0.04828757792711258,
-0.040304284542798996,
-0.03906138241291046,
-0.043997302651405334,
0.042971618473529816,
0.053838763386011124,
-0.02704038843512535,
0.032117657363414764,
0.0032387340907007456,
0.026380375027656555,
-0.001968943979591131,
0.05436421558260918,
-0.012210778892040253,
0.06776190549135208,
-0.02373209036886692,
0.049672018736600876,
-0.004906281363219023,
0.01843210868537426,
0.059896599501371384,
-0.03536013141274452,
-0.049304649233818054,
0.002593404147773981,
0.06101047620177269
] |
Dynamic Gaussians Mesh: Consistent Mesh Reconstruction from Dynamic Scenes | https://openreview.net/forum?id=LuGHbK8qTa | [
"Isabella Liu",
"Hao Su",
"Xiaolong Wang"
] | Poster | Modern 3D engines and graphics pipelines require mesh as a memory-efficient representation, which allows efficient rendering, geometry processing, texture editing, and many other downstream operations. However, it is still highly difficult to obtain high-quality mesh in terms of detailed structure and time consistency from dynamic observations. To this end, we introduce Dynamic Gaussians Mesh (DG-Mesh), a framework to reconstruct a high-fidelity and time-consistent mesh from dynamic input. Our work leverages the recent advancement in 3D Gaussian Splatting to construct the mesh sequence with temporal consistency from dynamic observations. Building on top of this representation, DG-Mesh recovers high-quality meshes from the Gaussian points and can track the mesh vertices over time, which enables applications such as texture editing on dynamic objects. We introduce the Gaussian-Mesh Anchoring, which encourages evenly distributed Gaussians, resulting better mesh reconstruction through mesh-guided densification and pruning on the deformed Gaussians. By applying cycle-consistent deformation between the canonical and the deformed space, we can project the anchored Gaussian back to the canonical space and optimize Gaussians across all time frames. During the evaluation on different datasets, DG-Mesh provides significantly better mesh reconstruction and rendering than baselines. | dynamic scene reconstruction, 4D reconstruction, 4D gaussian splatting, video reconstruction | DG-Mesh reconstructs high-fidelity, time-consistent meshes from dynamic observations and supports tracking and editing of dynamic mesh sequences. | 298 | 2404.12379 | [
-0.048756975680589676,
0.014118473045527935,
0.04749651253223419,
0.014111394062638283,
0.07121983915567398,
-0.0346684567630291,
-0.0279906764626503,
-0.07239817082881927,
0.023910285905003548,
-0.04982728883624077,
-0.04225942865014076,
-0.0025233242195099592,
0.006217539310455322,
0.11185873299837112,
-0.018778018653392792,
-0.035826511681079865,
0.06978428363800049,
0.08752980083227158,
-0.05035501345992088,
0.03727830573916435,
-0.0031650778837502003,
0.0014672797406092286,
-0.08277525752782822,
-0.024937551468610764,
0.08349724113941193,
-0.010063604451715946,
0.08600834012031555,
-0.03571179136633873,
0.050822850316762924,
-0.0045520407147705555,
0.022972241044044495,
-0.05686372146010399,
-0.08136073499917984,
-0.03235321119427681,
-0.03955948352813721,
0.007698632776737213,
-0.0008120467537082732,
0.024179572239518166,
-0.004238265100866556,
0.00576120987534523,
0.023881573230028152,
0.03162769600749016,
-0.04259449243545532,
-0.004215938970446587,
0.06012377515435219,
0.0490901805460453,
0.0028597950004041195,
-0.009579453617334366,
0.018618036061525345,
0.03475705534219742,
-0.07497464120388031,
-0.028141291812062263,
0.015219327993690968,
-0.017903108149766922,
0.005296729505062103,
0.0703306719660759,
0.022016450762748718,
-0.07074794918298721,
0.06530603021383286,
-0.025664953514933586,
0.034433070570230484,
-0.0005325505626387894,
0.05970869958400726,
0.02185017801821232,
-0.04808711260557175,
-0.0675441324710846,
0.06175661459565163,
-0.022184209898114204,
0.028969647362828255,
0.1138751208782196,
-0.0549737922847271,
0.014683176763355732,
0.02601749449968338,
-0.036286838352680206,
-0.009286101907491684,
0.05208687111735344,
-0.03601318970322609,
-0.046438004821538925,
-0.01989285834133625,
-0.023441966623067856,
0.06749297678470612,
0.04962175339460373,
0.030024712905287743,
0.0019290030468255281,
-0.0846463069319725,
0.03643428906798363,
-0.0303367767482996,
0.058581989258527756,
-0.028215359896421432,
-0.039360616356134415,
0.029713869094848633,
0.01114670094102621,
-0.013389261439442635,
-0.038654375821352005,
0.002392670139670372,
0.00041425719973631203,
0.03856921195983887,
-0.05145566910505295,
0.08467990905046463,
0.07071661204099655,
0.00040840229485183954,
-0.06141316145658493,
0.021678106859326363,
0.010035678744316101,
-0.0010738527635112405,
-0.010046559385955334,
0.05061706155538559,
0.07927952706813812,
-0.02115798555314541,
0.015068994835019112,
0.039899762719869614,
0.0046654632315039635,
-0.02481013536453247,
0.030499689280986786,
-0.04974927380681038,
0.01472205575555563,
-0.07010549306869507,
-0.030847584828734398,
-0.005288318730890751,
0.054891180247068405,
0.07957962900400162,
0.020255444571375847,
-0.02014242485165596,
0.058882031589746475,
0.03901216387748718,
-0.059516776353120804,
-0.03108970820903778,
3.5030381390639486e-34,
0.061605654656887054,
0.03353041410446167,
0.030925821512937546,
0.053833819925785065,
0.056829050183296204,
0.06218280270695686,
-0.017728624865412712,
0.007666989229619503,
-0.008981727994978428,
0.04301951825618744,
-0.14050784707069397,
-0.03466639295220375,
-0.061639536172151566,
0.05974430590867996,
-0.007634191308170557,
0.027915161103010178,
-0.08554752171039581,
0.026477916166186333,
-0.045746203511953354,
0.1020885780453682,
-0.011832079850137234,
-0.02721039019525051,
0.001567212282679975,
-0.07628130167722702,
-0.04762493446469307,
0.06525246798992157,
-0.005947290454059839,
-0.0014508109306916595,
0.005395563319325447,
-0.022678831592202187,
0.006999163422733545,
-0.014826443046331406,
-0.020048530772328377,
0.058277107775211334,
-0.029674068093299866,
-0.047539327293634415,
0.03589640557765961,
-0.08818304538726807,
-0.05778491497039795,
-0.029473138973116875,
0.01815820299088955,
0.0545831136405468,
-0.00855564046651125,
-0.0018289638683199883,
-0.07366001605987549,
0.02350619062781334,
0.07017253339290619,
0.03618931770324707,
-0.09292732924222946,
0.005819948855787516,
0.10302302986383438,
0.06401313096284866,
-0.058523718267679214,
-0.03954651579260826,
-0.10358356684446335,
-0.05795383080840111,
-0.02032911404967308,
-0.06105732545256615,
0.04789932072162628,
0.05656271055340767,
-0.005641263909637928,
0.03606735169887543,
-0.06160958111286163,
0.02174048125743866,
-0.07064189016819,
-0.006199959199875593,
0.07680077850818634,
0.0908375158905983,
0.019934656098484993,
0.018553052097558975,
-0.07625441253185272,
-0.018233556300401688,
0.016515003517270088,
-0.02140490524470806,
-0.0175553560256958,
-0.023693693801760674,
0.060896385461091995,
-0.036375466734170914,
0.028395358473062515,
-0.003008652478456497,
-0.03918827325105667,
-0.02137257531285286,
0.026772955432534218,
-0.04044307768344879,
-0.047546129673719406,
-0.10866790264844894,
0.020421376451849937,
0.0025831444654613733,
-0.016753682866692543,
-0.06041024625301361,
-0.018170416355133057,
-0.09479151666164398,
0.02553347498178482,
-0.02728213369846344,
-0.03019091859459877,
-6.89490435907636e-34,
0.05309862270951271,
0.02221807837486267,
-0.010649751871824265,
0.11403877288103104,
0.04878435656428337,
-0.02787075750529766,
-0.0720125064253807,
0.0004945429973304272,
0.014081715606153011,
-0.07997637242078781,
-0.015403720550239086,
-0.010506592690944672,
0.029104184359312057,
0.04638272896409035,
-0.017491646111011505,
0.014459237456321716,
-0.07203896343708038,
-0.10678748786449432,
-0.1320829838514328,
-0.08217393606901169,
0.013875944539904594,
-0.0283307246863842,
-0.00420650839805603,
-0.09499668329954147,
-0.06193509325385094,
0.1467626541852951,
-0.01721784472465515,
0.03736409917473793,
-0.006949041038751602,
0.0632549375295639,
-0.07910905033349991,
-0.020971858873963356,
0.015808982774615288,
-0.017326833680272102,
0.0003590497944969684,
0.03130913898348808,
0.06567596644163132,
0.035415638238191605,
0.015158209018409252,
-0.030954495072364807,
0.035423628985881805,
-0.02147071436047554,
-0.02887747995555401,
0.10818387567996979,
0.05035485327243805,
0.1045665591955185,
-0.018531637266278267,
0.0024431024212390184,
-0.03284353390336037,
0.074630968272686,
-0.06966021656990051,
0.0633685290813446,
-0.015006454661488533,
-0.012598254717886448,
0.000516061089001596,
-0.07269487529993057,
-0.10656191408634186,
0.03115743212401867,
-0.015567716211080551,
0.041819822043180466,
-0.051417041569948196,
-0.03748522698879242,
-0.03214813023805618,
0.03675125911831856,
0.054338615387678146,
0.03795555606484413,
0.015152070671319962,
-0.11922449618577957,
-0.02654704824090004,
0.1182108223438263,
-0.015185029245913029,
-0.05389297753572464,
-0.08133097738027573,
0.040530093014240265,
-0.012117106467485428,
-0.05252434313297272,
0.03995666280388832,
-0.033895716071128845,
0.05593330040574074,
0.04045713320374489,
0.012484250590205193,
-0.02881847321987152,
0.048063065856695175,
0.06723131984472275,
0.08781226724386215,
0.0823621079325676,
-0.04440528526902199,
0.03603683039546013,
0.04457490146160126,
0.051270756870508194,
0.0019354522228240967,
-0.017772791907191277,
-0.04133117198944092,
0.03457494452595711,
0.05070656165480614,
-3.96449983952607e-8,
-0.03357165679335594,
0.06237182021141052,
0.018980441614985466,
-0.052526455372571945,
0.023904869332909584,
-0.0743710920214653,
0.06853193044662476,
0.035067521035671234,
0.0225227028131485,
-0.13664507865905762,
0.053892772644758224,
-0.14106763899326324,
-0.0699821412563324,
-0.06718286871910095,
0.02109316550195217,
0.11858800798654556,
-0.08087015897035599,
-0.023611217737197876,
-0.029254503548145294,
-0.0028278473764657974,
-0.08742155134677887,
0.04332247003912926,
-0.03649112582206726,
0.031400810927152634,
-0.03869692608714104,
-0.006323348265141249,
-0.028463032096624374,
0.12759028375148773,
0.11547186970710754,
0.017507733777165413,
-0.011763439513742924,
0.013096340931952,
0.025941159576177597,
0.041564781218767166,
-0.10709793120622635,
-0.018181610852479935,
-0.04017692059278488,
0.0019504076335579157,
0.04038643836975098,
-0.09760300815105438,
0.019971372559666634,
0.07797157764434814,
0.015508858487010002,
-0.03776204586029053,
0.0365867055952549,
0.06898072361946106,
0.027181478217244148,
-0.05535399168729782,
0.04735769331455231,
0.035163331776857376,
-0.03338261693716049,
-0.016369327902793884,
-0.003569088876247406,
0.04598390311002731,
0.003936501685529947,
0.051899030804634094,
0.012109197676181793,
-0.0066155483946204185,
0.08364666253328323,
-0.023716585710644722,
-0.0753064751625061,
0.007123707327991724,
-0.01868363656103611,
-0.01681375503540039
] |
Robust Watermarking Using Generative Priors Against Image Editing: From Benchmarking to Advances | https://openreview.net/forum?id=16O8GCm8Wn | [
"Shilin Lu",
"Zihan Zhou",
"Jiayou Lu",
"Yuanzhi Zhu",
"Adams Wai-Kin Kong"
] | Poster | Current image watermarking methods are vulnerable to advanced image editing techniques enabled by large-scale text-to-image models. These models can distort embedded watermarks during editing, posing significant challenges to copyright protection. In this work, we introduce W-Bench, the first comprehensive benchmark designed to evaluate the robustness of watermarking methods against a wide range of image editing techniques, including image regeneration, global editing, local editing, and image-to-video generation. Through extensive evaluations of eleven representative watermarking methods against prevalent editing techniques, we demonstrate that most methods fail to detect watermarks after such edits. To address this limitation, we propose VINE, a watermarking method that significantly enhances robustness against various image editing techniques while maintaining high image quality. Our approach involves two key innovations: (1) we analyze the frequency characteristics of image editing and identify that blurring distortions exhibit similar frequency properties, which allows us to use them as surrogate attacks during training to bolster watermark robustness; (2) we leverage a large-scale pretrained diffusion model SDXL-Turbo, adapting it for the watermarking task to achieve more imperceptible and robust watermark embedding. Experimental results show that our method achieves outstanding watermarking performance under various image editing techniques, outperforming existing methods in both image quality and robustness. Code is available at https://github.com/Shilin-LU/VINE | AI Security, Watermark, Diffusion Model, Image Editing | We present the first comprehensive benchmark for evaluating the robustness of eleven watermarking methods against prevalent image editing techniques and propose a watermarking model based on SDXL-Turbo that remains robust to these editing methods. | 293 | 2410.18775 | [
-0.08075743913650513,
-0.034493256360292435,
0.05547540262341499,
0.06944208592176437,
0.08437135815620422,
0.02305050566792488,
-0.08324196189641953,
-0.06086483597755432,
0.018527302891016006,
-0.040063854306936264,
-0.018256234005093575,
0.009193908423185349,
0.05649273470044136,
0.016986975446343422,
-0.07365565001964569,
0.0011025299318134785,
-0.003043215489014983,
0.10409249365329742,
-0.11416629701852798,
-0.09131302684545517,
-0.0008733475115150213,
-0.08940614759922028,
0.04739925265312195,
-0.029882391914725304,
0.11308672279119492,
0.023765627294778824,
-0.016783762723207474,
0.011254378594458103,
0.03527238965034485,
-0.03639388456940651,
0.03139645233750343,
-0.04290255904197693,
0.0338563472032547,
0.02777712047100067,
-0.05700628459453583,
0.02035427838563919,
0.018061233684420586,
0.006096145138144493,
0.04532202333211899,
-0.07697324454784393,
0.03360112011432648,
0.0055092317052185535,
-0.05434859171509743,
0.030186541378498077,
0.019269747659564018,
-0.01853787899017334,
0.07206062227487564,
0.018986966460943222,
0.0019248160533607006,
0.042201507836580276,
0.0024769140873104334,
-0.038045454770326614,
-0.0605955608189106,
0.032649435102939606,
-0.05926899611949921,
-0.04085022956132889,
0.07844972610473633,
0.007229416631162167,
0.055052850395441055,
0.019319964572787285,
-0.01659739762544632,
-0.03394151106476784,
-0.0033291818108409643,
0.02306981571018696,
0.02687113545835018,
-0.0028144833631813526,
0.045459914952516556,
0.03225040063261986,
-0.009806432761251926,
-0.01762971468269825,
-0.02968326024711132,
0.020861301571130753,
-0.027669010683894157,
0.07426390051841736,
-0.021093834191560745,
0.061387401074171066,
-0.015462650917470455,
-0.004592700395733118,
-0.025405285879969597,
-0.07673326879739761,
0.09311234951019287,
-0.08479130268096924,
0.05672921612858772,
-0.006387020461261272,
0.10241439938545227,
-0.004864509683102369,
0.04703334718942642,
0.0343155674636364,
0.015477444045245647,
0.04105653986334801,
-0.10821469873189926,
0.009414110332727432,
0.04549284279346466,
-0.009418952278792858,
-0.005578272510319948,
0.009610211476683617,
0.005437518935650587,
-0.030624516308307648,
-0.01403910480439663,
0.052356619387865067,
-0.02236146479845047,
-0.08645236492156982,
-0.037714116275310516,
-0.00665302574634552,
0.1647002100944519,
-0.013665991835296154,
0.11388862133026123,
0.05477646738290787,
0.03564100340008736,
-0.059841904789209366,
0.05449343100190163,
0.03146322816610336,
-0.02034257911145687,
0.0016167322173714638,
0.05121393874287605,
-0.0383639819920063,
-0.008940760046243668,
-0.0312452781945467,
-0.03329400345683098,
-0.03195461630821228,
0.07044334709644318,
-0.01334074791520834,
-0.056001901626586914,
-0.02036149427294731,
-0.044488344341516495,
-0.03654390946030617,
-0.04860543832182884,
9.670742433047034e-34,
0.09836814552545547,
0.004543936345726252,
0.00037688258453272283,
-0.020665813237428665,
0.08253496140241623,
-0.0024720849469304085,
-0.07776084542274475,
-0.019986467435956,
-0.09371808171272278,
-0.07901909202337265,
0.038359757512807846,
-0.06526543945074081,
-0.047498274594545364,
0.11563250422477722,
0.05162820219993591,
0.00277883792296052,
-0.05203608423471451,
-0.0011819156352430582,
0.03142928332090378,
0.06558666378259659,
0.007766204886138439,
-0.059961769729852676,
-0.035838689655065536,
-0.02914929948747158,
-0.06243221461772919,
0.010815310291945934,
0.11692358553409576,
-0.07329927384853363,
0.04653305560350418,
-0.011787820607423782,
-0.10970046371221542,
-0.0423908494412899,
0.0967526063323021,
0.02885589189827442,
0.0071304962038993835,
-0.006571861449629068,
0.008617039769887924,
-0.04354165121912956,
-0.0009340766700915992,
0.018405180424451828,
-0.0155229102820158,
0.04557739943265915,
-0.06692555546760559,
-0.03507813811302185,
-0.09598471969366074,
0.00431698327884078,
0.016033707186579704,
0.035447634756565094,
-0.06455447524785995,
-0.014882915653288364,
0.05570339038968086,
0.08800014108419418,
0.03786765784025192,
-0.038581233471632004,
-0.10174804925918579,
0.021168997511267662,
0.08695640414953232,
-0.09254307299852371,
0.027664750814437866,
-0.007404208183288574,
0.023534530773758888,
0.069510817527771,
-0.05736558511853218,
0.06764563173055649,
-0.03409688174724579,
0.012682897038757801,
0.056069619953632355,
0.02989201433956623,
-0.0027180227916687727,
0.011982128955423832,
-0.0766931101679802,
-0.0012091582175344229,
-0.011404848657548428,
-0.002890117233619094,
0.010427288711071014,
-0.02657599374651909,
0.06014009565114975,
-0.016762809827923775,
-0.03370651975274086,
-0.0031335933599621058,
-0.07205034792423248,
-0.035727281123399734,
0.05276137590408325,
-0.06569962203502655,
-0.09405245631933212,
-0.0643070712685585,
-0.026259910315275192,
-0.035076942294836044,
-0.038621701300144196,
-0.03716301545500755,
0.019333740696310997,
-0.03267528861761093,
0.04855141043663025,
-0.0378817655146122,
-0.007973373867571354,
-3.198820430659881e-34,
-0.007829749025404453,
0.03084850311279297,
-0.03210579231381416,
0.071508027613163,
0.01596882939338684,
-0.04962524026632309,
-0.021327780559659004,
0.058600712567567825,
0.05086321011185646,
-0.1174788773059845,
-0.018615899607539177,
-0.015424397774040699,
-0.07014935463666916,
-0.01716448925435543,
-0.04426068812608719,
-0.056006669998168945,
0.006969136651605368,
0.0006340692052617669,
-0.09152395278215408,
-0.027315642684698105,
0.06632237881422043,
0.017477771267294884,
0.040455155074596405,
0.09747198969125748,
-0.02912861481308937,
0.06727896630764008,
-0.007370311301201582,
0.014221882447600365,
0.016790471971035004,
-0.07325109839439392,
-0.04772817716002464,
0.075198233127594,
0.07484285533428192,
-0.051276419311761856,
-0.054755937308073044,
0.04992618411779404,
0.06837189942598343,
-0.004790154751390219,
-0.0525672473013401,
0.11002097278833389,
0.0782894566655159,
-0.01548518892377615,
-0.07034874707460403,
0.0389520637691021,
0.018117057159543037,
0.017914853990077972,
-0.037411611527204514,
0.015366051346063614,
0.020920399576425552,
0.08523116260766983,
-0.00671515055000782,
-0.011762354522943497,
-0.08547093719244003,
0.05820996314287186,
0.008452104404568672,
-0.12294702231884003,
-0.05839845538139343,
-0.01641177199780941,
0.09103816747665405,
0.017199022695422173,
-0.007160142529755831,
-0.04162104055285454,
-0.05239999666810036,
-0.03485875576734543,
-0.028436405584216118,
0.00028282569837756455,
-0.011434252373874187,
0.04270630702376366,
-0.038461171090602875,
0.04064703732728958,
0.05928051471710205,
-0.03523758053779602,
0.004349885508418083,
0.049062579870224,
0.055489517748355865,
-0.04349454119801521,
-0.019221335649490356,
0.04845425486564636,
-0.02997392974793911,
-0.017637114971876144,
0.017002519220113754,
-0.02820511907339096,
0.0034584305249154568,
0.028181735426187515,
0.08154945075511932,
0.06064969673752785,
-0.019758710637688637,
0.017198801040649414,
-0.0051867226138710976,
0.053414344787597656,
-0.07091542333364487,
0.0524602085351944,
-0.03359885513782501,
0.02610483579337597,
-0.052968692034482956,
-3.58999976413088e-8,
-0.03893107920885086,
0.0072227418422698975,
-0.06288757920265198,
-0.0017143224831670523,
-0.044871218502521515,
-0.01508128922432661,
-0.0009951432002708316,
0.08925332874059677,
0.10106533765792847,
-0.1406882107257843,
0.0789138451218605,
-0.04086752608418465,
-0.10405687242746353,
-0.10907624661922455,
-0.088852159678936,
0.02301100268959999,
-0.01985645666718483,
0.07843825966119766,
-0.022122712805867195,
-0.0025905659422278404,
-0.10330521315336227,
0.05728086829185486,
-0.0028229416348040104,
-0.055153388530015945,
-0.0022248493041843176,
-0.017448848113417625,
0.03178627789020538,
0.05577101930975914,
0.06922261416912079,
0.04164934530854225,
-0.0008533628424629569,
0.03732063248753548,
0.07459288090467453,
-0.0017724926583468914,
-0.011010272428393364,
0.04956934228539467,
0.04864884540438652,
0.05502576008439064,
-0.024582698941230774,
0.05275608226656914,
-0.007091704290360212,
0.04478420689702034,
0.01736309379339218,
-0.03496977686882019,
0.0020342606585472822,
-0.00571074616163969,
0.09935744106769562,
-0.0100221186876297,
0.08040042221546173,
-0.052653729915618896,
0.03691839426755905,
-0.022882185876369476,
0.0025516627356410027,
0.047630518674850464,
0.023094341158866882,
0.0005129678756929934,
0.0354250892996788,
0.013570486567914486,
0.05848647281527519,
0.04380800575017929,
0.0683421641588211,
-0.029682151973247528,
0.09217382222414017,
0.030474882572889328
] |
Revisiting In-context Learning Inference Circuit in Large Language Models | https://openreview.net/forum?id=xizpnYNvQq | [
"Hakaze Cho",
"Mariko Kato",
"Yoshihiro Sakai",
"Naoya Inoue"
] | Poster | In-context Learning (ICL) is an emerging few-shot learning paradigm on Language Models (LMs) with inner mechanisms un-explored. There are already existing works describing the inner processing of ICL, while they struggle to capture all the inference phenomena in large language models. Therefore, this paper proposes a comprehensive circuit to model the inference dynamics and try to explain the observed phenomena of ICL. In detail, we divide ICL inference into 3 major operations: (1) Input Text Encode: LMs encode every input text (in the demonstrations and queries) into linear representation in the hidden states with sufficient information to solve ICL tasks. (2) Semantics Merge: LMs merge the encoded representations of demonstrations with their corresponding label tokens to produce joint representations of labels and demonstrations. (3) Feature Retrieval and Copy: LMs search the joint representations of demonstrations similar to the query representation on a task subspace, and copy the searched representations into the query. Then, language model heads capture these copied label representations to a certain extent and decode them into predicted labels. Through careful measurements, the proposed inference circuit successfully captures and unifies many fragmented phenomena observed during the ICL process, making it a comprehensive and practical explanation of the ICL inference process. Moreover, ablation analysis by disabling the proposed steps seriously damages the ICL performance, suggesting the proposed inference circuit is a dominating mechanism. Additionally, we confirm and list some bypass mechanisms that solve ICL tasks in parallel with the proposed circuit. | In-context Learning; Induction Circuit; Mechanistic Interpretability | We decompose In-context Learning into 3 operations and measure their operating dynamics to catch many inference phenomenon of ICL in Large Langauge Models. | 290 | 2410.04468 | [
-0.05430963262915611,
-0.056342896074056625,
0.06283171474933624,
0.057929106056690216,
0.11654988676309586,
0.02529185451567173,
0.08840116113424301,
-0.02519143372774124,
0.08187586814165115,
-0.006752255838364363,
-0.017712106928229332,
-0.06264950335025787,
0.08758043497800827,
0.0051611014641821384,
0.017948342487215996,
-0.02131347358226776,
0.05431964620947838,
0.03991052135825157,
-0.15826836228370667,
-0.05725478753447533,
0.06370875984430313,
-0.03643237054347992,
0.046862173825502396,
0.014055893756449223,
0.017924964427947998,
0.10260185599327087,
-0.015276228077709675,
-0.047580983489751816,
0.0456882044672966,
0.0038270181976258755,
0.020030567422509193,
0.020033378154039383,
-0.01943129114806652,
0.005653749220073223,
-0.036717142909765244,
0.0565948486328125,
-0.09949276596307755,
-0.0008371721487492323,
0.02765526995062828,
-0.05171508342027664,
-0.01701609417796135,
-0.01172857079654932,
0.020016685128211975,
-0.00003879162250086665,
0.0727301761507988,
0.04181642830371857,
-0.0075683025643229485,
-0.024510744959115982,
-0.017380129545927048,
0.0845838189125061,
-0.0973522961139679,
-0.04971294850111008,
0.05068711191415787,
0.008296860381960869,
-0.023703306913375854,
0.019774628803133965,
0.01713748089969158,
-0.0018466663314029574,
0.012790250591933727,
0.0025871002580970526,
-0.04550769552588463,
-0.11192376911640167,
-0.04214884713292122,
0.03389286249876022,
-0.049991171807050705,
-0.016267811879515648,
-0.03705030679702759,
0.05281505733728409,
0.0037005646154284477,
-0.020590804517269135,
0.010332225821912289,
0.035397425293922424,
0.005207246635109186,
0.01932254061102867,
0.03866101801395416,
0.0028373009990900755,
0.006688825786113739,
-0.09050172567367554,
0.038829050958156586,
-0.02893645688891411,
0.03372139111161232,
-0.03694746270775795,
0.02997833490371704,
-0.06882698833942413,
0.1267070770263672,
0.04304194822907448,
0.020902059972286224,
0.015599359758198261,
-0.002329241717234254,
-0.00026766606606543064,
-0.04120657965540886,
-0.13367077708244324,
-0.010349739342927933,
0.014341620728373528,
0.044957879930734634,
0.002253645798191428,
0.00772389629855752,
-0.0475698746740818,
0.027648916468024254,
0.07630811631679535,
0.028152888640761375,
0.0993054062128067,
0.013827105984091759,
-0.0516032800078392,
-0.026609638705849648,
-0.014635677449405193,
-0.021851301193237305,
0.02486705593764782,
0.04817601665854454,
-0.10632048547267914,
0.0025659832172095776,
0.04087040573358536,
0.009970368817448616,
0.09685950726270676,
0.04857219383120537,
-0.047137804329395294,
0.049478478729724884,
-0.10051153600215912,
0.007904469966888428,
0.03086145780980587,
-0.03265688940882683,
0.02526768110692501,
-0.08066192269325256,
-0.022757381200790405,
-0.017352450639009476,
-0.11039139330387115,
-0.017864419147372246,
9.824460359936126e-34,
0.11141308397054672,
0.01703178696334362,
0.011896907351911068,
0.0348348431289196,
0.013243368826806545,
0.013459969311952591,
0.03095540590584278,
0.026487872004508972,
0.024878449738025665,
-0.012281615287065506,
0.04246373102068901,
-0.03760522976517677,
-0.029552960768342018,
0.07531411200761795,
-0.017095200717449188,
0.04168213903903961,
-0.06956310570240021,
-0.011874248273670673,
-0.026667306199669838,
0.008238132111728191,
-0.0025558052584528923,
0.02895505167543888,
0.0023904202971607447,
-0.006437742616981268,
0.015635430812835693,
0.05188868194818497,
0.050482530146837234,
-0.031183259561657906,
-0.06132371351122856,
-0.0024122113827615976,
-0.08376438915729523,
0.004706603940576315,
-0.031208885833621025,
0.05765242502093315,
0.011426473967730999,
-0.006414541509002447,
0.04288443922996521,
-0.050719358026981354,
0.047315388917922974,
0.006342606153339148,
-0.03366581350564957,
-0.019100617617368698,
-0.006916986312717199,
-0.007421981543302536,
-0.13315676152706146,
-0.08012568950653076,
-0.062256451696157455,
0.02718181349337101,
-0.06675727665424347,
0.017898334190249443,
0.08355706185102463,
0.026057515293359756,
0.036689478904008865,
-0.08226928114891052,
0.047394365072250366,
0.0405450202524662,
-0.038369979709386826,
0.02949170209467411,
0.013941682875156403,
0.06730859726667404,
0.028721358627080917,
0.0017295933794230223,
0.011410217732191086,
0.09476211667060852,
0.050064414739608765,
0.14169526100158691,
-0.058990854769945145,
0.04477735236287117,
0.08820807933807373,
-0.013530410826206207,
-0.06082741916179657,
-0.009157293476164341,
-0.0279080867767334,
-0.030464213341474533,
0.11823608726263046,
0.0064050378277897835,
-0.020274734124541283,
-0.11835120618343353,
0.012349513359367847,
0.12082037329673767,
0.022892264649271965,
-0.08682398498058319,
0.01059760618954897,
0.020699093118309975,
-0.0432797372341156,
-0.07753445208072662,
0.006839613430202007,
-0.15371739864349365,
-0.027931878343224525,
-0.07948719710111618,
-0.05554609000682831,
-0.000030748298740945756,
-0.011843755841255188,
0.04189340025186539,
0.047566045075654984,
-1.8865688833919362e-33,
0.07445377111434937,
0.004185255616903305,
-0.028612419962882996,
-0.0034500875044614077,
-0.10039916634559631,
-0.03487524390220642,
0.049420036375522614,
-0.013308056630194187,
-0.04761625453829765,
-0.09097561240196228,
-0.028745850548148155,
0.012765966355800629,
0.005393850151449442,
0.013556236401200294,
-0.008128119632601738,
-0.026783954352140427,
0.031267281621694565,
0.0012653965968638659,
0.0589945949614048,
0.052654702216386795,
0.038042325526475906,
0.049483854323625565,
-0.06857886910438538,
0.02143162675201893,
-0.0398537740111351,
0.027928471565246582,
-0.025189733132719994,
0.058279309421777725,
0.05834967643022537,
0.020264310762286186,
-0.03442143648862839,
-0.04801731929183006,
0.031044185161590576,
-0.03202883154153824,
-0.08577010780572891,
0.07866200059652328,
0.05273943766951561,
-0.009909629821777344,
-0.06375086307525635,
0.03339746221899986,
0.04560059681534767,
-0.07027989625930786,
-0.001943615498021245,
0.025696324184536934,
-0.02527116984128952,
-0.020798252895474434,
-0.05320856720209122,
-0.0033943636808544397,
-0.010463660582900047,
0.0015805434668436646,
-0.0415182001888752,
0.040827926248311996,
-0.08126972615718842,
-0.014552474021911621,
-0.05786658823490143,
-0.06250189244747162,
-0.035723231732845306,
-0.029671531170606613,
0.07566182315349579,
-0.05307130143046379,
-0.019271792843937874,
-0.00020969139586668462,
0.017194267362356186,
-0.02039206773042679,
0.007622598670423031,
-0.005494722630828619,
-0.034845758229494095,
-0.019814861938357353,
-0.01584194228053093,
-0.03966386616230011,
0.06907561421394348,
0.015529594384133816,
0.023949595168232918,
-0.027025608345866203,
0.028623396530747414,
-0.03584377467632294,
-0.004083569161593914,
-0.05057722330093384,
-0.030614377930760384,
-0.04790417104959488,
0.014844685792922974,
-0.09342513978481293,
0.04610949009656906,
0.05909479409456253,
0.06941583752632141,
0.083197683095932,
-0.06115857884287834,
0.06091853231191635,
0.003213590709492564,
0.015604357235133648,
-0.02883588895201683,
0.03633439540863037,
0.020509518682956696,
0.050629932433366776,
-0.025949766859412193,
-4.441420742296032e-8,
-0.13157013058662415,
0.08545513451099396,
0.012656177394092083,
0.03869525343179703,
0.031787093728780746,
0.006575420964509249,
-0.011062316596508026,
0.03951955586671829,
0.0022726028691977262,
-0.04817807674407959,
0.04843098670244217,
0.009648099541664124,
0.012630682438611984,
-0.0672188475728035,
-0.029935581609606743,
0.09919331967830658,
0.057236816734075546,
0.007397190202027559,
0.03335258737206459,
-0.029266376048326492,
0.05777081102132797,
-0.011785490438342094,
0.03160812705755234,
0.007371262181550264,
0.013871555216610432,
-0.07122331857681274,
-0.012130786664783955,
0.1382555365562439,
0.05745020508766174,
-0.009871969930827618,
-0.05431576073169708,
0.004907150752842426,
-0.001334827160462737,
-0.037452489137649536,
0.029324917122721672,
0.07520690560340881,
0.04134424775838852,
-0.09038110822439194,
0.021163197234272957,
-0.06688837707042694,
-0.04615846648812294,
0.062374524772167206,
-0.06167000159621239,
0.026842569932341576,
0.0351339727640152,
0.01755688525736332,
0.06939307600259781,
-0.18144424259662628,
0.05806773528456688,
-0.06186292693018913,
0.005386419128626585,
-0.03400978818535805,
-0.042248647660017014,
0.09027078002691269,
0.042955074459314346,
0.04600612074136734,
-0.00010204349382547662,
-0.004059684928506613,
0.01367473229765892,
0.10194648057222366,
-0.01585329696536064,
0.021022187545895576,
-0.07238186150789261,
-0.004986600484699011
] |
FreeVS: Generative View Synthesis on Free Driving Trajectory | https://openreview.net/forum?id=dTGH9vUVdf | [
"Qitai Wang",
"Lue Fan",
"Yuqi Wang",
"Yuntao Chen",
"Zhaoxiang Zhang"
] | Poster | Existing reconstruction-based novel view synthesis methods for driving scenes focus on synthesizing camera views along the recorded trajectory of the ego vehicle.
Their image rendering performance will severely degrade on viewpoints falling out of the recorded trajectory, where camera rays are untrained.
We propose FreeVS, a novel fully generative approach that can synthesize camera views on free new trajectories in real driving scenes.
To control the generation results to be 3D consistent with the real scenes and accurate in viewpoint pose, we propose the pseudo-image representation of view priors to control the generation process.
Viewpoint translation simulation is applied on pseudo-images to simulate camera movement in each direction.
Once trained, FreeVS can be applied to any validation sequences without reconstruction process and synthesis views on novel trajectories.
Moreover, we propose two new challenging benchmarks tailored to driving scenes, which are novel camera synthesis and novel trajectory synthesis, emphasizing the freedom of viewpoints.
Given that no ground truth images are available on novel trajectories, we also propose to evaluate the consistency of images synthesized on novel trajectories with 3D perception models.
Experiments on the Waymo Open Dataset show that FreeVS has a strong image synthesis performance on both the recorded trajectories and novel trajectories.
The code is released. Project page: https://freevs24.github.io/. | Novel View Synthesis, Driving Scene, Free Trajectory, Image Generation | We propose FreeVS, a novel fully generative approach that can synthesize camera views on free new trajectories in real driving scenes. | 289 | 2410.18079 | [
-0.06354902684688568,
-0.0921761617064476,
-0.013642638921737671,
-0.025844484567642212,
0.09403453022241592,
0.011158490553498268,
-0.013608615845441818,
-0.04236263781785965,
0.06701873242855072,
-0.006591556593775749,
-0.01858251541852951,
-0.08727560937404633,
-0.009628462605178356,
0.04141012951731682,
-0.02025863155722618,
-0.037294477224349976,
0.04875855892896652,
0.1104709804058075,
-0.030956275761127472,
0.05011199787259102,
-0.02249925211071968,
-0.024940911680459976,
0.06140686199069023,
0.013462953269481659,
0.04576938599348068,
0.002875448204576969,
0.0920315831899643,
-0.06614967435598373,
-0.010591188445687294,
-0.046387627720832825,
0.048200950026512146,
0.026933113113045692,
-0.06618577241897583,
-0.03135101869702339,
-0.045472774654626846,
0.017360197380185127,
-0.012920566834509373,
0.02776467800140381,
-0.013993899337947369,
-0.05402552708983421,
-0.004430170636624098,
0.03281279653310776,
0.027717504650354385,
-0.007226270157843828,
0.0665893480181694,
0.06379929184913635,
0.0688539370894432,
-0.0238173920661211,
0.04142283275723457,
-0.004375569988042116,
-0.0678689107298851,
0.005600517615675926,
-0.011569754220545292,
-0.041658394038677216,
-0.017483849078416824,
0.04545959085226059,
-0.019730420783162117,
-0.06573215126991272,
0.016325125470757484,
0.03019927442073822,
0.04211105778813362,
-0.029459219425916672,
-0.03894638270139694,
-0.026484528556466103,
-0.054354358464479446,
-0.030902022495865822,
0.045707281678915024,
0.00643462548032403,
0.0417608842253685,
0.03449038416147232,
-0.06284800916910172,
0.03382527828216553,
0.049431029707193375,
-0.06323711574077606,
-0.005641303490847349,
-0.001028426573611796,
0.03963818401098251,
-0.04259376600384712,
-0.013937615789473057,
-0.10065227746963501,
0.08668702095746994,
0.023661823943257332,
0.01441213022917509,
0.09623276442289352,
0.0028786880429834127,
0.02550983428955078,
-0.03211931511759758,
0.10278841853141785,
0.02449186146259308,
0.06025606393814087,
-0.112421914935112,
-0.024513904005289078,
-0.04873337224125862,
-0.06773857772350311,
-0.009074744768440723,
0.017901618033647537,
0.007628313265740871,
-0.05881355330348015,
0.05844759941101074,
0.006627557333558798,
0.05402596667408943,
0.02825482375919819,
0.016476145014166832,
0.011442133225500584,
0.06065349653363228,
0.05001956596970558,
0.058157578110694885,
0.007361147087067366,
-0.029035788029432297,
0.03264206647872925,
0.09847366809844971,
-0.04800526425242424,
0.03582169488072395,
-0.02352799102663994,
-0.0341353602707386,
0.013110104016959667,
-0.040504999458789825,
-0.02916647493839264,
-0.025862907990813255,
0.02175876870751381,
-0.0725083202123642,
-0.006236657965928316,
0.04470428079366684,
0.09155367314815521,
0.023048074916005135,
-0.0991898626089096,
0.04807516559958458,
1.660235711335759e-33,
-0.050414077937603,
0.08025426417589188,
0.06062024086713791,
0.029747575521469116,
0.07700604200363159,
0.01541611086577177,
-0.05293804034590721,
0.009981656447052956,
-0.028111906722187996,
0.00085102254524827,
0.015878373757004738,
-0.06810719519853592,
-0.02055259235203266,
0.055077001452445984,
0.06816352903842926,
-0.04146867245435715,
-0.057060014456510544,
0.004436355549842119,
-0.049335625022649765,
0.08545409888029099,
0.026322025805711746,
-0.003675621934235096,
-0.03218249976634979,
0.024067115038633347,
-0.057740431278944016,
0.04379681870341301,
0.057579491287469864,
0.03490627557039261,
-0.012648047879338264,
-0.037676502019166946,
-0.04752586781978607,
0.00747217983007431,
0.025557082146406174,
0.024314573034644127,
-0.0032151935156434774,
0.014388966374099255,
-0.058910906314849854,
-0.04646507278084755,
-0.041851550340652466,
0.027204034850001335,
-0.02166788838803768,
0.09568869322538376,
-0.06424712389707565,
0.03528866171836853,
-0.08639803528785706,
-0.038219839334487915,
0.02017163299024105,
0.1662488728761673,
-0.09148132055997849,
0.045876119285821915,
0.027721306309103966,
-0.029123567044734955,
0.0010503515368327498,
-0.05817179009318352,
-0.13699829578399658,
-0.019367406144738197,
0.013161769136786461,
0.01065810862928629,
0.017369864508509636,
-0.04645238071680069,
0.013117354363203049,
-0.015518233180046082,
-0.07092542946338654,
0.015065116807818413,
-0.07748416811227798,
0.004478332120925188,
-0.01584705524146557,
0.05501662567257881,
0.07650677114725113,
0.07545796036720276,
-0.0024431501515209675,
-0.008644379675388336,
-0.015038732439279556,
-0.01326822116971016,
0.08633303642272949,
0.09436822682619095,
-0.0645822286605835,
-0.026826387271285057,
-0.015329360030591488,
0.08855924755334854,
-0.073358453810215,
0.06015841290354729,
0.001769677153788507,
0.0035593754146248102,
0.05281419679522514,
-0.04919248819351196,
-0.032772280275821686,
-0.04252472519874573,
0.002891042036935687,
-0.04444849118590355,
-0.08770787715911865,
-0.11753568053245544,
-0.05651121586561203,
-0.05519860237836838,
0.003597970586270094,
-1.356076548060492e-33,
0.05977465957403183,
0.006975003983825445,
-0.025484465062618256,
0.028265995904803276,
-0.039998460561037064,
-0.054803576320409775,
-0.009341687895357609,
0.07264354825019836,
0.06983676552772522,
-0.09897727519273758,
-0.08626643568277359,
0.004021981731057167,
0.06695222854614258,
0.04045133665204048,
-0.0009499368024989963,
-0.06319837272167206,
0.008102529682219028,
-0.04494665190577507,
-0.12412482500076294,
0.006672036834061146,
0.024952342733740807,
0.09273483604192734,
-0.02587597630918026,
0.03464258834719658,
-0.024335794150829315,
0.09829895943403244,
0.01675986312329769,
0.12216316908597946,
-0.0038641851861029863,
0.02089087851345539,
0.04476906731724739,
0.0018968157237395644,
0.04002610221505165,
-0.045922789722681046,
-0.08221272379159927,
0.06887190043926239,
0.04260265454649925,
-0.002076332224532962,
-0.055558811873197556,
0.05500493943691254,
-0.0012626296374946833,
-0.07550816982984543,
0.005633028689771891,
0.05766281113028526,
-0.009386827237904072,
0.02062157914042473,
0.0035611153580248356,
0.014684819616377354,
0.05002659186720848,
0.07389726489782333,
-0.08653520792722702,
0.07375038415193558,
-0.07393084466457367,
0.05199522525072098,
0.01807675138115883,
-0.09270484000444412,
-0.08947917073965073,
-0.04487365856766701,
0.10918519645929337,
-0.0644579604268074,
-0.016759317368268967,
-0.03704632446169853,
-0.06646519154310226,
-0.015110698528587818,
0.0020571956411004066,
-0.016839206218719482,
-0.0881403237581253,
-0.008210995234549046,
0.0465751588344574,
0.046928465366363525,
-0.007931799627840519,
0.017228825017809868,
-0.019102685153484344,
-0.0008080440456978977,
-0.009599917568266392,
-0.04405178502202034,
0.03587840497493744,
0.03322432190179825,
0.10031908005475998,
-0.049934931099414825,
0.07158517837524414,
-0.08396025747060776,
0.08731333166360855,
0.09285005182027817,
0.025195050984621048,
0.05548034608364105,
-0.03534223511815071,
-0.016640348359942436,
-0.026539117097854614,
0.10766222327947617,
0.010868237353861332,
0.03572944924235344,
-0.04649129882454872,
0.061251651495695114,
-0.07498695701360703,
-3.440881002347851e-8,
-0.0985562726855278,
0.04490463063120842,
0.049541302025318146,
0.026482511311769485,
-0.06695311516523361,
-0.018749872222542763,
0.004331073258072138,
-0.0038776055444031954,
-0.03804289549589157,
-0.06830675154924393,
0.03301713615655899,
-0.03664445877075195,
-0.02734377793967724,
0.005202135536819696,
-0.003507747547701001,
0.0443514846265316,
0.006275353487581015,
0.026701105758547783,
-0.06772678345441818,
0.033134255558252335,
-0.06435324251651764,
0.04563286155462265,
-0.030682343989610672,
-0.012083105742931366,
0.007013711147010326,
-0.014053554274141788,
-0.029645172879099846,
-0.03147035464644432,
0.12264860421419144,
-0.06661543250083923,
0.0010139428777620196,
0.010442795231938362,
0.07013425976037979,
-0.017705781385302544,
-0.02748446725308895,
0.005662263836711645,
-0.03277130797505379,
0.027015239000320435,
0.08432058990001678,
-0.094223752617836,
0.00047921721125021577,
0.07052404433488846,
0.003602274227887392,
0.032060544937849045,
-0.06863010674715042,
0.03054615668952465,
0.027405809611082077,
-0.07066991180181503,
-0.020755432546138763,
-0.01837879605591297,
-0.04031950980424881,
-0.023056035861372948,
0.004189267288893461,
0.1018279492855072,
0.05150222405791283,
0.0497976578772068,
0.03430210426449776,
-0.0788206085562706,
0.03335278108716011,
0.06687604635953903,
-0.02318875677883625,
0.05154845491051674,
0.004861047957092524,
-0.027959708124399185
] |
Enhancing End-to-End Autonomous Driving with Latent World Model | https://openreview.net/forum?id=fd2u60ryG0 | [
"Yingyan Li",
"Lue Fan",
"Jiawei He",
"Yuqi Wang",
"Yuntao Chen",
"Zhaoxiang Zhang",
"Tieniu Tan"
] | Poster | In autonomous driving, end-to-end planners directly utilize raw sensor data, enabling them to extract richer scene features and reduce information loss compared to traditional planners. This raises a crucial research question: how can we develop better scene feature representations to fully leverage sensor data in end-to-end driving? Self-supervised learning methods show great success in learning rich feature representations in NLP and computer vision. Inspired by this, we propose a novel self-supervised learning approach using the LAtent World model (LAW) for end-to-end driving. LAW predicts future latent scene features based on current features and ego trajectories. This self-supervised task can be seamlessly integrated into perception-free and perception-based frameworks, improving scene feature learning while optimizing trajectory prediction. LAW achieves state-of-the-art performance across multiple benchmarks, including real-world open-loop benchmark nuScenes, NAVSIM, and simulator-based closed-loop benchmark CARLA. The code will be released. | end-to-end autonomous driving, world model, self-supervised learning | The latent world model facilitates future prediction tasks, improving scene feature learning and trajectory prediction in end-to-end autonomous driving. | 288 | 2406.08481 | [
0.040144454687833786,
-0.10514836758375168,
0.03709731623530388,
0.036884505301713943,
0.13858821988105774,
0.026526622474193573,
0.005407629068940878,
0.004550078883767128,
-0.0026326696388423443,
-0.007338318042457104,
-0.03386919945478439,
-0.07828754931688309,
-0.03967836871743202,
0.11503791064023972,
0.017366105690598488,
0.014191080816090107,
0.10763602703809738,
0.04000813886523247,
-0.07828923314809799,
-0.051340073347091675,
0.0176597498357296,
0.06927122920751572,
0.06502731144428253,
0.06717783212661743,
0.028976840898394585,
0.0193132683634758,
0.0686323344707489,
-0.03604985028505325,
-0.05999596789479256,
-0.030160626396536827,
0.09258459508419037,
-0.03735065087676048,
-0.06629632413387299,
-0.016443198546767235,
-0.03823787719011307,
-0.031153904274106026,
-0.05034760758280754,
-0.026565777137875557,
-0.09669490158557892,
-0.04342975094914436,
0.031985681504011154,
-0.02768407203257084,
-0.020650092512369156,
0.09111592173576355,
0.06725593656301498,
0.02922583557665348,
0.024827228859066963,
-0.030099164694547653,
0.054751764982938766,
-0.023524058982729912,
-0.08156789839267731,
-0.010738889686763287,
-0.018764102831482887,
-0.03229636698961258,
-0.05145678669214249,
0.037591658532619476,
0.06618605554103851,
-0.01055253203958273,
0.02346746437251568,
0.03075108863413334,
0.08341866731643677,
-0.10156327486038208,
0.03216160833835602,
-0.03124913014471531,
-0.020976679399609566,
-0.04105192795395851,
0.027052411809563637,
0.015843048691749573,
0.007157472427934408,
-0.010501286014914513,
-0.02715459279716015,
0.049322210252285004,
0.026138337329030037,
-0.0214073546230793,
0.03616314008831978,
0.0230843648314476,
0.0386604480445385,
-0.006622246000915766,
0.05598810687661171,
-0.06800888478755951,
-0.05245649814605713,
0.023984447121620178,
0.02611618861556053,
0.06550336629152298,
0.09968042373657227,
-0.009987922385334969,
0.0052828192710876465,
0.11064908653497696,
-0.005646554287523031,
0.019101012498140335,
-0.05471225827932358,
-0.10186418890953064,
-0.039528075605630875,
-0.04983800649642944,
-0.031396135687828064,
-0.007003950886428356,
0.011794671416282654,
-0.13815699517726898,
0.010001402348279953,
0.016201021149754524,
-0.0227066483348608,
0.018508780747652054,
-0.03421599045395851,
0.056753970682621,
0.04845868796110153,
0.03385240584611893,
0.07088243216276169,
0.029101114720106125,
0.005671588703989983,
-0.007321716286242008,
0.09203951805830002,
-0.051189858466386795,
0.0328347273170948,
-0.03874439373612404,
0.06794631481170654,
0.024146290495991707,
0.0027693186420947313,
0.06921301782131195,
0.014777651987969875,
0.06676450371742249,
-0.11497893184423447,
-0.06521052867174149,
0.00629335455596447,
0.0190463624894619,
0.03862059488892555,
-0.025743016973137856,
0.012445339933037758,
1.4925282996459278e-33,
-0.05234913527965546,
0.008354519493877888,
0.06503926217556,
-0.013577170670032501,
0.04227731376886368,
-0.0168745219707489,
-0.058654509484767914,
0.05076999217271805,
-0.06811492890119553,
0.001926615834236145,
0.004794148728251457,
-0.07257257401943207,
-0.04756487160921097,
0.04510953277349472,
0.030467120930552483,
-0.024197392165660858,
-0.08172496408224106,
-0.014234357513487339,
-0.015371900051832199,
0.01925722323358059,
0.06251788884401321,
0.020668966695666313,
-0.0369935967028141,
-0.03005470521748066,
0.05840755254030228,
0.047543540596961975,
0.04395085945725441,
0.046848636120557785,
0.024785546585917473,
-0.010388962924480438,
-0.04549155756831169,
0.0915585607290268,
-0.019607137888669968,
0.05156613513827324,
-0.00930237676948309,
0.04482075944542885,
-0.07430680096149445,
-0.028704365715384483,
0.07511211931705475,
-0.017808781936764717,
0.008550155907869339,
0.025810284540057182,
-0.03446245193481445,
-0.03757818043231964,
-0.11933746188879013,
-0.09854944050312042,
0.04042080417275429,
0.05853428691625595,
-0.019377678632736206,
0.03334645926952362,
-0.0037713050842285156,
-0.05611772835254669,
-0.02007676288485527,
-0.08346506953239441,
-0.0931532233953476,
-0.010155554860830307,
0.009510188363492489,
0.018812235444784164,
0.024425450712442398,
0.016094189137220383,
0.021027345210313797,
-0.006888715084642172,
0.00008929253817768767,
0.01429446879774332,
-0.012532692402601242,
-0.03560926765203476,
0.07206772267818451,
0.07518719136714935,
0.0540359802544117,
0.021051565185189247,
-0.009370188228785992,
0.07576621323823929,
-0.01276825275272131,
-0.035631053149700165,
0.08996652811765671,
0.013444354757666588,
0.014527576044201851,
-0.09997590631246567,
0.03195524960756302,
0.04508392512798309,
-0.035301025956869125,
0.0332251712679863,
-0.010219139978289604,
-0.003766104346141219,
0.05585933104157448,
-0.016338467597961426,
-0.01956271566450596,
-0.08292269706726074,
-0.0056257485412061214,
0.05002264305949211,
-0.044341012835502625,
-0.08161324262619019,
-0.009779427200555801,
0.06543281674385071,
-0.07065344601869583,
-6.475616085220108e-34,
0.011666121892631054,
0.006520433817058802,
-0.059630148112773895,
0.04261622577905655,
0.022251440212130547,
-0.07416247576475143,
0.005883445031940937,
0.04553629830479622,
0.022494813427329063,
0.0028000161983072758,
-0.026385005563497543,
-0.057413049042224884,
0.08373351395130157,
0.03394371643662453,
0.046411219984292984,
-0.0054730940610170364,
0.0209855604916811,
-0.049465298652648926,
-0.02312599867582321,
-0.058025654405355453,
-0.02498607337474823,
0.01079277042299509,
-0.030912254005670547,
0.01091792993247509,
-0.03560134768486023,
0.028497876599431038,
-0.020269770175218582,
0.007239118218421936,
-0.019071925431489944,
-0.04079797491431236,
-0.008940876461565495,
-0.008206608705222607,
-0.00968853011727333,
-0.08142652362585068,
-0.06734582781791687,
0.042113255709409714,
-0.015822453424334526,
-0.09081501513719559,
-0.04876641929149628,
0.04119354858994484,
0.05429856479167938,
-0.09509603679180145,
-0.013670737855136395,
-0.033411845564842224,
-0.0677722916007042,
-0.07206293195486069,
-0.027816317975521088,
0.09579598158597946,
0.057105813175439835,
0.08890219777822495,
-0.0574345625936985,
0.037958916276693344,
-0.10615861415863037,
0.056639183312654495,
-0.04974769800901413,
-0.024998683482408524,
0.04258769750595093,
0.012309563346207142,
0.08705765753984451,
0.03601174056529999,
-0.021297186613082886,
0.013421602547168732,
0.04615680128335953,
-0.0017081735422834754,
0.006095069460570812,
-0.12992431223392487,
-0.07273518294095993,
0.019848160445690155,
0.011478662490844727,
-0.0208861343562603,
0.055597543716430664,
0.011341418139636517,
-0.008912458084523678,
0.014412989839911461,
0.0036806424614042044,
-0.09095417708158493,
0.04075484350323677,
-0.02315806970000267,
0.015578125603497028,
-0.04099821299314499,
0.10652672499418259,
-0.013710242696106434,
0.056280601769685745,
0.11752647161483765,
0.013267002999782562,
0.08860284835100174,
-0.005761058069765568,
0.006933674216270447,
0.01484000962227583,
0.08115941286087036,
-0.008371173404157162,
0.05463976785540581,
-0.08464834839105606,
0.08591918647289276,
-0.10981182754039764,
-3.4308005325556223e-8,
-0.1340726912021637,
0.034463997930288315,
0.06471333652734756,
-0.005099770147353411,
-0.09310042858123779,
-0.03772354871034622,
0.07026588916778564,
0.09446518123149872,
-0.032156702131032944,
0.05513101816177368,
0.1287691593170166,
-0.02786071039736271,
0.0006584288785234094,
0.01379278115928173,
-0.02438676357269287,
0.044874753803014755,
0.05684218555688858,
0.012114007025957108,
0.04778275638818741,
0.021589139476418495,
-0.02311309427022934,
-0.02828885056078434,
0.008591054007411003,
-0.06689129769802094,
0.061844512820243835,
-0.024869060143828392,
-0.034235648810863495,
0.10670402646064758,
0.03399433195590973,
-0.07097020000219345,
-0.042836613953113556,
-0.007540259975939989,
0.06599906831979752,
-0.00026841400540433824,
0.026914527639746666,
0.05858176574110985,
0.011054879054427147,
0.026103930547833443,
0.01553244236856699,
-0.009092343039810658,
-0.0016339041758328676,
0.11312537640333176,
-0.017203066498041153,
0.039608173072338104,
-0.014489695429801941,
0.0180524829775095,
0.0439189113676548,
-0.07682798057794571,
-0.006503579672425985,
-0.025823118165135384,
-0.011086205020546913,
-0.10659375786781311,
-0.010272239334881306,
0.06907793134450912,
0.07760581374168396,
0.02361014112830162,
0.016291620209813118,
-0.009309399873018265,
0.04168129712343216,
0.0791807696223259,
-0.052644163370132446,
-0.007037684321403503,
-0.02748890593647957,
0.003052725223824382
] |
Collapsed Language Models Promote Fairness | https://openreview.net/forum?id=kynD1UUk6q | [
"Jingxuan Xu",
"Wuyang Chen",
"Linyi Li",
"Yao Zhao",
"Yunchao Wei"
] | Poster | To mitigate societal biases implicitly encoded in recent successful pretrained language models, a diverse array of approaches have been proposed to encourage model fairness, focusing on prompting, data augmentation, regularized fine-tuning, and more. Despite the development, it is nontrivial to reach a principled understanding of fairness and an effective algorithm that can consistently debias language models. In this work, by rigorous evaluations of Neural Collapse -- a learning phenomenon happen in last-layer representations and classifiers in deep networks -- on fairness-related words, we find that debiased language models exhibit collapsed alignment between token representations and word embeddings. More importantly, this observation inspires us to design a principled fine-tuning method that can effectively improve fairness in a wide range of debiasing methods, while still preserving the performance of language models on standard natural language understanding tasks. We attach our code at https://github.com/Xujxyang/Fairness-NC-main | Neural Collapse, Fairness | null | 281 | 2410.04472 | [
-0.05949326604604721,
-0.07707859575748444,
0.09369606524705887,
0.012874653562903404,
0.031757891178131104,
0.031913287937641144,
-0.016941681504249573,
0.017663579434156418,
0.03281732276082039,
-0.009456007741391659,
-0.0574786402285099,
-0.014756561256945133,
0.03535863012075424,
0.006461342331022024,
-0.056830424815416336,
0.013497943058609962,
0.07477477937936783,
0.11786657571792603,
-0.16608723998069763,
-0.09335128962993622,
0.053391266614198685,
0.04423185810446739,
0.004408488981425762,
-0.00799982063472271,
-0.010444063693284988,
-0.05489359796047211,
-0.00793465692549944,
-0.05056466907262802,
0.029065098613500595,
-0.0039632790721952915,
0.05174902454018593,
-0.05943886190652847,
0.017878098413348198,
0.04449807107448578,
-0.013496773317456245,
0.03320362791419029,
-0.024331793189048767,
-0.04161536693572998,
0.019291263073682785,
-0.030128290876746178,
0.01693304441869259,
0.007229664362967014,
-0.016789771616458893,
0.030956661328673363,
0.10052268952131271,
0.0064423843286931515,
0.05989660322666168,
-0.018991684541106224,
-0.04448277875781059,
-0.04762038588523865,
0.00765637680888176,
0.009723972529172897,
0.05085946246981621,
0.053531281650066376,
-0.04588601365685463,
-0.02702130377292633,
-0.03700307011604309,
0.08178698271512985,
-0.05715425685048103,
-0.03619592264294624,
-0.05878588557243347,
-0.13646003603935242,
-0.05642526596784592,
0.019009927287697792,
0.001958640292286873,
-0.05024878308176994,
-0.024824926629662514,
0.07118507474660873,
-0.07049736380577087,
0.05380837991833687,
0.059905897825956345,
0.11547359079122543,
0.049641795456409454,
0.05248459056019783,
0.03387260437011719,
0.09413661807775497,
0.05747581273317337,
-0.049389515072107315,
0.09791655838489532,
-0.06065821647644043,
0.031872689723968506,
-0.026029519736766815,
0.07337424159049988,
-0.02055048570036888,
0.11671958118677139,
-0.0018410797929391265,
-0.015241713263094425,
0.018160605803132057,
-0.009440836496651173,
0.03520806133747101,
-0.05810777470469475,
-0.05181022360920906,
0.1014946922659874,
0.02941081114113331,
0.013655006885528564,
0.05780581757426262,
0.017981404438614845,
-0.0614335797727108,
-0.06401165574789047,
0.1268150508403778,
-0.031079992651939392,
0.05699962377548218,
-0.005639121867716312,
-0.046454355120658875,
0.04515164718031883,
-0.02056804485619068,
-0.00015268185234162956,
0.003150119911879301,
-0.017893044278025627,
-0.028257494792342186,
-0.020854441449046135,
0.02823825366795063,
-0.005680254194885492,
0.06340056657791138,
0.025352466851472855,
-0.007271360605955124,
0.021571964025497437,
-0.06371735036373138,
0.015803109854459763,
-0.0039295474998652935,
-0.07945399731397629,
0.0766640231013298,
-0.0031956962775439024,
0.0011858082143589854,
0.010391777381300926,
-0.00726371631026268,
-0.013963176868855953,
2.796806872808428e-34,
0.01410369761288166,
0.05318354070186615,
-0.010813286527991295,
0.013049927540123463,
0.06956912577152252,
-0.035151880234479904,
-0.016694450750947,
0.06443749368190765,
-0.05987277254462242,
0.06270389258861542,
-0.02615710161626339,
0.03483525663614273,
-0.0231702271848917,
0.04390346631407738,
-0.07101666182279587,
-0.029501263052225113,
-0.04645202308893204,
0.05849027261137962,
-0.03673945367336273,
-0.02429615519940853,
0.041530851274728775,
0.03011639229953289,
-0.02384008839726448,
-0.0279488954693079,
-0.07338184863328934,
-0.014261944219470024,
0.05712907388806343,
-0.04268396645784378,
-0.009630289860069752,
0.01752617582678795,
-0.1341467946767807,
-0.05055443197488785,
0.038005076348781586,
0.000526859366800636,
0.026585275307297707,
-0.040153201669454575,
0.009173683822154999,
0.06752189248800278,
-0.020211510360240936,
-0.034523818641901016,
0.018604807555675507,
0.029885929077863693,
0.03601442277431488,
-0.04460233822464943,
-0.014339893124997616,
-0.11061335355043411,
0.02848583273589611,
-0.0753534808754921,
0.00829320028424263,
-0.027853935956954956,
0.0733586773276329,
0.005592105910181999,
0.029260309413075447,
-0.032289769500494,
0.03383641317486763,
-0.029639534652233124,
0.001519997720606625,
0.040691081434488297,
0.007370744366198778,
-0.0021379876416176558,
-0.0024354555644094944,
0.0004233049403410405,
-0.02060868591070175,
0.027727527543902397,
0.05215700343251228,
0.018509477376937866,
-0.005566279403865337,
0.01756119169294834,
-0.015555652789771557,
-0.010853823274374008,
0.020688707008957863,
0.047638099640607834,
-0.06313365697860718,
-0.0036915037781000137,
-0.01996508240699768,
0.11840064078569412,
0.06856607645750046,
-0.05114774778485298,
0.0575188547372818,
-0.0017475426429882646,
0.04270339384675026,
-0.05842478573322296,
-0.0056121558882296085,
-0.054245200008153915,
-0.08322469145059586,
-0.0901726484298706,
0.06313139945268631,
-0.055933039635419846,
0.09051927924156189,
0.017799587920308113,
-0.05843198671936989,
-0.03247850760817528,
0.0342097245156765,
0.005192050244659185,
-0.0005230879760347307,
-7.152377111169789e-34,
-0.06672200560569763,
0.010071111842989922,
-0.07865292578935623,
0.0730932354927063,
-0.08992940187454224,
-0.026103021577000618,
0.018282758072018623,
-0.007098844740539789,
0.010679919272661209,
0.012428714893758297,
-0.006427675485610962,
-0.04901983216404915,
0.1101103201508522,
-0.0013209738535806537,
0.06863989681005478,
-0.06181579455733299,
0.018147066235542297,
0.03923248127102852,
0.03989937901496887,
0.05152609571814537,
0.03903130814433098,
0.10483801364898682,
-0.10910265147686005,
0.011349211446940899,
0.034713853150606155,
0.04436346888542175,
-0.041714299470186234,
0.04268181324005127,
0.04565933346748352,
-0.001155559322796762,
-0.10864467918872833,
-0.05899389460682869,
-0.00468880170956254,
-0.0324888713657856,
-0.06641308963298798,
0.015954094007611275,
0.007231551222503185,
-0.03255681321024895,
-0.029856210574507713,
0.008792501874268055,
0.0358516089618206,
-0.0109369782730937,
-0.16032616794109344,
0.028480954468250275,
-0.04595495015382767,
-0.016402963548898697,
-0.05554049834609032,
-0.04390227794647217,
0.01747046783566475,
-0.022102173417806625,
-0.018969450145959854,
-0.013540714979171753,
-0.033065564930438995,
0.1127127930521965,
-0.01306632999330759,
-0.07391221821308136,
0.05058525875210762,
-0.03307957202196121,
0.0553339459002018,
0.052759312093257904,
-0.03745752200484276,
-0.025001995265483856,
-0.02614276483654976,
-0.05440728738903999,
-0.004835667088627815,
-0.0892442986369133,
-0.04774727672338486,
0.021245663985610008,
0.012029941193759441,
-0.009949295781552792,
0.09056465327739716,
0.0017559631960466504,
-0.05020133778452873,
0.011924844235181808,
-0.02798118069767952,
0.01512843556702137,
0.06027422100305557,
-0.023626895621418953,
-0.04470136761665344,
-0.02953348122537136,
0.06050139293074608,
-0.0254301019012928,
0.030293183401226997,
-0.03750662878155708,
0.07984782010316849,
0.044941432774066925,
-0.010242193005979061,
0.0542304553091526,
-0.021917909383773804,
0.06766848266124725,
-0.052174799144268036,
-0.006378049496561289,
-0.007869412191212177,
0.09171162545681,
-0.10922089219093323,
-3.9402760165785367e-8,
-0.09714535623788834,
0.07207103073596954,
-0.027985794469714165,
0.14194123446941376,
-0.02930920012295246,
-0.12192334979772568,
-0.057937197387218475,
0.0266463290899992,
-0.030660340562462807,
-0.05352703481912613,
0.030692651867866516,
0.01869174651801586,
-0.04789890721440315,
-0.0496709868311882,
0.034134335815906525,
0.09932301938533783,
0.0007839618483558297,
0.05848340317606926,
0.04158482328057289,
0.04292943701148033,
0.03331977128982544,
0.05488315224647522,
-0.015546628274023533,
-0.07786067575216293,
0.022111713886260986,
-0.02966039441525936,
-0.0319492444396019,
0.05538394674658775,
0.056186240166425705,
-0.04918844997882843,
-0.021880345419049263,
0.07127072662115097,
-0.04350530356168747,
-0.1113998293876648,
0.014798300340771675,
0.1489851176738739,
-0.03171949461102486,
-0.025964850559830666,
0.008927357383072376,
0.03154205158352852,
0.07097519934177399,
0.11010000854730606,
-0.015270889736711979,
-0.007258142810314894,
0.03484467789530754,
-0.01517223846167326,
0.00041776415309868753,
0.030586687847971916,
0.012074639089405537,
-0.062064748257398605,
-0.035231560468673706,
-0.008335098624229431,
0.012394137680530548,
0.008681371808052063,
0.012568158097565174,
0.011262452229857445,
-0.0058759888634085655,
-0.025657998397946358,
-0.04886525869369507,
0.08635126799345016,
0.05393911525607109,
0.027306964620947838,
-0.05662539601325989,
-0.029428178444504738
] |
Long-horizon Visual Instruction Generation with Logic and Attribute Self-reflection | https://openreview.net/forum?id=EdMb9TqqDY | [
"Yucheng Suo",
"Fan Ma",
"Kaixin Shen",
"Linchao Zhu",
"Yi Yang"
] | Poster | Visual instructions for long-horizon tasks are crucial as they intuitively clarify complex concepts and enhance retention across extended steps.
Directly generating a series of images using text-to-image models without considering the context of previous steps results in inconsistent images, increasing cognitive load. Additionally, the generated images often miss objects or the attributes such as color, shape, and state of the objects are inaccurate.
To address these challenges, we propose LIGER, the first training-free framework for Long-horizon Instruction GEneration with logic and attribute self-Reflection. LIGER first generates a draft image for each step with the historical prompt and visual memory of previous steps. This step-by-step generation approach maintains consistency between images in long-horizon tasks. Moreover, LIGER utilizes various image editing tools to rectify errors including wrong attributes, logic errors, object redundancy, and identity inconsistency in the draft images. Through this self-reflection mechanism, LIGER improves the logic and object attribute correctness of the images.
To verify whether the generated images assist human understanding, we manually curated a new benchmark consisting of various long-horizon tasks. Human-annotated ground truth expressions reflect the human-defined criteria for how an image should appear to be illustrative.
Experiments demonstrate the visual instructions generated by LIGER are more comprehensive compared with baseline methods. The code and dataset will be available once accepted. | text to image generation, visual instruction generation | We propose a training-free framework for long-horizon visual instruction generation with logic and attribute self-reflection. | 274 | 2503.13500 | [
0.044909145683050156,
-0.019945790991187096,
0.030588964000344276,
0.060893166810274124,
0.05506681650876999,
-0.0196959488093853,
-0.04221004247665405,
-0.014663105830550194,
0.02534899115562439,
-0.022240810096263885,
-0.017371660098433495,
-0.06263923645019531,
0.03497070446610451,
0.015815747901797295,
0.017797021195292473,
-0.006360296625643969,
-0.020860010758042336,
0.008580563589930534,
-0.070266492664814,
-0.05291170999407768,
-0.0015867993934080005,
-0.03882897272706032,
0.05906336009502411,
0.018797019496560097,
-0.012931065633893013,
0.03877698630094528,
-0.00357391987927258,
-0.07235366851091385,
0.05976223200559616,
-0.09631723910570145,
0.044273268431425095,
-0.010701370425522327,
0.003451439319178462,
0.04397495836019516,
-0.027169791981577873,
0.12774620950222015,
-0.0024005465675145388,
0.061550986021757126,
-0.01923109032213688,
-0.08698184043169022,
-0.04917629808187485,
-0.035752832889556885,
0.029822155833244324,
0.03389377146959305,
0.06863825768232346,
0.031003491953015327,
-0.01721281372010708,
-0.08847792446613312,
0.008825833909213543,
0.03523997589945793,
-0.04567685350775719,
-0.01706976257264614,
-0.0387846864759922,
-0.049540892243385315,
0.020282559096813202,
0.1759774088859558,
0.0876060202717781,
0.021476250141859055,
-0.03165401890873909,
0.01371529046446085,
-0.08431010693311691,
0.002259212778881192,
-0.00073937454726547,
0.0010055243037641048,
0.019708625972270966,
0.00405070511624217,
0.07308126986026764,
-0.02007996290922165,
0.004412580281496048,
-0.02758452482521534,
-0.03993953391909599,
-0.024498386308550835,
0.0005887472652830184,
0.05169127508997917,
-0.023067651316523552,
0.09209880232810974,
0.007742815185338259,
-0.057854145765304565,
-0.0018680175999179482,
-0.07811066508293152,
-0.011549984104931355,
0.015980875119566917,
0.02425200864672661,
0.09492549300193787,
0.07888669520616531,
0.017934296280145645,
0.06588585674762726,
0.07609739154577255,
-0.009917137213051319,
0.030777715146541595,
-0.0045965975150465965,
-0.09320691972970963,
0.08008210361003876,
0.014137831516563892,
0.06462699174880981,
0.05117996782064438,
0.048701755702495575,
-0.13807545602321625,
-0.06938544660806656,
0.01630302518606186,
0.021700231358408928,
0.03644244372844696,
0.020319491624832153,
-0.006154294591397047,
0.02563297003507614,
-0.007564209401607513,
0.07979908585548401,
0.018785910680890083,
-0.04326687008142471,
-0.03893636539578438,
0.08835010975599289,
-0.021644627675414085,
-0.034519437700510025,
-0.039541225880384445,
0.032024357467889786,
0.005205572582781315,
-0.05608988553285599,
-0.05434316024184227,
-0.014528797008097172,
-0.01425034087151289,
0.01400693692266941,
0.01675373502075672,
-0.03676683083176613,
0.07358121871948242,
0.009729870595037937,
-0.15356992185115814,
-0.00857576634734869,
7.087788288366912e-34,
0.0578097440302372,
0.03482240438461304,
0.08621913939714432,
0.03581218048930168,
0.05528172850608826,
-0.020666124299168587,
-0.0019732785876840353,
0.024245090782642365,
-0.05976298823952675,
0.016050124540925026,
0.0654597282409668,
-0.03361228480935097,
-0.08523997664451599,
0.14088621735572815,
0.08173862844705582,
0.0018368788296356797,
-0.07931255549192429,
0.06796080619096756,
-0.06703847646713257,
0.01986115053296089,
0.004130074754357338,
-0.03604540601372719,
-0.034229177981615067,
-0.04283435642719269,
0.03736281022429466,
0.05961477756500244,
0.07712621986865997,
0.00037174284807406366,
-0.039990369230508804,
-0.040922705084085464,
-0.06147059425711632,
-0.025224898010492325,
0.051127541810274124,
0.02534748800098896,
-0.0005613668472506106,
-0.10407465696334839,
0.019438566640019417,
-0.06594404578208923,
0.025722846388816833,
0.008060663007199764,
-0.01300655398517847,
0.07328612357378006,
0.014242726378142834,
-0.005721770692616701,
-0.023918118327856064,
-0.03809584677219391,
0.003337431699037552,
0.050946421921253204,
-0.09446322172880173,
0.024049701169133186,
0.03140680864453316,
0.060815658420324326,
-0.04079281911253929,
-0.08489694446325302,
-0.08111881464719772,
-0.027544626966118813,
0.05007702857255936,
0.021791093051433563,
0.028266560286283493,
-0.0139364218339324,
0.09083978831768036,
0.08308467268943787,
-0.07337015867233276,
0.11658069491386414,
-0.01325186900794506,
0.008263712748885155,
0.016783619299530983,
0.05739586055278778,
0.04072093963623047,
-0.01533434633165598,
-0.06387310475111008,
-0.04697751998901367,
-0.01673482172191143,
-0.043569937348365784,
0.07923057675361633,
-0.03322143480181694,
0.035104747861623764,
-0.06443285942077637,
0.004325452260673046,
0.012042312882840633,
0.051253095269203186,
-0.018585877493023872,
-0.04805600643157959,
-0.07231657952070236,
-0.03869408369064331,
-0.06396707892417908,
0.06161167845129967,
-0.08026780933141708,
-0.07514394074678421,
-0.011910757049918175,
-0.0032559395767748356,
-0.04667813703417778,
-0.036145955324172974,
-0.018559347838163376,
-0.06422677636146545,
-7.748769680189246e-34,
0.0673319473862648,
-0.014907090924680233,
-0.05241899564862251,
0.03997444361448288,
0.021446598693728447,
-0.06333203613758087,
0.061354849487543106,
0.07406071573495865,
-0.001019746414385736,
-0.10045037418603897,
-0.07296943664550781,
-0.002221557544544339,
-0.09310660511255264,
0.0010969903087243438,
-0.033482104539871216,
-0.02208595722913742,
-0.035839419811964035,
0.007267257198691368,
-0.05318199843168259,
0.019299590960144997,
-0.017398174852132797,
0.06083213537931442,
-0.041186850517988205,
-0.004250004421919584,
-0.05674278363585472,
0.0757773146033287,
0.047317810356616974,
0.052252016961574554,
-0.022506237030029297,
0.04403601214289665,
0.01650206558406353,
0.0234783124178648,
0.07365737855434418,
-0.036614470183849335,
0.0029674668330699205,
-0.025024449452757835,
0.06293891370296478,
-0.03476174175739288,
-0.0744173675775528,
0.06771092861890793,
0.06614980846643448,
-0.05630289018154144,
-0.0222206711769104,
-0.022716600447893143,
-0.04694976285099983,
-0.03579239547252655,
-0.090243861079216,
-0.019501743838191032,
-0.012660903856158257,
0.056774768978357315,
-0.0590842105448246,
0.0029755448922514915,
-0.11150369793176651,
-0.040885698050260544,
-0.030902279540896416,
-0.09433650970458984,
-0.06081686168909073,
0.0022845531348139048,
0.07615397870540619,
0.008974510245025158,
-0.0561649464070797,
0.013601547107100487,
0.05187743157148361,
-0.06646157801151276,
0.01871512271463871,
-0.005272595677524805,
-0.07974949479103088,
-0.03361513093113899,
-0.026751188561320305,
0.04353465884923935,
0.017275266349315643,
-0.007141839247196913,
0.004658992402255535,
-0.020974064245820045,
0.042951446026563644,
-0.031279269605875015,
0.06031814590096474,
0.011744923889636993,
-0.012964272871613503,
-0.04695555940270424,
0.023409370332956314,
-0.04972173273563385,
0.02162376418709755,
0.14391092956066132,
0.012479710392653942,
0.0688781887292862,
-0.018605761229991913,
0.07072518765926361,
-0.007160193752497435,
0.04315483942627907,
-0.06219213083386421,
0.08987533301115036,
0.04105721041560173,
0.08124629408121109,
-0.09815411269664764,
-3.9609560076314665e-8,
-0.0698028951883316,
0.026455283164978027,
0.03817667067050934,
0.0844515860080719,
0.030214371159672737,
0.003914675675332546,
0.00954639632254839,
0.007387804798781872,
-0.014870742335915565,
-0.04224778339266777,
0.060690321028232574,
-0.027491699904203415,
-0.05600922182202339,
-0.052130624651908875,
0.019422950223088264,
0.07284121215343475,
-0.027403581887483597,
0.03984018787741661,
-0.0014923419803380966,
-0.026361750438809395,
0.02506055310368538,
0.017117522656917572,
-0.04706020653247833,
-0.02925248071551323,
-0.02389932982623577,
-0.006444117054343224,
-0.0379512645304203,
0.07947909832000732,
0.06466718018054962,
0.011816333048045635,
0.09606745094060898,
0.027737343683838844,
0.07690906524658203,
-0.005632678512483835,
0.06935065239667892,
0.010490917600691319,
0.019478704780340195,
0.00912060309201479,
0.02453543059527874,
-0.035956718027591705,
0.009816310368478298,
0.08740554004907608,
-0.00789661891758442,
0.012104583904147148,
0.04863946512341499,
-0.061961282044649124,
-0.0430455319583416,
-0.0832885280251503,
-0.039565641433000565,
-0.024686157703399658,
-0.04192374646663666,
-0.03825724497437477,
-0.05621761828660965,
0.10093210637569427,
0.06152611970901489,
-0.008690857328474522,
0.05776135250926018,
-0.025162391364574432,
0.04240083694458008,
0.12494493275880814,
0.059532806277275085,
0.016775313764810562,
0.007880289107561111,
-0.03879031538963318
] |
Learning Harmonized Representations for Speculative Sampling | https://openreview.net/forum?id=T9u56s7mbk | [
"Lefan Zhang",
"Xiaodan Wang",
"Yanhua Huang",
"Ruiwen Xu"
] | Poster | Speculative sampling is a promising approach to accelerate the decoding stage for Large Language Models (LLMs). Recent advancements that leverage target LLM's contextual information, such as hidden states and KV cache, have shown significant practical improvements. However, these approaches suffer from inconsistent context between training and decoding. We also observe another discrepancy between the training and decoding objectives in existing speculative sampling methods. In this work, we propose a solution named HArmonized Speculative Sampling (HASS) that learns harmonized representations to address these issues. HASS accelerates the decoding stage without adding inference overhead through harmonized objective distillation and harmonized context alignment. Experiments on four LLaMA models demonstrate that HASS achieves 2.81x-4.05x wall-clock time speedup ratio averaging across three datasets, surpassing EAGLE-2 by 8%-20%. The code is available at https://github.com/HArmonizedSS/HASS. | speculative sampling, large language model | null | 270 | 2408.15766 | [
-0.08258403092622757,
-0.049909524619579315,
-0.0013702090363949537,
-0.011012107133865356,
0.07720334082841873,
-0.0186776015907526,
-0.00441695423796773,
-0.03777069225907326,
0.04486531391739845,
-0.011537267826497555,
-0.09526045620441437,
-0.014046486467123032,
0.010064034722745419,
-0.037257857620716095,
0.047853998839855194,
0.021001579239964485,
0.1019754484295845,
0.046490442007780075,
-0.06298737227916718,
-0.04715116322040558,
-0.018216736614704132,
-0.028059333562850952,
0.04311267286539078,
0.020111337304115295,
0.09765906631946564,
0.004830571357160807,
-0.005312630906701088,
0.03545862436294556,
0.06227480620145798,
0.03864108398556709,
0.08843445777893066,
0.11795728653669357,
-0.044232334941625595,
-0.0536813884973526,
-0.05994676053524017,
0.05067969858646393,
0.033468712121248245,
0.023063920438289642,
0.07645788043737411,
-0.014400158077478409,
0.046176664531230927,
0.0520864762365818,
0.029869133606553078,
0.01862623170018196,
0.08699586242437363,
-0.05032657831907272,
0.0031640578527003527,
-0.036308709532022476,
-0.009932402521371841,
0.022912293672561646,
-0.11022144556045532,
-0.062029749155044556,
0.005360476206988096,
0.03381858766078949,
-0.04234745725989342,
0.03186100348830223,
0.03587473928928375,
0.051336660981178284,
-0.004684004466980696,
0.0055855801329016685,
-0.08711898326873779,
-0.09793582558631897,
-0.019474046304821968,
-0.0030671791173517704,
0.025973547250032425,
0.015096906572580338,
0.024515364319086075,
0.07733544707298279,
-0.05253656581044197,
-0.016425518319010735,
-0.026379862800240517,
0.028561905026435852,
0.003299306845292449,
0.06259670853614807,
-0.016734333708882332,
0.0787506029009819,
0.07309509068727493,
-0.030195143073797226,
0.08445332944393158,
-0.14014896750450134,
0.015522772446274757,
-0.030740439891815186,
0.02007743902504444,
-0.06961184740066528,
0.03168390691280365,
-0.044179897755384445,
0.009363078512251377,
0.030007919296622276,
0.024336911737918854,
0.005513996351510286,
-0.09953882545232773,
-0.10376892983913422,
0.052577026188373566,
0.019025620073080063,
0.0030327781569212675,
0.11759961396455765,
0.01727355271577835,
0.018752293661236763,
-0.01565425470471382,
0.12565523386001587,
-0.016197405755519867,
0.05853445827960968,
-0.004706245381385088,
-0.04902226850390434,
-0.08161039650440216,
-0.05948050692677498,
0.0015798575477674603,
0.07220739126205444,
0.034258559346199036,
-0.016805285587906837,
-0.0014519528485834599,
0.06350792199373245,
0.014946303330361843,
0.001994807505980134,
0.016228193417191505,
-0.05249430239200592,
0.07697956264019012,
-0.04800541698932648,
-0.029294125735759735,
0.07198363542556763,
-0.025365721434354782,
0.04427279159426689,
-0.052295755594968796,
0.04491138085722923,
0.013877758756279945,
-0.05635126680135727,
0.019176915287971497,
1.854880861771629e-33,
0.053539764136075974,
0.02472909912467003,
0.023895565420389175,
0.03098127990961075,
-0.024375585839152336,
0.005724014714360237,
0.030183840543031693,
0.0006268749129958451,
-0.021897993981838226,
0.010995755903422832,
0.05207374691963196,
-0.0040992917492985725,
-0.006867458578199148,
0.021892385557293892,
-0.07349786162376404,
-0.02800430916249752,
-0.04320799931883812,
0.0008982311119325459,
-0.03974566236138344,
-0.04143652319908142,
-0.013375473208725452,
0.0028728425968438387,
-0.06845293939113617,
-0.07774277776479721,
0.041215021163225174,
0.060002394020557404,
0.06425689160823822,
-0.013937885873019695,
-0.042320046573877335,
0.0614868588745594,
-0.11686613410711288,
-0.00017673791444394737,
-0.04403124377131462,
0.014649972319602966,
0.0536966472864151,
-0.011554527096450329,
-0.032467443495988846,
-0.0525902584195137,
0.01335895899683237,
-0.022353900596499443,
0.0417199581861496,
0.016244592145085335,
-0.022357875481247902,
0.0076461550779640675,
-0.10808425396680832,
-0.07988964021205902,
0.00126492022536695,
-0.08005766570568085,
-0.021248970180749893,
-0.0072894347831606865,
0.019371272996068,
0.010400527156889439,
-0.07464896887540817,
0.011277873069047928,
-0.05325733497738838,
-0.014792545698583126,
0.06434328109025955,
-0.040924739092588425,
0.03235439211130142,
0.04640237241983414,
-0.030656132847070694,
0.017956990748643875,
0.0005072400672361255,
-0.0356241874396801,
0.05130550637841225,
0.08911270648241043,
-0.036666832864284515,
0.047299765050411224,
0.01721409521996975,
0.05004270374774933,
-0.041199591010808945,
0.019005266949534416,
-0.0691022127866745,
0.0069936420768499374,
0.010896233841776848,
0.0076616196893155575,
0.06745783984661102,
0.020284559577703476,
0.00011741214984795079,
0.03403910622000694,
0.04267140105366707,
-0.04335341602563858,
-0.022597825154662132,
0.03434217348694801,
-0.06278244405984879,
-0.0760754644870758,
0.04922224208712578,
-0.07319192588329315,
-0.05892583355307579,
-0.04577731341123581,
-0.057508014142513275,
-0.03223009407520294,
-0.025991087779402733,
0.0031185271218419075,
-0.0248427651822567,
-2.4807638289883325e-33,
-0.0380076989531517,
0.0246918685734272,
-0.08348371833562851,
0.07937127351760864,
-0.04488377645611763,
-0.10065044462680817,
-0.030651982873678207,
0.04380129277706146,
-0.05075179040431976,
-0.053055983036756516,
0.0014167421031743288,
-0.08009794354438782,
0.015118839219212532,
-0.06281328201293945,
0.04986587539315224,
0.024263394996523857,
0.05511835962533951,
0.06992551684379578,
0.04232274368405342,
0.05101337283849716,
0.03642508015036583,
0.013391801156103611,
-0.0921243205666542,
0.045910052955150604,
0.04890623316168785,
0.011552417650818825,
0.004110014997422695,
0.06461489945650101,
0.05036133527755737,
-0.038106583058834076,
-0.08210266381502151,
-0.021928226575255394,
-0.00011247595102759078,
-0.050902388989925385,
-0.09645397216081619,
0.098843052983284,
0.015444054268300533,
-0.013689044862985611,
0.04560868814587593,
0.01733243837952614,
0.09258145838975906,
0.004944568034261465,
-0.027240335941314697,
0.04177355021238327,
-0.09764598309993744,
0.02108476124703884,
-0.0627380907535553,
-0.002775695873424411,
0.013322199694812298,
0.0568358451128006,
0.060884442180395126,
0.015099328942596912,
-0.03960781171917915,
0.07347176223993301,
-0.049668725579977036,
-0.12279467284679413,
-0.0402509905397892,
-0.046931494027376175,
0.03633502498269081,
0.012317565269768238,
-0.006025022827088833,
0.08131006360054016,
0.00010685213055694476,
-0.13296976685523987,
0.017007121816277504,
-0.07875923812389374,
-0.003557601710781455,
-0.05019906908273697,
0.013589136302471161,
0.00920909084379673,
0.09784749150276184,
-0.05001463368535042,
-0.05167354270815849,
0.025927061215043068,
-0.03371863812208176,
0.02836855687201023,
0.030633775517344475,
-0.012106964364647865,
-0.026417870074510574,
-0.021428357809782028,
0.029186278581619263,
0.06579350680112839,
0.01568344235420227,
0.032102007418870926,
0.12962162494659424,
0.0883842334151268,
-0.00024571543326601386,
0.07401593774557114,
-0.011582017876207829,
-0.01741643249988556,
-0.01675240509212017,
0.016094697639346123,
0.009121506474912167,
0.04190180450677872,
-0.02599179744720459,
-4.3219888112844274e-8,
-0.0790092945098877,
-0.002334162127226591,
-0.03791779279708862,
0.10178466141223907,
0.06751026958227158,
0.002267031231895089,
-0.038006555289030075,
-0.014219251461327076,
0.01987050473690033,
-0.022622674703598022,
0.17159292101860046,
-0.028474731370806694,
-0.07512897998094559,
-0.039348579943180084,
-0.02306324616074562,
0.04332081228494644,
-0.0028175124898552895,
-0.013744926080107689,
-0.038357656449079514,
0.013890142552554607,
0.026529476046562195,
0.05133497342467308,
0.004109723027795553,
-0.08235959708690643,
0.10583070665597916,
-0.02526118978857994,
-0.013252179138362408,
0.08437055349349976,
0.07880985736846924,
0.01545326691120863,
-0.08783749490976334,
0.018585000187158585,
-0.08986113220453262,
-0.061999183148145676,
-0.009434187784790993,
0.04360339045524597,
0.025844722986221313,
-0.0353216715157032,
0.023854395374655724,
-0.029900522902607918,
0.05357545241713524,
0.027207225561141968,
-0.14188958704471588,
0.03549933806061745,
0.013954502530395985,
-0.02938719652593136,
-0.0032687983475625515,
-0.0446704663336277,
-0.007682223804295063,
-0.041676945984363556,
0.062205128371715546,
-0.09987792372703552,
0.020402515307068825,
0.045792363584041595,
-0.011793624609708786,
0.03630188852548599,
-0.029857011511921883,
-0.004188667982816696,
0.028525901958346367,
0.06177423894405365,
0.058476224541664124,
0.008815135806798935,
-0.11041238158941269,
0.00712638720870018
] |
MDSGen: Fast and Efficient Masked Diffusion Temporal-Aware Transformers for Open-Domain Sound Generation | https://openreview.net/forum?id=yFEqYwgttJ | [
"Trung X. Pham",
"Tri Ton",
"Chang D. Yoo"
] | Poster | We introduce MDSGen, a novel framework for vision-guided open-domain sound generation optimized for model parameter size, memory consumption, and inference speed. This framework incorporates two key innovations: (1) a redundant video feature removal module that filters out unnecessary visual information, and (2) a temporal-aware masking strategy that leverages temporal context for enhanced audio generation accuracy. In contrast to existing resource-heavy Unet-based models, MDSGen employs denoising masked diffusion transformers, facilitating efficient generation without reliance on pre-trained diffusion models. Evaluated on the benchmark VGGSound dataset, our smallest model (5M parameters) achieves 97.9% alignment accuracy, using 172x fewer parameters, 371% less memory, and offering 36x faster inference than the current 860M-parameter state-of-the-art model (93.9% accuracy). The larger model (131M parameters) reaches nearly 99% accuracy while requiring 6.5x fewer parameters. These results highlight the scalability and effectiveness of our approach. The code is available at https://bit.ly/mdsgen. | vision-guided audio generation, fast inference, open-domain sound synthesis, masked diffusion models, temporal learning, visual sound source localization, generative AI | A novel approach is presented for highly efficient vision-guided sound synthesis using masked diffusion models | 265 | 2410.02130 | [
-0.024917827919125557,
-0.09605786949396133,
0.0018833595095202327,
-0.013367884792387486,
0.07087791711091995,
-0.004998037591576576,
-0.08932103216648102,
-0.1115841493010521,
-0.0011373587185516953,
-0.10388292372226715,
-0.042474742978811264,
-0.045568205416202545,
-0.07649704813957214,
-0.007001256570219994,
-0.036259815096855164,
-0.05692104250192642,
0.042071398347616196,
0.12425509095191956,
-0.06463997811079025,
-0.11029209196567535,
-0.011935043148696423,
0.06134548410773277,
0.0019776918925344944,
0.028919799253344536,
0.0999583899974823,
0.01625143550336361,
0.03548397496342659,
-0.027670327574014664,
0.0550478957593441,
-0.07062816619873047,
0.1123603954911232,
0.007016381714493036,
0.030180567875504494,
-0.03579961881041527,
-0.06040296331048012,
0.014224517159163952,
-0.03966950997710228,
-0.05534566566348076,
-0.03290678560733795,
-0.10958602279424667,
0.05652061477303505,
0.06578270345926285,
-0.0024441697169095278,
-0.0776873454451561,
-0.07129482924938202,
-0.035055093467235565,
-0.06954515725374222,
-0.029089782387018204,
-0.0007608566666021943,
0.006413106340914965,
-0.01624123938381672,
-0.03183432295918465,
-0.04911451041698456,
0.09559247642755508,
-0.10236477851867676,
0.05625220388174057,
0.04619458317756653,
0.0032491169404238462,
0.0033145046327263117,
-0.005508116912096739,
-0.05455949530005455,
-0.031320035457611084,
0.0006207082769833505,
0.002095534699037671,
-0.059564169496297836,
-0.04324990138411522,
0.05756477266550064,
-0.011384056881070137,
0.021194608882069588,
-0.014732781797647476,
-0.0861460417509079,
0.06520295143127441,
-0.011546391993761063,
0.04548197239637375,
0.006681615486741066,
0.0630715936422348,
0.0831170603632927,
-0.054798904806375504,
0.05285201221704483,
-0.038088683038949966,
0.10426255315542221,
-0.03279918432235718,
0.019398150965571404,
-0.04534856602549553,
0.03390372917056084,
-0.014860549941658974,
-0.0223239678889513,
0.010902775451540947,
-0.03175245597958565,
-0.03975173085927963,
-0.07850956171751022,
-0.034288328140974045,
-0.0012165625812485814,
0.010305038653314114,
0.04644208028912544,
0.02191009745001793,
0.007959846407175064,
0.009566937573254108,
-0.03037925809621811,
0.0933438092470169,
-0.004000309854745865,
-0.018356293439865112,
0.009569493122398853,
-0.0355830118060112,
0.08154909312725067,
-0.07447179406881332,
0.03670917823910713,
0.09496397525072098,
-0.03365185856819153,
-0.051991790533065796,
0.08752425760030746,
0.02275155484676361,
0.013267569243907928,
-0.03530063480138779,
0.0894991010427475,
0.04174249246716499,
-0.02588244155049324,
-0.038231708109378815,
-0.013705048710107803,
-0.0031348264310508966,
-0.036484986543655396,
-0.015187812969088554,
-0.07045567035675049,
0.014056315645575523,
0.029915934428572655,
0.022167855873703957,
-0.04414810612797737,
3.422714084596117e-33,
-0.006314214318990707,
0.02321580797433853,
0.017210615798830986,
-0.006472784094512463,
0.07715784013271332,
-0.0032349976245313883,
0.000117776173283346,
-0.007976561784744263,
0.051734112203121185,
0.03231565281748772,
-0.05730161815881729,
-0.03583891689777374,
-0.08402523398399353,
0.0113763976842165,
-0.015155439265072346,
-0.04480944573879242,
-0.024176323786377907,
0.10281361639499664,
-0.054497189819812775,
0.04906671121716499,
0.06521447747945786,
0.012689920142292976,
-0.028463950380682945,
-0.03389696776866913,
0.055917464196681976,
0.01549461204558611,
-0.007284122984856367,
-0.08083371818065643,
0.0016058399342000484,
0.031592126935720444,
-0.11180206388235092,
-0.03602617233991623,
-0.011685306206345558,
0.02590341866016388,
0.07183969765901566,
0.031651318073272705,
0.015547639690339565,
-0.035516731441020966,
0.008817980997264385,
-0.05814768373966217,
-0.007985956966876984,
0.044744934886693954,
-0.02997698076069355,
-0.05608947202563286,
-0.13715633749961853,
-0.018433302640914917,
-0.009279879741370678,
0.0951605811715126,
0.026928376406431198,
0.03589142858982086,
0.07853461056947708,
0.017781825736165047,
-0.033876918256282806,
-0.032476942986249924,
-0.02005663327872753,
-0.07924974709749222,
0.1101573035120964,
-0.003974769730120897,
0.03608252480626106,
-0.0064580910839140415,
0.07145414501428604,
0.00925888866186142,
-0.006147755775600672,
0.024958621710538864,
0.05937851592898369,
-0.04315125569701195,
0.03816781938076019,
0.04641389101743698,
-0.00015709052968304604,
-0.02186851017177105,
-0.043446335941553116,
0.020122116431593895,
0.003109588520601392,
-0.08056217432022095,
0.042855195701122284,
0.024590730667114258,
0.08402709662914276,
-0.029660116881132126,
0.04554062709212303,
0.06113778054714203,
-0.010249548591673374,
0.10421977192163467,
-0.061032745987176895,
-0.03835803642868996,
-0.004797841422259808,
0.06262541562318802,
0.0071140774525702,
-0.03699098154902458,
-0.036619141697883606,
-0.07801239937543869,
-0.023561470210552216,
0.038701821118593216,
-0.00047130274469964206,
-0.01588534004986286,
-0.013793093152344227,
-2.244499015364653e-33,
0.03155440092086792,
0.1373395323753357,
0.060367949306964874,
0.0781208798289299,
-0.009692597202956676,
-0.035086650401353836,
0.08431093394756317,
0.0889110341668129,
0.03290342539548874,
-0.07241953909397125,
0.03252119570970535,
-0.07023866474628448,
0.007828788831830025,
-0.06814725697040558,
0.00011384949903003871,
-0.02147756889462471,
0.002617706311866641,
-0.04769705981016159,
-0.0005706690135411918,
0.03781145438551903,
-0.042226579040288925,
0.058405693620443344,
-0.016168510541319847,
0.009132272563874722,
-0.10772610455751419,
0.03396952897310257,
-0.023233570158481598,
0.02347848378121853,
0.031817447394132614,
-0.031610243022441864,
-0.01253360603004694,
0.052153389900922775,
-0.020192183554172516,
-0.10525379329919815,
-0.03134293854236603,
0.06770669668912888,
0.06356912106275558,
0.03668901324272156,
0.019804518669843674,
0.016384460031986237,
0.010137359611690044,
0.0310472771525383,
-0.05849209800362587,
0.006960212718695402,
-0.008283610455691814,
-0.015464252792298794,
-0.028606107458472252,
0.06399442255496979,
0.027765052393078804,
0.012888462282717228,
0.02539701759815216,
0.06206109747290611,
0.010536342859268188,
-0.0060719773173332214,
-0.036702197045087814,
-0.057951316237449646,
-0.10111179202795029,
-0.005977059248834848,
0.05702074617147446,
0.013681619428098202,
-0.016762860119342804,
0.03243059292435646,
-0.07778377830982208,
-0.13266733288764954,
-0.012835117056965828,
0.07557197660207748,
-0.015077141113579273,
0.020788487046957016,
0.049154121428728104,
0.040803905576467514,
0.024169495329260826,
-0.024137919768691063,
-0.001732498873025179,
0.003210866590961814,
0.0014598803827539086,
-0.020675886422395706,
-0.054339077323675156,
-0.07866520434617996,
0.021461645141243935,
-0.0200085137039423,
0.011442898772656918,
-0.006410401780158281,
0.026351245120167732,
0.002754704561084509,
0.052831653505563736,
0.08100961893796921,
0.03158672899007797,
-0.055294353514909744,
-0.02150612138211727,
0.12031875550746918,
-0.10674386471509933,
0.0518314465880394,
0.017472123727202415,
0.06221688538789749,
-0.035957228392362595,
-4.622278382271361e-8,
-0.004835071042180061,
0.10354139655828476,
0.014848017133772373,
-0.017387842759490013,
-0.0569196492433548,
-0.093113973736763,
0.051979195326566696,
0.045159101486206055,
0.0940147265791893,
-0.008174066431820393,
0.0758449137210846,
-0.017526669427752495,
0.003179137362167239,
0.05371757596731186,
-0.07727444171905518,
0.07096938043832779,
-0.04287303611636162,
0.09184364974498749,
-0.04167517274618149,
-0.051598068326711655,
0.006535079795867205,
0.09185362607240677,
0.02667541801929474,
-0.04292341321706772,
0.07858558744192123,
0.011189746670424938,
-0.007901689037680626,
0.09590636938810349,
0.12752337753772736,
0.028444310650229454,
-0.034274157136678696,
0.059156037867069244,
-0.019244451075792313,
-0.008668720722198486,
0.024330221116542816,
0.04541526734828949,
-0.044216278940439224,
0.020122474059462547,
-0.0451052263379097,
-0.01537215244024992,
-0.05474957451224327,
0.037759535014629364,
-0.06726688146591187,
0.013761276379227638,
-0.08580776304006577,
-0.020299438387155533,
0.09665709733963013,
-0.0528363473713398,
-0.030560031533241272,
0.04989323019981384,
0.00032134063076227903,
0.011542191728949547,
-0.0223977230489254,
0.016319241374731064,
0.08055984973907471,
0.031187618151307106,
0.06036670506000519,
0.008513392880558968,
0.02931654267013073,
0.024108240380883217,
-0.06657905131578445,
-0.021714357659220695,
0.032712943851947784,
0.028362521901726723
] |
SONICS: Synthetic Or Not - Identifying Counterfeit Songs | https://openreview.net/forum?id=PY7KSh29Z8 | [
"Md Awsafur Rahman",
"Zaber Ibn Abdul Hakim",
"Najibul Haque Sarker",
"Bishmoy Paul",
"Shaikh Anowarul Fattah"
] | Poster | The recent surge in AI-generated songs presents exciting possibilities and challenges. These innovations necessitate the ability to distinguish between human-composed and synthetic songs to safeguard artistic integrity and protect human musical artistry. Existing research and datasets in fake song detection only focus on singing voice deepfake detection (SVDD), where the vocals are AI-generated but the instrumental music is sourced from real songs. However, these approaches are inadequate for detecting contemporary end-to-end artificial songs where all components (vocals, music, lyrics, and style) could be AI-generated. Additionally, existing datasets lack music-lyrics diversity, long-duration songs, and open-access fake songs. To address these gaps, we introduce SONICS, a novel dataset for end-to-end Synthetic Song Detection (SSD), comprising over 97k songs (4,751 hours) with over 49k synthetic songs from popular platforms like Suno and Udio. Furthermore, we highlight the importance of modeling long-range temporal dependencies in songs for effective authenticity detection, an aspect entirely overlooked in existing methods. To utilize long-range patterns, we introduce SpecTTTra, a novel architecture that significantly improves time and memory efficiency over conventional CNN and Transformer-based models. For long songs, our top-performing variant outperforms ViT by 8% in F1 score, is 38% faster, and uses 26% less memory, while also surpassing ConvNeXt with a 1% F1 score gain, 20% speed boost, and 67% memory reduction. | deepfake detection, fake song detection, synthetic song detection, efficient model, dataset, audio processing | We introduce SONICS, a large-scale dataset of end-to-end synthetic songs, propose SpecTTTra, an efficient model that captures long-range temporal patterns for effective fake song detection, and provide Human-AI benchmark for comprehensive analysis. | 261 | 2408.14080 | [
-0.11187560111284256,
-0.05647817626595497,
-0.00789497047662735,
-0.024257002398371696,
0.0851198360323906,
0.01820552721619606,
-0.023054547607898712,
-0.10199912637472153,
0.02710202895104885,
-0.10215191543102264,
-0.052479762583971024,
-0.05723346769809723,
0.050007522106170654,
-0.060849882662296295,
-0.030477076768875122,
-0.0030436990782618523,
0.045914556831121445,
0.08205803483724594,
-0.0482298918068409,
-0.029568573459982872,
0.014031512662768364,
0.015525012277066708,
-0.022638224065303802,
-0.04167146980762482,
0.0441817082464695,
0.039988599717617035,
-0.02223411202430725,
-0.013587580993771553,
0.00039277307223528624,
-0.05249901860952377,
0.07792859524488449,
0.08028187602758408,
0.045959822833538055,
-0.0355415903031826,
0.010257911868393421,
-0.012170125730335712,
-0.012401697225868702,
0.0018125961069017649,
-0.009349347092211246,
-0.08241936564445496,
0.04683678597211838,
0.03008217178285122,
-0.05946493148803711,
0.02656717412173748,
-0.049634575843811035,
0.00645790109410882,
-0.07113751024007797,
-0.023747742176055908,
0.029744978994131088,
0.07015445083379745,
-0.04926121234893799,
-0.019939295947551727,
0.018827468156814575,
0.09820111095905304,
-0.054168298840522766,
0.01034474279731512,
0.08083131909370422,
0.06507143378257751,
-0.021871475502848625,
0.07379847019910812,
0.014631261117756367,
-0.0842742919921875,
0.0309945959597826,
-0.08972545713186264,
0.0053794728592038155,
0.015307472087442875,
-0.03639381751418114,
0.02040235325694084,
0.04285403713583946,
0.028301335871219635,
0.035532642155885696,
0.07637787610292435,
-0.016489390283823013,
0.05354924127459526,
-0.003826156025752425,
0.025927219539880753,
-0.044137950986623764,
-0.028106965124607086,
-0.0160362608730793,
-0.11020436882972717,
0.03905119374394417,
-0.05509881302714348,
0.034908726811409,
-0.05658547952771187,
0.0738716572523117,
0.05224910378456116,
-0.004193543456494808,
0.030328422784805298,
-0.07960063219070435,
0.020161710679531097,
-0.11828305572271347,
-0.025507014244794846,
0.05091847479343414,
-0.07299939543008804,
0.0502818264067173,
0.05358658358454704,
-0.012091796845197678,
-0.026803605258464813,
0.0964295044541359,
0.0728907510638237,
-0.06519630551338196,
0.005626955535262823,
-0.03742149472236633,
-0.06611370295286179,
0.04639703780412674,
-0.06121582165360451,
0.060966722667217255,
0.009718202985823154,
0.02301555499434471,
-0.02875969372689724,
0.07050721347332001,
0.014009763486683369,
0.029059115797281265,
-0.027913618832826614,
0.07485363632440567,
0.05021189525723457,
0.024643424898386,
0.031291451305150986,
-0.03670763224363327,
0.019394157454371452,
0.007064636796712875,
0.008434426970779896,
-0.018769992515444756,
0.06022872030735016,
-0.039741434156894684,
0.01611504890024662,
-0.033859577029943466,
-2.739339663143863e-34,
0.051477279514074326,
0.006804718170315027,
0.015197676606476307,
-0.0002631841925904155,
0.04214203730225563,
-0.11825340986251831,
-0.08925464749336243,
0.0551513209939003,
-0.037280112504959106,
0.05341821536421776,
0.01707141101360321,
0.0034680108074098825,
-0.07778305560350418,
0.04847719892859459,
0.06698692589998245,
0.000627226079814136,
0.0016457976307719946,
-0.029898473992943764,
-0.005007781088352203,
-0.031936753541231155,
0.054235391318798065,
-0.0558028370141983,
0.07033758610486984,
-0.011793517507612705,
0.0669788345694542,
-0.006919339299201965,
0.013131281360983849,
-0.036314308643341064,
0.057083867490291595,
-0.006130068562924862,
-0.10994450747966766,
-0.017538100481033325,
0.09047483652830124,
0.04759906232357025,
-0.006274224724620581,
0.0112908398732543,
-0.05655352398753166,
-0.010694658383727074,
0.008688509464263916,
-0.0598331093788147,
-0.02013537846505642,
-0.0059492033906280994,
-0.04377765208482742,
-0.00920184701681137,
-0.07201400399208069,
0.0002434749621897936,
-0.02682916447520256,
0.05429130420088768,
0.06193133816123009,
0.029782822355628014,
0.0051134731620550156,
0.08494272083044052,
0.029293742030858994,
-0.0016545479884371161,
-0.0053697144612669945,
-0.038366351276636124,
0.0480213463306427,
-0.019301090389490128,
0.042826905846595764,
0.04703044518828392,
0.02427101507782936,
0.015820898115634918,
0.015124266967177391,
0.018392762169241905,
-0.054640304297208786,
0.10159610956907272,
0.07659423351287842,
0.015378041192889214,
0.04762013629078865,
0.03461328521370888,
-0.021502992138266563,
-0.0421023890376091,
-0.013576705940067768,
0.02609294094145298,
-0.06333758682012558,
-0.03972302004694939,
0.05633101984858513,
-0.036900702863931656,
0.04587117210030556,
0.01279184129089117,
-0.0036074514500796795,
0.03972547501325607,
-0.05013285204768181,
0.008637606166303158,
-0.03942553326487541,
0.026249466463923454,
-0.010278388857841492,
-0.10333243012428284,
-0.05262228474020958,
-0.007613415829837322,
-0.10973334312438965,
0.010569967329502106,
-0.07762088626623154,
0.02509617805480957,
-0.03488447889685631,
-4.8356958550024355e-34,
0.005807359702885151,
0.10006467252969742,
0.014933192171156406,
0.045750804245471954,
-0.012683959677815437,
-0.032730523496866226,
-0.03157468140125275,
0.06790643185377121,
0.030462950468063354,
0.05528165027499199,
-0.03276422619819641,
-0.05847072973847389,
0.04428714141249657,
-0.06944528222084045,
0.010203415527939796,
-0.04083796218037605,
0.004062978085130453,
0.04760398343205452,
-0.030328460037708282,
-0.012025617063045502,
0.010538502596318722,
0.020424993708729744,
-0.014290163293480873,
0.014103157445788383,
-0.0003310450119897723,
0.013226818293333054,
-0.03744128346443176,
0.058719929307699203,
0.027325410395860672,
-0.03726769983768463,
0.00901652779430151,
0.007089187391102314,
-0.10193479806184769,
-0.13087739050388336,
-0.05318332463502884,
-0.0462912917137146,
0.07040715962648392,
0.08696393668651581,
0.01501829270273447,
0.04724700748920441,
0.03975895419716835,
-0.0018592411652207375,
-0.10131453722715378,
-0.005354253575205803,
-0.04493871331214905,
-0.1021496057510376,
-0.06347319483757019,
0.11838015168905258,
0.035380080342292786,
-0.05205823481082916,
0.0848769024014473,
-0.032606340944767,
0.04244338348507881,
0.013632038608193398,
-0.024716651067137718,
-0.04957520216703415,
-0.03486473485827446,
-0.010300111956894398,
-0.000008833995707391296,
0.10551169514656067,
0.01961926557123661,
-0.07274169474840164,
-0.057607501745224,
-0.056881554424762726,
-0.03261307254433632,
-0.005485334899276495,
0.05061718821525574,
0.014116069301962852,
-0.018583863973617554,
0.058337047696113586,
0.09081006050109863,
-0.001259838929399848,
-0.018814507871866226,
0.08923008292913437,
-0.08851148933172226,
-0.029814010486006737,
-0.09042076766490936,
-0.019878502935171127,
-0.0004635080404113978,
0.0020201599691063166,
0.0018387268064543605,
0.015578878112137318,
0.051902905106544495,
0.05636276304721832,
0.11480981111526489,
0.1035115122795105,
-0.018412388861179352,
-0.020896850153803825,
-0.009013705886900425,
0.07660902291536331,
-0.024840502068400383,
0.08434107899665833,
-0.09641382843255997,
0.0649743527173996,
-0.042324554175138474,
-3.5166248579798776e-8,
-0.03922879323363304,
0.00722476514056325,
0.008738699369132519,
-0.0006146660889498889,
0.025960857048630714,
-0.018760420382022858,
0.02261151559650898,
-0.09765137732028961,
0.01446856465190649,
-0.07314454019069672,
0.05205812677741051,
-0.08125250786542892,
0.005810975097119808,
-0.03897117078304291,
-0.09902249276638031,
0.04499807208776474,
-0.019133565947413445,
0.037008222192525864,
-0.030530937016010284,
-0.0016434091376140714,
-0.021600019186735153,
0.09013436734676361,
0.0350467711687088,
-0.054654560983181,
0.04415483400225639,
-0.03679085522890091,
0.017444904893636703,
0.04824167862534523,
-0.016165781766176224,
0.07858745753765106,
0.015547781251370907,
-0.006413316819816828,
0.02879234403371811,
-0.06785370409488678,
0.02264322340488434,
0.09566739201545715,
-0.0045150392688810825,
0.004872325342148542,
-0.11440487951040268,
-0.038498908281326294,
-0.02798687480390072,
0.10457029938697815,
-0.05878731980919838,
0.004296590574085712,
-0.03602097928524017,
-0.13074876368045807,
0.047871191054582596,
-0.05751713737845421,
0.021237699314951897,
0.0450483039021492,
0.026142939925193787,
-0.09444782882928848,
-0.01475671585649252,
0.058262746781110764,
0.0489625409245491,
0.036665793508291245,
-0.004807949997484684,
0.0824509784579277,
0.06887838989496231,
0.06623047590255737,
0.025020984932780266,
-0.06724312901496887,
0.09464448690414429,
-0.015064104460179806
] |
Self-Correcting Decoding with Generative Feedback for Mitigating Hallucinations in Large Vision-Language Models | https://openreview.net/forum?id=tTBXePRKSx | [
"Ce Zhang",
"Zifu Wan",
"Zhehan Kan",
"Martin Q. Ma",
"Simon Stepputtis",
"Deva Ramanan",
"Russ Salakhutdinov",
"Louis-Philippe Morency",
"Katia P. Sycara",
"Yaqi Xie"
] | Poster | While recent Large Vision-Language Models (LVLMs) have shown remarkable performance in multi-modal tasks, they are prone to generating hallucinatory text responses that do not align with the given visual input, which restricts their practical applicability in real-world scenarios. In this work, inspired by the observation that the text-to-image generation process is the inverse of image-conditioned response generation in LVLMs, we explore the potential of leveraging text-to-image generative models to assist in mitigating hallucinations in LVLMs. We discover that generative models can offer valuable self-feedback for mitigating hallucinations at both the response and token levels. Building on this insight, we introduce self-correcting Decoding with Generative Feedback (DeGF), a novel training-free algorithm that incorporates feedback from text-to-image generative models into the decoding process to effectively mitigate hallucinations in LVLMs. Specifically, DeGF generates an image from the initial response produced by LVLMs, which acts as an auxiliary visual reference and provides self-feedback to verify and correct the initial response through complementary or contrastive decoding. Extensive experimental results validate the effectiveness of our approach in mitigating diverse types of hallucinations, consistently surpassing state-of-the-art methods across six benchmarks. Code is available at https://github.com/zhangce01/DeGF. | Large Vision-Language Models, Hallucinations, Generative Feedback | We introduce self-correcting Decoding with Generative Feedback (DeGF), a novel training-free algorithm that incorporates feedback from text-to-image generative models into the decoding process to effectively mitigate hallucinations in LVLMs. | 258 | 2502.06130 | [
-0.017682336270809174,
-0.03589547425508499,
0.07707074284553528,
0.029477378353476524,
0.06817027926445007,
0.0032583808060735464,
-0.012092653661966324,
-0.017084771767258644,
0.07236719876527786,
-0.05685347691178322,
-0.070391446352005,
-0.06810342520475388,
0.046815454959869385,
0.009264274500310421,
-0.036686234176158905,
0.03340030834078789,
0.07339909672737122,
0.08553775399923325,
-0.05837855860590935,
-0.04249994456768036,
0.07295969873666763,
0.015167820267379284,
0.038269560784101486,
-0.03256724402308464,
0.08374185115098953,
0.04234332963824272,
-0.042910490185022354,
-0.03932387754321098,
0.008827472105622292,
-0.049939192831516266,
0.033161137253046036,
0.06365318596363068,
-0.04246332123875618,
-0.0006648471462540329,
-0.0582062229514122,
0.0824858546257019,
-0.06994898617267609,
0.03535253554582596,
0.020318152382969856,
-0.06876007467508316,
-0.02817576937377453,
0.051973260939121246,
0.029070306569337845,
0.016358517110347748,
-0.0035925228148698807,
-0.036665670573711395,
-0.03954729437828064,
0.017058532685041428,
0.010615203529596329,
-0.040626004338264465,
-0.03086029924452305,
0.013965739868581295,
-0.005130671430379152,
0.023529747501015663,
-0.052007611840963364,
0.04443046823143959,
0.06694461405277252,
0.021899176761507988,
-0.020070092752575874,
-0.002231900580227375,
-0.04341471195220947,
-0.03258159011602402,
-0.025184564292430878,
-0.030588608235120773,
-0.0003874040558002889,
-0.008112103678286076,
0.037091657519340515,
0.02136826328933239,
-0.04589751735329628,
-0.008496458642184734,
-0.013995300978422165,
-0.019903011620044708,
-0.010271004401147366,
0.03165971860289574,
0.059490352869033813,
0.05526513233780861,
-0.04509831964969635,
-0.09871800243854523,
0.056963808834552765,
-0.0641617402434349,
0.052327439188957214,
0.049294617027044296,
0.10785480588674545,
-0.0065139043144881725,
0.10272880643606186,
0.008767602033913136,
0.006150749046355486,
0.08671814203262329,
0.0332995280623436,
0.0576557032763958,
-0.09411101788282394,
-0.11523015797138214,
-0.0022836450953036547,
0.006656295154243708,
0.004391312599182129,
0.039006076753139496,
0.017210841178894043,
-0.08473161607980728,
-0.017547527328133583,
0.027779042720794678,
0.04236055538058281,
0.0236008632928133,
-0.010893512517213821,
-0.04716448858380318,
0.03381853923201561,
-0.017032992094755173,
0.12171871960163116,
0.09490416944026947,
-0.0006543440977111459,
-0.07514125108718872,
0.04863635078072548,
0.011659554205834866,
-0.019397921860218048,
0.006362729240208864,
-0.029554402455687523,
-0.04039974883198738,
-0.006697501987218857,
0.0027234700974076986,
0.022612862288951874,
0.0031371808145195246,
0.013241739012300968,
0.0007140440284274518,
-0.01705886609852314,
-0.012293681502342224,
0.0016203642589971423,
-0.10916417092084885,
-0.04910645633935928,
-6.0064519534004515e-34,
0.005795069504529238,
0.03057745099067688,
0.025735577568411827,
0.06260233372449875,
0.08126769214868546,
-0.012262755073606968,
-0.10098595172166824,
-0.026386303827166557,
0.0012594693107530475,
-0.0663764625787735,
-0.01949380338191986,
-0.0711384043097496,
-0.0381011962890625,
0.10547104477882385,
0.01154532190412283,
0.03132441267371178,
-0.08274023979902267,
0.0213491078466177,
-0.054292645305395126,
-0.011613648384809494,
-0.0044744545593857765,
0.012939205393195152,
-0.037750113755464554,
0.03862110525369644,
-0.0020584037993103266,
0.03545771539211273,
0.07136601954698563,
-0.06069376319646835,
-0.042194221168756485,
0.009486109018325806,
-0.14556875824928284,
0.037420470267534256,
0.07604323327541351,
-0.007786826230585575,
0.014327206648886204,
0.005837030243128538,
0.030695289373397827,
-0.02100410684943199,
0.01047348789870739,
-0.01390100084245205,
0.025574736297130585,
0.08167436718940735,
-0.02967817708849907,
-0.002126471372321248,
-0.08244768530130386,
-0.01964060589671135,
0.04156770557165146,
0.07293505221605301,
-0.14036943018436432,
-0.04355170577764511,
0.042579494416713715,
0.04170818626880646,
-0.01633232831954956,
-0.03640875965356827,
-0.031482040882110596,
-0.03404929116368294,
0.04357103258371353,
-0.007680495735257864,
0.047585342079401016,
-0.027564607560634613,
0.09831909090280533,
0.07649088650941849,
0.009334912523627281,
0.07919551432132721,
0.006847416050732136,
-0.04112815484404564,
0.03335394337773323,
-0.03613807633519173,
-0.01055154949426651,
-0.07016687095165253,
-0.09056565910577774,
0.01691443659365177,
-0.024479864165186882,
-0.022750474512577057,
0.018756598234176636,
0.011394177563488483,
0.029347926378250122,
-0.04010115563869476,
-0.013064777478575706,
0.04716647043824196,
0.015253402292728424,
-0.029909297823905945,
-0.03208833187818527,
-0.05842873081564903,
0.009056233800947666,
-0.0802450031042099,
0.0178061556071043,
-0.08213472366333008,
-0.1027376726269722,
-0.0644221305847168,
-0.03514361381530762,
-0.0506981760263443,
0.04920341074466705,
-0.031868383288383484,
-0.000799016619566828,
1.821597674802213e-35,
0.003923408687114716,
0.007190392818301916,
0.0332200825214386,
0.02355198934674263,
-0.020515242591500282,
-0.02237103506922722,
0.07275407016277313,
0.017501026391983032,
0.04362264275550842,
-0.07546314597129822,
-0.051368966698646545,
0.026995405554771423,
-0.009140018373727798,
0.007087960373610258,
0.012283643707633018,
-0.08737334609031677,
0.07011093944311142,
0.016903795301914215,
-0.09655049443244934,
0.021907832473516464,
-0.00557975796982646,
0.11949238926172256,
-0.06467097252607346,
0.0005863001570105553,
-0.019082769751548767,
0.0869256854057312,
-0.012245922349393368,
0.07009091973304749,
0.033723052591085434,
-0.010304130613803864,
-0.004640964791178703,
0.07567212730646133,
0.019082970917224884,
-0.02213616669178009,
-0.037316132336854935,
0.09328100085258484,
0.026785753667354584,
-0.032300129532814026,
-0.11234387755393982,
0.05953159183263779,
0.08622872829437256,
-0.05674789473414421,
-0.07622435688972473,
0.056583937257528305,
-0.01391567662358284,
-0.08840222656726837,
-0.01575777865946293,
-0.05115924030542374,
0.04698468744754791,
0.09687478840351105,
-0.024267328903079033,
0.014514644630253315,
-0.1282079517841339,
0.014267036691308022,
-0.08558063209056854,
-0.12771931290626526,
-0.10371797531843185,
0.013683965429663658,
0.08581607788801193,
0.0027426297310739756,
-0.05520612746477127,
-0.002715297508984804,
0.03432415798306465,
-0.027304893359541893,
-0.0201727282255888,
-0.01247645728290081,
-0.04937456548213959,
-0.000679365242831409,
0.10798000544309616,
0.01623949408531189,
0.07785575091838837,
0.009781720116734505,
0.02744326926767826,
0.10643334686756134,
0.055825285613536835,
-0.049196623265743256,
0.00502537377178669,
-0.004860784858465195,
-0.024866562336683273,
-0.10775300860404968,
0.02530619502067566,
-0.04218827188014984,
0.017328452318906784,
0.12987671792507172,
0.09275827556848526,
-0.010131040588021278,
-0.008338558487594128,
0.05021681264042854,
-0.028718376532197,
0.02361001819372177,
-0.015118785202503204,
0.012529496103525162,
0.03240691497921944,
0.06749094277620316,
-0.0007952930172905326,
-3.417778771108715e-8,
-0.06153742969036102,
-0.02262897416949272,
-0.014503846876323223,
0.04470302537083626,
0.023187464103102684,
-0.0511796735227108,
-0.01281272154301405,
0.04253862425684929,
-0.004259904380887747,
-0.1030871570110321,
0.07871107757091522,
-0.04022669047117233,
-0.03368276730179787,
-0.11366888880729675,
-0.0034494439605623484,
0.03850822150707245,
0.0025481409393250942,
0.08873911947011948,
-0.031307440251111984,
-0.02597237192094326,
0.03161100670695305,
-0.009901908226311207,
-0.03149377554655075,
-0.06716553121805191,
0.04853218421339989,
0.015924077481031418,
-0.02789515256881714,
-0.0028762263245880604,
0.013413319364190102,
-0.056066710501909256,
0.10441847890615463,
0.053985681384801865,
0.010348407551646233,
-0.005930867046117783,
-0.04438818618655205,
0.030994269996881485,
-0.005818159785121679,
-0.0071721505373716354,
0.0657474547624588,
-0.09875203669071198,
0.06232328712940216,
0.059435825794935226,
-0.04190594702959061,
0.02268231101334095,
-0.009912284091114998,
-0.021632816642522812,
0.0783432275056839,
-0.09868171066045761,
0.023835988715291023,
-0.006520661059767008,
-0.035789694637060165,
0.015324782580137253,
0.023527977988123894,
0.09523534774780273,
0.03427966311573982,
-0.016793617978692055,
0.09398943185806274,
0.015632204711437225,
0.04028685763478279,
0.0618019625544548,
0.018784038722515106,
0.040488164871931076,
-0.04122033715248108,
-0.023942595347762108
] |
Latent Space Chain-of-Embedding Enables Output-free LLM Self-Evaluation | https://openreview.net/forum?id=jxo70B9fQo | [
"Yiming Wang",
"Pei Zhang",
"Baosong Yang",
"Derek F. Wong",
"Rui Wang"
] | Poster | LLM self-evaluation relies on the LLM's own ability to estimate response correctness, which can greatly improve its deployment reliability.
In this research track, we propose the Chain-of-Embedding (CoE) in the latent space to enable LLMs to perform output-free self-evaluation. CoE consists of all progressive hidden states produced during the inference time, which can be treated as the latent thinking path of LLMs. We find that when LLMs respond correctly and incorrectly, their CoE features differ, these discrepancies assist us in estimating LLM response correctness. Experiments in four diverse domains and seven LLMs fully demonstrate the effectiveness of our method. Meanwhile, its label-free design intent without any training and millisecond-level computational cost ensure real-time feedback in large-scale scenarios.
More importantly, we provide interesting insights into LLM response correctness from the perspective of hidden state changes inside LLMs. | Large Language Models, Label-free Self-Evaluation, AI Reliability | We propose the Chain-of-Embedding method for LLM self-evaluation, which enables output-free response correctness estimation during inference time. | 255 | 2410.13640 | [
-0.040152084082365036,
-0.05596331134438515,
0.027447445318102837,
0.041255705058574677,
0.0840960443019867,
0.0043960679322481155,
-0.07141955196857452,
-0.04050102084875107,
0.062345243990421295,
-0.01868479512631893,
-0.07019779831171036,
-0.057827871292829514,
0.05630278214812279,
-0.028552120551466942,
-0.018200574442744255,
0.006299148313701153,
0.15574948489665985,
0.05110696330666542,
-0.021034136414527893,
-0.08390878885984421,
0.057282958179712296,
0.046113982796669006,
0.0777275338768959,
-0.023014193400740623,
0.037151891738176346,
0.00028979041962884367,
-0.0031456276774406433,
0.03228822723031044,
0.016634907573461533,
-0.07288740575313568,
0.06770392507314682,
0.04760534688830376,
-0.03770298510789871,
-0.017760135233402252,
-0.04427279531955719,
0.03853195533156395,
-0.05847655609250069,
0.016926808282732964,
-0.08890356123447418,
-0.004779335577040911,
-0.02075769565999508,
0.0394572950899601,
0.019872764125466347,
0.06183179095387459,
0.0012782621197402477,
-0.06254405528306961,
0.04287462309002876,
-0.06888782232999802,
-0.10679664462804794,
0.040695589035749435,
-0.01106581836938858,
0.03166871890425682,
-0.042069390416145325,
0.014675244688987732,
-0.04996931552886963,
0.04493854194879532,
0.06322088092565536,
0.08928349614143372,
-0.05808113515377045,
0.018887335434556007,
-0.01646333746612072,
-0.13294143974781036,
-0.0013899145415052772,
-0.0008390832226723433,
0.07663377374410629,
0.03181777894496918,
0.012573287822306156,
0.09878288954496384,
-0.08630126714706421,
-0.0026823317166417837,
-0.04874374717473984,
0.028732623904943466,
-0.06789124757051468,
0.01773565262556076,
0.017388485372066498,
0.05675018951296806,
0.027511758729815483,
-0.013202047906816006,
0.12553562223911285,
0.009410033002495766,
0.00011290571274003014,
0.020078424364328384,
-0.06095467135310173,
0.037047129124403,
0.12237266451120377,
0.015289037488400936,
0.023891571909189224,
0.0965672954916954,
-0.0035034441389143467,
0.026682475581765175,
0.02858632616698742,
-0.09978773444890976,
-0.044833071529865265,
0.03130844235420227,
0.012639312073588371,
0.004452001769095659,
0.05209129676222801,
-0.11030200123786926,
0.010136103257536888,
0.06391257792711258,
0.023859692737460136,
0.10189981758594513,
0.035933997482061386,
-0.022066842764616013,
-0.0010814436245709658,
-0.036239903420209885,
0.04791436344385147,
0.1099105179309845,
0.018870240077376366,
-0.06826458126306534,
-0.010210045613348484,
0.043187808245420456,
0.030329011380672455,
0.06893359869718552,
0.025301869958639145,
-0.09657341241836548,
0.061846084892749786,
-0.066077820956707,
0.049247466027736664,
-0.02670396864414215,
-0.03629949316382408,
-0.00002708491047087591,
0.0627107098698616,
0.008819992654025555,
0.06201425939798355,
-0.022325560450553894,
-0.08078128844499588,
3.270340996712763e-33,
0.010319484397768974,
-0.012402222491800785,
0.028767500072717667,
0.03884120285511017,
0.03081675060093403,
0.011197947897017002,
-0.006790450774133205,
0.057804174721241,
-0.025631479918956757,
0.030500348657369614,
-0.017572877928614616,
-0.014330016449093819,
-0.0005302820354700089,
0.07197460532188416,
0.03237201273441315,
0.06292958557605743,
-0.06493062525987625,
0.03591202571988106,
-0.0880102589726448,
-0.020174192264676094,
-0.019410641863942146,
0.02371932938694954,
0.024235330522060394,
-0.10093741863965988,
0.06351152062416077,
-0.023232020437717438,
0.06551705300807953,
-0.008851996622979641,
-0.08663688600063324,
-0.02014704793691635,
-0.14295220375061035,
0.034260839223861694,
0.02080322615802288,
0.01441258005797863,
-0.016272127628326416,
-0.024108197540044785,
0.032648514956235886,
-0.06328921020030975,
0.009719112887978554,
-0.026431109756231308,
-0.010796607472002506,
0.01700376160442829,
0.036071114242076874,
-0.010459298267960548,
-0.1072007343173027,
-0.07950873672962189,
0.0043075913563370705,
-0.016092177480459213,
-0.017088955268263817,
0.004968115594238043,
0.0351785384118557,
0.05649576336145401,
-0.0022711744531989098,
-0.06801023334264755,
-0.05386187508702278,
-0.03589804470539093,
-0.009976853616535664,
0.04607203230261803,
0.026321643963456154,
0.03668829798698425,
-0.01797882840037346,
0.03729041665792465,
-0.006625803653150797,
0.016226567327976227,
-0.020936625078320503,
0.04962402954697609,
-0.006755400914698839,
0.0321301631629467,
0.08172151446342468,
-0.010991583578288555,
-0.029802381992340088,
0.04551270976662636,
-0.07689046859741211,
-0.05493803322315216,
0.0498896948993206,
-0.07223247736692429,
-0.00793777871876955,
-0.10904555767774582,
0.05140967667102814,
-0.009157238528132439,
0.08133197575807571,
-0.0030548162758350372,
-0.08301857858896255,
0.02863243967294693,
0.008752788417041302,
-0.10308124125003815,
0.015178973786532879,
-0.1199789047241211,
-0.044323138892650604,
0.050000179558992386,
-0.014296608977019787,
-0.058167245239019394,
0.027813833206892014,
0.06394242495298386,
-0.025092916563153267,
-2.394963577966852e-33,
-0.10024505108594894,
-0.022298207506537437,
-0.005104518495500088,
0.09596721827983856,
0.0023646424524486065,
-0.04522966966032982,
0.027971191331744194,
0.057799868285655975,
0.021420611068606377,
-0.015091568231582642,
-0.03311988338828087,
-0.03589668497443199,
-0.06472982466220856,
0.06746600568294525,
0.04487534984946251,
-0.0018330864841118455,
-0.035910554230213165,
-0.04444153234362602,
0.04985007643699646,
-0.04663006216287613,
0.04629415646195412,
0.06122596934437752,
-0.014420358464121819,
0.02949153073132038,
-0.04014824703335762,
0.07099315524101257,
0.020844753831624985,
0.01697446219623089,
-0.00567104434594512,
-0.03186478465795517,
0.004446069244295359,
-0.025505537167191505,
-0.027211204171180725,
0.01843707263469696,
-0.0102446423843503,
0.054016463458538055,
0.03751816228032112,
0.02878432720899582,
-0.029498804360628128,
0.0005412563914433122,
0.04979772865772247,
-0.026536084711551666,
-0.09324994683265686,
0.0315282829105854,
-0.053116828203201294,
-0.06102928891777992,
-0.06608276814222336,
0.005115930922329426,
0.084364153444767,
0.047987330704927444,
-0.06001664698123932,
0.029510291293263435,
-0.05437278002500534,
-0.009375398978590965,
-0.037228140980005264,
-0.09739154577255249,
-0.023780696094036102,
-0.02918390929698944,
0.054843828082084656,
0.04261897876858711,
-0.06097932532429695,
-0.006773779168725014,
0.033620934933423996,
-0.13462050259113312,
0.03508258983492851,
-0.035909503698349,
-0.0170432161539793,
0.026283003389835358,
0.036249980330467224,
0.008578522130846977,
0.07569153606891632,
0.047489844262599945,
-0.04427579790353775,
-0.014846471138298512,
0.10657678544521332,
-0.023341041058301926,
0.016358477994799614,
-0.05833229422569275,
-0.03072008490562439,
-0.05346708372235298,
0.0373212955892086,
0.01996631920337677,
-0.022402141243219376,
0.025510340929031372,
-0.013831420801579952,
0.025115426629781723,
0.060978785157203674,
0.07048272341489792,
0.006935388781130314,
-0.0032136535737663507,
0.02585977129638195,
0.09220847487449646,
-0.05233647674322128,
0.08241427689790726,
-0.04611658304929733,
-4.381824325605521e-8,
-0.061501480638980865,
-0.017119599506258965,
0.05591993033885956,
0.08778790384531021,
-0.029187172651290894,
-0.013745483011007309,
0.07273316383361816,
0.053206365555524826,
-0.02050679363310337,
-0.07144403457641602,
0.10374100506305695,
-0.052820928394794464,
-0.07656533271074295,
-0.04237092286348343,
0.03883035480976105,
-0.0005499320686794817,
-0.021892234683036804,
0.022869303822517395,
-0.05534550920128822,
-0.02643311396241188,
-0.00995480827987194,
0.0681421309709549,
0.018515361472964287,
-0.03722880408167839,
0.017302967607975006,
-0.06276644766330719,
0.011278154328465462,
0.1190665140748024,
0.05849673971533775,
-0.008589311502873898,
0.0009659386123530567,
0.01594114676117897,
-0.014398068189620972,
0.023128950968384743,
-0.013389700092375278,
0.04335886240005493,
0.024664057418704033,
0.0012994464486837387,
0.07979010045528412,
-0.016506655141711235,
-0.06297286599874496,
0.05601631477475166,
-0.06252725422382355,
0.04921352118253708,
-0.009600852616131306,
-0.010043037123978138,
-0.05005629360675812,
-0.09932293742895126,
-0.022711826488375664,
-0.09615638852119446,
0.03530585765838623,
-0.054540783166885376,
-0.06633388251066208,
0.08123916387557983,
0.09749168902635574,
0.0221138596534729,
0.022207584232091904,
-0.037660710513591766,
0.03474392369389534,
0.08094219863414764,
0.01957603357732296,
-0.026369687169790268,
-0.040109168738126755,
-0.006407239940017462
] |
FasterCache: Training-Free Video Diffusion Model Acceleration with High Quality | https://openreview.net/forum?id=W49UjcpGxx | [
"Zhengyao Lv",
"Chenyang Si",
"Junhao Song",
"Zhenyu Yang",
"Yu Qiao",
"Ziwei Liu",
"Kwan-Yee K. Wong"
] | Poster | In this paper, we present \textbf{\textit{FasterCache}}, a novel training-free strategy designed to accelerate the inference of video diffusion models with high-quality generation. By analyzing existing cache-based methods, we observe that \textit{directly reusing adjacent-step features degrades video quality due to the loss of subtle variations}. We further perform a pioneering investigation of the acceleration potential of classifier-free guidance (CFG) and reveal significant redundancy between conditional and unconditional features within the same timestep. Capitalizing on these observations, we introduce FasterCache to substantially accelerate diffusion-based video generation. Our key contributions include a dynamic feature reuse strategy that preserves both feature distinction and temporal continuity, and CFG-Cache which optimizes the reuse of conditional and unconditional outputs to further enhance inference speed without compromising video quality. We empirically evaluate FasterCache on recent video diffusion models. Experimental results show that FasterCache can significantly accelerate video generation (\eg 1.67$\times$ speedup on Vchitect-2.0) while keeping video quality comparable to the baseline, and consistently outperform existing methods in both inference speed and video quality. \textit{Our code will be made public upon publication.} | Efficient Video Synthesis | null | 248 | 2410.19355 | [
0.004174129106104374,
-0.1472194939851761,
0.018823817372322083,
0.03909976780414581,
0.15135663747787476,
-0.0454607754945755,
-0.07205406576395035,
-0.10382075607776642,
0.02519093081355095,
-0.07166139781475067,
-0.05234129726886749,
-0.027153493836522102,
-0.0794384777545929,
0.017029903829097748,
-0.020466094836592674,
-0.041074302047491074,
0.08905702084302902,
0.07620342820882797,
-0.1304255872964859,
-0.054903313517570496,
-0.0036496513057500124,
-0.11435125023126602,
0.02125912718474865,
0.014213385060429573,
0.036253489553928375,
0.019464822486042976,
0.025865763425827026,
-0.007717103231698275,
0.07384578883647919,
-0.018563318997621536,
0.022341972216963768,
-0.008699585683643818,
-0.04848441481590271,
-0.040873438119888306,
-0.039700206369161606,
0.0018110951641574502,
-0.010644939728081226,
-0.021088290959596634,
-0.04499315470457077,
-0.03338264673948288,
0.022389693185687065,
0.05172232165932655,
-0.009790079668164253,
0.05798829346895218,
0.0324011966586113,
0.05334935337305069,
0.031933367252349854,
-0.015846189111471176,
-0.02269834466278553,
-0.02854060009121895,
-0.018102513626217842,
-0.02193208411335945,
-0.09883467108011246,
0.03360961377620697,
-0.05840017646551132,
0.07488931715488434,
0.019439397379755974,
0.07498285919427872,
0.020763739943504333,
-0.023437025025486946,
-0.030924316495656967,
-0.10674144327640533,
0.04434961825609207,
-0.0012177543248981237,
-0.04875698313117027,
-0.0458793006837368,
0.08900165557861328,
0.030741654336452484,
0.039069488644599915,
0.06671608239412308,
-0.10504287481307983,
0.08016224950551987,
-0.01576947048306465,
0.04802420362830162,
-0.0431235246360302,
0.019246922805905342,
0.06996307522058487,
-0.0018584069330245256,
0.03551705181598663,
-0.10383623838424683,
0.024711698293685913,
-0.0476321205496788,
0.07087530940771103,
-0.00980309210717678,
0.07316373288631439,
-0.04245863854885101,
0.0249718576669693,
-0.0009492794633843005,
-0.03723588585853577,
0.00948377512395382,
-0.049247656017541885,
0.05115184932947159,
-0.01111238356679678,
-0.08810009062290192,
-0.014985854737460613,
0.005068033933639526,
-0.011334809474647045,
-0.08690700680017471,
-0.018846316263079643,
0.06960806250572205,
-0.027891801670193672,
-0.0008108990732580423,
-0.023941664025187492,
0.004960923921316862,
0.09828749299049377,
-0.012264171615242958,
0.0756593570113182,
0.09095527231693268,
0.03481091931462288,
0.04003262519836426,
0.04990502446889877,
0.09864698350429535,
0.016910912469029427,
-0.05210096761584282,
0.05466515198349953,
0.053770579397678375,
-0.09238385409116745,
-0.024183010682463646,
0.021271947771310806,
0.07305727154016495,
0.00008736721792956814,
-0.057449210435152054,
-0.09256613999605179,
-0.02569163590669632,
0.04062803462147713,
-0.07223174721002579,
-0.012748584151268005,
2.0889148075226384e-33,
-0.005553470458835363,
0.0074731456115841866,
0.03982774168252945,
0.013664277270436287,
0.03856964781880379,
0.03129618242383003,
0.03642946854233742,
0.03933003544807434,
-0.014416265301406384,
-0.012672039680182934,
0.00927466619759798,
-0.10499916970729828,
-0.08238770067691803,
0.11275821179151535,
0.02567685954272747,
-0.02201341651380062,
-0.08802011609077454,
0.012908147647976875,
-0.010967669077217579,
0.02146483212709427,
0.048111505806446075,
-0.03270329535007477,
-0.036293670535087585,
-0.044382818043231964,
-0.02989554964005947,
-0.0005180310690775514,
0.05849539861083031,
-0.022076819092035294,
0.011854512616991997,
-0.008094294928014278,
-0.07975786924362183,
-0.022305108606815338,
-0.0029851242434233427,
0.005973974708467722,
0.02332693338394165,
0.011189957149326801,
-0.04638190194964409,
-0.03806901350617409,
0.009533639997243881,
0.014094692654907703,
-0.019489558413624763,
0.0496690571308136,
-0.01705721952021122,
-0.05496058613061905,
-0.11958827823400497,
-0.0420527383685112,
0.0396285355091095,
0.01697886735200882,
-0.04315856844186783,
0.020769821479916573,
0.06153373420238495,
-0.018748747184872627,
0.01884276792407036,
-0.09277927130460739,
-0.06900590658187866,
0.001130375312641263,
0.09224622696638107,
-0.005494584795087576,
-0.011366584338247776,
-0.0011479011736810207,
0.024771487340331078,
0.04409495368599892,
-0.07151756435632706,
0.0657266303896904,
-0.03968116268515587,
-0.02090572938323021,
0.033594582229852676,
0.1389923095703125,
-0.0016435562865808606,
-0.04004674032330513,
-0.035787433385849,
0.02289285697042942,
0.00843734573572874,
-0.13483741879463196,
0.08155257254838943,
0.0021495406981557608,
0.03352446109056473,
-0.009656686335802078,
0.022850044071674347,
0.009558571502566338,
-0.005092794541269541,
-0.034566234797239304,
0.0241318978369236,
-0.08740013092756271,
0.020111309364438057,
0.0057913996279239655,
-0.003974479623138905,
-0.0721357986330986,
-0.04866405576467514,
-0.02022392489016056,
-0.02916652150452137,
0.01495340745896101,
0.07953392714262009,
0.02477584034204483,
0.02428530342876911,
-1.9121571913574296e-33,
0.036977335810661316,
0.06383583694696426,
0.039252378046512604,
0.07986719161272049,
-0.029940921813249588,
0.017972702160477638,
-0.010539533570408821,
-0.013671222142875195,
0.04073483124375343,
-0.15013863146305084,
-0.000420377153204754,
-0.06059592217206955,
-0.047531694173812866,
-0.028246868401765823,
-0.03403973579406738,
0.004503978881984949,
-0.05238356441259384,
-0.05636370927095413,
-0.057489167898893356,
0.02492458000779152,
0.025490019470453262,
0.023646898567676544,
-0.011148589663207531,
-0.004393219947814941,
-0.0636674240231514,
0.014628896489739418,
0.013325135223567486,
0.11904262006282806,
0.06807775050401688,
-0.06442711502313614,
0.0045696343295276165,
0.014274034649133682,
-0.01134960725903511,
-0.07791770249605179,
-0.06033928692340851,
0.0103827603161335,
0.06940633058547974,
-0.020855609327554703,
-0.04043465480208397,
0.16303983330726624,
0.06916936486959457,
-0.013577860780060291,
-0.04298022389411926,
0.05728686973452568,
0.0014018936781212687,
-0.0008488871389999986,
-0.04596920683979988,
-0.006846987642347813,
0.03349073603749275,
0.018078459426760674,
0.015323851257562637,
0.054733093827962875,
-0.12543132901191711,
0.06854760646820068,
0.0188088808208704,
-0.041632696986198425,
-0.13538393378257751,
0.008143543265759945,
0.0566730871796608,
0.02235989272594452,
-0.002864375477656722,
-0.011383706703782082,
-0.06527642160654068,
-0.07120757550001144,
0.06817234307527542,
-0.007882579229772091,
-0.04578716680407524,
0.006946692708879709,
0.037987224757671356,
0.07114487886428833,
0.04683684930205345,
0.012987201102077961,
-0.0450015515089035,
-0.050737760961055756,
-0.011781536974012852,
-0.03520403429865837,
0.09572166949510574,
-0.0544755719602108,
0.019745195284485817,
-0.053107935935258865,
-0.021092409268021584,
-0.0071210782043635845,
0.03249073028564453,
-0.017233768478035927,
0.02064797282218933,
0.07864511013031006,
0.007412218023091555,
0.04575968161225319,
0.03873822093009949,
0.07230836898088455,
-0.03726330026984215,
0.031857606023550034,
0.03449755907058716,
0.025557665154337883,
-0.04936586320400238,
-4.089911698201831e-8,
-0.08039973676204681,
0.03968146815896034,
-0.014749105088412762,
0.02778814174234867,
-0.007709470111876726,
-0.047674279659986496,
0.06383625417947769,
0.11403720825910568,
0.041991181671619415,
-0.07067059725522995,
0.056516360491514206,
0.053313348442316055,
-0.014312850311398506,
0.004525873344391584,
0.024423804134130478,
0.05799265578389168,
-0.019115492701530457,
-0.013009062968194485,
0.008735645562410355,
0.044888537377119064,
-0.023113645613193512,
-0.012468777596950531,
0.061708610504865646,
-0.042699944227933884,
0.012734021991491318,
-0.01898571290075779,
0.03523312881588936,
0.0625719428062439,
0.1304614543914795,
-0.02868342213332653,
-0.06529858708381653,
0.027332870289683342,
-0.016268141567707062,
-0.05395900830626488,
0.07808077335357666,
0.09095953404903412,
-0.014075500890612602,
0.0025925058871507645,
0.001571035012602806,
0.023037446662783623,
-0.006691626738756895,
0.03375832736492157,
0.01952635496854782,
0.01295652985572815,
-0.04143306612968445,
0.010421790182590485,
0.050794851034879684,
-0.05363842844963074,
-0.009227891452610493,
-0.02305060811340809,
-0.025031859055161476,
-0.010110235773026943,
-0.04406822472810745,
0.029578879475593567,
0.07358372956514359,
-0.011065381579101086,
0.05210259184241295,
0.020633889362215996,
0.032681904733181,
0.08954530954360962,
0.027991419658064842,
-0.027561409398913383,
0.02415599673986435,
-0.028857747092843056
] |
Correlation and Navigation in the Vocabulary Key Representation Space of Language Models | https://openreview.net/forum?id=VipcVxaTnG | [
"Letian Peng",
"Chenyang An",
"Jingbo Shang"
] | Poster | Language model (LM) decoding is based on the next-token prediction (NTP) probability distribution. For neural LMs (e.g., Transformer-based), NTP distribution is
essentially a softmax-regularized dot product between an encoded input context
(query) and fixed vocabulary representations (keys). In this paper, we study the
effect of the key distribution on the NTP distribution, with a focus on whether
the similarity between keys will trigger spurious correlations in NTP. Through
knowledge-probing tasks, we show that in the NTP distribution, the few top-ranked
tokens are typically accurate. However, the middle-ranked prediction is highly biased
towards the tokens that are distributionally (not necessarily semantically) similar to
these top ones. For instance, if “P” is predicted as the top-1 token, “A”-“Z” will all
be ranked high in NTP, no matter whether they can lead to correct decoding results.
This hurts the sampling diversity and makes the sampling of correct, long-tail
results hopeless and noisy. We attempt to alleviate this issue via a novel in-context
method that iteratively pushes the query representation away from explored regions.
Specifically, we include the explored decoding results in the context and prompt
the LM to generate something else, which encourages the LM to produce a query
representation that has small dot products with explored keys. Experiments on
knowledge-probing tasks show that our method leads to efficient navigation away
from explored keys to correct new keys. We further extend our method to open-ended and chain-of-thought (for reasoning) generation. Experiment results show
that ICN contributes to better generation diversity and improved self-consistency
voting performance. Finally, we discuss potential training issues caused by the
fixed key space together with the challenges and possible ways to address them in
future research. | Language Modeling, Next Token Prediction, Spurious Correlation, Generation Diversity | null | 246 | 2410.02284 | [
-0.06273062527179718,
-0.15393270552158356,
0.04814109578728676,
0.0225613284856081,
0.036164335906505585,
0.0413610078394413,
0.05369409918785095,
0.005476670805364847,
0.06776588410139084,
-0.02202834002673626,
-0.06193164363503456,
-0.014218988828361034,
0.08213983476161957,
0.0816769003868103,
-0.035086359828710556,
-0.014565734192728996,
0.08331012725830078,
0.07422485202550888,
-0.08194750547409058,
-0.07061076909303665,
0.05258093401789665,
0.013872462324798107,
0.046313632279634476,
-0.020819012075662613,
0.039486270397901535,
-0.004641097970306873,
-0.0019397910218685865,
-0.00734341936185956,
0.033979423344135284,
0.015573779121041298,
0.011447620578110218,
0.023673400282859802,
-0.026501145213842392,
0.014515532180666924,
-0.026966113597154617,
0.002655325224623084,
-0.08614786714315414,
-0.012027549557387829,
0.09571675956249237,
-0.07190459221601486,
0.03279970958828926,
0.04026103392243385,
0.017797859385609627,
0.05107307434082031,
0.08391623198986053,
0.062267620116472244,
0.025893257930874825,
-0.00044812963460572064,
-0.07922688871622086,
0.048145975917577744,
-0.08829726278781891,
0.03349130228161812,
0.00736286211758852,
0.027153976261615753,
-0.0922166034579277,
0.05328972265124321,
-0.06102252006530762,
0.052955225110054016,
0.001141781802289188,
-0.005275289062410593,
-0.023160375654697418,
-0.11413630843162537,
-0.04037684202194214,
0.03228753060102463,
-0.042057208716869354,
-0.036432743072509766,
-0.01863708533346653,
0.09004001319408417,
-0.02223576232790947,
0.029247352853417397,
0.029966721311211586,
0.06819268316030502,
-0.0055812024511396885,
0.04955706000328064,
0.018974540755152702,
0.06089210882782936,
0.050741877406835556,
-0.07608609646558762,
0.04184279218316078,
-0.06421233713626862,
0.006102896295487881,
0.06532277166843414,
0.09414973109960556,
-0.015359566546976566,
0.18721693754196167,
-0.0019348623463883996,
0.044669825583696365,
0.022434335201978683,
-0.037864167243242264,
-0.022708240896463394,
-0.053709130734205246,
-0.10374628752470016,
0.08159126341342926,
0.020555846393108368,
0.042449671775102615,
0.08673161268234253,
0.05253133922815323,
-0.016628947108983994,
0.032962918281555176,
0.06516582518815994,
-0.026861529797315598,
0.0797293484210968,
-0.002547007752582431,
-0.11456462740898132,
-0.02521475777029991,
-0.021388689056038857,
-0.025564784184098244,
0.0061311740428209305,
-0.005583790596574545,
-0.054302554577589035,
-0.015174830332398415,
0.09432560950517654,
0.023513173684477806,
0.025281568989157677,
-0.011677470058202744,
0.020778058096766472,
0.0036457667592912912,
-0.05565866827964783,
0.03221026435494423,
0.10851557552814484,
-0.07745189964771271,
-0.005521272774785757,
-0.05718226730823517,
-0.010786968283355236,
-0.04832084849476814,
0.029129335656762123,
0.052889201790094376,
-9.199001856382665e-34,
0.00508875073865056,
0.02627411298453808,
-0.030580837279558182,
-0.06757799535989761,
0.04907914623618126,
-0.006216051522642374,
0.0343848317861557,
0.07483299821615219,
-0.03218446671962738,
0.017209628596901894,
-0.027175234630703926,
0.03711846470832825,
-0.036533210426568985,
-0.006730007007718086,
-0.0749250277876854,
0.03766962140798569,
-0.03577305004000664,
0.04797815904021263,
-0.11240898072719574,
-0.07510469853878021,
0.010448729619383812,
0.02973964437842369,
-0.04610205814242363,
-0.04767368361353874,
-0.03301563486456871,
0.03403288498520851,
0.025132449343800545,
-0.0473795123398304,
-0.06306137144565582,
0.025640245527029037,
-0.10744983702898026,
-0.034614335745573044,
-0.03488728776574135,
-0.01587170921266079,
0.0003670367877930403,
-0.06104711443185806,
0.023464273661375046,
-0.017976924777030945,
-0.04636001959443092,
-0.07749057561159134,
-0.03731396421790123,
0.03156913444399834,
-0.011458859778940678,
0.009301124140620232,
-0.05819724127650261,
-0.050007157027721405,
0.007970956154167652,
-0.033512115478515625,
-0.02671639434993267,
-0.03951895609498024,
0.05649265646934509,
0.004136154428124428,
-0.02890026941895485,
-0.03518254682421684,
0.03448699414730072,
-0.01971466839313507,
0.018391385674476624,
0.013021762482821941,
-0.010890305042266846,
0.07893446087837219,
-0.041218992322683334,
0.005078188609331846,
0.05897849053144455,
0.06522267311811447,
0.10687123984098434,
0.007903315126895905,
-0.04321088269352913,
0.030803192406892776,
0.07385087013244629,
0.0006205664249137044,
0.05592011287808418,
0.02290387824177742,
-0.011510750278830528,
-0.06571438163518906,
0.020836200565099716,
0.04161274805665016,
0.06587224453687668,
-0.02088065817952156,
-0.029630789533257484,
0.021012334153056145,
0.037921931594610214,
-0.09077614545822144,
0.017421849071979523,
-0.019667599350214005,
-0.04575536027550697,
-0.01390868704766035,
0.05281413346529007,
-0.10378403961658478,
0.006284364499151707,
-0.03953227773308754,
0.008437272161245346,
-0.03163227066397667,
0.00864899531006813,
-0.012238098308444023,
-0.03548133373260498,
8.557009290100919e-34,
-0.024769745767116547,
-0.005495786666870117,
0.014974820427596569,
0.015062142163515091,
-0.07339634746313095,
-0.084962397813797,
-0.03475189954042435,
0.05451960116624832,
-0.045952461659908295,
0.001569927902892232,
-0.04846934229135513,
-0.10106601566076279,
0.10275305807590485,
0.025463519617915154,
0.10571084171533585,
-0.0006884942413307726,
0.05417797714471817,
0.02394382283091545,
0.07688245922327042,
0.0764307901263237,
0.03344373404979706,
0.05103855952620506,
-0.153199702501297,
0.07401878386735916,
-0.004672294948250055,
0.040708430111408234,
-0.029193159192800522,
0.05007077008485794,
-0.03063909336924553,
0.012615513056516647,
-0.002468058140948415,
0.014581812545657158,
-0.01425644289702177,
-0.06373363733291626,
-0.10127323865890503,
-0.022409310564398766,
0.010712338611483574,
-0.04044733941555023,
-0.07424286007881165,
0.07817048579454422,
0.06014096364378929,
-0.030011121183633804,
0.008509758859872818,
-0.02816818095743656,
-0.04396655783057213,
-0.019004052504897118,
-0.05068591609597206,
-0.04277588054537773,
0.023930078372359276,
0.02514597959816456,
0.05102698132395744,
0.038821734488010406,
-0.0460067056119442,
0.025147823616862297,
-0.04999018460512161,
-0.036949459463357925,
-0.07383529841899872,
0.004796199966222048,
0.038436174392700195,
-0.0180276557803154,
0.018293442204594612,
-0.07855481654405594,
0.06947972625494003,
-0.04040759056806564,
0.03446357324719429,
-0.04321455955505371,
-0.021516896784305573,
-0.0383058600127697,
0.002862424124032259,
-0.024828176945447922,
0.01957009546458721,
0.008495406247675419,
0.02152075804769993,
0.03559035807847977,
-0.012963797897100449,
-0.0024862405844032764,
0.07239768654108047,
-0.09081640094518661,
-0.027126355096697807,
-0.04867652803659439,
-0.026304040104150772,
-0.003734336234629154,
0.01946287229657173,
0.055716291069984436,
0.016299284994602203,
0.057400189340114594,
0.04003758355975151,
0.040756870061159134,
0.05662749335169792,
0.0037128934636712074,
0.03524356707930565,
-0.040998488664627075,
-0.047041624784469604,
0.05056925117969513,
-0.019970089197158813,
-4.892011418178299e-8,
-0.08002523332834244,
0.09506005048751831,
-0.0666903406381607,
0.07683931291103363,
0.011421795934438705,
-0.03833118826150894,
-0.008867845870554447,
0.05963124334812164,
-0.011759424582123756,
-0.029748504981398582,
0.11396200954914093,
-0.013732834719121456,
-0.05606246739625931,
-0.04173337295651436,
0.033860355615615845,
0.0789361521601677,
-0.014077575877308846,
-0.004143639467656612,
0.007591313682496548,
-0.013828079216182232,
0.06856858730316162,
0.10703551024198532,
0.04806647449731827,
0.011650856584310532,
0.003687641816213727,
0.01296754740178585,
-0.014910810627043247,
0.05697263032197952,
0.11129661649465561,
-0.06943240761756897,
-0.019553005695343018,
-0.008818899281322956,
-0.00202360050752759,
-0.06830661743879318,
-0.08907292783260345,
0.10826945304870605,
-0.000047544297558488324,
-0.0588567741215229,
-0.04388678818941116,
0.02152278833091259,
0.07703518867492676,
-0.006976477801799774,
-0.06425772607326508,
0.07254879176616669,
-0.07291560620069504,
0.015978530049324036,
0.06806659698486328,
-0.05631804093718529,
0.07800795137882233,
-0.06500986218452454,
-0.04443524032831192,
-0.052736274898052216,
-0.001521441270597279,
-0.002407446736469865,
0.0625915601849556,
0.0456501804292202,
-0.03551352769136429,
-0.0312507264316082,
0.009529362432658672,
0.005760885309427977,
0.011558739468455315,
0.06438761949539185,
-0.07787863910198212,
0.05487717688083649
] |
Reconstructive Visual Instruction Tuning | https://openreview.net/forum?id=8q9NOMzRDg | [
"Haochen Wang",
"Anlin Zheng",
"Yucheng Zhao",
"Tiancai Wang",
"Zheng Ge",
"Xiangyu Zhang",
"Zhaoxiang Zhang"
] | Poster | This paper introduces reconstructive visual instruction tuning (ROSS), a family of Large Multimodal Models (LMMs) that exploit vision-centric supervision signals. In contrast to conventional visual instruction tuning approaches that exclusively supervise text outputs, ROSS prompts LMMs to supervise visual outputs via reconstructing input images. By doing so, it capitalizes on the inherent richness and detail present within input images themselves, which are often lost in pure text supervision. However, producing meaningful feedback from natural images is challenging due to the heavy spatial redundancy of visual signals. To address this issue, ROSS employs a denoising objective to reconstruct latent representations of input images, avoiding directly regressing exact raw RGB values. This intrinsic activation design inherently encourages LMMs to maintain image detail, thereby enhancing their fine-grained comprehension capabilities and reducing hallucinations. Empirically, ROSS consistently brings significant improvements across different visual encoders and language models. In comparison with extrinsic assistance state-of-the-art alternatives that aggregate multiple visual experts, ROSS delivers competitive performance with a single SigLIP visual encoder, demonstrating the efficacy of our vision-centric supervision tailored for visual outputs. The code will be made publicly available upon acceptance. | Large Multimodal Models, Multimodal Comprehension | We design a vision-centric reconstructive supervision that boosts multimodal comprehension capabilities. | 244 | 2410.09575 | [
0.0468163900077343,
-0.12059292942285538,
0.04068972170352936,
0.010717430151998997,
0.06715170294046402,
-0.0030380822718143463,
-0.0663384422659874,
-0.010046316310763359,
0.012721650302410126,
-0.09652794897556305,
-0.07418894022703171,
-0.061025816947221756,
0.030989887192845345,
0.0466054268181324,
-0.003217388642951846,
0.04182440787553787,
0.04702555388212204,
0.08362771570682526,
-0.0883435383439064,
-0.079575315117836,
0.06910789012908936,
-0.015046299435198307,
0.045671116560697556,
-0.026934096589684486,
0.034164879471063614,
0.04802245646715164,
-0.023996148258447647,
-0.04732099175453186,
0.0037094694562256336,
-0.07694218307733536,
0.0033850325271487236,
0.01919560320675373,
0.0022715802770107985,
0.04189290478825569,
-0.03499994054436684,
0.11099684238433838,
-0.03673179820179939,
0.043553564697504044,
-0.03280284255743027,
-0.05788545310497284,
-0.007902337238192558,
0.0967629998922348,
-0.012638632208108902,
0.03257114440202713,
0.1067219078540802,
-0.05502661317586899,
-0.01041092537343502,
-0.04800250753760338,
-0.004172204993665218,
0.0028704863507300615,
-0.053065571933984756,
-0.008415396325290203,
0.007054484449326992,
0.021664805710315704,
-0.0316438302397728,
0.09242268651723862,
0.09581632912158966,
-0.023447807878255844,
-0.014217077754437923,
-0.0035899141803383827,
-0.041513893753290176,
-0.028932755813002586,
0.015506016090512276,
-0.03840095177292824,
0.015211766585707664,
-0.03653554245829582,
0.04833894595503807,
0.001449620584025979,
-0.010000564157962799,
-0.02865365333855152,
-0.0339069627225399,
-0.014453930780291557,
-0.05200770124793053,
0.019034354016184807,
0.01898001879453659,
0.054810333997011185,
0.0022228704765439034,
-0.04807198420166969,
0.050503458827733994,
-0.06957543641328812,
0.050491753965616226,
0.03832076117396355,
0.035358816385269165,
-0.015579772181808949,
0.10540810972452164,
0.004580845590680838,
0.04622859135270119,
0.018604649230837822,
0.01790257915854454,
0.008831187151372433,
-0.05939647555351257,
-0.09194731712341309,
0.011729782447218895,
0.007427204865962267,
0.08789212256669998,
0.014171506278216839,
0.021696388721466064,
-0.07921460270881653,
-0.034698400646448135,
0.008705473504960537,
0.05773312970995903,
-0.0014066167641431093,
0.007959690876305103,
-0.021556105464696884,
-0.013725939206779003,
-0.027393238618969917,
0.13277478516101837,
0.054543159902095795,
0.0034784101881086826,
-0.07069864869117737,
0.03986577317118645,
0.020428087562322617,
0.0007695247186347842,
0.04592988267540932,
0.03609883412718773,
0.019549177959561348,
0.01955629326403141,
0.014728453941643238,
0.0833098366856575,
0.02262214571237564,
0.03211568668484688,
-0.03094295784831047,
-0.043068256229162216,
-0.04208650812506676,
0.055338162928819656,
-0.04790972173213959,
-0.052163511514663696,
2.8233792899507505e-33,
0.02611909992992878,
0.04934828728437424,
0.04043884575366974,
0.029011467471718788,
0.08314841985702515,
-0.005002864636480808,
-0.0510764978826046,
0.07096710056066513,
-0.055090125650167465,
0.008281688205897808,
-0.012471223250031471,
-0.025584295392036438,
-0.06085342913866043,
0.13957007229328156,
0.02453901246190071,
-0.03835171461105347,
-0.076116643846035,
0.04754342883825302,
-0.03412222862243652,
0.01885928027331829,
-0.04425477981567383,
0.046566661447286606,
-0.05434364080429077,
-0.06851884722709656,
-0.01752266101539135,
0.008138909935951233,
0.06165033206343651,
-0.04532465711236,
-0.0038874391466379166,
-0.03736444190144539,
-0.11380268633365631,
0.04886500909924507,
0.012948541902005672,
0.009224123321473598,
0.003162622917443514,
-0.06953588873147964,
0.06320175528526306,
-0.036904290318489075,
0.037733715027570724,
-0.010684379376471043,
0.04620908573269844,
0.04546467214822769,
-0.02481823042035103,
-0.049698829650878906,
-0.08191751688718796,
0.007109487429261208,
0.02658160962164402,
0.05784889683127403,
-0.08093653619289398,
0.006101361010223627,
0.021422989666461945,
0.05876646935939789,
-0.0054035489447414875,
-0.0949031412601471,
0.02812054380774498,
-0.06779782474040985,
0.07129096239805222,
0.005532542709261179,
0.03904595971107483,
-0.01175531093031168,
0.040773242712020874,
0.07471027970314026,
-0.009748437441885471,
0.12616674602031708,
0.022282173857092857,
-0.025587160140275955,
0.05204265937209129,
0.053956303745508194,
0.018578344956040382,
-0.018590454012155533,
-0.1250504106283188,
0.023688551038503647,
0.01615855284035206,
-0.07743572443723679,
-0.003335515735670924,
-0.04927044361829758,
0.06427113711833954,
-0.024328624829649925,
0.07792861759662628,
0.0235569030046463,
0.03291729465126991,
-0.014425430446863174,
-0.03796173632144928,
-0.065919890999794,
-0.05710506811738014,
-0.055248457938432693,
0.024002712219953537,
-0.08368582278490067,
0.03706273064017296,
-0.03812766447663307,
-0.02427811734378338,
-0.0920608788728714,
-0.009756588377058506,
-0.061376046389341354,
-0.05338618904352188,
-1.6832809107416453e-33,
0.03337009251117706,
0.028794944286346436,
-0.029266677796840668,
0.03007056936621666,
-0.010288096964359283,
-0.046941984444856644,
0.0672508105635643,
-0.012077515944838524,
0.04924400523304939,
-0.10981693863868713,
-0.027063695713877678,
0.016535719856619835,
-0.012283154763281345,
0.012337906286120415,
0.02114613726735115,
0.00011689248640323058,
-0.007035309914499521,
0.005585618782788515,
-0.0621846579015255,
-0.05452696606516838,
-0.0029712840914726257,
0.09104140102863312,
-0.011553457006812096,
0.028637779876589775,
-0.06464274227619171,
0.09077313542366028,
0.0012846012832596898,
0.04825720936059952,
-0.04473581537604332,
-0.044911373406648636,
0.01386093907058239,
-0.0005555697716772556,
0.03166850283741951,
-0.0037028752267360687,
-0.034848008304834366,
0.03799344226717949,
0.07278627157211304,
-0.0752594843506813,
-0.09239181876182556,
0.10293653607368469,
0.07629410177469254,
0.05483086407184601,
-0.012292101047933102,
0.034815870225429535,
-0.010137777775526047,
-0.04822153598070145,
-0.05370518937706947,
-0.06109776347875595,
0.016037369146943092,
0.10358927398920059,
-0.05900859832763672,
0.03100685589015484,
-0.11196031421422958,
-0.03995669633150101,
-0.07347848266363144,
-0.1189638003706932,
-0.05197468027472496,
0.021434089168906212,
0.09526621550321579,
0.009718986228108406,
-0.03668706491589546,
-0.0016481649363413453,
-0.05464960262179375,
-0.0800340473651886,
-0.005175430793315172,
-0.04133172333240509,
-0.0695900246500969,
-0.010593796148896217,
0.0801122635602951,
0.03147828206419945,
0.011795631609857082,
-0.012706318870186806,
-0.00027217177557758987,
0.020166879519820213,
0.035219136625528336,
0.010934152640402317,
0.007293303031474352,
-0.03659447282552719,
-0.04819294810295105,
-0.031562965363264084,
0.055826716125011444,
-0.046292271465063095,
0.014349755831062794,
0.10214440524578094,
0.06526339799165726,
0.09336620569229126,
-0.03703540936112404,
0.0489020012319088,
0.04632740840315819,
-0.016106411814689636,
-0.04558272659778595,
-0.020175235345959663,
0.017233075574040413,
0.06460514664649963,
-0.028355196118354797,
-4.019784682895988e-8,
-0.03511432185769081,
0.020665939897298813,
-0.0018238420598208904,
-0.014768748544156551,
0.004066499415785074,
-0.09977409243583679,
-0.02163519337773323,
0.032802484929561615,
0.021279875189065933,
-0.04191334918141365,
0.07804414629936218,
-0.08876229077577591,
-0.04943966120481491,
-0.11100368201732635,
0.11175911128520966,
0.07076272368431091,
0.0514669194817543,
0.06557808071374893,
0.014849239028990269,
0.005225587170571089,
0.07715538889169693,
0.05921676382422447,
-0.0513748899102211,
0.03165418654680252,
-0.018065953627228737,
0.026606708765029907,
-0.026539819315075874,
0.10026399791240692,
0.05444956570863724,
0.02104428969323635,
0.11063473671674728,
0.019163236021995544,
0.08499301224946976,
-0.04871509596705437,
0.0009399534319527447,
-0.024683557450771332,
0.04204089939594269,
0.0037890514358878136,
0.013009616173803806,
-0.036713216453790665,
-0.033146586269140244,
0.029091639444231987,
-0.030557122081518173,
0.05285436287522316,
-0.03992072865366936,
-0.03755366429686546,
-0.008829159662127495,
-0.1299649178981781,
0.01213394571095705,
-0.024607714265584946,
-0.05767018347978592,
-0.02867661602795124,
0.0016088343691080809,
0.0915508046746254,
0.07233802229166031,
-0.05040130019187927,
0.03078291565179825,
0.04241257533431053,
-0.018308347091078758,
0.10003568232059479,
-0.015413668937981129,
0.05254596471786499,
0.016174357384443283,
-0.009597329422831535
] |
BLEND: Behavior-guided Neural Population Dynamics Modeling via Privileged Knowledge Distillation | https://openreview.net/forum?id=jE5ZbtMtcU | [
"Zhengrui Guo",
"Fangxu Zhou",
"Wei Wu",
"Qichen Sun",
"Lishuang Feng",
"Jinzhuo Wang",
"Hao Chen"
] | Poster | Modeling the nonlinear dynamics of neuronal populations represents a key pursuit in computational neuroscience. Recent research has increasingly focused on jointly modeling neural activity and behavior to unravel their interconnections. Despite significant efforts, these approaches often necessitate either intricate model designs or oversimplified assumptions. Given the frequent absence of perfectly paired neural-behavioral datasets in real-world scenarios when deploying these models, a critical yet understudied research question emerges: how to develop a model that performs well using only neural activity as input at inference, while benefiting from the insights gained from behavioral signals during training?
To this end, we propose **BLEND**, the **B**ehavior-guided neura**L** population dynamics mod**E**lling framework via privileged k**N**owledge **D**istillation. By considering behavior as privileged information, we train a teacher model that takes both behavior observations (privileged features) and neural activities (regular features) as inputs. A student model is then distilled using only neural activity. Unlike existing methods, our framework is model-agnostic and avoids making strong assumptions about the relationship between behavior and neural activity. This allows BLEND to enhance existing neural dynamics modeling architectures without developing specialized models from scratch. Extensive experiments across neural population activity modeling and transcriptomic neuron identity prediction tasks demonstrate strong capabilities of BLEND, reporting over 50% improvement in behavioral decoding and over 15% improvement in transcriptomic neuron identity prediction after behavior-guided distillation. Furthermore, we empirically explore various behavior-guided distillation strategies within the BLEND framework and present a comprehensive analysis of effectiveness and implications for model performance. Code will be made available at https://github.com/dddavid4real/BLEND. | Computational Neuroscience, Neural Dynamics Modeling, Behavior as Guidance, Privileged Knowledge Distillation | null | 241 | 2410.13872 | [
-0.04360972344875336,
-0.1443307250738144,
0.10524692386388779,
0.03799671307206154,
-0.029544807970523834,
0.018665218725800514,
-0.006244027055799961,
-0.04646363854408264,
-0.013420160859823227,
-0.02022918500006199,
0.007609718479216099,
-0.048031777143478394,
0.028359487652778625,
0.07402239739894867,
-0.00561848608776927,
0.009144866839051247,
0.04517770931124687,
0.09305805712938309,
-0.1445319801568985,
0.028567761182785034,
0.002422117395326495,
-0.056035805493593216,
-0.03561081364750862,
0.03510470688343048,
0.01381436362862587,
0.05296984314918518,
0.03586261346936226,
-0.06013457104563713,
0.050474729388952255,
-0.06676959991455078,
0.056434597820043564,
-0.06022762879729271,
0.01298672053962946,
0.06263306736946106,
-0.06051415950059891,
0.025076374411582947,
-0.08429587632417679,
0.010795645415782928,
-0.05311459302902222,
-0.05667699873447418,
-0.0017445151461288333,
0.014321906492114067,
0.002006227383390069,
0.04049728065729141,
0.07705996930599213,
0.03178664669394493,
-0.025094693526625633,
-0.04260871186852455,
0.0005832818569615483,
-0.015078936703503132,
-0.027135370299220085,
-0.0033655145671218634,
0.024350950494408607,
0.0633174404501915,
0.02805415913462639,
0.05928628891706467,
-0.040822625160217285,
-0.03376388922333717,
0.024690672755241394,
0.03256450220942497,
-0.038886163383722305,
-0.03525114431977272,
-0.07193440943956375,
-0.03416634723544121,
-0.0465092658996582,
-0.001693970407359302,
0.003202039049938321,
0.08993836492300034,
0.04248659685254097,
0.02409215271472931,
0.025056391954421997,
0.017955230548977852,
-0.0032656369730830193,
-0.06014834716916084,
0.05094055086374283,
-0.006507071200758219,
0.008028052747249603,
0.004178641829639673,
0.08685386180877686,
-0.11994701623916626,
0.016878746449947357,
0.015825316309928894,
0.020608527585864067,
-0.04961644858121872,
0.09048839658498764,
0.04056628420948982,
0.011885659769177437,
0.059504441916942596,
0.07113941758871078,
0.004696655087172985,
-0.06036040931940079,
-0.008895903825759888,
-0.02266021817922592,
-0.017739344388246536,
0.0370892658829689,
0.036236654967069626,
0.001640691771171987,
-0.054721370339393616,
0.015575182624161243,
0.11192116886377335,
0.0006459425203502178,
0.14559665322303772,
-0.00029817770700901747,
0.011274234391748905,
-0.02875605598092079,
-0.06727530062198639,
0.05104617774486542,
0.05838994309306145,
0.022151468321681023,
-0.028928101062774658,
0.041844286024570465,
0.002512646373361349,
0.0003197219630237669,
0.08352681994438171,
0.04495343565940857,
-0.05090665817260742,
0.06549999862909317,
-0.02151687629520893,
0.017680425196886063,
0.025613002479076385,
0.008332116529345512,
-0.06068628281354904,
-0.027290429919958115,
0.04450557008385658,
0.01446805614978075,
-0.028261104598641396,
-0.1128539890050888,
1.2428122490644913e-33,
0.03947770968079567,
-0.013399395160377026,
0.0011625532060861588,
0.01310587115585804,
0.00891089253127575,
-0.053545355796813965,
0.00026958007947541773,
-0.02286304160952568,
-0.03998757153749466,
0.03202689066529274,
-0.08675112575292587,
0.008785296231508255,
-0.016098514199256897,
0.13565842807292938,
0.00952991470694542,
0.05832352489233017,
-0.05319744348526001,
0.03838405758142471,
-0.0517776757478714,
0.0012968924129381776,
0.09218939393758774,
0.006204245612025261,
-0.029375584796071053,
-0.061600856482982635,
-0.005314144771546125,
-0.054236747324466705,
-0.0075583732686936855,
-0.009876777417957783,
0.011412283405661583,
0.025921549648046494,
-0.04684091731905937,
-0.017294999212026596,
-0.07830099016427994,
0.05062709003686905,
0.01985870860517025,
-0.004265228286385536,
0.04097144678235054,
-0.034933965653181076,
-0.002527510514482856,
-0.03247646614909172,
0.016777966171503067,
0.014820883981883526,
0.027991367504000664,
-0.02490673027932644,
-0.09824023395776749,
-0.032903555780649185,
0.08349381387233734,
0.06103645637631416,
-0.05787248536944389,
-0.05097564309835434,
-0.007833221927285194,
0.00493447994813323,
0.010722563602030277,
-0.1593569815158844,
-0.06921017169952393,
0.03505809232592583,
0.0026897715870290995,
0.03129018098115921,
0.0073667531833052635,
0.02988295443356037,
-0.02011096477508545,
0.017391428351402283,
0.016863400116562843,
0.007452055811882019,
0.07767606526613235,
0.03797367587685585,
-0.13969852030277252,
0.0554770790040493,
0.08276970684528351,
-0.06343931704759598,
-0.03779125586152077,
0.01504658255726099,
-0.035817887634038925,
-0.06601493805646896,
0.04151838645339012,
-0.05357787385582924,
0.08937830477952957,
-0.04989968612790108,
-0.011907391250133514,
0.0065717753022909164,
0.06852822005748749,
-0.08229007571935654,
-0.11848969012498856,
0.005803493317216635,
-0.02324540726840496,
0.01351562887430191,
0.03878709673881531,
-0.09825079888105392,
-0.06985703110694885,
-0.026787888258695602,
-0.023977043107151985,
-0.03472590073943138,
-0.009789466857910156,
-0.002806359203532338,
0.009112135507166386,
-2.2276115696471525e-33,
0.02284429222345352,
0.040597639977931976,
-0.016915589570999146,
0.019634094089269638,
-0.008088368922472,
0.026547083631157875,
0.0202162005007267,
0.03417205065488815,
-0.036289460957050323,
-0.09385056048631668,
0.014403410255908966,
-0.039872851222753525,
0.0689040869474411,
0.023019809275865555,
-0.006950079929083586,
-0.02000485360622406,
-0.07196221500635147,
0.05551578849554062,
0.07152079045772552,
0.056301869451999664,
-0.09329115599393845,
0.037462323904037476,
-0.08381659537553787,
0.012924530543386936,
0.017263473942875862,
0.040702901780605316,
-0.04331647977232933,
0.1058797612786293,
-0.05281602218747139,
0.03950590640306473,
-0.03095732443034649,
-0.011526543647050858,
-0.05853483825922012,
-0.11945073306560516,
-0.08270757645368576,
0.1066896989941597,
0.02256707102060318,
-0.019517775624990463,
-0.04385846108198166,
-0.049411457031965256,
-0.015215583145618439,
-0.006658571772277355,
0.018144192174077034,
0.055402275174856186,
0.03531531244516373,
0.008392863906919956,
-0.06966336816549301,
0.051079247146844864,
-0.061485420912504196,
0.019551638513803482,
0.05438804253935814,
0.0592021569609642,
-0.08712923526763916,
-0.009002975188195705,
-0.05778272822499275,
-0.020048838108778,
0.08285582810640335,
-0.048114120960235596,
-0.010313184931874275,
-0.007094960659742355,
-0.05449962988495827,
-0.07806432247161865,
-0.062322892248630524,
0.01602238230407238,
-0.0193473007529974,
-0.02997502125799656,
-0.026423942297697067,
0.07711859792470932,
0.028981592506170273,
-0.036053288727998734,
-0.004952550865709782,
-0.015035178512334824,
0.05895008146762848,
-0.07909849286079407,
0.056270696222782135,
-0.04971921443939209,
0.019368426874279976,
-0.10130124539136887,
-0.004370962735265493,
0.013389074243605137,
-0.0053809406235814095,
-0.025262385606765747,
0.01150659378618002,
-0.00899689644575119,
0.004180461168289185,
0.05883214995265007,
0.0425553061068058,
0.04286815598607063,
-0.030258366838097572,
0.008308952674269676,
-0.06527026742696762,
0.008228585124015808,
-0.023115096613764763,
0.0673450231552124,
-0.06907949596643448,
-4.927569108303942e-8,
-0.044167373329401016,
0.1156601533293724,
0.05675541236996651,
0.07033518701791763,
0.06384708732366562,
-0.055318813771009445,
0.04813592508435249,
0.0014317099703475833,
-0.05118416249752045,
0.00577900093048811,
0.11215851455926895,
-0.0038759573362767696,
0.007407235912978649,
-0.029471388086676598,
0.06096864864230156,
0.09520260989665985,
0.0652858316898346,
0.036990076303482056,
-0.05191435292363167,
0.01756606437265873,
0.10283299535512924,
-0.03329112380743027,
0.05677115544676781,
0.020170263946056366,
0.10707918554544449,
-0.05242462456226349,
-0.036186788231134415,
0.08210819214582443,
0.022784003987908363,
0.005170770455151796,
-0.014914670027792454,
0.0038129601161926985,
0.03267880901694298,
-0.02092088758945465,
0.011705677025020123,
0.04634115844964981,
0.03859815001487732,
-0.06873635947704315,
-0.037520237267017365,
-0.024946197867393494,
-0.0021742803510278463,
0.042448315769433975,
-0.03214951977133751,
0.0508204847574234,
0.0100023802369833,
0.013543220236897469,
0.02632695436477661,
-0.14327506721019745,
0.08646794408559799,
0.09203651547431946,
0.01710236445069313,
-0.009619706310331821,
0.020720113068819046,
0.04950197786092758,
0.03790252283215523,
0.03446616977453232,
-0.05932524800300598,
-0.019198821857571602,
0.05030495673418045,
0.020795144140720367,
0.0009037487325258553,
0.06709590554237366,
-0.09728120267391205,
-0.01711903139948845
] |
COMBO: Compositional World Models for Embodied Multi-Agent Cooperation | https://openreview.net/forum?id=YXRyYkb1im | [
"Hongxin Zhang",
"Zeyuan Wang",
"Qiushi Lyu",
"Zheyuan Zhang",
"Sunli Chen",
"Tianmin Shu",
"Behzad Dariush",
"Kwonjoon Lee",
"Yilun Du",
"Chuang Gan"
] | Poster | In this paper, we investigate the problem of embodied multi-agent cooperation, where decentralized agents must cooperate given only egocentric views of the world. To effectively plan in this setting, in contrast to learning world dynamics in a single-agent scenario, we must simulate world dynamics conditioned on an arbitrary number of agents' actions given only partial egocentric visual observations of the world. To address this issue of partial observability, we first train generative models to estimate the overall world state given partial egocentric observations. To enable accurate simulation of multiple sets of actions on this world state, we then propose to learn a compositional world model for multi-agent cooperation by factorizing the naturally composable joint actions of multiple agents and compositionally generating the video conditioned on the world state. By leveraging this compositional world model, in combination with Vision Language Models to infer the actions of other agents, we can use a tree search procedure to integrate these modules and facilitate online cooperative planning. We evaluate our methods on three challenging benchmarks with 2-4 agents. The results show our compositional world model is effective and the framework enables the embodied agents to cooperate efficiently with different agents across various tasks and an arbitrary number of agents, showing the promising future of our proposed methods. More videos can be found at \url{https://embodied-agi.cs.umass.edu/combo/}. | Embodied AI; Multi-agent Planning; Compositional World Model | We learn a compositional world model for multi-agent cooperation by factorizing the joint actions of multiple agents and compositionally generating the video. In combination with VLMs, a tree search procedure enables online cooperative planning. | 238 | 2404.10775 | [
0.04787420853972435,
-0.12421083450317383,
0.031078245490789413,
-0.031587131321430206,
0.0651206225156784,
0.01952383667230606,
0.006260995287448168,
-0.04598339647054672,
0.02308168075978756,
-0.01415137480944395,
-0.004556484520435333,
-0.09444078803062439,
0.02281491458415985,
0.09764669090509415,
0.019046464934945107,
0.01144876517355442,
0.10179277509450912,
0.08769004791975021,
-0.09970254451036453,
-0.03933168202638626,
-0.0015249507268890738,
-0.03728405013680458,
0.0368218868970871,
-0.010669293813407421,
0.0059687611646950245,
0.00928697269409895,
0.01545128133147955,
-0.09532469511032104,
0.020045556128025055,
-0.011193429119884968,
0.04057934507727623,
-0.008399429731070995,
-0.012500773184001446,
-0.007406565360724926,
-0.09703956544399261,
0.07980380952358246,
-0.05827796831727028,
-0.022638779133558273,
-0.017260083928704262,
-0.08282588422298431,
0.01449884194880724,
0.05462677776813507,
0.024000979959964752,
-0.051786262542009354,
0.0025423523038625717,
0.062356650829315186,
-0.047106023877859116,
-0.02214195393025875,
0.02062763087451458,
-0.00019301078282296658,
-0.10557303577661514,
-0.09548752009868622,
-0.025295719504356384,
-0.00655472744256258,
0.009043162688612938,
-0.0028625898994505405,
0.0057596140541136265,
-0.06667760759592056,
0.07042375206947327,
-0.03316103667020798,
0.020437637344002724,
-0.055423304438591,
0.024754442274570465,
0.015802763402462006,
0.015737324953079224,
-0.0708397924900055,
0.060493551194667816,
0.052376702427864075,
-0.061941638588905334,
-0.0286695659160614,
-0.013446014374494553,
-0.04305991902947426,
-0.0021880532149225473,
-0.08952149003744125,
0.05933757126331329,
0.009128566831350327,
-0.0023214868269860744,
-0.10728409886360168,
0.03755289688706398,
-0.10257502645254135,
0.016480350866913795,
0.08440463989973068,
0.009585065767168999,
0.01739652082324028,
0.09575437009334564,
-0.004473117645829916,
-0.021060941740870476,
0.10471105575561523,
0.03592761978507042,
-0.0035797487944364548,
-0.11451181769371033,
0.034560427069664,
0.019030863419175148,
-0.01788230426609516,
0.012743903324007988,
0.045719295740127563,
-0.020642079412937164,
-0.07461178302764893,
-0.0013531680451706052,
0.07575293630361557,
-0.02971736714243889,
0.030959751456975937,
0.05448593944311142,
-0.059910908341407776,
0.002251124009490013,
0.02367980405688286,
0.01262563094496727,
0.01985880173742771,
0.01822151243686676,
-0.05126743018627167,
0.004539798013865948,
-0.02804991602897644,
-0.0014697759179398417,
0.041143834590911865,
-0.019848104566335678,
-0.004049865063279867,
0.07599176466464996,
-0.016037534922361374,
-0.001920136739499867,
0.007415484171360731,
0.06521546095609665,
-0.036321330815553665,
0.11252624541521072,
0.036956459283828735,
0.09528027474880219,
-0.050193339586257935,
0.0482943058013916,
-2.1685354480076476e-33,
0.11272844672203064,
-0.002149026608094573,
0.04370596259832382,
0.10442084819078445,
0.10080739855766296,
0.03453870117664337,
0.026275794953107834,
0.029863711446523666,
-0.04148795083165169,
0.01792723499238491,
0.011319762095808983,
-0.01087371539324522,
-0.01749015972018242,
0.13009852170944214,
0.0008988612680695951,
-0.026632079854607582,
-0.0028841460589319468,
0.01482726912945509,
-0.016563300043344498,
0.0507659912109375,
0.07687725126743317,
0.051276590675115585,
-0.017097625881433487,
-0.03719892352819443,
-0.028779050335288048,
0.013407548889517784,
0.046131063252687454,
-0.00005269235771265812,
0.03399686515331268,
-0.010294744744896889,
-0.01953437551856041,
0.004306304734200239,
-0.0922553539276123,
0.07545983046293259,
0.030727213248610497,
0.010455816052854061,
-0.08658445626497269,
-0.034189026802778244,
-0.042016737163066864,
-0.03034159354865551,
0.004968483466655016,
0.04074348509311676,
-0.06505132466554642,
-0.037886664271354675,
-0.06406664848327637,
-0.08462563902139664,
0.02375439926981926,
0.10856029391288757,
-0.031214486807584763,
-0.01262901071459055,
-0.007944367825984955,
0.05195692181587219,
0.0325915701687336,
-0.12229283899068832,
-0.05018902197480202,
0.06325313448905945,
0.03203856199979782,
0.06117912009358406,
0.02584495022892952,
-0.023376114666461945,
-0.049956947565078735,
0.027484415099024773,
0.00031227461295202374,
0.08611708134412766,
0.03955184668302536,
0.058016661554574966,
-0.04185599461197853,
0.03764154389500618,
0.09292244911193848,
-0.011275170370936394,
-0.00750192254781723,
0.057155411690473557,
0.010432546958327293,
-0.05562906712293625,
0.049264512956142426,
-0.00612297048792243,
0.06545022875070572,
-0.11148529499769211,
-0.04392317682504654,
0.10902200639247894,
-0.05750548094511032,
0.023728320375084877,
0.024639122188091278,
-0.016712211072444916,
0.022803666070103645,
-0.032076429575681686,
0.01750575564801693,
-0.02673248201608658,
-0.03201409429311752,
0.045449841767549515,
-0.06993036717176437,
-0.028082583099603653,
-0.05310722813010216,
0.0257515050470829,
0.024433167651295662,
-6.22923442079205e-35,
-0.012628896161913872,
-0.014175365678966045,
0.03963921219110489,
0.04175427556037903,
0.004821307957172394,
-0.10580769181251526,
-0.059188853949308395,
-0.023663705214858055,
-0.05156811699271202,
-0.03126152232289314,
-0.041620805859565735,
0.03654227405786514,
0.05012039840221405,
0.005622330587357283,
0.026140255853533745,
0.005777816288173199,
0.043493591248989105,
-0.020161041989922523,
-0.03486824780702591,
0.014916935004293919,
0.021692976355552673,
0.07830227166414261,
-0.05990270525217056,
0.016299018636345863,
-0.03223753720521927,
0.08508793264627457,
-0.011482419446110725,
0.034821730107069016,
0.010163603350520134,
0.0841292217373848,
0.0030784369446337223,
-0.08518005907535553,
-0.030584579333662987,
-0.04171008616685867,
-0.088051438331604,
0.047480568289756775,
-0.043291542679071426,
0.05245120823383331,
-0.09257560223340988,
0.07568752020597458,
0.02316810004413128,
-0.07405073940753937,
-0.09753530472517014,
0.031047556549310684,
-0.05022873729467392,
-0.03928322345018387,
-0.027275366708636284,
0.0031651807948946953,
-0.032771605998277664,
0.02807268686592579,
-0.05936263129115105,
0.04861362650990486,
-0.09841100126504898,
-0.039694517850875854,
-0.019212866201996803,
-0.05535193905234337,
-0.000007963457392179407,
0.03344331309199333,
0.14585350453853607,
-0.10722483694553375,
-0.04028228670358658,
-0.01230890303850174,
0.024076806381344795,
0.04189153388142586,
-0.002448238432407379,
-0.02629641816020012,
-0.051967136561870575,
0.01758803240954876,
0.06181845813989639,
-0.001588943530805409,
0.005783035885542631,
0.03476998209953308,
-0.03774905204772949,
0.002226622076705098,
0.10392570495605469,
0.05364758521318436,
0.035198237746953964,
-0.014520159922540188,
0.0468275211751461,
0.009436863474547863,
0.006598134990781546,
-0.02214806154370308,
0.04062613099813461,
0.02839859388768673,
0.01247867289930582,
0.05040193349123001,
0.02259148843586445,
0.11691287904977798,
-0.017070716246962547,
0.020949430763721466,
-0.018873512744903564,
0.044232986867427826,
0.007431436330080032,
0.026458872482180595,
-0.07546168565750122,
-3.9833885523421486e-8,
-0.10688088834285736,
0.011928725987672806,
0.07412733882665634,
-0.0004074759781360626,
-0.04545135423541069,
-0.013096853159368038,
0.01948961243033409,
-0.04403095319867134,
-0.006529606878757477,
-0.0015614276053383946,
0.06721930205821991,
-0.0015232287114486098,
0.017772531136870384,
-0.05156214162707329,
-0.015023257583379745,
-0.010017712600529194,
-0.020171985030174255,
-0.04373230040073395,
-0.03573659062385559,
0.06508474051952362,
-0.0318586528301239,
-0.026408154517412186,
0.0055108442902565,
-0.01505021657794714,
-0.051260221749544144,
0.013397383503615856,
-0.04738223925232887,
0.018061785027384758,
0.0524587407708168,
0.009492986835539341,
0.038272224366664886,
0.01906122826039791,
0.005866663996130228,
-0.0769982561469078,
-0.07116983830928802,
0.011023621074855328,
-0.0000839423737488687,
-0.06226198375225067,
-0.01599961519241333,
-0.05775901675224304,
0.030305905267596245,
0.0622810535132885,
-0.012201623991131783,
0.007317512761801481,
0.07554513961076736,
0.027711650356650352,
-0.03150574117898941,
-0.1818585842847824,
0.036860737949609756,
-0.012893031351268291,
-0.08341392874717712,
-0.059577953070402145,
-0.07497967779636383,
0.08431036025285721,
0.042404141277074814,
-0.03248436748981476,
0.041341278702020645,
-0.07791220396757126,
0.11575477570295334,
0.04731256142258644,
-0.0218595489859581,
-0.008674022741615772,
-0.08536239713430405,
-0.041836779564619064
] |
Image-level Memorization Detection via Inversion-based Inference Perturbation | https://openreview.net/forum?id=vwOq7twk7L | [
"Yue Jiang",
"Haokun Lin",
"Yang Bai",
"Bo Peng",
"Zhili Liu",
"Yueming Lyu",
"Yong Yang",
"Xingzheng",
"Jing Dong"
] | Poster | Recent studies have discovered that widely used text-to-image diffusion models can replicate training samples during image generation, a phenomenon known as memorization. Existing detection methods primarily focus on identifying memorized prompts. However, in real-world scenarios, image owners may need to verify whether their proprietary or personal images have been memorized by the model, even in the absence of paired prompts or related metadata. We refer to this challenge as image-level memorization detection, where current methods relying on original prompts fall short. In this work, we uncover two characteristics of memorized images after perturbing the inference procedure: lower similarity of the original images and larger magnitudes of TCNP.
Building on these insights, we propose Inversion-based Inference Perturbation (IIP), a new framework for image-level memorization detection. Our approach uses unconditional DDIM inversion to derive latent codes that contain core semantic information of original images and optimizes random prompt embeddings to introduce effective perturbation. Memorized images exhibit distinct characteristics within the proposed pipeline, providing a robust basis for detection. To support this task, we construct a comprehensive setup for the image-level memorization detection, carefully curating datasets to simulate realistic memorization scenarios. Using this setup, we evaluate our IIP framework across three different memorization settings, demonstrating its state-of-the-art performance in identifying memorized images in various settings, even in the presence of data augmentation attacks. | Text-to-image diffusion model, data memorization detection, DDIM Inversion | we propose a simple yet effective image-level memorization detection method, namely Inversion-based Inference Perturbation (IIP). | 233 | null | [
-0.03829881176352501,
-0.07742072641849518,
0.009454324841499329,
0.03364749997854233,
0.11860799789428711,
-0.05498914793133736,
0.018265048041939735,
-0.04970491677522659,
0.04999364912509918,
-0.059274978935718536,
0.07096246629953384,
0.027668504044413567,
0.06625758111476898,
0.05744602158665657,
-0.029005007818341255,
0.0033416112419217825,
0.023064134642481804,
0.0983656644821167,
-0.06785570830106735,
-0.06104324012994766,
-0.060483381152153015,
-0.0045815445482730865,
0.0925949290394783,
-0.004749822895973921,
0.02850372903048992,
0.12159988284111023,
0.018876798450946808,
-0.016573216766119003,
-0.0006457488634623587,
-0.006397831719368696,
0.04396801069378853,
0.017427543178200722,
0.015826936811208725,
0.07530844211578369,
-0.020763985812664032,
0.04113864153623581,
-0.07539614289999008,
0.005026268772780895,
-0.010892636142671108,
-0.027224382385611534,
0.005637016613036394,
0.0031154954340308905,
0.0023950347676873207,
0.05571800097823143,
0.08006877452135086,
-0.00054793746676296,
-0.03609360381960869,
-0.05313714221119881,
-0.020084531977772713,
0.03957753628492355,
0.01017000898718834,
0.04124685749411583,
-0.09632816165685654,
0.02980680763721466,
-0.055071476846933365,
0.0016239379765465856,
0.07555758953094482,
0.013289562426507473,
-0.03321920335292816,
0.09934967011213303,
-0.029906122013926506,
-0.11523222923278809,
-0.017480796203017235,
-0.03993857279419899,
-0.0011666849022731185,
0.014377645216882229,
0.042224377393722534,
0.0469977892935276,
-0.012726170010864735,
-0.009659347124397755,
-0.041319675743579865,
0.09926209598779678,
-0.10550715029239655,
0.027684928849339485,
-0.01442612987011671,
0.021292567253112793,
-0.019234027713537216,
-0.042135369032621384,
0.0531960129737854,
-0.14046545326709747,
0.03682808578014374,
-0.036088600754737854,
0.101472407579422,
-0.02821158990263939,
0.10767392814159393,
0.023580675944685936,
0.047141820192337036,
0.07465580105781555,
-0.029195738956332207,
-0.0081314817070961,
-0.052193813025951385,
-0.04797894507646561,
-0.06632602959871292,
-0.013033919967710972,
0.030204370617866516,
0.04580547288060188,
-0.0036609929520636797,
-0.06196243315935135,
0.010609254240989685,
0.09792450070381165,
-0.02128709852695465,
0.0386849008500576,
-0.04656439647078514,
-0.006254022475332022,
0.044650495052337646,
-0.01597125269472599,
0.0839277058839798,
0.01891942135989666,
0.05240928754210472,
-0.03913222253322601,
0.06459391862154007,
0.0132610397413373,
0.03656471520662308,
0.06831884384155273,
0.06944530457258224,
-0.11169803142547607,
0.07720020413398743,
-0.011684962548315525,
0.008091765455901623,
0.019437747076153755,
-0.031944774091243744,
-0.005603511352092028,
-0.05727149173617363,
-0.021112171933054924,
-0.02776893600821495,
-0.05383610725402832,
-0.012382600456476212,
3.178997371472194e-33,
0.05382084101438522,
0.030573464930057526,
0.07890672981739044,
0.05864093825221062,
0.0484766848385334,
0.013732556253671646,
-0.023020491003990173,
-0.016099821776151657,
-0.06580580770969391,
-0.03781568259000778,
0.029149619862437248,
-0.02862388826906681,
-0.053539421409368515,
0.06579970568418503,
0.006462350953370333,
-0.05696100741624832,
-0.0797174945473671,
0.022628212347626686,
0.021775346249341965,
-0.038370441645383835,
0.05273967236280441,
-0.017409255728125572,
-0.017152419313788414,
-0.042572952806949615,
0.0049033937975764275,
-0.08326001465320587,
0.06506188213825226,
-0.03451068699359894,
0.0393730103969574,
0.0017385632963851094,
-0.12541280686855316,
-0.004198242910206318,
-0.0010199605021625757,
0.0212097205221653,
0.02840787172317505,
-0.02691779099404812,
0.03410522639751434,
-0.06598823517560959,
-0.013096841052174568,
-0.04329167306423187,
-0.01718626730144024,
0.017408890649676323,
-0.04608343169093132,
-0.03881869092583656,
-0.1724139302968979,
-0.02327488735318184,
0.022501077502965927,
0.052410755306482315,
-0.0050953226163983345,
-0.03835827857255936,
0.02100236341357231,
0.009914295747876167,
-0.024349600076675415,
-0.0971641018986702,
-0.03177407383918762,
-0.08342832326889038,
0.01587398163974285,
-0.049541622400283813,
0.04758310317993164,
0.0010617920197546482,
0.07968883216381073,
0.02620805613696575,
-0.04011378809809685,
0.07227052003145218,
-0.028575943782925606,
-0.02730594202876091,
-0.01697828434407711,
0.05069863051176071,
0.020529940724372864,
0.05138013884425163,
0.00474016647785902,
0.0014364714734256268,
-0.0958896353840828,
-0.07561689615249634,
0.07387374341487885,
-0.009071397595107555,
0.027844004333019257,
-0.004574866034090519,
0.02785342000424862,
0.02672433666884899,
-0.036831293255090714,
-0.09120139479637146,
0.04954683780670166,
-0.04801739752292633,
-0.08929362148046494,
0.00916111096739769,
0.040775906294584274,
-0.15115676820278168,
-0.0736699104309082,
0.014545031823217869,
-0.018352320417761803,
-0.019857438281178474,
-0.007965534925460815,
0.007232792675495148,
0.03255881369113922,
-2.6203297222484858e-33,
-0.008535689674317837,
0.06602045893669128,
0.002105481456965208,
0.03561526536941528,
-0.046866655349731445,
-0.06229047104716301,
0.07357566058635712,
0.07840991020202637,
0.013846619054675102,
-0.10733316838741302,
0.014757860451936722,
0.03382795304059982,
-0.04009750857949257,
-0.020823663100600243,
0.0004409952671267092,
-0.04341046139597893,
-0.001427054638043046,
0.03735538199543953,
-0.04119303077459335,
0.011040077544748783,
-0.027552472427487373,
0.05254485085606575,
-0.0023811019491404295,
0.052089352160692215,
-0.049346257001161575,
0.09553806483745575,
-0.0017138137482106686,
0.034686218947172165,
0.02045929618179798,
-0.055355675518512726,
-0.034813292324543,
0.06167755275964737,
-0.0040805889293551445,
0.0014499766984954476,
-0.07501470297574997,
0.03514840453863144,
0.06424108892679214,
0.00588851748034358,
-0.0500233955681324,
0.1178443506360054,
0.039470624178647995,
-0.023483309894800186,
-0.06813947856426239,
0.02485605701804161,
-0.02007470838725567,
-0.13408859074115753,
-0.03338165953755379,
0.049739908427000046,
0.06374572962522507,
0.00950610637664795,
-0.014846544712781906,
0.04462122544646263,
-0.07972173392772675,
0.05096510052680969,
-0.05688697472214699,
-0.04271289333701134,
-0.07626395672559738,
-0.030095433816313744,
0.07416731119155884,
0.003088252851739526,
0.006418505217880011,
-0.049068957567214966,
-0.012105055153369904,
-0.1053103432059288,
-0.028701908886432648,
-0.07808878272771835,
-0.0028176731429994106,
0.0024066150654107332,
0.00450929207727313,
0.0808672308921814,
0.0206539835780859,
-0.01117631234228611,
0.003774154232814908,
0.011276296339929104,
0.03653749078512192,
-0.054000772535800934,
0.019756190478801727,
0.020579874515533447,
-0.005273174960166216,
-0.056591033935546875,
0.05445575341582298,
0.030130676925182343,
-0.0031698469538241625,
0.07253927737474442,
0.004197974223643541,
0.043650124222040176,
0.014344907365739346,
-0.018005698919296265,
-0.006781819276511669,
0.04775913432240486,
-0.04750305041670799,
-0.005498882848769426,
0.0007307790801860392,
0.06678258627653122,
-0.03209361433982849,
-4.141070064633823e-8,
-0.11037851870059967,
0.003112110309302807,
0.03189468756318092,
-0.0005108917248435318,
0.04666278138756752,
0.021321870386600494,
0.023397142067551613,
0.0748535767197609,
0.055222269147634506,
-0.08569666743278503,
0.08815058320760727,
-0.01157736498862505,
-0.05362844094634056,
-0.08370497077703476,
-0.030792349949479103,
0.11699081212282181,
0.023669539019465446,
-0.004400934558361769,
0.008793029002845287,
-0.009914779104292393,
-0.07180222123861313,
0.025808727368712425,
0.05755816027522087,
-0.02982216142117977,
-0.012400571256875992,
-0.010746853426098824,
0.01835261844098568,
0.10408896952867508,
0.09306018799543381,
-0.03239898383617401,
0.01441413164138794,
-0.029357073828577995,
0.025003328919410706,
-0.020694484934210777,
0.041969992220401764,
0.07980015873908997,
0.07200737297534943,
-0.054183900356292725,
-0.018545513972640038,
-0.0971788838505745,
-0.01857980340719223,
-0.01829264871776104,
-0.012252621352672577,
0.05401758849620819,
0.040591467171907425,
0.00848421547561884,
0.0845458135008812,
-0.03529142960906029,
-0.02230888046324253,
-0.08451544493436813,
0.04015219584107399,
-0.026350755244493484,
0.0038134281057864428,
0.12833818793296814,
0.011261329986155033,
-0.03508802130818367,
0.022191453725099564,
0.020458931103348732,
0.036569107323884964,
0.08427704870700836,
0.06126418337225914,
0.026687536388635635,
0.010945800691843033,
-0.01982320472598076
] |
SaRA: High-Efficient Diffusion Model Fine-tuning with Progressive Sparse Low-Rank Adaptation | https://openreview.net/forum?id=wGVOxplEbf | [
"Teng Hu",
"Jiangning Zhang",
"Ran Yi",
"Hongrui Huang",
"Yabiao Wang",
"Lizhuang Ma"
] | Poster | The development of diffusion models has led to significant progress in image and video generation tasks, with pre-trained models like the Stable Diffusion series playing a crucial role.
However, a key challenge remains in downstream task applications: how to effectively and efficiently adapt pre-trained diffusion models to new tasks.
Inspired by model pruning which lightens large pre-trained models by removing unimportant parameters, we propose a novel model fine-tuning method to make full use of these ineffective parameters and enable the pre-trained model with new task-specified capabilities.
In this work, we first investigate the importance of parameters in pre-trained diffusion models and discover that parameters with the smallest absolute values do not contribute to the generation process due to training instabilities.
Based on this observation, we propose a fine-tuning method termed SaRA that re-utilizes these temporarily ineffective parameters, equating to optimizing a sparse weight matrix to learn the task-specific knowledge.
To mitigate potential overfitting, we propose a nuclear-norm-based low-rank sparse training scheme for efficient fine-tuning.
Furthermore, we design a new progressive parameter adjustment strategy to make full use of the finetuned parameters.
Finally, we propose a novel unstructural backpropagation strategy, which significantly reduces memory costs during fine-tuning.
Our method enhances the generative capabilities of pre-trained models in downstream applications and outperforms existing fine-tuning methods in maintaining model's generalization ability. Source code is available at https://sjtuplayer.github.io/projects/SaRA. | Diffusion Model, Fine-tuning | null | 229 | 2409.06633 | [
-0.004624522291123867,
-0.10197281092405319,
-0.012160707265138626,
0.06280233711004257,
0.05855255946516991,
0.0006234790198504925,
-0.03730176016688347,
-0.04799044877290726,
0.0034701083786785603,
-0.026104046031832695,
0.005476509686559439,
0.015064435079693794,
-0.0036356139462441206,
0.0589720755815506,
-0.061396412551403046,
-0.03494244068861008,
0.05188213288784027,
0.06541583687067032,
-0.10650192946195602,
-0.035939715802669525,
-0.01323711033910513,
-0.08524466305971146,
0.043314721435308456,
-0.02663448452949524,
0.013956788927316666,
-0.004726296290755272,
0.034303050488233566,
-0.06478825211524963,
0.02982906810939312,
-0.049001794308423996,
0.018247151747345924,
-0.017554335296154022,
-0.012052665464580059,
0.0022758229169994593,
-0.03536976873874664,
0.05119898542761803,
-0.0519503578543663,
0.03770563751459122,
-0.0047327810898423195,
-0.03369031101465225,
0.026595454663038254,
0.02122301422059536,
-0.005113997031003237,
0.04501854255795479,
0.035409700125455856,
0.007900003343820572,
0.0069986311718821526,
-0.07969150692224503,
0.03360738605260849,
-0.009490416385233402,
-0.029470764100551605,
-0.008516634814441204,
-0.0697794258594513,
0.041437502950429916,
0.007105809170752764,
0.01924583688378334,
0.04206777736544609,
0.0037401376757770777,
0.03160252422094345,
-0.046290427446365356,
-0.036088064312934875,
-0.12782874703407288,
-0.005175671074539423,
-0.03732411935925484,
0.00521803880110383,
-0.009719009511172771,
0.05425914376974106,
0.030343983322381973,
-0.03244910389184952,
-0.009525567293167114,
-0.046783626079559326,
0.08265054225921631,
-0.03972090408205986,
0.03789985552430153,
-0.01682381145656109,
0.0017706135986372828,
0.05399574711918831,
-0.02064843662083149,
-0.003335404908284545,
-0.00023958533711265773,
0.061498355120420456,
0.01875923201441765,
0.036592140793800354,
-0.04584348574280739,
0.11193767935037613,
0.014511785469949245,
-0.0073926146142184734,
0.02302146516740322,
0.05206627771258354,
-0.010736774653196335,
-0.06112399324774742,
0.05061611533164978,
-0.07415267080068588,
0.008485350757837296,
-0.02513911761343479,
0.051055584102869034,
-0.021101515740156174,
-0.0712149515748024,
-0.03884981945157051,
0.10419930517673492,
0.04627058655023575,
-0.016361117362976074,
0.0028589172288775444,
0.062462761998176575,
0.05172253027558327,
-0.024929121136665344,
0.05722372978925705,
0.07320212572813034,
0.017188098281621933,
-0.03632234036922455,
0.06416884809732437,
0.03796302154660225,
0.011652877554297447,
0.05836102366447449,
0.015086502768099308,
-0.02778809890151024,
0.013562318868935108,
-0.02350909821689129,
-0.02000809647142887,
0.08075560629367828,
0.03850047290325165,
-0.01343077328056097,
-0.0459185466170311,
-0.022420044988393784,
0.015972968190908432,
-0.017143549397587776,
-0.0515441857278347,
2.2766545255781526e-34,
-0.003921108786016703,
0.03350141644477844,
0.08887488394975662,
-0.010878476314246655,
0.031106851994991302,
-0.029142921790480614,
0.06888924539089203,
0.011224400252103806,
-0.048709336668252945,
-0.0682259351015091,
-0.04945943132042885,
-0.05291764438152313,
-0.12423167377710342,
0.09917530417442322,
-0.011468720622360706,
-0.06297429651021957,
0.004916965961456299,
0.07169967144727707,
0.04450342431664467,
-0.017358001321554184,
0.04119707643985748,
-0.07207851856946945,
-0.022602934390306473,
-0.07106564939022064,
0.003602556884288788,
-0.010751385241746902,
0.09560774266719818,
-0.09864989668130875,
-0.01941986009478569,
0.01226998120546341,
-0.12503647804260254,
-0.03538494184613228,
0.002608382375910878,
0.02364846132695675,
0.0203896164894104,
-0.0333656370639801,
-0.05544712021946907,
0.0005142676527611911,
0.041820183396339417,
-0.027163660153746605,
-0.005656244233250618,
0.07317036390304565,
-0.0007364754565060139,
-0.012186489999294281,
-0.10600513219833374,
-0.03634984418749809,
0.06333336234092712,
0.03666340187191963,
-0.05340718850493431,
-0.012706881389021873,
0.012262815609574318,
-0.03915490582585335,
-0.012738817371428013,
-0.08141884952783585,
-0.05881797522306442,
-0.02128482051193714,
0.10166101157665253,
-0.054191820323467255,
-0.025607407093048096,
-0.025386299937963486,
0.08346147835254669,
0.02104390412569046,
-0.0998183935880661,
0.08924537897109985,
0.060952357947826385,
-0.06466525793075562,
0.0008956400561146438,
0.09222954511642456,
0.04175494983792305,
-0.0028533218428492546,
-0.033861130475997925,
0.010399208404123783,
0.020441396161913872,
-0.10390900075435638,
0.08581546694040298,
-0.01687650941312313,
0.10611946135759354,
-0.010434653609991074,
-0.025736648589372635,
0.03554350882768631,
-0.00356836779974401,
0.006283993367105722,
0.00783631019294262,
-0.13188676536083221,
-0.03694542124867439,
0.005564521532505751,
0.01626930758357048,
-0.05098315328359604,
-0.026879403740167618,
-0.077816441655159,
-0.003018657909706235,
-0.013033676892518997,
-0.06933724880218506,
-0.003367957891896367,
-0.009857484139502048,
1.003644607242754e-33,
0.027835369110107422,
0.06503771990537643,
0.0033887254539877176,
0.04809578135609627,
0.034347813576459885,
0.02649839222431183,
0.008527779020369053,
0.03138225898146629,
0.010233107022941113,
-0.0772915631532669,
-0.029431307688355446,
-0.07657050341367722,
-0.09768617153167725,
-0.07843462377786636,
0.03959237039089203,
-0.021800361573696136,
-0.08318211883306503,
0.013725470751523972,
-0.024632181972265244,
0.0019426060607656837,
-0.0519489049911499,
0.09566876292228699,
-0.06233232840895653,
0.0666079968214035,
-0.05895989015698433,
0.010304151102900505,
-0.004985925275832415,
0.11360703408718109,
-0.016926726326346397,
-0.01554456539452076,
-0.036439355462789536,
-0.005455103237181902,
0.018349161371588707,
-0.0008877734653651714,
-0.057517919689416885,
0.09352502971887589,
0.03948570042848587,
0.010032908990979195,
-0.023932915180921555,
0.18689970672130585,
0.033620644360780716,
0.01343533955514431,
-0.08847326040267944,
0.034656230360269547,
-0.027386488392949104,
-0.027266239747405052,
-0.03751577436923981,
-0.05966872721910477,
-0.024409977719187737,
0.03850873187184334,
0.01224337238818407,
0.05055582895874977,
-0.06886123865842819,
0.06254586577415466,
-0.016100957989692688,
-0.06925215572118759,
-0.06876317411661148,
0.0028066281229257584,
0.0953335091471672,
-0.04240274429321289,
0.027510255575180054,
-0.026045195758342743,
-0.017468465492129326,
-0.1003398448228836,
-0.0274299755692482,
-0.017344266176223755,
-0.03535902500152588,
0.035417743027210236,
-0.002526288153603673,
0.1278027892112732,
-0.07771485298871994,
0.04262266308069229,
0.049503061920404434,
-0.028177503496408463,
0.00883140042424202,
-0.08022940903902054,
-0.009800326079130173,
-0.05390526354312897,
0.019133467227220535,
-0.03500601649284363,
-0.037029266357421875,
0.008037767373025417,
-0.009725870564579964,
0.024299394339323044,
0.06195443123579025,
0.1420806497335434,
0.08362693339586258,
0.06468924880027771,
0.022250913083553314,
0.008060467429459095,
-0.02545214258134365,
-0.008416499942541122,
0.09671617299318314,
0.012158963829278946,
-0.11182212084531784,
-4.37923368679094e-8,
-0.10156015306711197,
0.036868151277303696,
-0.030423037707805634,
0.039455581456422806,
0.04330817982554436,
-0.019006073474884033,
0.04644690826535225,
0.11194499582052231,
0.04062264412641525,
-0.03549810126423836,
0.03963884711265564,
-0.010417790152132511,
-0.011308584362268448,
-0.017946628853678703,
-0.01645427942276001,
0.07001163810491562,
0.014294746331870556,
0.06799013167619705,
0.007601873949170113,
0.005896002985537052,
0.008553788997232914,
0.07019839435815811,
-0.0072709862142801285,
-0.006580069661140442,
0.07740672677755356,
-0.04220644384622574,
-0.0194866843521595,
0.08809009939432144,
0.1529984176158905,
-0.0022490506526082754,
-0.021920517086982727,
-0.020843008533120155,
0.03181528300046921,
-0.039628490805625916,
0.007081088609993458,
0.09659366309642792,
0.07355712354183197,
-0.023390933871269226,
-0.06974699348211288,
-0.03436332195997238,
0.026637040078639984,
0.10076875239610672,
0.02790376916527748,
0.012741846032440662,
-0.029810886830091476,
0.031000729650259018,
-0.008143806830048561,
-0.03301854804158211,
0.038735531270504,
-0.02402939274907112,
0.049803487956523895,
0.0017020530067384243,
-0.05699671432375908,
0.09426259249448776,
0.03387715294957161,
-0.032736267894506454,
0.069184310734272,
-0.017587633803486824,
0.018958451226353645,
0.06455014646053314,
0.04077916219830513,
0.008747001178562641,
-0.03987163305282593,
0.06362276524305344
] |
Microcanonical Langevin Ensembles: Advancing the Sampling of Bayesian Neural Networks | https://openreview.net/forum?id=QMtrW8Ej98 | [
"Emanuel Sommer",
"Jakob Robnik",
"Giorgi Nozadze",
"Uros Seljak",
"David Rügamer"
] | Poster | Despite recent advances, sampling-based inference for Bayesian Neural Networks (BNNs) remains a significant challenge in probabilistic deep learning. While sampling-based approaches do not require a variational distribution assumption, current state-of-the-art samplers still struggle to navigate the complex and highly multimodal posteriors of BNNs. As a consequence, sampling still requires considerably longer inference times than non-Bayesian methods even for small neural networks, despite recent advances in making software implementations more efficient. Besides the difficulty of finding high-probability regions, the time until samplers provide sufficient exploration of these areas remains unpredictable. To tackle these challenges, we introduce an ensembling approach that leverages strategies from optimization and a recently proposed sampler called Microcanonical Langevin Monte Carlo (MCLMC) for efficient, robust and predictable sampling performance. Compared to approaches based on the state-of-the-art No-U-Turn Sampler, our approach delivers substantial speedups up to an order of magnitude, while maintaining or improving predictive performance and uncertainty quantification across diverse tasks and data modalities. The suggested Microcanonical Langevin Ensembles and modifications to MCLMC additionally enhance the method's predictability in resource requirements, facilitating easier parallelization. All in all, the proposed method offers a promising direction for practical, scalable inference for BNNs. | Sampling, Bayesian Neural Networks | null | 228 | 2502.06335 | [
-0.10474172234535217,
-0.10178516805171967,
0.09674450755119324,
0.03872086480259895,
0.06441423296928406,
-0.0032669214997440577,
-0.049879055470228195,
-0.02853715419769287,
0.0016564792022109032,
-0.027327295392751694,
-0.07028193026781082,
-0.09224487096071243,
-0.022521566599607468,
-0.01209744531661272,
-0.06940236687660217,
-0.002743507269769907,
0.02200319617986679,
0.12414999306201935,
-0.04041740670800209,
0.049925897270441055,
0.022135743871331215,
0.051092516630887985,
0.03386159986257553,
-0.0019329054048284888,
0.0621134415268898,
-0.025563158094882965,
0.07283433526754379,
-0.01951850950717926,
-0.052440982311964035,
-0.004791165702044964,
0.06552384793758392,
-0.023262472823262215,
-0.06732067465782166,
-0.05570950731635094,
0.07242553681135178,
0.005395312327891588,
-0.09093248099088669,
0.051635272800922394,
0.02346988208591938,
0.004373619798570871,
-0.015530209988355637,
-0.008346891961991787,
0.05713439732789993,
0.027283277362585068,
0.0694381445646286,
0.007660205010324717,
0.012014538049697876,
-0.008805285207927227,
0.08627265691757202,
0.06263448297977448,
0.010201284661889076,
0.006735912058502436,
-0.0507328137755394,
0.0175627414137125,
0.03293982893228531,
0.008122337982058525,
0.0066246832720935345,
-0.045330896973609924,
0.05935616418719292,
-0.020211825147271156,
0.010088526643812656,
-0.1021999716758728,
0.004820664878934622,
-0.002276236657053232,
0.006600037217140198,
0.009436890482902527,
0.044145356863737106,
0.0938272774219513,
0.052927516400814056,
0.03462952375411987,
0.04484465345740318,
0.09606075286865234,
-0.0684249997138977,
0.0519869327545166,
0.03512407839298248,
0.03411184623837471,
0.028325051069259644,
-0.022059619426727295,
0.053858812898397446,
-0.04957333579659462,
-0.005663873162120581,
0.056508492678403854,
0.027333347126841545,
-0.026498006656765938,
0.08220169693231583,
0.028855960816144943,
-0.0219979640096426,
0.10723701119422913,
0.03843282535672188,
0.01874164678156376,
-0.05750398337841034,
0.037869516760110855,
-0.03172723203897476,
-0.11448082327842712,
0.006188100203871727,
-0.012656954117119312,
0.012018864043056965,
0.003514531534165144,
0.010381261818110943,
0.07210471481084824,
0.007250952534377575,
0.05324863642454147,
0.09498068690299988,
-0.040303103625774384,
0.016461392864584923,
-0.10500799119472504,
0.11246354132890701,
0.01882150024175644,
0.052530255168676376,
-0.05908779427409172,
0.05258167162537575,
0.014093713834881783,
0.037249498069286346,
-0.009284979663789272,
0.028587374836206436,
-0.021751517429947853,
0.05393661558628082,
0.010586345568299294,
-0.033135175704956055,
0.08064182102680206,
-0.10675369203090668,
-0.009552321396768093,
-0.10009030252695084,
0.0024895183742046356,
0.02698269486427307,
0.05254719778895378,
-0.042969610542058945,
3.58215556694776e-33,
0.060245003551244736,
-0.0550830103456974,
0.07973232865333557,
0.07393261790275574,
0.04282699152827263,
-0.037171073257923126,
0.0011710847029462457,
-0.07224749028682709,
-0.07728897780179977,
0.017044734209775925,
-0.030502740293741226,
0.018281694501638412,
-0.10203175991773605,
0.08807491511106491,
0.03608754649758339,
0.009946743957698345,
-0.053633712232112885,
0.013760094530880451,
-0.00971678551286459,
-0.03128602355718613,
0.027263719588518143,
-0.024635540321469307,
-0.02045200765132904,
-0.02032071352005005,
-0.049112387001514435,
-0.016005657613277435,
0.04924319311976433,
0.05265530198812485,
-0.02220679633319378,
-0.001007339102216065,
-0.08076062798500061,
-0.0001714999380055815,
-0.04101438447833061,
0.012251365929841995,
0.09525252133607864,
-0.04108477756381035,
0.06545467674732208,
0.0026401702780276537,
-0.01532797422260046,
-0.0110741276293993,
0.015558941289782524,
0.039793115109205246,
0.01844109781086445,
0.003852431196719408,
-0.06956443190574646,
-0.07296272367238998,
0.028423039242625237,
-0.06156864017248154,
0.005134048871695995,
-0.07134577631950378,
0.04339185729622841,
-0.03493001312017441,
-0.00026096415240317583,
-0.027095681056380272,
0.007939673960208893,
0.011775771155953407,
0.07122475653886795,
-0.00513480743393302,
0.026697786524891853,
0.02727041393518448,
-0.027583783492445946,
0.052369408309459686,
-0.042055245488882065,
0.044074803590774536,
0.06806619465351105,
0.0622076578438282,
-0.05814168602228165,
0.09556359797716141,
-0.018051080405712128,
-0.05028926953673363,
0.02428334578871727,
0.009876164607703686,
0.020046737045049667,
-0.08632208406925201,
0.06698011606931686,
0.06052303686738014,
-0.015437871217727661,
0.004262048285454512,
-0.03653745725750923,
0.05961987376213074,
-0.04102357104420662,
0.002289008116349578,
-0.11539650708436966,
0.03201431408524513,
-0.006941254250705242,
0.04489225521683693,
0.03258581832051277,
-0.0359296016395092,
-0.004739537835121155,
-0.001192372990772128,
-0.04551839455962181,
-0.07448404282331467,
-0.0033164573833346367,
-0.052737537771463394,
-0.12100973725318909,
-2.913542645394697e-33,
-0.03967577591538429,
-0.01553463563323021,
0.013124437071383,
0.05529584363102913,
-0.061318594962358475,
-0.031601328402757645,
-0.015389661304652691,
0.006126314867287874,
0.005416582804173231,
-0.1115422397851944,
-0.04146742448210716,
0.0043309214524924755,
0.06330006569623947,
-0.03652181103825569,
0.05783335864543915,
0.012524711899459362,
-0.04856698215007782,
0.055896978825330734,
0.005833753850311041,
-0.03811876103281975,
0.01728009805083275,
0.02281961217522621,
0.032558873295784,
-0.023214835673570633,
-0.08664877712726593,
0.05033568665385246,
-0.05625426024198532,
0.11554394662380219,
-0.02878921665251255,
-0.06210397183895111,
-0.08181556314229965,
-0.002300511347129941,
-0.011910472065210342,
0.000013365243830776308,
-0.0655624195933342,
0.07900885492563248,
0.05462828278541565,
0.04167531058192253,
-0.015307902358472347,
0.00923942681401968,
0.028619414195418358,
-0.021533412858843803,
-0.09547058492898941,
0.01641702465713024,
0.006371005903929472,
-0.05843738839030266,
0.03386306390166283,
-0.009273100644350052,
0.0012864903546869755,
0.05111243948340416,
-0.013866081833839417,
0.04688858240842819,
-0.09482787549495697,
0.04961186274886131,
-0.05171900615096092,
-0.028519190847873688,
-0.019487794488668442,
-0.03305630385875702,
0.1145583987236023,
0.050692763179540634,
-0.10917990654706955,
-0.0401233471930027,
-0.013080446049571037,
-0.03180590271949768,
-0.03380931168794632,
0.02194991521537304,
-0.00722398329526186,
0.056773170828819275,
0.057554692029953,
0.0016178896185010672,
-0.023357398808002472,
-0.0414619967341423,
0.08160136640071869,
-0.004262702539563179,
0.009990163147449493,
-0.04589500278234482,
0.01120521780103445,
-0.01265637669712305,
-0.005784571636468172,
-0.04245372861623764,
0.020640132948756218,
-0.030971763655543327,
-0.03831580653786659,
0.013907486572861671,
0.11510473489761353,
0.05349985137581825,
0.03922848030924797,
-0.04143963381648064,
0.05104554817080498,
-0.02072123996913433,
-0.02059387043118477,
0.04794943705201149,
0.08658505976200104,
0.05899647995829582,
-0.10530250519514084,
-3.9542179308682535e-8,
-0.015204771421849728,
0.1232302263379097,
0.042417481541633606,
0.008302243426442146,
0.0619981624186039,
-0.027469782158732414,
-0.04406704008579254,
0.039087068289518356,
0.016895011067390442,
-0.0008317687315866351,
0.18061324954032898,
-0.037061672657728195,
-0.01446149405092001,
-0.06889832764863968,
0.06959628313779831,
0.0264789666980505,
0.05786965787410736,
0.02676970511674881,
-0.03381567820906639,
-0.07284539937973022,
-0.005332488100975752,
0.05651086941361427,
0.05054686591029167,
-0.0006515038548968732,
0.033907316625118256,
-0.03385651111602783,
-0.0032980761025100946,
0.0039419508539140224,
0.01632489077746868,
-0.02462674304842949,
-0.07229834794998169,
0.036944687366485596,
-0.08376037329435349,
0.035610929131507874,
0.026041800156235695,
0.041838664561510086,
-0.07997889816761017,
0.02308107726275921,
0.030399102717638016,
-0.06946499645709991,
-0.08994461596012115,
0.004613885190337896,
-0.1264129877090454,
0.02665881998836994,
-0.057352446019649506,
0.005725238937884569,
-0.010688722133636475,
0.007782370317727327,
0.04939016327261925,
0.018754268065094948,
-0.020035624504089355,
-0.049908801913261414,
0.01503897737711668,
0.027899665758013725,
-0.018096838146448135,
0.08330543339252472,
0.059861455112695694,
-0.019743017852306366,
-0.02108778804540634,
0.025159161537885666,
0.00970868207514286,
-0.03961404785513878,
-0.15784770250320435,
-0.005519051104784012
] |
Posterior-Mean Rectified Flow: Towards Minimum MSE Photo-Realistic Image Restoration | https://openreview.net/forum?id=hPOt3yUXii | [
"Guy Ohayon",
"Tomer Michaeli",
"Michael Elad"
] | Poster | Photo-realistic image restoration algorithms are typically evaluated by distortion measures (e.g., PSNR, SSIM) and by perceptual quality measures (e.g., FID, NIQE), where the desire is to attain the lowest possible distortion without compromising on perceptual quality. To achieve this goal, current methods commonly attempt to sample from the posterior distribution, or to optimize a weighted sum of a distortion loss (e.g., MSE) and a perceptual quality loss (e.g., GAN). Unlike previous works, this paper is concerned specifically with the *optimal* estimator that minimizes the MSE under a constraint of perfect perceptual index, namely where the distribution of the reconstructed images is equal to that of the ground-truth ones. A recent theoretical result shows that such an estimator can be constructed by optimally transporting the posterior mean prediction (MMSE estimate) to the distribution of the ground-truth images. Inspired by this result, we introduce Posterior-Mean Rectified Flow (PMRF), a simple yet highly effective algorithm that approximates this optimal estimator. In particular, PMRF first predicts the posterior mean, and then transports the result to a high-quality image using a rectified flow model that approximates the desired optimal transport map. We investigate the theoretical utility of PMRF and demonstrate that it consistently outperforms previous methods on a variety of image restoration tasks. | Image restoration, blind face image restoration, inverse problems, perception-distortion tradeoff, image processing, computer vision, machine learning, generative models, optimal transport, rectified flow, flow matching | We propose Posterior-Mean Rectified Flow (PMRF), a simple image restoration algorithm. Unlike previous works, our method approximates the optimal estimator that minimizes the MSE under a perfect perceptual index constraint. | 227 | 2410.00418 | [
-0.07866748422384262,
-0.04406097158789635,
0.11202972382307053,
0.009486313909292221,
-0.01098256278783083,
-0.011582502163946629,
-0.0009759714594110847,
-0.036265525966882706,
0.06453978270292282,
0.008394181728363037,
-0.06377974897623062,
-0.021409042179584503,
0.10487624257802963,
-0.0035570135805755854,
-0.09321975708007812,
-0.028351398184895515,
0.04749808460474014,
0.19877851009368896,
-0.03941607475280762,
-0.021598147228360176,
0.06922866404056549,
0.029480231925845146,
-0.006505708675831556,
-0.08615384250879288,
0.0940779596567154,
-0.013330444693565369,
-0.011626817286014557,
-0.04660097882151604,
0.0036698286421597004,
-0.027927102521061897,
-0.02080349065363407,
0.004015068523585796,
0.014626058749854565,
-0.045209888368844986,
-0.03562883287668228,
0.044111672788858414,
0.02211834117770195,
0.000867866154294461,
-0.009201093576848507,
0.0076169203966856,
-0.0291335079818964,
0.08565191924571991,
-0.0034519294276833534,
-0.09757447987794876,
0.02608466148376465,
0.03377998620271683,
0.034406501799821854,
-0.030137808993458748,
0.019710935652256012,
0.03606712073087692,
0.036812372505664825,
0.027345286682248116,
-0.09761273860931396,
0.052620939910411835,
0.01746014878153801,
0.020967736840248108,
0.045073043555021286,
-0.05005944147706032,
0.017401713877916336,
0.10599660128355026,
0.01275177113711834,
-0.01788448728621006,
-0.027703355997800827,
-0.0005275086732581258,
0.07517068088054657,
-0.02036266028881073,
0.0451723076403141,
-0.024733873084187508,
0.0015666844556108117,
-0.0014239453012123704,
-0.054517630487680435,
0.02900509350001812,
-0.04620068520307541,
-0.016060229390859604,
-0.09194575250148773,
0.01032556127756834,
0.0729527398943901,
0.006156108807772398,
0.002054492011666298,
0.0218874029815197,
0.08746246248483658,
0.02238951250910759,
0.04640545696020126,
0.0155675383284688,
0.04095227271318436,
0.021005066111683846,
-0.03336546570062637,
0.043893881142139435,
0.042179208248853683,
0.06989418715238571,
-0.11981751769781113,
-0.005832841619849205,
-0.15026763081550598,
0.028514893725514412,
0.06262214481830597,
-0.019550420343875885,
-0.04461396858096123,
-0.06519930064678192,
0.03739186003804207,
0.06931375712156296,
0.06382795423269272,
0.06104627624154091,
0.03341180458664894,
-0.10568603873252869,
0.03808588162064552,
0.00030378898372873664,
0.08159539103507996,
0.057878702878952026,
0.016296543180942535,
0.00856734998524189,
-0.016592422500252724,
0.03203847259283066,
0.045687101781368256,
-0.057712048292160034,
-0.0009192884317599237,
0.005684574134647846,
0.04249719902873039,
0.05636230856180191,
-0.054226264357566833,
-0.05839569866657257,
-0.07673259824514389,
-0.00999018456786871,
-0.006133414804935455,
-0.023814743384718895,
0.07088030874729156,
-0.06384840607643127,
0.022894347086548805,
8.695016000307302e-34,
-0.06713993847370148,
0.07933083176612854,
0.07663581520318985,
-0.0018269444117322564,
0.05717236176133156,
0.06039373576641083,
-0.015732916072010994,
-0.038297735154628754,
0.032137513160705566,
-0.03342142701148987,
-0.028828876093029976,
-0.027390381321310997,
-0.13389508426189423,
-0.005449704360216856,
0.022737272083759308,
-0.039014849811792374,
-0.041866593062877655,
0.05293168127536774,
-0.037630535662174225,
0.01790885254740715,
-0.010399213060736656,
-0.012495897710323334,
0.0650157555937767,
-0.03270859271287918,
-0.03748571500182152,
0.03350980952382088,
0.11845637857913971,
0.007517705671489239,
-0.005319864489138126,
-0.03583597391843796,
-0.07979007810354233,
-0.01787744089961052,
0.035605765879154205,
-0.026188330724835396,
0.010809053666889668,
-0.08975247293710709,
0.02177843078970909,
-0.01261516660451889,
-0.07917633652687073,
0.04295487701892853,
-0.026523452252149582,
0.16728869080543518,
-0.044703807681798935,
-0.014746151864528656,
-0.0315193273127079,
-0.09176506847143173,
0.014574766159057617,
0.03581513464450836,
-0.006345424335449934,
-0.026190608739852905,
0.009331036359071732,
-0.02178848162293434,
0.0028615561313927174,
0.0002676434814929962,
-0.04532966762781143,
-0.03458838537335396,
0.027584457769989967,
0.0353291817009449,
0.05023260414600372,
0.018083594739437103,
0.06999290734529495,
0.009272849187254906,
-0.01874050870537758,
0.043255168944597244,
0.04018482565879822,
0.025189081206917763,
-0.02273721434175968,
0.059731464833021164,
-0.07896026223897934,
-0.012652968987822533,
0.009383094497025013,
0.018722418695688248,
-0.025835875421762466,
-0.04902368038892746,
0.05116685852408409,
-0.009433155879378319,
-0.006171246990561485,
0.05704108998179436,
0.048556629568338394,
0.001049988903105259,
-0.07683710008859634,
0.03757963329553604,
-0.06029787287116051,
-0.07252252101898193,
0.006548079662024975,
0.04740138724446297,
0.02985353395342827,
-0.02813541330397129,
-0.031163817271590233,
-0.08647149056196213,
-0.058655593544244766,
-0.016318775713443756,
0.005524985957890749,
-0.017496280372142792,
0.018161404877901077,
4.419029025384014e-34,
-0.04633454605937004,
0.08753413707017899,
0.0029055210761725903,
0.06119032949209213,
-0.05194447934627533,
-0.10975423455238342,
-0.034330859780311584,
0.056143153458833694,
0.03339507058262825,
-0.054116830229759216,
-0.0890159159898758,
-0.02872704528272152,
-0.009606683626770973,
-0.023633617907762527,
-0.022565849125385284,
-0.00044764234917238355,
0.025015750899910927,
-0.030536556616425514,
-0.11368118226528168,
-0.015362820588052273,
0.04774988815188408,
0.08052460849285126,
-0.01018866803497076,
0.06737393140792847,
-0.009249438531696796,
0.08711560070514679,
-0.03470969572663307,
0.12470267713069916,
-0.005153100471943617,
-0.09825169295072556,
0.07028679549694061,
-0.05249714106321335,
-0.02521994337439537,
0.035123132169246674,
-0.0729190930724144,
0.025373466312885284,
0.068260058760643,
0.0786561518907547,
-0.06920809298753738,
0.12803441286087036,
0.03640657290816307,
0.03577790409326553,
-0.02155827358365059,
-0.005304398015141487,
0.030230246484279633,
-0.024297310039401054,
-0.026378177106380463,
-0.07013051211833954,
0.06561236828565598,
-0.007483689580112696,
-0.05050855502486229,
0.019256552681326866,
-0.1090376228094101,
0.097575843334198,
0.0016286657191812992,
-0.07956696301698685,
-0.10103029012680054,
-0.02440829388797283,
0.07353387773036957,
0.02937564067542553,
-0.03272281214594841,
-0.060291044414043427,
-0.07121020555496216,
0.013218769803643227,
-0.03364962339401245,
0.06890616565942764,
0.007017273455858231,
0.06975096464157104,
-0.010545395314693451,
0.11343305557966232,
-0.05576537549495697,
0.0062943315133452415,
0.027225494384765625,
0.007551081478595734,
0.08621137589216232,
0.005082876421511173,
0.0134104173630476,
0.06004328280687332,
0.03507707267999649,
-0.04644934833049774,
0.04934028908610344,
-0.05528376251459122,
0.012536744587123394,
0.012686169706285,
0.10424969345331192,
-0.018206167966127396,
-0.02040017582476139,
-0.08652330189943314,
0.03601134195923805,
0.038922879844903946,
-0.049011483788490295,
0.001613992964848876,
0.0012415419332683086,
-0.04182413965463638,
-0.05047440901398659,
-4.2079847162312944e-8,
-0.03956429660320282,
-0.01226135529577732,
-0.013238722458481789,
-0.02973732352256775,
0.029999826103448868,
-0.07166632264852524,
-0.060879409313201904,
0.011574974283576012,
0.0012167936656624079,
-0.17152847349643707,
0.028956057503819466,
-0.049538467079401016,
-0.03102377988398075,
-0.05025973170995712,
-0.02615797333419323,
0.03544430434703827,
0.0004908807459287345,
0.05518220365047455,
0.05054730549454689,
0.05412351340055466,
-0.027889704331755638,
0.011113437823951244,
0.015161736868321896,
0.00516451196745038,
0.005978755187243223,
-0.009693311527371407,
0.052143361419439316,
-0.05048659071326256,
0.04907279834151268,
0.03692416846752167,
0.034762777388095856,
0.00112333451397717,
0.08553169667720795,
-0.00887066125869751,
-0.029626239091157913,
-0.010339454747736454,
0.023492716252803802,
0.0405661016702652,
-0.012168942019343376,
-0.04245563596487045,
-0.027418067678809166,
0.05668524280190468,
-0.002393510192632675,
-0.005720148328691721,
0.03359756991267204,
0.04820821061730385,
0.07911404967308044,
-0.024484751746058464,
-0.050771575421094894,
0.0021050269715487957,
0.0633283331990242,
-0.023432103917002678,
0.014193574897944927,
0.004589807707816362,
0.019708117470145226,
-0.046771299093961716,
0.07569263130426407,
-0.04424411803483963,
0.024875467643141747,
0.032671332359313965,
0.0021220026537775993,
0.027903230860829353,
-0.025924140587449074,
0.016565391793847084
] |
Deep Weight Factorization: Sparse Learning Through the Lens of Artificial Symmetries | https://openreview.net/forum?id=vNdOHr7mn5 | [
"Chris Kolb",
"Tobias Weber",
"Bernd Bischl",
"David Rügamer"
] | Poster | Sparse regularization techniques are well-established in machine learning, yet their application in neural networks remains challenging due to the non-differentiability of penalties like the $L_1$ norm, which is incompatible with stochastic gradient descent. A promising alternative is shallow weight factorization, where weights are decomposed into two factors, allowing for smooth optimization of $L_1$-penalized neural networks by adding differentiable $L_2$ regularization to the factors.
In this work, we introduce deep weight factorization, extending previous shallow approaches to more than two factors. We theoretically establish equivalence of our deep factorization with non-convex sparse regularization and analyze its impact on training dynamics and optimization. Due to the limitations posed by standard training practices, we propose a tailored initialization scheme and identify important learning rate requirements necessary for training factorized networks.
We demonstrate the effectiveness of our deep weight factorization through experiments on various architectures and datasets, consistently outperforming its shallow counterpart and widely used pruning methods. | Sparsity, Regularization, Neural Networks, Overparametrization | We propose a deep weight factorization for sparse neural networks that enables smooth optimization of non-convex sparse regularization. | 226 | 2502.02496 | [
-0.050138186663389206,
-0.07549560070037842,
-0.0005999761051498353,
0.05638668313622475,
0.06275979429483414,
0.03727652132511139,
-0.03526777774095535,
-0.06337234377861023,
-0.07445676624774933,
-0.07825285196304321,
0.044177230447530746,
0.0004916813923045993,
-0.021733062341809273,
0.07596897333860397,
-0.05962434038519859,
0.015513991937041283,
0.06835249811410904,
0.10427631437778473,
-0.14083029329776764,
-0.005290739703923464,
-0.013762922957539558,
-0.04228249192237854,
0.030952386558055878,
-0.014697945676743984,
0.10030586272478104,
-0.02815288119018078,
0.0009668894344940782,
-0.025556102395057678,
-0.02737673744559288,
-0.01776278205215931,
0.03437720239162445,
-0.005796033889055252,
-0.027301037684082985,
-0.028192436322569847,
-0.010652460157871246,
0.05989601090550423,
0.00008599755528848618,
0.042537812143564224,
-0.06071365252137184,
0.017463816329836845,
0.031902387738227844,
0.0008457016083411872,
0.012219981290400028,
0.05760957673192024,
0.03827887028455734,
0.01670145057141781,
0.044573210179805756,
-0.010792165994644165,
0.03369919955730438,
0.046061575412750244,
0.014232700690627098,
-0.004544935654848814,
-0.06240042671561241,
0.043698422610759735,
-0.024083051830530167,
-0.0061548128724098206,
0.026304008439183235,
0.059796303510665894,
-0.08157255500555038,
-0.03308740630745888,
-0.02627251297235489,
-0.09821215271949768,
-0.023853735998272896,
-0.004532576072961092,
-0.026615915820002556,
0.00441717216745019,
0.01981566660106182,
-0.03158010169863701,
-0.013241628184914589,
0.06556582450866699,
-0.019854726269841194,
0.08110634237527847,
-0.050662294030189514,
0.04314080998301506,
0.06602704524993896,
0.010331852361559868,
-0.012324506416916847,
-0.04295092076063156,
0.013877013698220253,
-0.05202978104352951,
0.007925717160105705,
-0.003922400064766407,
-0.021561209112405777,
0.010314185172319412,
0.07918388396501541,
-0.028051014989614487,
0.01998242363333702,
-0.01091438066214323,
0.0068241627886891365,
-0.014057813212275505,
0.005960424896329641,
0.03351226821541786,
-0.04179530218243599,
0.014116469770669937,
0.003879602300003171,
-0.0042328499257564545,
0.03181491419672966,
-0.06374205648899078,
-0.07007506489753723,
0.112300343811512,
-0.04715961590409279,
-0.012743434868752956,
0.03457044064998627,
0.13068944215774536,
0.05374809354543686,
-0.018565239384770393,
0.04348205775022507,
0.07297138124704361,
0.037518132477998734,
-0.051926322281360626,
0.09951600432395935,
0.03989729285240173,
0.048784662038087845,
0.0271746926009655,
0.04231216013431549,
-0.06079056113958359,
0.06513160467147827,
-0.05062979832291603,
0.01274250540882349,
0.048548001796007156,
-0.023165885359048843,
0.008067785762250423,
-0.035584867000579834,
-0.03758322820067406,
0.017460046336054802,
0.06994081288576126,
-0.07985507696866989,
5.070124968893413e-33,
0.068194180727005,
0.05234755948185921,
0.06661146879196167,
-0.05073359236121178,
-0.03294749557971954,
-0.032447781413793564,
-0.01267253141850233,
0.03589944541454315,
-0.04497575759887695,
0.0416945144534111,
-0.054179102182388306,
-0.008870833553373814,
-0.048075467348098755,
0.0773673951625824,
0.0037224472034722567,
-0.04612058401107788,
0.013161536306142807,
0.052870217710733414,
0.10127180814743042,
-0.09282131493091583,
0.033793337643146515,
0.019044628366827965,
-0.06739803403615952,
-0.02943420223891735,
-0.03500398248434067,
-0.07514327764511108,
0.1112557128071785,
-0.04809930920600891,
-0.012887104414403439,
0.009500103071331978,
-0.11609597504138947,
0.01752598024904728,
0.01498368289321661,
-0.011805206537246704,
0.04475116357207298,
-0.02799524925649166,
0.06937344372272491,
-0.017891166731715202,
0.07285226881504059,
-0.026592032983899117,
0.03788316994905472,
0.007840159349143505,
-0.0006448269123211503,
-0.0731133297085762,
-0.07007730007171631,
-0.12160804867744446,
0.042127981781959534,
0.057092826813459396,
-0.04397334158420563,
-0.056398823857307434,
0.009454694576561451,
0.005617637187242508,
0.0034996464382857084,
0.01802317425608635,
-0.029027367010712624,
-0.031946662813425064,
0.07130096852779388,
-0.022339947521686554,
0.042045366019010544,
0.020531855523586273,
-0.025008229538798332,
0.04148489609360695,
-0.021530985832214355,
0.04845108836889267,
-0.03437810763716698,
0.014742107130587101,
0.05246894806623459,
0.0318092405796051,
-0.00960446521639824,
-0.07109981030225754,
-0.04215874522924423,
0.02594059146940708,
-0.012125162407755852,
-0.07321752607822418,
0.06900405138731003,
0.0342632457613945,
0.08703800290822983,
-0.07817020267248154,
-0.011782154440879822,
-0.02021222561597824,
-0.04979287460446358,
0.11014673858880997,
-0.03129122033715248,
0.010821372270584106,
-0.07397683709859848,
0.014733439311385155,
0.04630398377776146,
-0.10110119730234146,
-0.007423341739922762,
-0.031201262027025223,
-0.05713655427098274,
-0.030626937747001648,
-0.026810675859451294,
-0.05078477784991264,
0.026915667578577995,
-2.7755205068249596e-33,
-0.07228053361177444,
0.015698401257395744,
-0.05163828283548355,
0.05478698015213013,
0.09153594076633453,
-0.05047166720032692,
-0.04423440247774124,
0.002878497587516904,
0.05019931122660637,
-0.04400064796209335,
0.05230215564370155,
-0.05518445000052452,
-0.008559558540582657,
-0.0675506591796875,
0.1008281335234642,
0.06262814253568649,
-0.10199779272079468,
0.08892383426427841,
-0.039802033454179764,
0.017105622217059135,
0.03791987895965576,
0.04323280230164528,
-0.07610617578029633,
0.06422805041074753,
-0.01801123656332493,
-0.007623062003403902,
-0.06278431415557861,
0.1044665053486824,
0.022548753768205643,
-0.014697070233523846,
-0.06525366753339767,
-0.009709762409329414,
-0.07362523674964905,
-0.05751599743962288,
-0.034605298191308975,
0.0877644494175911,
-0.022013992071151733,
0.018524013459682465,
-0.043839581310749054,
0.010110083036124706,
0.03221482038497925,
0.006927057635039091,
-0.08431161940097809,
0.0013945169048383832,
0.04000254347920418,
-0.06507426500320435,
0.005015549715608358,
-0.06299083679914474,
-0.010996772907674313,
0.03583575785160065,
-0.06000576168298721,
-0.038877226412296295,
-0.018708812072873116,
0.07235759496688843,
-0.0742519348859787,
-0.03527960926294327,
0.07677607983350754,
0.03610663488507271,
0.10346127301454544,
0.019194554537534714,
0.011678767390549183,
-0.03243359550833702,
-0.05431603640317917,
-0.01854756288230419,
-0.036706067621707916,
-0.04936989024281502,
0.08372092992067337,
0.02778088115155697,
-0.00033518439158797264,
0.0757087990641594,
0.004606424830853939,
0.039882462471723557,
0.04639581963419914,
0.09125054627656937,
-0.048449672758579254,
0.010016769170761108,
0.08122825622558594,
-0.00013003800995647907,
-0.027408653870224953,
0.05066784471273422,
0.09272224456071854,
-0.008592202328145504,
0.0005005723796784878,
-0.025310367345809937,
0.08898113667964935,
0.09838204830884933,
0.060447074472904205,
0.04637623578310013,
0.007591953966766596,
0.005226948298513889,
-0.027047116309404373,
-0.017238561064004898,
0.032933928072452545,
0.02633543685078621,
-0.057383663952350616,
-3.788012392647033e-8,
-0.10816291719675064,
0.0341535322368145,
-0.021276865154504776,
0.03830208256840706,
0.027192478999495506,
-0.05953264608979225,
0.003229563357308507,
0.08721271902322769,
-0.012037575244903564,
-0.04014802724123001,
0.08302640169858932,
0.015430847182869911,
-0.06936215609312057,
-0.09419398009777069,
-0.0408858098089695,
0.16237986087799072,
0.018067609518766403,
0.0022696650121361017,
-0.013832229189574718,
-0.006995012052357197,
0.00003137533713015728,
0.05497776344418526,
0.014274079352617264,
-0.04510481283068657,
0.06599777191877365,
-0.07883420586585999,
0.007946847938001156,
0.08823069930076599,
0.1419273465871811,
0.031745780259370804,
-0.05345063656568527,
0.023653777316212654,
0.020323077216744423,
-0.09321881830692291,
-0.038800813257694244,
0.11738774180412292,
0.050207268446683884,
0.0035483846440911293,
-0.06552958488464355,
-0.07191809266805649,
0.008205470629036427,
0.04118351265788078,
-0.016332609578967094,
-0.014489148743450642,
0.0010874661384150386,
0.014571214094758034,
-0.048210933804512024,
-0.06056320294737816,
0.023740366101264954,
-0.051267802715301514,
0.02339882403612137,
0.005554709117859602,
0.021311704069375992,
0.08705843240022659,
0.04833059385418892,
0.0224466510117054,
0.0367971733212471,
-0.04055984318256378,
0.016416184604167938,
0.04799915477633476,
-0.006421805825084448,
-0.000410116626881063,
-0.02839096449315548,
0.04456642270088196
] |
MotionClone: Training-Free Motion Cloning for Controllable Video Generation | https://openreview.net/forum?id=aY3L65HgHJ | [
"Pengyang Ling",
"Jiazi Bu",
"Pan Zhang",
"Xiaoyi Dong",
"Yuhang Zang",
"Tong Wu",
"Huaian Chen",
"Jiaqi Wang",
"Yi Jin"
] | Poster | Motion-based controllable video generation offers the potential for creating captivating visual content. Existing methods typically necessitate model training to encode particular motion cues or incorporate fine-tuning to inject certain motion patterns, resulting in limited flexibility and generalization. In this work, we propose MotionClone, a training-free framework that enables motion cloning from reference videos to versatile motion-controlled video generation, including text-to-video and image-to-video. Based on the observation that the dominant components in temporal-attention maps drive motion synthesis, while the rest mainly capture noisy or very subtle motions, MotionClone utilizes sparse temporal attention weights as motion representations for motion guidance, facilitating diverse motion transfer across varying scenarios. Meanwhile, MotionClone allows for the direct extraction of motion representation through a single denoising step, bypassing the cumbersome inversion processes and thus promoting both efficiency and flexibility.
Extensive experiments demonstrate that MotionClone exhibits proficiency in both global camera motion and local object motion, with notable superiority in terms of motion fidelity, textual alignment, and temporal consistency. | controllable video image, viusal representation, text-to-video generation | Extracting compact motion representation from given reference videos facilitates controlllable video generation | 220 | 2406.05338 | [
-0.010658656246960163,
-0.12445670366287231,
0.030932772904634476,
-0.012434491887688637,
0.06160883978009224,
0.003967768512666225,
-0.06023966521024704,
-0.059698957949876785,
0.033663682639598846,
-0.06953360140323639,
-0.022556288167834282,
-0.03586342930793762,
-0.02990753762423992,
0.016693297773599625,
-0.0353160984814167,
0.043237701058387756,
0.10594421625137329,
0.09487565606832504,
-0.03774882107973099,
0.0035777867306023836,
0.054906465113162994,
-0.09747131168842316,
0.08085347712039948,
0.033316198736429214,
0.04841550067067146,
0.04073960334062576,
0.00960402749478817,
-0.04021778702735901,
0.014752130955457687,
0.007595684379339218,
-0.0014977784594520926,
0.04165704920887947,
-0.02232767827808857,
0.007226705085486174,
-0.06936459243297577,
0.10284195840358734,
-0.011750404722988605,
0.012967953458428383,
-0.011777525767683983,
-0.06624852865934372,
0.020066993311047554,
0.04594339802861214,
0.07817903906106949,
0.0442688949406147,
0.037659503519535065,
0.06545945256948471,
0.05358375981450081,
-0.06593655794858932,
0.0061186314560472965,
0.05977820232510567,
-0.0038866011891514063,
-0.043781984597444534,
0.027191689237952232,
-0.03980217128992081,
0.011090230196714401,
0.050562284886837006,
0.02851070836186409,
-0.041652850806713104,
0.030454060062766075,
-0.004496327601373196,
0.03445035591721535,
-0.0033805514685809612,
0.014965960755944252,
-0.01297624222934246,
-0.03641873598098755,
-0.07242752611637115,
0.09290117770433426,
0.010126397013664246,
-0.026202792301774025,
-0.012105792760848999,
-0.06605136394500732,
-0.014941373839974403,
0.041015103459358215,
0.05183569714426994,
-0.017495974898338318,
0.04786721616983414,
0.015537775121629238,
-0.06078583002090454,
0.0006132081034593284,
-0.06793366372585297,
0.0916089341044426,
-0.058751918375492096,
-0.0003940254682675004,
0.03834228590130806,
0.0984916165471077,
0.034701790660619736,
0.047007206827402115,
0.08641557395458221,
-0.009120726957917213,
0.029265202581882477,
-0.09552972763776779,
-0.055245012044906616,
0.07341711968183517,
-0.1264772117137909,
0.039500124752521515,
0.03737194463610649,
-0.05171140655875206,
-0.025002775713801384,
-0.058138180524110794,
0.009975497610867023,
0.01975363679230213,
0.02182612195611,
-0.011791800148785114,
-0.00018296949565410614,
0.009367863647639751,
-0.01900339312851429,
0.10311023145914078,
0.05670721083879471,
-0.0029254334513098,
0.023504218086600304,
0.018894342705607414,
0.01720660738646984,
-0.010392936877906322,
-0.0018937067361548543,
0.07999680936336517,
0.03812769055366516,
-0.06794801354408264,
-0.031945809721946716,
-0.029209725558757782,
0.037118423730134964,
0.009515848010778427,
0.008477587252855301,
-0.08288977295160294,
-0.007581502199172974,
0.029337190091609955,
-0.08442309498786926,
0.020997818559408188,
2.9787531945065634e-34,
0.08856528252363205,
0.060556765645742416,
0.0708797499537468,
0.031228717416524887,
0.03279666602611542,
-0.0064179133623838425,
0.019646162167191505,
0.022685423493385315,
-0.05345197767019272,
-0.002582128392532468,
0.02877713367342949,
-0.082295261323452,
-0.002652033930644393,
0.13295798003673553,
-0.011067266575992107,
0.011423485353589058,
-0.0036395802162587643,
0.006102422717958689,
0.05491311475634575,
0.036378975957632065,
-0.021733615547418594,
-0.0006259111105464399,
-0.016011111438274384,
-0.02372855506837368,
-0.05493800342082977,
-0.03316619619727135,
-0.018708111718297005,
-0.005326955579221249,
0.05417303740978241,
-0.028889194130897522,
-0.11212010681629181,
0.022712484002113342,
0.0038489389698952436,
-0.020546812564134598,
0.06235470250248909,
-0.04469020292162895,
-0.03597297891974449,
-0.09227557480335236,
0.05202694982290268,
-0.016957884654402733,
-0.01682543009519577,
0.030962781980633736,
-0.06484205275774002,
0.010983590967953205,
-0.07347600907087326,
-0.05366496741771698,
0.05268644541501999,
0.1451871395111084,
-0.13595321774482727,
0.09456443786621094,
0.042532406747341156,
0.026600373908877373,
0.011988191865384579,
-0.09467235207557678,
-0.04729573801159859,
-0.027103716507554054,
0.10299177467823029,
0.011668582446873188,
-0.019731886684894562,
0.04291921854019165,
0.005620426964014769,
0.06155809387564659,
-0.02372218854725361,
0.023278089240193367,
0.0427994541823864,
0.0793573409318924,
-0.009889369830489159,
0.07634874433279037,
0.05803375318646431,
0.004142987076193094,
0.028200624510645866,
-0.007233412470668554,
-0.1101190373301506,
0.021700430661439896,
0.09402165561914444,
0.025275202468037605,
0.04555227980017662,
-0.06528545916080475,
-0.015188797377049923,
0.031187880784273148,
-0.08030533790588379,
-0.02915647067129612,
-0.012449249625205994,
0.0037552721332758665,
-0.02422216907143593,
-0.04501046612858772,
0.047366876155138016,
-0.07521653175354004,
0.0006928499205969274,
-0.03696268051862717,
-0.01894264481961727,
-0.07410725951194763,
-0.05663431063294411,
-0.016082091256976128,
0.00777586130425334,
1.1902145706659558e-33,
0.06226003170013428,
0.07175498455762863,
-0.054186295717954636,
0.013890102505683899,
-0.028801657259464264,
0.013739277608692646,
0.015781037509441376,
-0.04222382605075836,
0.05573592334985733,
-0.018100351095199585,
-0.035565730184316635,
-0.07065550982952118,
-0.023090824484825134,
-0.07241115719079971,
-0.05498481169342995,
0.004431250039488077,
0.02021213248372078,
-0.04535192251205444,
-0.07640627771615982,
-0.04821845144033432,
0.0293593667447567,
0.10833190381526947,
0.04372932016849518,
0.019286615774035454,
-0.08286526054143906,
0.0733572468161583,
0.03662175312638283,
0.11093650758266449,
0.08007357269525528,
-0.02229343354701996,
0.016185792163014412,
0.023051518946886063,
0.05478987842798233,
-0.045406222343444824,
-0.09532836079597473,
0.09422894567251205,
0.08080901950597763,
0.017855055630207062,
-0.03242427483201027,
0.04027346521615982,
0.0012433864176273346,
-0.04937988147139549,
-0.013449495658278465,
0.05182580277323723,
-0.033416371792554855,
0.061518825590610504,
-0.028746576979756355,
-0.011214460246264935,
-0.027967169880867004,
0.07174283266067505,
-0.03993740677833557,
0.04204372316598892,
-0.06423157453536987,
-0.0478145107626915,
-0.006135226227343082,
-0.07785864919424057,
-0.08276426047086716,
-0.002203018171712756,
0.09404858201742172,
-0.08987649530172348,
-0.0017416805494576693,
0.05222310870885849,
-0.06394106149673462,
-0.03340880572795868,
0.0018364860443398356,
0.0007381794857792556,
-0.01893172413110733,
-0.02451462857425213,
0.012426551431417465,
0.025363408029079437,
0.05253983289003372,
-0.009082757867872715,
0.023827334865927696,
0.003287829691544175,
-0.009636588394641876,
-0.026603642851114273,
0.08737865090370178,
-0.034035198390483856,
-0.035582367330789566,
-0.11765872687101364,
0.03676149249076843,
-0.0513148233294487,
0.030752763152122498,
0.06345192342996597,
0.016345275565981865,
0.09975427389144897,
0.006055436562746763,
-0.01404005941003561,
-0.00556401489302516,
0.028431765735149384,
-0.06369755417108536,
0.06292327493429184,
-0.06373321264982224,
0.05231325328350067,
-0.05223144590854645,
-3.504037948687255e-8,
-0.0918547585606575,
0.076858289539814,
-0.0005309726693667471,
0.03227788954973221,
-0.07122784107923508,
-0.08927962183952332,
0.05475272238254547,
0.06279987096786499,
0.05279441922903061,
-0.04744357243180275,
0.03951320797204971,
-0.006527309771627188,
0.03190718963742256,
-0.05941450968384743,
-0.029036128893494606,
0.0022671548649668694,
0.030728166922926903,
0.016018277034163475,
-0.022957323119044304,
0.021790670230984688,
-0.027210181578993797,
0.002569274976849556,
-0.05066331475973129,
-0.09741666913032532,
0.018526587635278702,
-0.04254189133644104,
-0.00036178709706291556,
0.04120213910937309,
0.10832870751619339,
-0.0672326609492302,
0.06712453812360764,
0.0004926094552502036,
0.03464227914810181,
-0.06489071249961853,
-0.06532146781682968,
0.00034981753560714424,
-0.004284579306840897,
-0.024388659745454788,
-0.014289931394159794,
-0.033184755593538284,
-0.019420651718974113,
0.08858288824558258,
0.002393702743574977,
0.04399710148572922,
-0.08934750407934189,
-0.010208294726908207,
0.009704369120299816,
-0.1303693801164627,
0.039835527539253235,
-0.039312802255153656,
-0.0022409826051443815,
-0.016406770795583725,
0.046649154275655746,
0.06790641695261002,
0.057680822908878326,
0.037356939166784286,
0.07479079067707062,
-0.027443597093224525,
0.0005439117085188627,
0.06677617877721786,
-0.029598534107208252,
0.024652646854519844,
0.019610045477747917,
-0.029738083481788635
] |
Learning View-invariant World Models for Visual Robotic Manipulation | https://openreview.net/forum?id=vJwjWyt4Ed | [
"Jing-Cheng Pang",
"Nan Tang",
"Kaiyuan Li",
"Yuting Tang",
"Xin-Qiang Cai",
"Zhen-Yu Zhang",
"Gang Niu",
"Masashi Sugiyama",
"Yang Yu"
] | Poster | Robotic manipulation tasks often rely on visual inputs from cameras to perceive the environment. However, previous approaches still suffer from performance degradation when the camera’s viewpoint changes during manipulation. In this paper, we propose ReViWo (Representation learning for View-invariant World model), leveraging multi-view data to learn robust representations for control under viewpoint disturbance. ReViWo utilizes an autoencoder framework to reconstruct target images by an architecture that combines view-invariant representation (VIR) and view-dependent representation. To train ReViWo, we collect multi-view data in simulators with known view labels, meanwhile, ReViWo is simutaneously trained on Open X-Embodiment datasets without view labels. The VIR is then used to train a world model on pre-collected manipulation data and a policy through interaction with the world model. We evaluate the effectiveness of ReViWo in various viewpoint disturbance scenarios, including control under novel camera positions and frequent camera shaking, using the Meta-world & PandaGym environments. Besides, we also conduct experiments on real world ALOHA robot. The results demonstrate that ReViWo maintains robust performance under viewpoint disturbance, while baseline methods suffer from significant performance degradation. Furthermore, we show that the VIR captures task-relevant state information and remains stable for observations from novel viewpoints, validating the efficacy of the ReViWo approach. | Robotic manipulation, reinforcement learning, world model | We study robust robotic manipulation under viewpoint disturbance by learning view-invariant representation and world model. | 219 | null | [
-0.028574740514159203,
-0.11529526859521866,
-0.0306907519698143,
-0.020860755816102028,
0.0656604915857315,
0.007478538900613785,
0.00025750728673301637,
-0.06187191233038902,
0.00953441672027111,
-0.03211458772420883,
0.04183020815253258,
-0.08207468688488007,
-0.007356571964919567,
0.04171357303857803,
-0.029143987223505974,
0.005203232169151306,
0.12330985814332962,
0.0546448715031147,
-0.11871105432510376,
0.04006483778357506,
-0.014668084681034088,
-0.013323522172868252,
0.07986926287412643,
0.03151065111160278,
0.0008877472137100995,
0.05308472365140915,
0.02737489342689514,
-0.07162953913211823,
0.01815958134829998,
-0.09331537038087845,
0.0022809188812971115,
-0.06033068150281906,
-0.06399837136268616,
-0.022054186090826988,
-0.046415943652391434,
0.07312297075986862,
-0.03184233233332634,
0.019556889310479164,
-0.008609974756836891,
-0.062117788940668106,
-0.05394720286130905,
0.07560831308364868,
0.03629359230399132,
-0.030584178864955902,
0.05598810687661171,
0.0827733501791954,
0.05504925176501274,
-0.027615750208497047,
0.05091656744480133,
0.07185998558998108,
-0.07799790799617767,
-0.05574677512049675,
0.04218214377760887,
-0.013385890983045101,
0.002791499951854348,
0.030514655634760857,
0.03274845331907272,
-0.05894012004137039,
0.06373067945241928,
0.02113165892660618,
0.03660346940159798,
0.016837717965245247,
0.020897649228572845,
-0.002208450809121132,
-0.08555152267217636,
-0.03271128237247467,
0.05420267581939697,
-0.020557625219225883,
-0.030431125313043594,
-0.033691681921482086,
-0.03096211887896061,
0.015254197642207146,
0.04364047944545746,
-0.06479418277740479,
0.005651612766087055,
-0.014171316288411617,
-0.03595711663365364,
-0.08735983818769455,
0.04375661537051201,
-0.057971108704805374,
0.08950182795524597,
0.046084754168987274,
0.015080437995493412,
0.046081431210041046,
0.060274772346019745,
-0.044619668275117874,
-0.006503118202090263,
0.06890154629945755,
0.05991555377840996,
0.00436160946264863,
0.0011654035188257694,
-0.0513453371822834,
-0.004914274904876947,
-0.03823975846171379,
0.033725377172231674,
-0.04407081380486488,
0.04566728696227074,
-0.055827103555202484,
-0.02497405931353569,
0.03604516014456749,
-0.009351368993520737,
0.029298026114702225,
0.010034818202257156,
0.023412948474287987,
0.0386730320751667,
0.011638741940259933,
0.011426066979765892,
0.0067497738637030125,
-0.0340888686478138,
-0.005231671035289764,
0.01566244661808014,
-0.04628641530871391,
-0.02375376783311367,
0.024276698008179665,
0.027281682938337326,
-0.019036488607525826,
-0.005930790212005377,
-0.022357899695634842,
-0.0517786480486393,
0.02435852214694023,
0.020443851128220558,
-0.0695825144648552,
0.025790447369217873,
0.08759934455156326,
0.08349315822124481,
-0.03145059943199158,
-0.010037393309175968,
5.407673957368291e-34,
0.0788259208202362,
0.009182046167552471,
0.011947072111070156,
-0.02678939513862133,
0.07116223871707916,
-0.01886339671909809,
0.003684057854115963,
0.07565329968929291,
-0.038620371371507645,
-0.020071083679795265,
0.020189844071865082,
-0.06838983297348022,
-0.010870710015296936,
0.07705586403608322,
0.029381783679127693,
0.03873587027192116,
-0.02947981469333172,
0.027548709884285927,
-0.07049354165792465,
0.08276346325874329,
0.03078615851700306,
-0.04245813935995102,
-0.04470895230770111,
-0.010892681777477264,
-0.0628734827041626,
0.03919878602027893,
0.027092238888144493,
0.06278149783611298,
-0.03637300431728363,
-0.018130440264940262,
-0.03575774282217026,
0.020950553938746452,
-0.019630754366517067,
0.04553063213825226,
-0.03290928900241852,
0.013240154832601547,
-0.022344494238495827,
-0.05012301728129387,
-0.005730613134801388,
-0.019717393442988396,
-0.0001953709579538554,
0.10499674826860428,
-0.03449004516005516,
-0.04116562381386757,
-0.046016618609428406,
-0.02271527424454689,
0.053185418248176575,
0.07264286279678345,
-0.16823852062225342,
-0.0164028313010931,
0.012657476589083672,
0.014629589393734932,
-0.05322505533695221,
-0.06437502056360245,
-0.08122332394123077,
-0.025562265887856483,
0.014806428924202919,
0.023845020681619644,
-0.0027798085939139128,
0.004076887853443623,
-0.006190622225403786,
0.057850051671266556,
-0.023167584091424942,
0.08237869292497635,
-0.04585800692439079,
0.037768442183732986,
0.05709289014339447,
0.10224112868309021,
0.0633217915892601,
0.022734161466360092,
-0.07405894994735718,
0.0075157321989536285,
-0.014897650107741356,
-0.07568781077861786,
0.05619898438453674,
0.01928720809519291,
0.007048534229397774,
-0.07985330373048782,
0.025625381618738174,
0.11355067789554596,
-0.03732534125447273,
0.018427329137921333,
0.10012561827898026,
0.024309396743774414,
-0.026081178337335587,
-0.06989613175392151,
0.019307108595967293,
-0.04900761693716049,
-0.0027165778446942568,
-0.003173917531967163,
-0.07667017728090286,
-0.09963196516036987,
-0.045140590518713,
-0.04043387994170189,
0.013733803294599056,
-9.824115058470572e-34,
0.04416991397738457,
0.011752894148230553,
0.023782875388860703,
0.0033580458257347345,
-0.0844150260090828,
-0.10198388993740082,
-0.030613619834184647,
0.07105343788862228,
0.01200483925640583,
-0.13260996341705322,
-0.052187729626894,
-0.04053264483809471,
0.00876424740999937,
0.04627440124750137,
0.033351290971040726,
-0.015151922591030598,
-0.027312584221363068,
-0.012864249758422375,
-0.04828205704689026,
0.003997734282165766,
0.0588236041367054,
0.0576808787882328,
0.011343874037265778,
-0.031039858236908913,
-0.061646103858947754,
0.06172811612486839,
0.028110506013035774,
0.0561414435505867,
0.06370428204536438,
0.027894064784049988,
-0.017287490889430046,
-0.026210177689790726,
0.02887021005153656,
0.019407672807574272,
-0.08232239633798599,
0.061723388731479645,
-0.07070626318454742,
-0.020388808101415634,
-0.009050814434885979,
0.08030866086483002,
0.046420324593782425,
-0.019905854016542435,
-0.02000787854194641,
0.061772994697093964,
-0.04331241175532341,
0.03963112458586693,
-0.05661589652299881,
0.05500982701778412,
0.015176535584032536,
-0.0066102296113967896,
-0.10839201509952545,
0.0853705108165741,
-0.08023432642221451,
-0.024497803300619125,
-0.039367545396089554,
-0.021920081228017807,
-0.019163351505994797,
-0.028684722259640694,
0.061792340129613876,
-0.0008163077873177826,
-0.04457971453666687,
-0.0054829660803079605,
-0.008578798733651638,
0.06628129631280899,
-0.024912437424063683,
0.022334059700369835,
-0.063247911632061,
-0.009209866635501385,
0.09832272678613663,
0.022404199466109276,
0.09680920839309692,
0.016909735277295113,
0.02435222640633583,
-0.00978780910372734,
0.035248130559921265,
0.005547321867197752,
-0.012089528143405914,
-0.013496294617652893,
0.027796704322099686,
-0.01716586947441101,
0.050178274512290955,
-0.059102728962898254,
0.10395887494087219,
0.12045147269964218,
0.02583238296210766,
0.14111681282520294,
-0.039594415575265884,
-0.000010494478374312166,
-0.006575413979589939,
0.0126483840867877,
-0.012939798645675182,
-0.03410915285348892,
-0.010723782703280449,
0.06518397480249405,
-0.023342683911323547,
-3.70466963772742e-8,
-0.1369323432445526,
0.007202318869531155,
0.10447532683610916,
0.0015684723621234298,
-0.06105540320277214,
0.021256569772958755,
0.030070921406149864,
0.03619816526770592,
-0.035132117569446564,
0.003918093163520098,
0.07024945318698883,
-0.026986807584762573,
-0.027818601578474045,
-0.046564314514398575,
-0.05132301524281502,
0.06492321193218231,
0.04955792799592018,
0.0353962779045105,
0.005082166288048029,
0.04095014929771423,
0.026538491249084473,
-0.058243121951818466,
0.10549935698509216,
-0.06920433044433594,
-0.0014236500719562173,
-0.03572482243180275,
-0.08458204567432404,
0.04542391374707222,
0.05077073723077774,
-0.049135059118270874,
0.06509868055582047,
0.01270644087344408,
0.045369479805231094,
-0.00546950614079833,
-0.020950162783265114,
-0.0014740086626261473,
-0.04384240135550499,
-0.09743326902389526,
0.023073265329003334,
-0.04009954258799553,
-0.00778998201712966,
0.07690019905567169,
-0.015335372649133205,
0.017726948484778404,
-0.01571124605834484,
0.06530062854290009,
0.0761251375079155,
-0.16124509274959564,
0.026857368648052216,
-0.01938541978597641,
-0.036229223012924194,
-0.016466286033391953,
0.024080021306872368,
0.11709605157375336,
0.1078081950545311,
0.048612117767333984,
0.029220853000879288,
-0.12270253896713257,
-0.0001835927541833371,
0.1032637283205986,
-0.026338139548897743,
0.017911085858941078,
-0.05050744488835335,
-0.02877914160490036
] |
Recovery of Causal Graph Involving Latent Variables via Homologous Surrogates | https://openreview.net/forum?id=fGhr39bqZa | [
"Xiu-Chuan Li",
"Jun Wang",
"Tongliang Liu"
] | Poster | Causal discovery with latent variables is an important and challenging problem. To identify latent variables and infer their causal relations, most existing works rely on the assumption that latent variables have pure children. Considering that this assumption is potentially restrictive in practice and not strictly necessary in theory, in this paper, by introducing the concept of homologous surrogate, we eliminate the need for pure children in the context of causal discovery with latent variables. The homologous surrogate fundamentally differs from the pure child in the sense that the latter is characterized by having strictly restricted parents while the former allows for much more flexible parents. We formulate two assumptions involving homologous surrogates and develop theoretical results under each assumption. Under the weaker assumption, our theoretical results imply that we can determine each variable's ancestors, that is, partially recover the causal graph. The stronger assumption further enables us to determine each variable's parents exactly, that is, fully recover the causal graph. Building on these theoretical results, we derive an algorithm that fully leverages the properties of homologous surrogates for causal graph recovery. Also, we validate its efficacy through experiments. Our work broadens the applicability of causal discovery. Our code is available at: https://github.com/XiuchuanLi/ICLR2025-CDHS | causal discovery, latent variables | null | 217 | null | [
-0.07063890993595123,
0.0025940989144146442,
0.022541102021932602,
0.04471895098686218,
0.07182972878217697,
0.05045120045542717,
-0.026142289862036705,
-0.09929481893777847,
0.03679429739713669,
0.05160730332136154,
0.05067743733525276,
0.00926840491592884,
0.028070557862520218,
0.06411422044038773,
0.03463158756494522,
0.02578604966402054,
0.011773698963224888,
0.03343519940972328,
-0.014605221338570118,
-0.07006203383207321,
-0.01289801299571991,
0.004059462808072567,
-0.06889946013689041,
-0.04141688346862793,
-0.0059043038636446,
0.025601990520954132,
0.0141758369281888,
0.0406636968255043,
0.030862372368574142,
0.0316588431596756,
0.03677145764231682,
0.05416116863489151,
-0.06752830743789673,
-0.07687622308731079,
-0.0035670932848006487,
0.08482540398836136,
0.035939134657382965,
0.02211274579167366,
0.0391550213098526,
0.07091963291168213,
0.03544369712471962,
0.04027419537305832,
-0.05310681089758873,
0.05500681325793266,
0.013227641582489014,
0.004699639044702053,
-0.02807411551475525,
0.004262231756001711,
-0.057916220277547836,
-0.019743476063013077,
-0.07069521397352219,
-0.01637287251651287,
0.0008041187538765371,
0.01442626491189003,
0.1016751304268837,
-0.005227388348430395,
0.006733357906341553,
-0.06997346878051758,
-0.01377171091735363,
0.014783875085413456,
-0.04728797823190689,
-0.025837460532784462,
-0.0863838866353035,
-0.08729379624128342,
0.0902642235159874,
0.01142100989818573,
-0.009244057349860668,
0.04287738725543022,
-0.0036675287410616875,
0.15757712721824646,
-0.03159959241747856,
0.028129417449235916,
-0.09798487275838852,
0.03322150558233261,
0.02183438278734684,
0.07399998605251312,
0.02645893394947052,
0.0197300985455513,
-0.016313740983605385,
-0.08619837462902069,
-0.002291227923706174,
0.016585009172558784,
-0.05228247866034508,
0.03781091421842575,
0.02777574211359024,
-0.03319082781672478,
0.0032693922985345125,
-0.002279489766806364,
0.05565598979592323,
0.08862870931625366,
-0.019100142642855644,
-0.03634581342339516,
0.12134502828121185,
0.020650310441851616,
0.057840947061777115,
-0.004747722297906876,
0.058356206864118576,
-0.00782626960426569,
0.09028413146734238,
0.060745514929294586,
0.013731730170547962,
0.021225513890385628,
0.02647947333753109,
0.04434777423739433,
-0.07116061449050903,
-0.02933577634394169,
-0.02954735979437828,
0.046529363840818405,
-0.0195853840559721,
-0.04017938673496246,
-0.06727404147386551,
-0.03660839796066284,
0.005651469342410564,
0.05801109969615936,
0.027590706944465637,
-0.09976835548877716,
0.04152561351656914,
-0.038416966795921326,
-0.04268123209476471,
-0.04684258997440338,
0.03760264441370964,
0.02772638015449047,
0.06913158297538757,
-0.03278530016541481,
0.03655070438981056,
-0.024182818830013275,
-0.06091248244047165,
2.75867438708073e-33,
0.05244004726409912,
0.06174607574939728,
0.0644516870379448,
0.015696821734309196,
-0.040835388004779816,
0.06769628077745438,
-0.050852809101343155,
-0.02007175050675869,
0.002868480747565627,
0.021557016298174858,
-0.05653459206223488,
0.016961989924311638,
-0.057164452970027924,
0.0057057845406234264,
0.015435054898262024,
0.0435635931789875,
-0.07218902558088303,
0.08658039569854736,
-0.04498356208205223,
-0.017702439799904823,
-0.00914083607494831,
0.01912555657327175,
-0.032711926847696304,
-0.056917328387498856,
0.07858740538358688,
-0.034054625779390335,
0.05888552591204643,
-0.058963343501091,
-0.009524662047624588,
-0.01718207448720932,
-0.07926720380783081,
0.043697092682123184,
0.05705425515770912,
0.08101626485586166,
-0.06305911391973495,
0.004943297244608402,
0.0038254554383456707,
-0.11026634275913239,
-0.05129322409629822,
-0.00576883926987648,
-0.0029759458266198635,
-0.0319877527654171,
0.040227487683296204,
-0.004887289367616177,
-0.06938359886407852,
-0.030807601287961006,
0.006577113643288612,
0.01623499020934105,
-0.036843523383140564,
-0.01685836724936962,
0.005900106392800808,
0.0341314971446991,
0.004418868571519852,
-0.13811129331588745,
-0.09724634885787964,
0.0665731430053711,
-0.05659027025103569,
0.0347197987139225,
-0.008597902953624725,
0.010528863407671452,
0.03930182754993439,
-0.05494461953639984,
-0.021891869604587555,
0.002167021157220006,
-0.016850922256708145,
0.06376825273036957,
-0.06095680221915245,
-0.04456368833780289,
0.12556466460227966,
0.007638060953468084,
-0.05850652605295181,
0.06015123426914215,
-0.1026037260890007,
-0.08836112916469574,
0.05232004448771477,
-0.009266001172363758,
-0.02707306481897831,
-0.042026154696941376,
0.0566839762032032,
-0.0716080516576767,
-0.05136646330356598,
-0.11526183038949966,
-0.007802294567227364,
-0.012465286068618298,
-0.012397178448736668,
-0.05290907248854637,
-0.023100921884179115,
-0.03535591438412666,
-0.055802468210458755,
-0.07046527415513992,
-0.044733013957738876,
-0.052229832857847214,
0.01087066251784563,
0.01173623651266098,
0.06794627755880356,
-4.6808229130781644e-33,
-0.05371618643403053,
-0.03685886040329933,
0.09731393307447433,
-0.0024909237399697304,
0.0390767864882946,
-0.07003878057003021,
0.01345743890851736,
-0.10158205032348633,
-0.03505053371191025,
-0.020460408180952072,
-0.030275078490376472,
-0.04431125521659851,
0.024604901671409607,
-0.0352940671145916,
-0.04348692670464516,
0.09438874572515488,
-0.032748762518167496,
-0.0238999854773283,
-0.018267987295985222,
-0.006063803099095821,
0.05602860078215599,
0.0879552885890007,
-0.01980474404990673,
0.01241385005414486,
0.09204234182834625,
0.01615646295249462,
0.02257697843015194,
0.06157462298870087,
-0.0007808668306097388,
0.05514610558748245,
-0.013603013940155506,
-0.08428993821144104,
-0.053655192255973816,
-0.018144141882658005,
0.029355963692069054,
0.10676655918359756,
0.0316060408949852,
0.05999639257788658,
-0.07671231776475906,
-0.023768771439790726,
-0.012230039574205875,
0.0006138425669632852,
-0.0632343590259552,
0.030419936403632164,
0.020975317806005478,
0.03257092088460922,
-0.033522140234708786,
0.16857928037643433,
0.0297910925000906,
0.035031091421842575,
-0.05296815559267998,
0.053044356405735016,
0.04292510449886322,
-0.005150850862264633,
-0.07268549501895905,
-0.01211797446012497,
0.0912456065416336,
-0.0925195962190628,
0.039431244134902954,
0.01710495352745056,
-0.0739651694893837,
-0.0467039979994297,
0.005079437512904406,
0.026566078886389732,
-0.01789001375436783,
-0.1340286284685135,
0.002411790192127228,
-0.030853550881147385,
-0.0023827003315091133,
-0.025011222809553146,
0.0323353037238121,
0.06283073127269745,
-0.03020266629755497,
-0.08024869859218597,
0.11243800818920135,
-0.006939391605556011,
-0.06619612872600555,
0.07071740925312042,
-0.039445217698812485,
-0.005452896002680063,
0.03935695067048073,
0.009047260507941246,
0.11178244650363922,
-0.03636471927165985,
-0.04205168038606644,
0.0006567009259015322,
0.058479685336351395,
0.002748847473412752,
0.0906379297375679,
-0.0007016173331066966,
-0.08243901282548904,
-0.0184448454529047,
-0.07198149710893631,
0.049557603895664215,
-0.10955654084682465,
-3.7988638013075615e-8,
-0.05530618876218796,
-0.0017878268845379353,
0.031785428524017334,
0.014336303807795048,
0.07479895651340485,
-0.05190063267946243,
0.05061623826622963,
0.0519469790160656,
-0.025446712970733643,
0.03894231095910072,
0.029466457664966583,
-0.027841312810778618,
0.01926221139729023,
-0.011082520708441734,
0.0074665239080786705,
-0.08570195734500885,
0.015064435079693794,
-0.014877359382808208,
-0.03866555914282799,
0.021290045231580734,
0.040256477892398834,
0.012086241506040096,
-0.005656120367348194,
0.11237670481204987,
0.10816669464111328,
-0.11039873212575912,
-0.04376792535185814,
0.021501237526535988,
0.02970241941511631,
0.006463134661316872,
0.01092088595032692,
0.022479631006717682,
0.019798025488853455,
0.043873026967048645,
-0.00503568397834897,
0.057064153254032135,
0.023544499650597572,
0.006092756520956755,
-0.04802093654870987,
-0.021301476284861565,
-0.01633555442094803,
0.025376206263899803,
-0.05930064618587494,
0.0101051926612854,
0.08756764978170395,
-0.01505025289952755,
-0.02641415037214756,
-0.01866215653717518,
-0.03916327282786369,
-0.019562415778636932,
-0.07015510648488998,
-0.07574675977230072,
-0.023406177759170532,
-0.010548007674515247,
0.0876537412405014,
0.03777231648564339,
0.008369948714971542,
0.01921629160642624,
0.017741777002811432,
-0.006324137095361948,
0.08836062997579575,
0.03274667263031006,
0.043348442763090134,
-0.01077206525951624
] |
Reconstruction-Guided Policy: Enhancing Decision-Making through Agent-Wise State Consistency | https://openreview.net/forum?id=Y8L5RB4GWb | [
"Liang Qifan",
"Yixiang Shan",
"Haipeng Liu",
"Zhengbang Zhu",
"Ting Long",
"Weinan Zhang",
"Yuan Tian"
] | Poster | An important challenge in multi-agent reinforcement learning is partial observability, where agents cannot access the global state of the environment during execution and can only receive observations within their field of view. To address this issue, previous works typically use the dimensional-wise state, which is obtained by applying MLP or dimensional-based attention on the global state, for decision-making during training and relying on a reconstructed dimensional-wise state during execution. However, dimensional-wise states tend to divert agent attention to specific features, neglecting potential dependencies between agents, making it difficult to make optimal decisions. Moreover, the inconsistency between the states used in training and execution further increases additional errors. To resolve these issues, we propose a method called Reconstruction-Guided Policy (RGP) to reconstruct the agent-wise state, which represents the information of inter-agent relationships, as input for decision-making during both training and execution. This not only preserves the potential dependencies between agents but also ensures consistency between the states used in training and execution. We conducted extensive experiments on both discrete and continuous action environments to evaluate RGP, and the results demonstrates its superior effectiveness. Our code is public in https://anonymous.4open.science/r/RGP-9F79 | multi-agent reinforcement learning, partial observability, cooperation, centralized training distributed execution, global state | null | 216 | null | [
-0.02437528409063816,
-0.04718403145670891,
-0.01334739476442337,
0.018320592120289803,
0.03911220282316208,
0.041587889194488525,
-0.012255114503204823,
-0.1092505156993866,
-0.06826606392860413,
-0.029331816360354424,
-0.04103416949510574,
0.005279945209622383,
0.043612074106931686,
0.055209770798683167,
0.008166572079062462,
0.019356079399585724,
0.11044556647539139,
0.06590928137302399,
-0.08975378423929214,
-0.008488813415169716,
0.02698354609310627,
0.035902414470911026,
0.020597446709871292,
0.010745858773589134,
-0.055439967662096024,
0.02769242599606514,
0.04922831803560257,
-0.07246194034814835,
0.016656525433063507,
-0.01400432363152504,
0.03802859038114548,
-0.06289320439100266,
-0.0613107867538929,
-0.07661035656929016,
-0.08743076771497726,
0.11919518560171127,
-0.17152999341487885,
0.05186395347118378,
-0.022755101323127747,
0.007543773390352726,
-0.0056854672729969025,
0.04506189376115799,
-0.07030665129423141,
-0.012807800434529781,
0.043405164033174515,
-0.00976711418479681,
0.010091585107147694,
-0.0739816203713417,
0.049364011734724045,
-0.055363889783620834,
-0.040712494403123856,
-0.047636523842811584,
0.028962308540940285,
0.006156800780445337,
0.0364651195704937,
0.030014144256711006,
0.008136632852256298,
0.017058905214071274,
-0.0393829308450222,
-0.04117267206311226,
-0.023096170276403427,
-0.03123386576771736,
-0.04997313767671585,
-0.006523475516587496,
-0.03980051353573799,
0.006184806581586599,
0.013372426852583885,
0.016134003177285194,
-0.005934186279773712,
-0.003054641420021653,
-0.05311490595340729,
0.027895431965589523,
0.020925110206007957,
-0.07966513186693192,
-0.003678165143355727,
0.038171060383319855,
-0.013730792328715324,
-0.0394444577395916,
0.05655742064118385,
0.020399227738380432,
-0.025116387754678726,
-0.008466525003314018,
-0.040546730160713196,
0.027941560372710228,
0.12366832047700882,
-0.027158496901392937,
0.021787455305457115,
0.030682964250445366,
0.058448899537324905,
0.03847828134894371,
-0.0025417364668101072,
-0.01735926978290081,
0.027273299172520638,
0.034691594541072845,
0.011795218102633953,
0.010610642842948437,
-0.001203707535751164,
-0.07070145010948181,
-0.01576940342783928,
0.05675293877720833,
-0.01750030368566513,
0.029103832319378853,
-0.023052185773849487,
0.0337369330227375,
-0.0018733034376055002,
-0.051829226315021515,
0.042733561247587204,
0.0109560526907444,
0.05564075708389282,
-0.10711973160505295,
0.032317884266376495,
-0.0008293664432130754,
0.04689263924956322,
0.051970649510622025,
0.036437276750802994,
0.06463941931724548,
-0.03342544287443161,
-0.0029010362923145294,
-0.027453284710645676,
0.042644333094358444,
0.051345426589250565,
-0.09331753849983215,
0.08025582879781723,
0.04898465797305107,
0.07283860445022583,
-0.03582582250237465,
0.05859397351741791,
-4.564863334105423e-34,
0.04080667346715927,
-0.01434997282922268,
-0.0026672626845538616,
0.0063727498054504395,
0.01437077485024929,
-0.0027976855635643005,
0.08255033195018768,
-0.02086741104722023,
0.01927647739648819,
0.0011493327328935266,
-0.0002595675177872181,
0.032430391758680344,
-0.00259950733743608,
0.1095401793718338,
0.0025124801322817802,
0.06208348274230957,
-0.04059111699461937,
0.11720630526542664,
-0.023511545732617378,
-0.042318008840084076,
0.03554233908653259,
0.014037192799150944,
0.003190892282873392,
-0.06708831340074539,
0.0335177518427372,
-0.02023494802415371,
0.038792580366134644,
0.022654293105006218,
-0.01729993335902691,
0.0026565149892121553,
-0.010214134119451046,
0.03933165222406387,
-0.0138362105935812,
0.1131768599152565,
-0.03480015695095062,
-0.003513007890433073,
-0.055942948907613754,
-0.046564143151044846,
-0.02042059600353241,
-0.0003165057860314846,
0.011700334027409554,
-0.00036229280522093177,
0.06388567388057709,
-0.006717788055539131,
-0.038960326462984085,
-0.07072187960147858,
-0.03440678119659424,
-0.02554813027381897,
-0.07913484424352646,
-0.01671304553747177,
0.026697427034378052,
0.046381834894418716,
0.09497998654842377,
-0.10567884147167206,
-0.07577970623970032,
0.03512207418680191,
0.003969189710915089,
0.10891271382570267,
-0.0348622128367424,
0.007256619166582823,
0.00934248324483633,
0.02099364437162876,
0.0020816705655306578,
0.13448882102966309,
-0.009491406381130219,
0.0773465558886528,
-0.07585891336202621,
0.08103279024362564,
0.07427485287189484,
-0.013078098185360432,
-0.039037469774484634,
0.024841856211423874,
0.058023978024721146,
-0.044406238943338394,
0.09467139095067978,
-0.06704802066087723,
0.07667461037635803,
-0.07908732444047928,
0.025596395134925842,
-0.02495676279067993,
-0.010631138458848,
0.023184441030025482,
-0.07285165041685104,
-0.026072975248098373,
-0.011766093783080578,
-0.037834711372852325,
-0.038296669721603394,
-0.05277639254927635,
0.018137967213988304,
0.014193762093782425,
-0.118413046002388,
-0.011628868989646435,
-0.004357588943094015,
0.007283266633749008,
0.014775499701499939,
-6.011050156693082e-34,
0.06117808073759079,
-0.004383749328553677,
0.003410191973671317,
-0.0004096107149962336,
-0.028063800185918808,
-0.026342792436480522,
-0.06879239529371262,
-0.01939457096159458,
0.010317794047296047,
-0.09205403923988342,
-0.10604273527860641,
-0.003909276332706213,
0.06884755194187164,
0.02137826569378376,
-0.06223807483911514,
-0.0007693599909543991,
-0.047675251960754395,
-0.03017452359199524,
-0.033274587243795395,
0.00227944552898407,
0.0790068581700325,
0.025847550481557846,
-0.07024726271629333,
0.02697952650487423,
-0.0047364914789795876,
0.07255508005619049,
0.008225401863455772,
0.05870275944471359,
0.0655846819281578,
0.002814652631059289,
-0.002401722129434347,
-0.07234897464513779,
-0.020785244181752205,
0.003878014162182808,
-0.07036539912223816,
0.06240735948085785,
0.07711493968963623,
-0.007375325541943312,
-0.10500785708427429,
0.1635713428258896,
0.036455921828746796,
-0.09449239075183868,
-0.05549003183841705,
0.022310806438326836,
0.02361602522432804,
-0.0023158870171755552,
-0.012620019726455212,
0.014851258136332035,
-0.03604951128363609,
0.007756899110972881,
-0.06707718968391418,
0.052389293909072876,
-0.07882463186979294,
-0.009334088303148746,
-0.050797875970602036,
-0.01740761660039425,
0.040609125047922134,
0.04856608062982559,
0.08294777572154999,
-0.052632007747888565,
-0.01571846939623356,
-0.018192186951637268,
0.02558966539800167,
0.06983105093240738,
-0.03685556724667549,
0.05192650482058525,
0.011511708609759808,
0.03586055338382721,
0.05424799397587776,
0.010234521701931953,
-0.017557041719555855,
0.03369599208235741,
-0.02078552171587944,
-0.019476426765322685,
0.10725144296884537,
0.0008399474318139255,
-0.02769896388053894,
-0.08161323517560959,
-0.018407266587018967,
0.016777437180280685,
0.05323516204953194,
0.010031446814537048,
0.033209703862667084,
0.056928377598524094,
0.02514181286096573,
0.13541431725025177,
0.05558948218822479,
0.058272674679756165,
0.050120074301958084,
-0.09697915613651276,
-0.06710763275623322,
0.01835823617875576,
-0.05060974881052971,
0.009922225959599018,
-0.05416058376431465,
-4.8958462173231965e-8,
-0.09015001356601715,
0.04497285932302475,
0.056507986038923264,
0.051456283777952194,
0.029564151540398598,
-0.08538413047790527,
-0.04418337345123291,
0.009608891792595387,
-0.029377000406384468,
-0.051411744207143784,
0.0597582682967186,
-0.016833480447530746,
0.02002466470003128,
-0.08698032796382904,
-0.014989604242146015,
-0.026281632483005524,
-0.01580110751092434,
-0.07059238851070404,
-0.02105214074254036,
0.0034288419410586357,
0.009852261282503605,
-0.03285139054059982,
0.03545569255948067,
0.019884658977389336,
0.05802656337618828,
-0.0552961491048336,
-0.07204022258520126,
0.08255217224359512,
0.03001324087381363,
0.034445442259311676,
0.06965674459934235,
0.027406154200434685,
0.12043771892786026,
0.027932122349739075,
-0.03805020451545715,
0.0480935275554657,
0.0679350420832634,
-0.058225370943546295,
0.019669873639941216,
-0.08767969161272049,
-0.03979665786027908,
0.05261209234595299,
0.04394439235329628,
0.023950446397066116,
0.026627788320183754,
0.017581649124622345,
-0.03482680767774582,
-0.08690556138753891,
0.05189033970236778,
-0.10969536006450653,
-0.011506099253892899,
-0.026042390614748,
-0.04590381681919098,
0.07478729635477066,
0.11613393574953079,
-0.03291596844792366,
0.020553791895508766,
-0.0979633554816246,
0.052654117345809937,
0.05746636539697647,
-0.040376707911491394,
0.018659619614481926,
-0.07950928062200546,
-0.05329173058271408
] |
Efficient and Trustworthy Causal Discovery with Latent Variables and Complex Relations | https://openreview.net/forum?id=BZYIEw4mcY | [
"Xiu-Chuan Li",
"Tongliang Liu"
] | Poster | Most traditional causal discovery methods assume that all task-relevant variables are observed, an assumption often violated in practice. Although some recent works allow the presence of latent variables, they typically assume the absence of certain special causal relations to ensure a degree of simplicity, which might also be invalid in real-world scenarios. This paper tackles a challenging and important setting where latent and observed variables are interconnected through complex causal relations. Under a pure children assumption ensuring that latent variables leave adequate footprints in observed variables, we develop novel theoretical results, leading to an efficient causal discovery algorithm which is the first one capable of handling the setting with both latent variables and complex relations within polynomial time. Our algorithm first sequentially identifies latent variables from leaves to roots and then sequentially infers causal relations from roots to leaves. Moreover, we prove trustworthiness of our algorithm, meaning that when the assumption is invalid, it can raise an error signal rather than draw an incorrect causal conclusion, thus preventing potential damage to downstream tasks. We demonstrate the efficacy of our algorithm through experiments. Our work significantly enhances efficiency and reliability of causal discovery in complex systems. Our code is available at: https://github.com/XiuchuanLi/ICLR2025-ETCD | causal discovery, latent variables, complex causal relations | null | 215 | null | [
-0.1308130919933319,
0.01706836186349392,
-0.008636187762022018,
0.09397360682487488,
0.06671731919050217,
-0.0017778980545699596,
-0.03313201665878296,
-0.08944062143564224,
0.029440265148878098,
0.05575906112790108,
0.011381768621504307,
-0.04625000059604645,
0.0642777681350708,
0.07486475259065628,
-0.05279238522052765,
0.01239338330924511,
0.054338470101356506,
0.06022267043590546,
-0.03564823791384697,
-0.11411884427070618,
-0.060078930109739304,
0.00885127205401659,
-0.05787377431988716,
0.0009546878281980753,
0.023114053532481194,
0.015295803546905518,
0.04724566265940666,
-0.009857979603111744,
-0.0218780729919672,
0.02801666036248207,
0.0192523505538702,
0.03646400570869446,
-0.03544233739376068,
-0.10448530316352844,
0.010185766965150833,
0.10175315290689468,
0.022705214098095894,
0.006777484901249409,
-0.02253562957048416,
-0.02240012027323246,
0.046481627970933914,
0.02988331764936447,
-0.0017607114277780056,
0.04799490422010422,
-0.00955290999263525,
0.011716675944626331,
0.007261727005243301,
-0.02558199129998684,
-0.024439264088869095,
-0.05046391859650612,
-0.10107523202896118,
-0.027418462559580803,
0.035930510610342026,
0.04067257419228554,
0.053036414086818695,
-0.06262088567018509,
0.05295131355524063,
-0.03508710861206055,
-0.061232790350914,
0.017459310591220856,
0.033515676856040955,
-0.07681961357593536,
-0.08708448708057404,
-0.09119860827922821,
0.010734631679952145,
0.030774183571338654,
-0.02267391048371792,
0.0357649102807045,
-0.022352227941155434,
0.05488291755318642,
-0.037097979336977005,
0.08124189078807831,
-0.10111578553915024,
0.07503058761358261,
-0.011745846830308437,
0.044647570699453354,
-0.012545025907456875,
-0.058039747178554535,
-0.006261305883526802,
-0.06088286638259888,
-0.02191716618835926,
0.047966547310352325,
-0.03527253866195679,
0.01611572876572609,
0.09035663306713104,
-0.0199343990534544,
0.06655745953321457,
0.05480807647109032,
0.0600699745118618,
0.10490936040878296,
0.02429160289466381,
-0.03355145454406738,
0.12403720617294312,
0.03395361453294754,
0.06511814892292023,
0.023720864206552505,
0.06635113060474396,
-0.024096474051475525,
0.033484093844890594,
0.07100333273410797,
0.02344328723847866,
0.051494162529706955,
0.029087720438838005,
-0.0008547967299818993,
-0.03653312847018242,
-0.0071008880622684956,
0.004263427574187517,
0.021191079169511795,
-0.023517442867159843,
-0.0934785008430481,
-0.06546270102262497,
-0.048059094697237015,
0.005612221080809832,
0.013599495403468609,
0.03318651765584946,
-0.029360009357333183,
0.06166955456137657,
0.02105131186544895,
-0.007706173695623875,
-0.023735320195555687,
0.044916652143001556,
0.014098882675170898,
0.05448204651474953,
0.014750741422176361,
0.09901081770658493,
-0.03152070939540863,
-0.059523001313209534,
1.9384245309815152e-33,
0.058173768222332,
0.047637104988098145,
0.02158227562904358,
0.031089112162590027,
-0.0017126691527664661,
0.052459120750427246,
-0.01639481447637081,
0.01196976751089096,
0.00596997095271945,
0.01988084986805916,
-0.02225225605070591,
-0.02324730157852173,
-0.07663944363594055,
-0.03998119384050369,
0.03401697054505348,
0.03392241522669792,
-0.05723978206515312,
0.0638885647058487,
0.007874454371631145,
0.00714763393625617,
0.048685669898986816,
-0.11392207443714142,
-0.05388527736067772,
-0.07585214078426361,
0.08354686945676804,
-0.014122708700597286,
0.08236834406852722,
-0.06951922923326492,
0.029956838116049767,
-0.0029795316513627768,
-0.036635369062423706,
0.006429871544241905,
0.04124561697244644,
0.07543709129095078,
-0.07304441928863525,
-0.049855832010507584,
-0.03759794309735298,
-0.10025656223297119,
0.013059236109256744,
-0.005625000223517418,
0.046184178441762924,
-0.02219545841217041,
0.041231125593185425,
0.06311652064323425,
-0.06349465250968933,
-0.03921256959438324,
-0.02628343366086483,
0.045774467289447784,
0.017656436190009117,
-0.0018733888864517212,
0.02901201695203781,
0.08310145139694214,
0.05980284512042999,
-0.13789871335029602,
-0.09986413270235062,
0.05435997620224953,
-0.04386935010552406,
0.03871482238173485,
-0.03208194673061371,
0.029088502749800682,
0.034128084778785706,
-0.016163932159543037,
-0.058195434510707855,
0.034776486456394196,
-0.0343501940369606,
0.0338783860206604,
-0.06998252123594284,
-0.014433843083679676,
0.07138048857450485,
0.0183867197483778,
-0.060445405542850494,
-0.0025041094049811363,
-0.08169197291135788,
-0.081627756357193,
0.06791336089372635,
-0.018529623746871948,
0.018984515219926834,
-0.0655699223279953,
0.021994678303599358,
-0.07683766633272171,
-0.054790470749139786,
-0.04057183116674423,
-0.009181138128042221,
0.0625067949295044,
-0.031335193663835526,
-0.06032560020685196,
-0.07040468603372574,
-0.076761394739151,
-0.0919729471206665,
-0.021066855639219284,
-0.04034241661429405,
-0.009573409333825111,
0.0641314685344696,
0.08508988469839096,
0.0303695946931839,
-2.8175728988706654e-33,
-0.012422987259924412,
-0.045124128460884094,
0.08936619758605957,
-0.007855214178562164,
0.0420415997505188,
-0.06208021193742752,
-0.04849998280405998,
-0.10619541257619858,
0.016482269391417503,
-0.036281269043684006,
-0.026521911844611168,
0.00964286457747221,
-0.030957521870732307,
0.015668602660298347,
0.01369473896920681,
0.04999929666519165,
0.06094147264957428,
0.003407295560464263,
0.013939671218395233,
-0.00953708402812481,
0.014443568885326385,
0.077011339366436,
-0.036784302443265915,
-0.014696033671498299,
0.03405323252081871,
0.0040078978054225445,
0.06008356437087059,
-0.02897869609296322,
-0.04137974604964256,
0.08300958573818207,
-0.0553634874522686,
-0.01869775727391243,
-0.03635234385728836,
-0.00440613180398941,
-0.047425612807273865,
0.07503484934568405,
0.013807781040668488,
-0.008604868315160275,
-0.0946885272860527,
0.0028845302294939756,
0.07835888117551804,
-0.004999751690775156,
-0.0558682344853878,
-0.009541266597807407,
-0.021075546741485596,
-0.024083618074655533,
-0.04220881685614586,
0.15041932463645935,
-0.0001968669384950772,
0.0698845311999321,
-0.016259005293250084,
0.01991146244108677,
-0.007006316911429167,
0.002403066260740161,
-0.015523374080657959,
-0.00042977582779712975,
0.056028593331575394,
-0.063215471804142,
0.0675688162446022,
0.022111766040325165,
-0.11010722815990448,
-0.04338342696428299,
0.013304084539413452,
0.034109584987163544,
-0.019359968602657318,
-0.09487045556306839,
-0.012198706157505512,
0.005797643214464188,
0.056626033037900925,
-0.030355924740433693,
0.1001499593257904,
0.07480019330978394,
-0.0786733627319336,
-0.04817749187350273,
0.043472666293382645,
-0.021337663754820824,
0.015653487294912338,
0.0067183515056967735,
-0.0627046674489975,
-0.0010979627259075642,
0.07229141891002655,
0.05876582860946655,
0.10569404065608978,
-0.07748092710971832,
-0.04207858815789223,
0.028223667293787003,
0.09524786472320557,
-0.00042128574568778276,
0.06664010137319565,
0.03228988125920296,
-0.08247244358062744,
0.019205335527658463,
-0.04085999354720116,
0.02754267305135727,
-0.10341330617666245,
-3.7276912223660474e-8,
-0.024774562567472458,
0.028091032058000565,
-0.024658270180225372,
-0.008219115436077118,
0.05127742886543274,
-0.08486954122781754,
-0.0006848638877272606,
0.07668296247720718,
-0.033109381794929504,
0.010893597267568111,
0.057394880801439285,
-0.018044983968138695,
0.04887803643941879,
-0.004221641458570957,
0.04849538207054138,
-0.022449472919106483,
0.005523370113223791,
-0.05998161435127258,
-0.0804511085152626,
0.0028543216176331043,
0.08845221251249313,
0.040469810366630554,
-0.031874239444732666,
0.08126457780599594,
0.08490268141031265,
-0.11680406332015991,
-0.04329601675271988,
0.07155062258243561,
0.06991955637931824,
0.033573854714632034,
0.03602801635861397,
0.00974566861987114,
-0.002260385314002633,
0.005965883377939463,
-0.014683766290545464,
0.07351163029670715,
0.011849121190607548,
-0.03974618762731552,
-0.03149247542023659,
-0.04448404535651207,
-0.016000712290406227,
0.022678006440401077,
-0.04645301774144173,
0.02994333766400814,
0.051791854202747345,
-0.03270256146788597,
-0.04950105771422386,
-0.05187692120671272,
-0.0055422051809728146,
0.02378186397254467,
-0.028838051483035088,
-0.05074078217148781,
0.004013491328805685,
0.020261075347661972,
0.09481371939182281,
0.048952944576740265,
0.02029803767800331,
-0.0016537829069420695,
0.02660326287150383,
0.03880837932229042,
0.07726123183965683,
0.043243277817964554,
0.012202771380543709,
0.0034713824279606342
] |
DeepGate4: Efficient and Effective Representation Learning for Circuit Design at Scale | https://openreview.net/forum?id=b10lRabU9W | [
"Ziyang Zheng",
"Shan Huang",
"Jianyuan Zhong",
"Zhengyuan Shi",
"Guohao Dai",
"Ningyi Xu",
"Qiang Xu"
] | Poster | Circuit representation learning has become pivotal in electronic design automation, enabling critical tasks such as testability analysis, logic reasoning, power estimation, and SAT solving. However, existing models face significant challenges in scaling to large circuits due to limitations like over-squashing in graph neural networks and the quadratic complexity of transformer-based models. To address these issues, we introduce \textbf{DeepGate4}, a scalable and efficient graph transformer specifically designed for large-scale circuits. DeepGate4 incorporates several key innovations: (1) an update strategy tailored for circuit graphs, which reduce memory complexity to sub-linear and is adaptable to any graph transformer; (2) a GAT-based sparse transformer with global and local structural encodings for AIGs; and (3) an inference acceleration CUDA kernel that fully exploit the unique sparsity patterns of AIGs. Our extensive experiments on the ITC99 and EPFL benchmarks show that DeepGate4 significantly surpasses state-of-the-art methods, achieving 15.5\% and 31.1\% performance improvements over the next-best models. Furthermore, the Fused-DeepGate4 variant reduces runtime by 35.1\% and memory usage by 46.8\%, making it highly efficient for large-scale circuit analysis. These results demonstrate the potential of DeepGate4 to handle complex EDA tasks while offering superior scalability and efficiency. | circuit representation learning, graph transformer | null | 212 | 2502.01681 | [
-0.04765821620821953,
0.00926669780164957,
0.012534569017589092,
0.02863817848265171,
0.03156997635960579,
-0.061405349522829056,
-0.11618003249168396,
0.023054014891386032,
-0.050948500633239746,
-0.010661561973392963,
-0.025926638394594193,
-0.028975456953048706,
-0.03217039257287979,
0.06696520000696182,
0.002238425426185131,
0.07545207440853119,
-0.00244159041903913,
0.05463657155632973,
-0.07603208720684052,
-0.11287587881088257,
0.016555476933717728,
0.006865460891276598,
-0.025203192606568336,
-0.0712728425860405,
0.05578303709626198,
0.031688570976257324,
0.051791995763778687,
-0.06416375935077667,
-0.004817141219973564,
-0.05322391167283058,
0.02702372521162033,
-0.047647420316934586,
-0.08230390399694443,
0.04598042741417885,
-0.008128475397825241,
0.051726870238780975,
-0.014272617176175117,
-0.007016213610768318,
0.08980808407068253,
-0.04412330314517021,
0.02908685989677906,
-0.020975394174456596,
0.03532956540584564,
0.05032194033265114,
0.08778968453407288,
0.034585125744342804,
-0.008382663130760193,
-0.03392684459686279,
0.059912145137786865,
-0.03853674978017807,
0.05112941563129425,
-0.012029926292598248,
-0.0004810276732314378,
0.019471120089292526,
-0.009143399074673653,
0.0018524210900068283,
0.02956438809633255,
-0.06013198569417,
0.005704203154891729,
0.04486221447587013,
0.011291885748505592,
-0.08233223855495453,
0.05026594549417496,
0.015989404171705246,
-0.10632267594337463,
0.05855603516101837,
0.10806017369031906,
-0.01569555513560772,
-0.02935834415256977,
0.021794063970446587,
0.003771387040615082,
0.05701086297631264,
-0.05505344271659851,
0.05293886736035347,
0.07654958963394165,
0.028910905122756958,
0.06103779748082161,
-0.059834301471710205,
0.055208563804626465,
-0.025733569636940956,
0.00019852844707202166,
0.03676426410675049,
0.0015004286542534828,
0.013268690556287766,
0.14415478706359863,
-0.04773244261741638,
-0.03472485765814781,
0.02049698494374752,
-0.027826478704810143,
-0.0005238950252532959,
0.011283292435109615,
-0.05048360675573349,
-0.010402468033134937,
0.016508251428604126,
-0.038841888308525085,
-0.02400106191635132,
0.007747587747871876,
-0.06140130013227463,
-0.09927797317504883,
0.09534913301467896,
-0.007134077604860067,
0.05669833719730377,
0.04664873704314232,
0.002119137439876795,
-0.027712270617485046,
0.019780512899160385,
0.018075764179229736,
0.08373548835515976,
-0.019996928051114082,
-0.08915309607982635,
0.0266396664083004,
0.0822870209813118,
0.016108421608805656,
0.06675645709037781,
0.01634182035923004,
-0.02365552820265293,
-0.014803755097091198,
-0.047416217625141144,
-0.005712132900953293,
0.045421961694955826,
-0.07885794341564178,
0.08540353178977966,
-0.06176561862230301,
0.05266736447811127,
-0.05440409854054451,
-0.01558163482695818,
-0.1180514246225357,
4.897436399891972e-34,
-0.02493276074528694,
0.1046397015452385,
0.05177478864789009,
-0.01838202215731144,
0.10344067215919495,
0.06251391768455505,
0.013439273461699486,
0.09379121661186218,
-0.03149668127298355,
0.04747851565480232,
-0.04479016363620758,
0.0070436131209135056,
-0.03702092170715332,
0.12257048487663269,
0.0026100720278918743,
0.008242578245699406,
-0.05068458989262581,
-0.015862731263041496,
0.027798784896731377,
-0.1262747347354889,
0.04094262793660164,
-0.034597672522068024,
0.019693341106176376,
0.017148515209555626,
-0.0118675846606493,
-0.032418619841337204,
0.041471004486083984,
0.03357608616352081,
-0.014871377497911453,
-0.02400045655667782,
-0.13760077953338623,
-0.01999255083501339,
0.02861376293003559,
0.012732470408082008,
0.015547473914921284,
-0.050200894474983215,
0.062436360865831375,
-0.09300145506858826,
0.03411322087049484,
-0.025070201605558395,
-0.03329421579837799,
0.02819838747382164,
0.017931899055838585,
-0.0859924778342247,
-0.08709216862916946,
0.01452069915831089,
0.004003418143838644,
-0.03303425386548042,
-0.00022198830265551805,
-0.04268937557935715,
-0.014663666486740112,
-0.04622635617852211,
-0.035775814205408096,
0.0134861059486866,
0.0835317000746727,
-0.03881630301475525,
0.0645226389169693,
0.06616393476724625,
0.12170577794313431,
0.09110705554485321,
-0.02895670011639595,
0.029700225219130516,
-0.075310617685318,
0.06871777027845383,
-0.0195083636790514,
0.08961591124534607,
0.0013720616698265076,
0.01821124739944935,
-0.02131885290145874,
-0.03022618032991886,
0.01136757992208004,
0.034605156630277634,
-0.029429800808429718,
-0.06848403066396713,
0.03938070684671402,
-0.02492816373705864,
0.044412851333618164,
-0.04639294743537903,
-0.03501289710402489,
0.012640847824513912,
-0.05056983605027199,
-0.014404607005417347,
0.005605264566838741,
0.025268858298659325,
-0.018982121720910072,
-0.06235417351126671,
0.03389719873666763,
-0.03535541892051697,
0.04861069470643997,
-0.06055457517504692,
-0.04487883672118187,
-0.018710972741246223,
0.023842185735702515,
-0.03396522253751755,
-0.00486858980730176,
-3.4591155171389853e-34,
-0.023527806624770164,
0.07384177297353745,
-0.0016603203257545829,
0.07528483122587204,
-0.011155335232615471,
-0.09427153319120407,
-0.02471407875418663,
-0.024184800684452057,
-0.007532662246376276,
-0.03144189342856407,
0.003930563572794199,
-0.02478131465613842,
0.029569391161203384,
-0.09272998571395874,
0.08685434609651566,
0.03720010071992874,
-0.10264725983142853,
-0.020965851843357086,
-0.01043369434773922,
-0.010030322708189487,
-0.00960441306233406,
0.12571610510349274,
-0.06407018005847931,
-0.03636348620057106,
-0.007373271510004997,
0.022935397922992706,
-0.04775144159793854,
0.02640704996883869,
0.04454287886619568,
0.03315865248441696,
-0.007180061191320419,
0.003083796938881278,
-0.02167184092104435,
0.03982705995440483,
0.005396687425673008,
0.0587092824280262,
0.0708402767777443,
0.007502058520913124,
0.028493378311395645,
-0.007023600861430168,
0.023508232086896896,
-0.010944119654595852,
-0.08153137564659119,
0.041935596615076065,
-0.04573383554816246,
0.07561811059713364,
0.015877775847911835,
0.004695336800068617,
-0.028165005147457123,
0.051341619342565536,
-0.032145559787750244,
0.03483257070183754,
0.0033356393687427044,
-0.0014687824295833707,
-0.020716141909360886,
-0.09590044617652893,
0.02005055919289589,
0.012340722605586052,
0.069086454808712,
-0.05303516983985901,
-0.05972952023148537,
-0.057624172419309616,
0.033964406698942184,
-0.10125845670700073,
-0.03850486874580383,
-0.04596700146794319,
-0.05504949763417244,
0.032457612454891205,
0.015007131733000278,
0.11035393923521042,
-0.04593578726053238,
-0.03580698370933533,
-0.03791796416044235,
-0.03541785478591919,
-0.041143447160720825,
-0.012635847553610802,
-0.009018616750836372,
-0.032782427966594696,
-0.008781696669757366,
0.0004818706074729562,
0.00029092023032717407,
-0.0029502902179956436,
-0.030884847044944763,
-0.012467559427022934,
0.02458846941590309,
0.019634796306490898,
0.04054943844676018,
0.09220897406339645,
0.07070652395486832,
0.05982028320431709,
-0.09457695484161377,
0.12954849004745483,
-0.02714090421795845,
0.04910949990153313,
-0.052122652530670166,
-3.582556118431057e-8,
-0.05708492174744606,
0.060947220772504807,
0.008101564832031727,
-0.0253764521330595,
0.05273917317390442,
-0.1190403401851654,
0.07787109911441803,
0.10002423822879791,
-0.03002888150513172,
0.04526034742593765,
0.1271565705537796,
-0.050722576677799225,
-0.0006017257692292333,
-0.06705118715763092,
0.017705615609884262,
0.09152216464281082,
0.06307928264141083,
0.0047799223102629185,
0.06079641729593277,
0.02463267743587494,
-0.08040604740381241,
0.02973385900259018,
-0.006581651978194714,
0.006249553058296442,
0.03917575627565384,
-0.13120435178279877,
-0.005165662616491318,
0.03892034292221069,
0.03657979145646095,
0.02898334339261055,
-0.06213732436299324,
0.05625235289335251,
0.07993076741695404,
-0.06174834817647934,
0.05635333061218262,
0.07627935707569122,
0.054376546293497086,
-0.005641116760671139,
0.04227902367711067,
0.0051842969842255116,
-0.03652834892272949,
0.047982342541217804,
-0.09210094064474106,
0.00040820351568982005,
-0.11033560335636139,
-0.039916738867759705,
-0.03459257632493973,
-0.013531678356230259,
0.04437782242894173,
0.019623616710305214,
-0.002911405172199011,
0.0028911209665238857,
-0.05397928133606911,
0.005396187771111727,
0.006518543232232332,
0.05632641911506653,
-0.03432067856192589,
-0.028695104643702507,
0.031016089022159576,
0.05416007339954376,
0.0024121443275362253,
-0.023275483399629593,
-0.0523461252450943,
-0.006612295284867287
] |
TGB-Seq Benchmark: Challenging Temporal GNNs with Complex Sequential Dynamics | https://openreview.net/forum?id=8e2LirwiJT | [
"Lu Yi",
"Jie Peng",
"Yanping Zheng",
"Fengran Mo",
"Zhewei Wei",
"Yuhang Ye",
"Yue Zixuan",
"Zengfeng Huang"
] | Poster | Future link prediction is a fundamental challenge in various real-world dynamic systems. To address this, numerous temporal graph neural networks (temporal GNNs) and benchmark datasets have been developed. However, these datasets often feature excessive repeated edges and lack complex sequential dynamics, a key characteristic inherent in many real-world applications such as recommender systems and "Who-To-Follow" on social networks. This oversight has led existing methods to inadvertently downplay the importance of learning sequential dynamics, focusing primarily on predicting repeated edges.
In this study, we demonstrate that existing methods, such as GraphMixer and DyGFormer, are inherently incapable of learning simple sequential dynamics, such as "a user who has followed OpenAI and Anthropic is more likely to follow AI at Meta next." Motivated by this issue, we introduce the Temporal Graph Benchmark with Sequential Dynamics (TGB-Seq), a new benchmark carefully curated to minimize repeated edges, challenging models to learn sequential dynamics and generalize to unseen edges. TGB-Seq comprises large real-world datasets spanning diverse domains, including e-commerce interactions, movie ratings, business reviews, social networks, citation networks and web link networks. Benchmarking experiments reveal that current methods usually suffer significant performance degradation and incur substantial training costs on TGB-Seq, posing new challenges and opportunities for future research. TGB-Seq datasets, leaderboards, and example codes are available at https://tgb-seq.github.io/. | datasets and benchmarks, temporal graph learning | We introduce TGB-Seq, a novel temporal graph benchmark with diverse real-world datasets designed to emphasize sequential dynamics over repeated edges, demonstrating that existing temporal GNNs struggle to learn complex sequential dynamics. | 206 | 2502.02975 | [
-0.07562977820634842,
-0.11638462543487549,
0.019406799226999283,
0.040918439626693726,
0.03736627474427223,
0.04223206266760826,
-0.015275479294359684,
-0.038247428834438324,
0.01746128872036934,
-0.08872951567173004,
-0.038496773689985275,
-0.04009421542286873,
-0.04654959589242935,
0.10276743024587631,
-0.022903291508555412,
0.006997587624937296,
0.08958913385868073,
-0.026372743770480156,
-0.030168741941452026,
-0.09562815725803375,
-0.0066826604306697845,
0.045822907239198685,
0.032208703458309174,
-0.002585877198725939,
0.0578555166721344,
-0.024387730285525322,
-0.03884478658437729,
0.0036309852730482817,
0.03538329154253006,
0.01042740698903799,
0.11629214137792587,
-0.013473257422447205,
-0.052421439439058304,
0.010869559831917286,
-0.0680290162563324,
-0.024587932974100113,
-0.11252622306346893,
-0.005724108777940273,
0.03343420475721359,
-0.04761167988181114,
0.062519870698452,
-0.0023851317819207907,
0.02130761183798313,
0.047097690403461456,
0.05797599256038666,
0.0008026036666706204,
0.0020664480980485678,
-0.013571653515100479,
-0.015146908350288868,
0.044174037873744965,
-0.11127403378486633,
-0.003759770654141903,
-0.011279412545263767,
0.07647863030433655,
0.03259720280766487,
0.029783494770526886,
-0.0017989912303164601,
0.042300786823034286,
0.062329601496458054,
-0.018566083163022995,
0.04513958469033241,
-0.13963158428668976,
0.02866695635020733,
-0.028779275715351105,
-0.026822812855243683,
0.06348337978124619,
-0.04297687113285065,
0.054131537675857544,
0.008514107204973698,
0.10224758833646774,
0.046822838485240936,
0.03442296013236046,
-0.06850963830947876,
0.06906028091907501,
0.0007958101341500878,
0.05114152282476425,
0.027766456827521324,
-0.036772094666957855,
-0.008242004550993443,
-0.03628513216972351,
-0.017452267929911613,
-0.056365687400102615,
0.0087819155305624,
0.02837236598134041,
0.07378586381673813,
-0.0214818324893713,
-0.01264919899404049,
-0.0009667930426076055,
-0.056588150560855865,
0.04279543459415436,
-0.01938575506210327,
0.05298062041401863,
0.03233812004327774,
-0.02299412339925766,
0.08010903745889664,
0.054477792233228683,
0.024145755916833878,
-0.03477789834141731,
0.009011786431074142,
0.06431743502616882,
-0.043406952172517776,
0.03018280304968357,
0.006909796968102455,
0.06744207441806793,
0.012985077686607838,
-0.039722930639982224,
0.019599460065364838,
0.07136504352092743,
0.007923190481960773,
-0.06412848085165024,
-0.003228902816772461,
0.09270644187927246,
0.08471807092428207,
0.010109036229550838,
0.07264938950538635,
-0.04028935357928276,
0.001239124801941216,
0.02500559762120247,
0.07379698753356934,
0.1187053695321083,
0.010789892636239529,
0.038217999041080475,
0.0023936042562127113,
-0.019313761964440346,
-0.02063506655395031,
0.02677883580327034,
-0.018158582970499992,
2.2578458353632786e-33,
0.0723993107676506,
-0.008018378168344498,
0.0677623525261879,
-0.0576932467520237,
0.008564048446714878,
0.026265207678079605,
-0.032097626477479935,
0.04094597324728966,
-0.04969950392842293,
0.015848087146878242,
-0.10881426930427551,
-0.023488998413085938,
-0.04096168652176857,
0.10222069174051285,
-0.0314955972135067,
-0.01600295677781105,
-0.027542907744646072,
-0.0073800403624773026,
0.017194153741002083,
-0.048160627484321594,
0.0544520728290081,
-0.04104403033852577,
-0.0031661028042435646,
0.022665830329060555,
0.02276664786040783,
0.017681073397397995,
0.08441201597452164,
0.02720707841217518,
0.11058861762285233,
0.011342587880790234,
-0.05516721308231354,
0.03625837713479996,
-0.036902572959661484,
0.03654829040169716,
0.027861714363098145,
-0.012195535935461521,
-0.025713907554745674,
-0.04046250879764557,
-0.008796933107078075,
-0.08238253742456436,
0.00012655457248911262,
0.03387316316366196,
-0.05351311340928078,
-0.017655078321695328,
-0.10478807240724564,
-0.08392013609409332,
0.02143222652375698,
-0.01658780500292778,
-0.058020103722810745,
-0.0020312019623816013,
0.017022110521793365,
-0.02600310742855072,
0.02989753521978855,
-0.03108765184879303,
-0.031262002885341644,
-0.06592635065317154,
0.049724601209163666,
0.011388341896235943,
0.007720346562564373,
0.056020937860012054,
0.047117412090301514,
0.002100097481161356,
-0.012743230909109116,
-0.01876126416027546,
-0.02267502062022686,
0.04820496588945389,
-0.03150662034749985,
0.12784063816070557,
0.010371636599302292,
-0.06907596439123154,
0.031420301645994186,
0.0927983820438385,
0.015809668228030205,
-0.07269943505525589,
0.04023043438792229,
0.03174344822764397,
-0.01605316810309887,
-0.13459385931491852,
-0.02402905747294426,
0.036624759435653687,
-0.0659460797905922,
-0.06759754568338394,
-0.0024307200219482183,
0.0302291102707386,
-0.06199907138943672,
-0.05478893220424652,
0.0684804618358612,
0.007832206785678864,
0.033608436584472656,
-0.035638995468616486,
-0.05042952671647072,
-0.030214635655283928,
0.0056740036234259605,
0.11164027452468872,
0.01998456008732319,
-5.3521566448759e-34,
0.012890402227640152,
0.029441479593515396,
-0.001291069551371038,
0.05780516192317009,
0.06782753020524979,
-0.07665415853261948,
-0.07726456224918365,
-0.01573963277041912,
0.02820657566189766,
-0.034139249473810196,
0.042380914092063904,
-0.09688913822174072,
0.0546974278986454,
-0.02260034717619419,
0.06717265397310257,
0.00046403234591707587,
-0.01454772986471653,
-0.018325427547097206,
-0.06384854018688202,
0.023493286222219467,
0.035768039524555206,
0.0163738913834095,
-0.12734302878379822,
-0.03698921576142311,
0.008200711570680141,
0.03482302650809288,
0.050296444445848465,
0.018388260155916214,
-0.02148960530757904,
0.016939515247941017,
-0.0011880892561748624,
-0.05837029963731766,
-0.06842318177223206,
-0.0423671118915081,
-0.022773487493395805,
0.09089680016040802,
0.0288094449788332,
-0.07204536348581314,
0.019830428063869476,
0.021203091368079185,
0.08183275908231735,
-0.027010610327124596,
-0.06263069063425064,
0.00983654148876667,
-0.012509094551205635,
0.09054142236709595,
-0.07469774037599564,
0.08330724388360977,
-0.06791815161705017,
0.049371007829904556,
-0.03090430609881878,
0.10194029659032822,
-0.03141398727893829,
-0.047835733741521835,
-0.04787919670343399,
-0.07111731916666031,
0.03452754020690918,
0.05242788419127464,
0.0337364599108696,
0.006394430994987488,
-0.009130027145147324,
-0.01736212521791458,
0.00426330603659153,
-0.029942847788333893,
-0.05002473667263985,
-0.011936483904719353,
0.005091970320791006,
-0.030176246538758278,
0.005740285851061344,
-0.021592983976006508,
0.011340518482029438,
-0.010373742319643497,
-0.08786002546548843,
-0.045220982283353806,
0.006430540233850479,
-0.030959757044911385,
-0.05201779678463936,
0.0011341875651851296,
0.0034975355956703424,
-0.04954095929861069,
-0.01377745158970356,
0.012480129487812519,
0.04005279019474983,
-0.014558534137904644,
0.028627993538975716,
0.09459046274423599,
-0.015145715326070786,
0.08395648747682571,
0.06562402099370956,
0.07885954529047012,
-0.05157994106411934,
-0.034289244562387466,
-0.07223431766033173,
0.04091696813702583,
-0.059840790927410126,
-3.686410110503857e-8,
-0.17246320843696594,
0.005122397094964981,
0.044302355498075485,
0.05446047708392143,
0.05724145472049713,
0.032777249813079834,
0.09756188839673996,
0.11608803272247314,
0.027378151193261147,
0.00795036368072033,
0.13394804298877716,
-0.029753029346466064,
-0.0209952499717474,
-0.06300386041402817,
-0.056170910596847534,
-0.023019790649414062,
0.06864909827709198,
-0.08258925378322601,
0.06136568263173103,
-0.01247607171535492,
-0.0075978166423738,
-0.005914038047194481,
0.06285319477319717,
-0.015981711447238922,
0.010332399047911167,
-0.045015621930360794,
-0.03929479047656059,
0.07020166516304016,
0.10255859047174454,
0.04345417022705078,
-0.03843820467591286,
-0.0042076692916452885,
0.06133553758263588,
-0.0869402289390564,
-0.07959005236625671,
0.06350237131118774,
-0.009802352637052536,
-0.045775312930345535,
-0.023981556296348572,
-0.009637759998440742,
0.03011825866997242,
0.05881831422448158,
-0.008895469829440117,
-0.002189353108406067,
-0.027651552110910416,
-0.010919428430497646,
-0.01944563165307045,
-0.03227001428604126,
0.05599196255207062,
-0.03553251177072525,
0.0002516184176784009,
-0.10999930649995804,
0.005478809121996164,
0.06002631410956383,
0.08991950750350952,
0.032448552548885345,
-0.0014031893806532025,
-0.006175357848405838,
0.05339614674448967,
0.04813450574874878,
0.009359442628920078,
-0.017172815278172493,
-0.12359848618507385,
0.0013143616961315274
] |
ContraDiff: Planning Towards High Return States via Contrastive Learning | https://openreview.net/forum?id=XMOaOigOQo | [
"Yixiang Shan",
"Zhengbang Zhu",
"Ting Long",
"Liang Qifan",
"Yi Chang",
"Weinan Zhang",
"Liang Yin"
] | Poster | The performance of offline reinforcement learning (RL) is sensitive to the proportion of high-return trajectories in the offline dataset. However, in many simulation environments and real-world scenarios, there are large ratios of low-return trajectories rather than high-return trajectories, which makes learning an efficient policy challenging. In this paper, we propose a method called Contrastive Diffuser (ContraDiff) to make full use of low-return trajectories and improve the performance of offline RL algorithms. Specifically, ContraDiff groups the states of trajectories in the offline dataset into high-return states and low-return states and treats them as positive and negative samples correspondingly. Then, it designs a contrastive mechanism to pull the planned trajectory of an agent toward high-return states and push them away from low-return states. Through the contrast mechanism, trajectories with low returns can serve as negative examples for policy learning, guiding the agent to avoid areas associated with low returns and achieve better performance. Through the contrast mechanism, trajectories with low returns provide a ``counteracting force'' guides the agent to avoid areas associated with low returns and achieve better performance.
Experiments on 27 sub-optimal datasets demonstrate the effectiveness of our proposed method. Our code is publicly available at https://github.com/Looomo/contradiff. | Offline Reinforcement Learning, Decision Making, Diffusion Models, Machine Learning | We propose CDiffuser, which devises a return contrast mechanism and enhances the trajectory generation-based diffusion RL by pulling the states in generated trajectories towards high-return states while pushing them away from low-return states. | 201 | null | [
-0.040179673582315445,
-0.08255962282419205,
0.04816703498363495,
0.065725177526474,
0.039106570184230804,
-0.03372414782643318,
-0.05302456021308899,
-0.02833130583167076,
0.0309399776160717,
0.0274607315659523,
-0.04524083808064461,
0.00755373714491725,
0.0005315114394761622,
0.024341288954019547,
-0.02205362357199192,
0.040541138499975204,
0.11696323752403259,
0.07195556163787842,
-0.05965714901685715,
-0.041487764567136765,
-0.02688680961728096,
0.01790984347462654,
0.03875751420855522,
-0.03789042681455612,
-0.0017468113219365478,
0.02796665020287037,
-0.0118568055331707,
-0.043385230004787445,
0.018020981922745705,
-0.037661705166101456,
0.08701842278242111,
0.007048366125673056,
-0.052235197275877,
-0.09208831191062927,
-0.043311670422554016,
0.08220034092664719,
-0.12451816350221634,
-0.0153509471565485,
-0.020244300365447998,
0.05812516063451767,
0.010212398134171963,
0.058259736746549606,
-0.025623150169849396,
-0.021018026396632195,
0.05790221318602562,
-0.011370865628123283,
0.0025982202496379614,
-0.014998116530478,
0.061880044639110565,
-0.03707735612988472,
-0.04824095964431763,
-0.00020987263997085392,
-0.03441888093948364,
0.04868542402982712,
-0.008513841778039932,
0.046714019030332565,
0.037849970161914825,
0.013809210620820522,
-0.02195350080728531,
-0.10139918327331543,
-0.017922451719641685,
-0.08156169950962067,
-0.019461825489997864,
-0.09508099406957626,
-0.06535454839468002,
0.007151056546717882,
-0.018908526748418808,
0.05909731984138489,
0.028489669784903526,
0.04911042004823685,
-0.03646719083189964,
0.045005641877651215,
0.0159920584410429,
-0.005792137235403061,
0.03216215968132019,
-0.012111375108361244,
0.006265671923756599,
0.034683361649513245,
0.033355917781591415,
-0.1029442548751831,
0.05517546087503433,
-0.05949262157082558,
-0.017409613355994225,
0.03184929117560387,
0.0709676593542099,
-0.01835671439766884,
-0.031168226152658463,
0.021469544619321823,
0.14443396031856537,
-0.008199432864785194,
0.010606887750327587,
0.07784556597471237,
0.0471574030816555,
-0.02124619483947754,
-0.09796421229839325,
0.011428441852331161,
0.019481156021356583,
-0.03298192471265793,
-0.025406453758478165,
0.10777455568313599,
0.0004410180845297873,
0.025947431102395058,
-0.02129979431629181,
0.018048105761408806,
-0.003094463376328349,
-0.08953983336687088,
0.05856573209166527,
-0.019332069903612137,
0.03262842074036598,
-0.05552041530609131,
0.04592271149158478,
0.0069314781576395035,
0.03765324130654335,
0.03810746595263481,
-0.002266890602186322,
0.0016108579002320766,
-0.012631147168576717,
-0.006525327451527119,
-0.04958643391728401,
0.02796786092221737,
0.059354230761528015,
-0.07335089147090912,
0.03430267423391342,
0.06264091283082962,
0.04674152657389641,
-0.10027885437011719,
-0.024835238233208656,
3.522897060012885e-33,
0.04464908689260483,
-0.06867999583482742,
0.038399528712034225,
-0.006817579735070467,
0.03856266289949417,
-0.036543671041727066,
0.02919987589120865,
-0.08978842198848724,
0.001098365057259798,
0.024120062589645386,
0.04988908767700195,
0.036590754985809326,
-0.05176416039466858,
0.11292031407356262,
0.03347529098391533,
-0.11119253188371658,
-0.015505442395806313,
0.018911991268396378,
-0.011119074188172817,
-0.020545082166790962,
0.01062087807804346,
-0.008731141686439514,
-0.0007491763681173325,
-0.09897324442863464,
-0.007161312736570835,
0.03471406549215317,
-0.013520677573978901,
-0.0027933453675359488,
0.026901789009571075,
-0.002758418908342719,
0.005564217455685139,
0.005418934393674135,
-0.030845189467072487,
0.08240152895450592,
0.021007956936955452,
0.0377909392118454,
-0.10492559522390366,
-0.045615632086992264,
-0.009038195013999939,
0.014645732939243317,
-0.05551372468471527,
0.027302784845232964,
-0.019357362762093544,
0.025012092664837837,
-0.04834913834929466,
-0.1270027607679367,
0.061582066118717194,
-0.05332313850522041,
-0.04769288748502731,
-0.0023494164925068617,
0.02059028670191765,
-0.007533687632530928,
0.006823247764259577,
-0.1396399587392807,
-0.06112530827522278,
0.022555461153388023,
0.0018497823039069772,
0.08521492779254913,
-0.038435354828834534,
0.051092151552438736,
-0.00024518341524526477,
-0.03181079775094986,
-0.007072258275002241,
0.07954314351081848,
-0.03820139914751053,
0.0870048776268959,
-0.05549580231308937,
0.05006197839975357,
-0.00021256635955069214,
0.0007434476865455508,
0.010453691706061363,
0.09502721577882767,
0.05091877281665802,
-0.06695817410945892,
0.13733267784118652,
-0.02956756390631199,
0.055351853370666504,
-0.005467424634844065,
0.004960222635418177,
-0.042672593146562576,
-0.0027740944642573595,
-0.0002785375399980694,
-0.09471546858549118,
-0.05091989040374756,
-0.030290137976408005,
-0.040075890719890594,
0.005116819404065609,
-0.06356292963027954,
-0.049524836242198944,
-0.021657995879650116,
-0.16060519218444824,
-0.031700074672698975,
-0.01084986887872219,
0.026190921664237976,
0.08970765769481659,
-4.872688570387317e-33,
0.028273984789848328,
-0.040527161210775375,
0.001334036816842854,
0.047678519040346146,
-0.0642167255282402,
0.04602234065532684,
0.05296409875154495,
-0.023044947534799576,
0.04346923157572746,
0.019558683037757874,
-0.04343356564640999,
0.06324063986539841,
0.04004652425646782,
-0.012459833174943924,
0.024109849706292152,
-0.06122470647096634,
0.08327072858810425,
0.02129337191581726,
0.010287772864103317,
-0.011121989227831364,
-0.004382343031466007,
0.06832291930913925,
-0.05007011815905571,
0.03329514339566231,
-0.0405319519340992,
0.006200625095516443,
-0.024807587265968323,
0.04796258360147476,
-0.054436150938272476,
-0.042146410793066025,
-0.053831979632377625,
-0.008455599658191204,
-0.00036892565549351275,
-0.041763607412576675,
-0.07872585952281952,
0.12351083755493164,
-0.025241751223802567,
0.024297839030623436,
-0.12972702085971832,
0.11613766849040985,
0.0567421019077301,
-0.07619497179985046,
-0.038161877542734146,
-0.01682952791452408,
0.00018063631432596594,
-0.012551471590995789,
-0.09155789762735367,
0.019599545747041702,
0.03263228014111519,
-0.002985605038702488,
-0.0051392861641943455,
0.0530472993850708,
-0.03869757801294327,
0.0174604170024395,
-0.03283815085887909,
-0.024032993242144585,
-0.006601231638342142,
0.009632445871829987,
0.05778126418590546,
0.031447868794202805,
-0.020040396600961685,
-0.022975381463766098,
-0.04937158152461052,
0.010229388251900673,
-0.02373964712023735,
-0.010020959191024303,
-0.04914121702313423,
-0.00644471263512969,
0.08138160407543182,
-0.0263576190918684,
0.018708599731326103,
0.028535282239317894,
0.05185862258076668,
-0.023355761542916298,
0.08606461435556412,
-0.037158600986003876,
-0.036889683455228806,
-0.0073487525805830956,
-0.03510091081261635,
0.0543951541185379,
0.01754891313612461,
0.01853886991739273,
0.002465377561748028,
0.036115922033786774,
0.02360718883574009,
0.05960839241743088,
-0.0673825740814209,
0.025145461782813072,
0.04522215574979782,
-0.0023484758567065,
0.010634585283696651,
-0.023284710943698883,
0.09870542585849762,
0.040610216557979584,
-0.058771342039108276,
-5.058703322902147e-8,
-0.09953707456588745,
-0.011570006608963013,
0.17437945306301117,
-0.02805587463080883,
0.06735430657863617,
-0.007800532970577478,
-0.021764375269412994,
-0.01898832805454731,
-0.028616227209568024,
-0.0421425886452198,
0.09066256135702133,
-0.008647648617625237,
0.0021212869323790073,
-0.10245591402053833,
0.008926166221499443,
0.10258172452449799,
0.026341943070292473,
-0.07829313725233078,
0.000023522277842857875,
0.03442459926009178,
0.060941245406866074,
-0.017132433131337166,
-0.005389428231865168,
-0.030703529715538025,
0.07417654991149902,
-0.019471606239676476,
-0.024722004309296608,
0.01596582680940628,
0.023578770458698273,
-0.03279906511306763,
0.10149607062339783,
-0.018994653597474098,
0.08310927450656891,
-0.012279066257178783,
-0.0092172771692276,
0.0755942165851593,
0.07535868883132935,
-0.04351773113012314,
0.01898997277021408,
0.030900152400135994,
-0.018999241292476654,
0.0611790269613266,
-0.03580436855554581,
0.014122288674116135,
-0.0057485103607177734,
0.017188485711812973,
-0.0660637766122818,
-0.11068352311849594,
0.06199190020561218,
-0.0978011041879654,
0.017047205939888954,
-0.07818296551704407,
-0.023335281759500504,
0.05635452643036842,
0.09100506454706192,
-0.059621311724185944,
0.006739449687302113,
-0.021789666265249252,
0.008222689852118492,
0.08016793429851532,
0.04568440094590187,
0.040897026658058167,
-0.091991126537323,
0.030110163614153862
] |
FreeCG: Free the Design Space of Clebsch-Gordan Transform for Machine Learning Force Fields | https://openreview.net/forum?id=sfi2j1Ot6j | [
"Shihao Shao",
"Haoran Geng",
"Zun Wang",
"Qinghua Cui"
] | Poster | Machine Learning Force Fields (MLFFs) are of great importance for chemistry, physics, materials science, and many other related fields. The Clebsch–Gordan transform (CG transform) effectively encodes many-body interactions and is thus an important building block for many models of MLFFs. However, the permutation-equivariance requirement of MLFFs limits the design space of CG transform, that is, intensive CG transform has to be conducted for each neighboring edge and the operations should be performed in the same manner for all edges. Freeing up the design space can greatly improve the model's expressiveness while simultaneously decreasing computational demands. To reach this goal, we utilize a mathematical proposition, invariance transitivity, to show that implementing the CG transform layer on the permutation-invariant abstract edges allows complete freedom in the design of the layer without compromising the overall permutation equivariance. Developing on this free design space, we further propose group CG transform with sparse path, abstract edges shuffling, and attention enhancer to form a powerful and efficient CG transform layer. Our method, known as FreeCG, achieves state-of-the-art (SOTA) results in force prediction for MD17, rMD17, MD22, and is well extended to property prediction in QM9 datasets with several improvements greater than 15% and the maximum beyond 20%. The extensive real-world applications showcase high practicality. FreeCG introduces a novel paradigm for carrying out efficient and expressive CG transform in future geometric network designs. To demonstrate this, the recent SOTA, QuinNet, is also enhanced under our paradigm. Code: https://github.com/ShihaoShao-GH/FreeCG. | Machine Learning Force Fields, Equivariant Graph Neural Network, Clebsch-Gordan Transform | We present a novel SoTA MLFFs model with high practicality through freeing the design space of CG transform, together with a new paradigm for guiding the future designs of EGNNs. | 198 | 2407.02263 | [
-0.11424741894006729,
-0.0057839821092784405,
0.043357450515031815,
0.02430596947669983,
0.034477632492780685,
0.003907307982444763,
-0.07147438824176788,
-0.07970769703388214,
0.05336187779903412,
0.025602130219340324,
-0.03821352496743202,
-0.052993688732385635,
-0.0881117433309555,
-0.0071249548345804214,
-0.020930657163262367,
0.033784348517656326,
-0.023082619532942772,
0.12061578780412674,
-0.10536881536245346,
-0.005685216281563044,
-0.027983475476503372,
-0.030239365994930267,
0.009029651060700417,
-0.009745532646775246,
-0.009034569375216961,
-0.021237825974822044,
0.0356588177382946,
0.013227859511971474,
0.028138792142271996,
-0.07010329514741898,
0.050005245953798294,
0.021574966609477997,
-0.06368125975131989,
0.016843290999531746,
-0.10125114768743515,
0.09558672457933426,
-0.007825931534171104,
-0.022886674851179123,
-0.030885115265846252,
0.001640387694351375,
-0.02924889512360096,
0.013678085058927536,
0.046554166823625565,
0.07587793469429016,
0.08737153559923172,
0.04765244573354721,
0.04733065888285637,
-0.0751950591802597,
0.011507514864206314,
-0.02195279486477375,
0.08165756613016129,
-0.03797074034810066,
-0.0888410434126854,
0.03368283435702324,
0.009463254362344742,
-0.01132676750421524,
0.015496106818318367,
-0.051081690937280655,
0.0010692729847505689,
-0.04063943028450012,
0.01667073182761669,
-0.02565811201930046,
0.016036467626690865,
0.029131028801202774,
-0.03729928284883499,
0.007240673527121544,
0.08711649477481842,
0.06452616304159164,
-0.010572646744549274,
-0.043466921895742416,
-0.0153424721211195,
0.02493622526526451,
0.0010057331528514624,
0.032060008496046066,
0.05483859404921532,
0.1496443897485733,
0.025112004950642586,
-0.0013758041895925999,
0.07264389097690582,
-0.03879944235086441,
-0.00032988854218274355,
0.010065507143735886,
0.008689484558999538,
0.013927487656474113,
0.04946598410606384,
-0.011257575824856758,
-0.10212547332048416,
0.06228715181350708,
0.0025207295548170805,
0.03356866165995598,
0.03946078196167946,
0.04024522006511688,
-0.01669437624514103,
-0.08323369920253754,
-0.06808074563741684,
0.04035857692360878,
0.004052973352372646,
-0.049095235764980316,
-0.043408069759607315,
0.07022291421890259,
-0.04766925424337387,
0.018609359860420227,
0.0606260672211647,
0.023477822542190552,
0.019574463367462158,
-0.011107832193374634,
-0.0012235856847837567,
-0.010967443697154522,
-0.03915814310312271,
-0.06804464757442474,
0.04488414153456688,
0.06406985223293304,
-0.04388904199004173,
-0.016343969851732254,
-0.039610233157873154,
-0.020537998527288437,
-0.02743726223707199,
-0.010760008357465267,
0.09867648035287857,
0.07781565934419632,
-0.02934630587697029,
0.031796421855688095,
-0.044528018683195114,
0.05139390379190445,
0.005502522923052311,
-0.002547742100432515,
-0.09552334994077682,
3.506991152078321e-33,
0.0002543120353948325,
0.10959366708993912,
0.11655046045780182,
-0.015607557259500027,
0.039763953536748886,
-0.08136757463216782,
0.030900022014975548,
-0.009312552399933338,
-0.038253240287303925,
0.03199121728539467,
-0.10940613597631454,
0.014438463374972343,
0.00790323130786419,
-0.013468057848513126,
-0.033101070672273636,
-0.027352649718523026,
-0.002960133133456111,
-0.04220132157206535,
-0.018314365297555923,
-0.0700555145740509,
0.0005312677239999175,
0.015659378841519356,
-0.007899659685790539,
0.010675420984625816,
-0.03218710422515869,
0.07197469472885132,
0.012266844511032104,
-0.00871956069022417,
-0.02853279747068882,
-0.00033054756931960583,
-0.06674455851316452,
0.06978941708803177,
-0.022585593163967133,
0.07621851563453674,
-0.05170710012316704,
-0.009991835802793503,
-0.001281483331695199,
-0.10303623229265213,
0.044273167848587036,
-0.02130512148141861,
0.06594197452068329,
-0.002430408727377653,
0.030763883143663406,
-0.055637091398239136,
-0.031928595155477524,
0.0069917719811201096,
0.07191403955221176,
0.059905439615249634,
0.017867611721158028,
0.011909335851669312,
0.03231512010097504,
-0.02403363399207592,
0.013770059682428837,
-0.017806129530072212,
0.03089987114071846,
0.0050870198756456375,
0.06796897202730179,
0.06581000983715057,
0.07883060723543167,
0.032899778336286545,
-0.0023123733699321747,
0.10703692585229874,
-0.024591756984591484,
0.032856155186891556,
0.09771861881017685,
0.007358100730925798,
-0.06233673170208931,
0.017247233539819717,
0.018192516639828682,
-0.04392601177096367,
-0.022822950035333633,
0.026293044909834862,
-0.0163189098238945,
-0.0014595896936953068,
0.04044889286160469,
0.0022383341565728188,
-0.00898695643991232,
-0.1257784366607666,
-0.014710331335663795,
-0.01691361330449581,
-0.11753512173891068,
0.06139296293258667,
0.027485059574246407,
-0.04125995561480522,
-0.02275022864341736,
-0.0497557707130909,
0.005390730686485767,
-0.029268106445670128,
-0.032904837280511856,
-0.09053442627191544,
-0.02813497558236122,
-0.03848649188876152,
0.029079284518957138,
-0.005898348521441221,
0.012408547103404999,
-2.949827768281681e-33,
-0.04938279837369919,
0.016073770821094513,
-0.004980289842933416,
0.09968066215515137,
0.08591128140687943,
-0.04914357513189316,
-0.006924556102603674,
-0.047359783202409744,
0.015358411706984043,
0.026930643245577812,
0.05870276317000389,
-0.07049375027418137,
-0.0145250353962183,
-0.029534514993429184,
0.08170393109321594,
0.0020518871024250984,
-0.08227535337209702,
-0.043811287730932236,
-0.04873158410191536,
0.018703892827033997,
0.001349089783616364,
0.13258115947246552,
0.020232629030942917,
0.030506813898682594,
-0.026279769837856293,
-0.014573143795132637,
-0.03971534222364426,
-0.01910403184592724,
0.04822734370827675,
0.1022229865193367,
-0.004066715016961098,
-0.034002941101789474,
-0.021656595170497894,
0.015740366652607918,
0.02412065491080284,
-0.03449402377009392,
-0.0008125703898258507,
0.1101626306772232,
0.06971576064825058,
0.06368895620107651,
-0.010994688607752323,
-0.041029609739780426,
-0.07150277495384216,
0.07675106823444366,
-0.015016043558716774,
0.005193878896534443,
-0.0040945373475551605,
-0.05501639097929001,
-0.02140536718070507,
-0.0046266657300293446,
-0.056490495800971985,
0.08761502802371979,
-0.023615967482328415,
-0.0733676329255104,
-0.03443021699786186,
0.009639590978622437,
-0.01636979915201664,
-0.07843199372291565,
0.16464117169380188,
-0.012772975489497185,
-0.03229790925979614,
-0.04437241703271866,
-0.012151521630585194,
-0.05263429135084152,
0.034404657781124115,
0.019441045820713043,
-0.03037518449127674,
-0.006774062290787697,
-0.036609500646591187,
0.015337027609348297,
-0.03672600910067558,
-0.02052626758813858,
0.0026381835341453552,
-0.013909692876040936,
0.01639169082045555,
-0.06506048887968063,
0.10743989795446396,
-0.07211817800998688,
0.016125284135341644,
0.009941293857991695,
0.012919806875288486,
-0.040778178721666336,
-0.0017612603260204196,
0.08259505033493042,
-0.022189848124980927,
-0.014642360620200634,
0.018881745636463165,
0.0554945282638073,
0.02583480253815651,
-0.01096807699650526,
-0.04656028375029564,
0.1380074918270111,
0.09067413210868835,
0.09117047488689423,
0.021344276145100594,
-4.7166846428581266e-8,
-0.039047081023454666,
0.048541657626628876,
0.05884025990962982,
0.001930267782881856,
0.0651206225156784,
0.022620460018515587,
-0.011065421625971794,
0.025360435247421265,
-0.03616054728627205,
0.0118778171017766,
0.061805617064237595,
-0.003046865575015545,
-0.020152555778622627,
-0.013891070149838924,
-0.07491577416658401,
0.10472825169563293,
-0.046695511788129807,
-0.020120147615671158,
-0.03803504258394241,
-0.02793850004673004,
-0.0756024569272995,
0.011194949969649315,
0.012224245816469193,
-0.014736626297235489,
0.0747741162776947,
-0.1324523240327835,
-0.10708130151033401,
-0.034899357706308365,
0.10271035879850388,
0.029851101338863373,
-0.04268179088830948,
-0.03414607048034668,
0.08528758585453033,
0.08380023390054703,
-0.01607418619096279,
0.10122484713792801,
-0.0001595758367329836,
0.005852322559803724,
-0.011638223193585873,
0.0662931576371193,
-0.013456971384584904,
0.06468325108289719,
-0.05275854840874672,
0.008221468888223171,
-0.05417002737522125,
-0.0017429623985663056,
0.01666375808417797,
-0.10792454332113266,
-0.01558354590088129,
0.06734826415777206,
0.07128431648015976,
-0.06444039195775986,
-0.06086564436554909,
0.02169463410973549,
0.03028869442641735,
0.12422343343496323,
0.023244047537446022,
-0.010581557638943195,
0.09480348974466324,
-0.02502768114209175,
-0.045382991433143616,
-0.022174254059791565,
-0.06230475753545761,
-0.04516167938709259
] |
Diffusion2: Dynamic 3D Content Generation via Score Composition of Video and Multi-view Diffusion Models | https://openreview.net/forum?id=fectsEG2GU | [
"Zeyu Yang",
"Zijie Pan",
"Chun Gu",
"Li Zhang"
] | Poster | Recent advancements in 3D generation are predominantly propelled by improvements in 3D-aware image diffusion models. These models are pretrained on Internet-scale image data and fine-tuned on massive 3D data, offering the capability of producing highly consistent multi-view images. However, due to the scarcity of synchronized multi-view video data, it remains challenging to adapt this paradigm to 4D generation directly. Despite that, the available video and 3D data are adequate for training video and multi-view diffusion models separately that can provide satisfactory dynamic and geometric priors respectively. To take advantage of both, this paper presents Diffusion$^2$, a novel framework for dynamic 3D content creation that reconciles the knowledge about geometric consistency and temporal smoothness from these models to directly sample dense multi-view multi-frame images which can be employed to optimize continuous 4D representation. Specifically, we design a simple yet effective denoising strategy via score composition of pretrained video and multi-view diffusion models based on the probability structure of the target image array. To alleviate the potential conflicts between two heterogeneous scores, we further introduce variance-reducing sampling via interpolated steps, facilitating smooth and stable generation. Owing to the high parallelism of the proposed image generation process and the efficiency of the modern 4D reconstruction pipeline, our framework can generate 4D content within few minutes. Notably, our method circumvents the reliance on expensive and hard-to-scale 4D data, thereby having the potential to benefit from the scaling of the foundation video and multi-view diffusion models. Extensive experiments demonstrate the efficacy of our proposed framework in generating highly seamless and consistent 4D assets under various types of conditions. | 4D generation, diffusion model, generative model, video diffusion | null | 197 | 2404.02148 | [
0.04165571182966232,
-0.13849368691444397,
0.03196641430258751,
-0.019060304388403893,
0.04241970181465149,
-0.00551924854516983,
-0.09500306844711304,
-0.11782793700695038,
0.06482097506523132,
-0.04566327854990959,
-0.058089565485715866,
0.0006260118680074811,
0.00788270402699709,
0.08068112283945084,
-0.04883226379752159,
-0.009373518638312817,
0.035650983452796936,
0.07575458288192749,
-0.13046860694885254,
-0.019427327439188957,
-0.02983478643000126,
-0.07042043656110764,
-0.03682698309421539,
-0.04333391785621643,
0.07118421047925949,
0.052754174917936325,
0.08937029540538788,
-0.06632199138402939,
0.036635514348745346,
0.004495371598750353,
0.05537981539964676,
-0.0654531717300415,
-0.006315750535577536,
0.0031038173474371433,
-0.054239001125097275,
-0.024266252294182777,
-0.01886313408613205,
0.03810850530862808,
-0.015351954847574234,
0.008277441374957561,
0.0220942422747612,
0.07417450100183487,
0.0023789124097675085,
0.01974950172007084,
0.019241085276007652,
0.028545282781124115,
-0.033997997641563416,
-0.06639373302459717,
0.04803748056292534,
0.005477914121001959,
-0.030667085200548172,
-0.020000247284770012,
-0.0697568953037262,
0.031182408332824707,
-0.007631110493093729,
0.007847917266190052,
0.03472106531262398,
0.0013384437188506126,
0.02423287369310856,
-0.004690560512244701,
0.029212741181254387,
-0.054702028632164,
0.057604607194662094,
0.03419823199510574,
0.023551715537905693,
0.01637674868106842,
0.05575732886791229,
-0.004679114557802677,
-0.02946104295551777,
0.012263317592442036,
-0.10896048694849014,
0.052053969353437424,
-0.01129227690398693,
-0.06757784634828568,
-0.043104905635118484,
-0.025739595293998718,
0.04189314693212509,
-0.10483239591121674,
-0.03917897492647171,
-0.07066621631383896,
0.09781843423843384,
0.06605479121208191,
0.05877043679356575,
-0.010422985069453716,
-0.003453510580584407,
0.035570595413446426,
0.03297528252005577,
0.021709226071834564,
-0.026687057688832283,
-0.01829398050904274,
-0.08339481055736542,
0.10397159308195114,
-0.10257008671760559,
0.024599192664027214,
-0.0024811665061861277,
-0.014686102978885174,
0.020168932154774666,
-0.08084534853696823,
0.0766340047121048,
0.057914283126592636,
0.027390627190470695,
-0.015591185539960861,
-0.02131575159728527,
0.01087479293346405,
0.07521738111972809,
-0.03924698755145073,
0.09835923463106155,
0.0752977654337883,
-0.04831120744347572,
0.06601724773645401,
0.039219509810209274,
0.04577748849987984,
0.02203226462006569,
0.056399859488010406,
0.008487576618790627,
0.0023548933677375317,
0.03956330195069313,
-0.0015842728316783905,
0.056784119457006454,
0.057425688952207565,
0.039514392614364624,
-0.03552680090069771,
0.008749177679419518,
-0.05531278997659683,
0.017428545281291008,
-0.0025124179664999247,
0.008385522291064262,
6.811640788129756e-34,
-0.009292863309383392,
-0.023189697414636612,
0.055687226355075836,
0.058131661266088486,
0.07328495383262634,
-0.005671323277056217,
0.005120831076055765,
-0.04648474231362343,
-0.010702612809836864,
-0.06104588881134987,
-0.03691841661930084,
-0.06332335621118546,
-0.03352741152048111,
0.11365800350904465,
0.01687547378242016,
-0.044491302222013474,
-0.03655341640114784,
0.02876829169690609,
0.03200142830610275,
0.09769832342863083,
0.04080440476536751,
-0.0313737690448761,
-0.017062608152627945,
0.028970349580049515,
-0.07208868861198425,
-0.01466762088239193,
0.05262482166290283,
-0.015148979611694813,
0.03421392664313316,
-0.005157530773431063,
-0.11785390973091125,
-0.06467639654874802,
-0.04662664234638214,
-0.008145102299749851,
0.04597996547818184,
0.0007905065431259573,
-0.07423880696296692,
0.000529663113411516,
-0.06812219321727753,
-0.04841972887516022,
-0.002576494589447975,
0.05260472372174263,
-0.0259278304874897,
-0.08106965571641922,
-0.13571013510227203,
0.015330192632973194,
0.037667110562324524,
0.036504052579402924,
-0.07046105712652206,
-0.031238626688718796,
0.09893379360437393,
0.0034933786373585463,
0.026746787130832672,
-0.0933464765548706,
-0.08947806060314178,
-0.04939983785152435,
0.07286620140075684,
-0.08927036821842194,
0.06646668910980225,
0.005746284034103155,
0.009405730292201042,
0.015949217602610588,
-0.08197754621505737,
0.04860461503267288,
0.044722672551870346,
-0.04677432030439377,
0.05845845490694046,
0.07735671103000641,
0.04301663488149643,
0.005402930546551943,
-0.00008001860260264948,
0.00008484131831210107,
0.003016403643414378,
-0.09873149544000626,
0.05297936126589775,
-0.02739296481013298,
0.03162020072340965,
0.030597569420933723,
0.05710897594690323,
0.023383285850286484,
-0.09203203022480011,
-0.029933609068393707,
0.013938793912529945,
-0.11364282667636871,
0.00024100167502183467,
0.002422092715278268,
-0.03166060149669647,
-0.0341138057410717,
-0.021655194461345673,
-0.0829976499080658,
-0.044773053377866745,
-0.03432907536625862,
-0.025443583726882935,
-0.03790321946144104,
0.06148335710167885,
-3.818846175849376e-34,
0.03367920592427254,
0.08173392713069916,
-0.008506262674927711,
0.05405505374073982,
0.01155930757522583,
-0.011952094733715057,
-0.05030836910009384,
0.03456185758113861,
0.07941602915525436,
-0.0955570638179779,
0.0000626865730737336,
-0.018157819285988808,
-0.005338194314390421,
0.010036157444119453,
-0.0013725931057706475,
-0.03865886107087135,
0.029037369415163994,
-0.07254989445209503,
-0.10090532898902893,
0.013202120549976826,
0.05059317499399185,
-0.017398832365870476,
-0.03844403475522995,
0.021870708093047142,
-0.018503040075302124,
0.08609391748905182,
0.010053721256554127,
0.05093706026673317,
0.025099551305174828,
-0.011566072702407837,
0.04226883128285408,
-0.003740738146007061,
0.0446942001581192,
-0.05358561500906944,
-0.1252860724925995,
0.03342241421341896,
0.0384770967066288,
0.0012711252784356475,
-0.07543972134590149,
0.06670372188091278,
-0.0020826482214033604,
-0.007277453783899546,
-0.016968337818980217,
0.10327965021133423,
0.03869881108403206,
0.009044223465025425,
0.00289614824578166,
0.01181831769645214,
-0.005077360663563013,
0.06857020407915115,
-0.04276435822248459,
0.0970497727394104,
-0.07842153310775757,
0.05648968368768692,
0.02574210986495018,
-0.08238375931978226,
-0.11504270881414413,
-0.026544280350208282,
0.08032289892435074,
0.012953810393810272,
0.019826669245958328,
-0.06647779792547226,
-0.048825252801179886,
-0.007981174625456333,
-0.0409311018884182,
-0.00820100400596857,
-0.06235824525356293,
-0.014734518714249134,
-0.01927534118294716,
0.1359744817018509,
-0.07116705924272537,
0.005519145168364048,
-0.01348959282040596,
-0.05635637417435646,
-0.0034683796111494303,
-0.11317500472068787,
0.006621445529162884,
0.04007101431488991,
0.07922551035881042,
0.002525679999962449,
0.020443623885512352,
-0.03585762158036232,
0.0434829443693161,
0.05204838886857033,
0.08387531340122223,
-0.027740314602851868,
0.024679020047187805,
0.002064025728031993,
0.02294505573809147,
0.03332517668604851,
0.011554612778127193,
0.011014857329428196,
-0.04377417638897896,
0.008992305025458336,
-0.06899716705083847,
-4.079145554669594e-8,
-0.024095332249999046,
0.0372643880546093,
0.010780672542750835,
0.017364459112286568,
-0.05745875835418701,
-0.03012032061815262,
0.05175212770700455,
0.06743402034044266,
0.06541916728019714,
-0.0755288153886795,
0.0513538159430027,
-0.060466282069683075,
-0.06660471111536026,
-0.015630638226866722,
0.001063890173099935,
0.09138703346252441,
-0.02915247157216072,
-0.027400510385632515,
0.014844666235148907,
0.04941927269101143,
-0.07285401970148087,
0.022502893581986427,
-0.019064968451857567,
-0.005282481200993061,
0.04080483317375183,
0.09869496524333954,
0.02714548073709011,
-0.012427683919668198,
0.1215767189860344,
-0.04155736789107323,
0.012613090686500072,
0.015611249953508377,
0.007170334458351135,
-0.07194650173187256,
-0.05868038162589073,
0.016287168487906456,
0.051916249096393585,
0.014524560421705246,
-0.007764311973005533,
-0.06715292483568192,
-0.017455730587244034,
0.05379315838217735,
0.048263322561979294,
0.001738873077556491,
0.015338611789047718,
0.047094784677028656,
0.019965842366218567,
-0.007649593520909548,
0.027653543278574944,
0.00029528484446927905,
-0.01756739430129528,
-0.05870106443762779,
0.0008430647430941463,
0.09681858867406845,
0.11694345623254776,
-0.013980083167552948,
0.03588775917887688,
0.008112339302897453,
0.08238394558429718,
0.026607584208250046,
0.01726512238383293,
-0.010278066620230675,
0.025529323145747185,
0.033306967467069626
] |
Understanding and Mitigating Hallucination in Large Vision-Language Models via Modular Attribution and Intervention | https://openreview.net/forum?id=Bjq4W7P2Us | [
"Tianyun Yang",
"Ziniu Li",
"Juan Cao",
"Chang Xu"
] | Poster | Large Vision-Language Models (LVLMs) exhibit impressive capabilities in complex visual tasks but are prone to hallucination, especially in open-ended generation tasks. This paper explores why LVLMs tend to hallucinate and how to mitigate it. First, we conduct causal mediation analysis through counterfactual edits on specific modules in LVLMs. Our results disclose that Multi-Head Attention (MHA) modules contribute more to the probability of generating hallucination words than multi-layer perceptron modules. We then identify specific heads that are responsible for hallucination, referred to as hallucination heads. Second, we examine the behavior of hallucination heads. We find that they are concentrated in the middle and deeper layers, displaying a strong attention bias toward text tokens. Further, we show that the attention patterns of certain hallucination heads exhibit greater similarity to the base language model and change slowly during the instruction tuning process. Finally, we propose two simple yet effective methods to mitigate hallucination: one is training-free and can be applied directly during decoding, while the other involves fine-tuning. Both methods are targeted for hallucination heads to reduce their reliance on text tokens. Notably, our methods achieve up to 1.7x reduction in hallucination rate for the LLaVA-v1.5-7B model in COCO captioning task, outperforming existing baselines. Overall, our findings suggest that hallucinations in LVLMs are likely to stem from certain modules, and targeted interventions can effectively mitigate these issues. | Large Vision-Language Models, Hallucination, Interpretability | null | 196 | null | [
0.02645725943148136,
-0.04423798248171806,
0.04316321015357971,
-0.007223233114928007,
0.10042992234230042,
0.012760825455188751,
-0.011756205931305885,
-0.015236620791256428,
0.09590451419353485,
-0.03504267707467079,
-0.048408765345811844,
-0.05953662842512131,
0.053966253995895386,
-0.02840273827314377,
-0.000840287480968982,
-0.003363952273502946,
0.035623013973236084,
0.09207102656364441,
-0.09173222631216049,
-0.08611585199832916,
0.09086611866950989,
-0.013917816802859306,
0.046466581523418427,
-0.014541020616889,
0.04940135404467583,
0.0173829086124897,
-0.04899511858820915,
-0.05129081755876541,
0.031161507591605186,
-0.07292135804891586,
0.01936298981308937,
0.06523095816373825,
-0.050622738897800446,
0.026687776669859886,
-0.060202647000551224,
0.08366044610738754,
-0.06628957390785217,
0.0221350509673357,
0.016982201486825943,
-0.11677758395671844,
-0.03167939931154251,
0.03877495229244232,
0.008013946935534477,
0.045930761843919754,
0.07450821995735168,
0.004409214947372675,
0.032850123941898346,
-0.012984012253582478,
-0.07957277446985245,
-0.06442838907241821,
-0.06349582225084305,
-0.08723078668117523,
0.008368687704205513,
-0.012293982319533825,
-0.06257949024438858,
0.02346634306013584,
0.046917565166950226,
0.06169414520263672,
0.008137591183185577,
0.011506704613566399,
-0.12259954959154129,
-0.03292763978242874,
-0.06254412233829498,
-0.057750578969717026,
-0.029105128720402718,
-0.003176364814862609,
0.0551963746547699,
0.04340801388025284,
-0.03927648812532425,
0.0060370005667209625,
0.055088646709918976,
0.03940451517701149,
-0.00910067930817604,
0.04282986372709274,
0.09266893565654755,
0.029138796031475067,
-0.01580808497965336,
-0.08260582387447357,
0.05022747069597244,
-0.10722120106220245,
0.004025226924568415,
0.02505345828831196,
0.04998878762125969,
-0.00954904779791832,
0.1000295877456665,
-0.0022090699058026075,
-0.0028009095694869757,
0.03528808802366257,
-0.012830188497900963,
0.06925062090158463,
-0.07809540629386902,
-0.13666631281375885,
0.028058309108018875,
-0.02914409153163433,
0.01540989987552166,
0.028217796236276627,
-0.042735256254673004,
-0.05407560244202614,
-0.012943445704877377,
0.03500982001423836,
0.030725568532943726,
0.012073975056409836,
0.009975791908800602,
-0.006094915326684713,
-0.0032537870574742556,
-0.03878364711999893,
0.08396939188241959,
0.1092606708407402,
-0.04707493260502815,
-0.06390899419784546,
-0.0027054196689277887,
0.0418856181204319,
-0.007999099791049957,
-0.007887644693255424,
-0.0028028753586113453,
-0.02672584541141987,
-0.004208416678011417,
-0.04295070096850395,
0.03762703016400337,
0.058158472180366516,
0.04800526052713394,
0.002518136752769351,
0.025053516030311584,
0.003954789601266384,
-0.010307848453521729,
-0.1505073457956314,
-0.06955481320619583,
1.9854703862953156e-33,
0.061607517302036285,
0.03718893975019455,
-0.016489721834659576,
0.04541446268558502,
0.06836078315973282,
-0.0039041941054165363,
0.014379043132066727,
-0.046946074813604355,
0.013423493131995201,
-0.03733992949128151,
-0.008927030488848686,
-0.02244671806693077,
-0.07424349337816238,
0.06779210269451141,
0.030157502740621567,
-0.0006513220141641796,
-0.06859760731458664,
0.05177857726812363,
-0.09532643109560013,
0.03710765019059181,
0.015215723775327206,
0.010169432498514652,
-0.01622905023396015,
-0.0021202010102570057,
-0.025603018701076508,
0.03247424215078354,
0.04464932531118393,
-0.11294377595186234,
-0.06477508693933487,
0.05141323432326317,
-0.057340193539857864,
0.029598120599985123,
0.019907744601368904,
0.04435298219323158,
0.03776198625564575,
-0.02765042521059513,
-0.004566782154142857,
-0.05059776455163956,
0.046569500118494034,
-0.032234374433755875,
-0.03770836815237999,
0.061189986765384674,
-0.008954712189733982,
-0.013048929162323475,
-0.08423863351345062,
-0.007445198018103838,
0.030329111963510513,
0.06371481716632843,
-0.09361516684293747,
-0.039784226566553116,
0.06892412155866623,
0.07084136456251144,
0.03223121538758278,
-0.038240645080804825,
-0.013608680106699467,
0.018295040354132652,
0.022161593660712242,
-0.013484601862728596,
0.0019422079203650355,
-0.001540257828310132,
0.04810222238302231,
0.09358786046504974,
-0.05967611446976662,
0.13765913248062134,
0.031400565057992935,
-0.009318269789218903,
-0.015872854739427567,
0.01418656762689352,
0.03722403198480606,
-0.04384545236825943,
-0.06258299946784973,
-0.018766144290566444,
-0.04002350941300392,
-0.021509408950805664,
-0.0021154319401830435,
-0.04401549696922302,
0.008923408575356007,
-0.023420171812176704,
0.023962553590536118,
0.03139794245362282,
0.0641484186053276,
-0.03274316340684891,
-0.013482957147061825,
-0.05117983743548393,
-0.029912548139691353,
-0.06423888355493546,
-0.02347003109753132,
-0.11262773722410202,
-0.08071599900722504,
-0.032200638204813004,
0.05530565232038498,
-0.03157626464962959,
0.05876370519399643,
-0.01753097213804722,
-0.07357220351696014,
-2.7709549969689613e-33,
-0.02370213344693184,
-0.030243510380387306,
0.06525909155607224,
-0.012241031974554062,
-0.06712502986192703,
-0.007976250723004341,
0.052272021770477295,
-0.028622128069400787,
0.01885993406176567,
-0.02651287242770195,
-0.030091362074017525,
-0.02974598854780197,
-0.029615528881549835,
0.04271392896771431,
0.03585829958319664,
-0.09225431084632874,
0.11332552134990692,
0.08121395856142044,
0.008004659786820412,
0.021936042234301567,
-0.016429642215371132,
0.11030029505491257,
-0.038006797432899475,
-0.05801453813910484,
-0.003250598208978772,
0.03089299611747265,
-0.07348406314849854,
0.05661651864647865,
0.05825033411383629,
-0.0025096938479691744,
-0.009884705767035484,
0.06239106506109238,
0.011303100734949112,
-0.04116169735789299,
-0.03759453818202019,
0.0346764475107193,
0.0011156137334182858,
0.021115489304065704,
-0.08053756505250931,
0.0431034117937088,
0.10160811245441437,
-0.027955234050750732,
-0.027620304375886917,
0.012182127684354782,
-0.010182393714785576,
-0.07364866882562637,
-0.10856356471776962,
-0.006649462040513754,
-0.0013630051398649812,
0.009285197593271732,
-0.05884276330471039,
-0.00885865930467844,
-0.021197034046053886,
-0.03679298236966133,
-0.10091907531023026,
-0.09781455248594284,
-0.008597996085882187,
0.006969774141907692,
0.14158318936824799,
-0.04062987491488457,
-0.03275016322731972,
-0.004120407160371542,
-0.0023116932716220617,
-0.03806670010089874,
0.011510458774864674,
-0.020234044641256332,
-0.03816983103752136,
-0.007954804226756096,
0.0861578956246376,
-0.02819451317191124,
0.07539331167936325,
-0.007416595239192247,
-0.02376280166208744,
0.046035051345825195,
-0.02065112441778183,
-0.019674301147460938,
-0.01490929163992405,
-0.04183461517095566,
-0.02427104115486145,
-0.07597452402114868,
-0.02436683140695095,
-0.0191592238843441,
-0.006002533715218306,
0.05470125377178192,
0.05428008735179901,
0.03741711005568504,
0.028428420424461365,
0.10665613412857056,
0.00846792384982109,
-0.024326706305146217,
-0.024049142375588417,
0.017301851883530617,
0.0025384589098393917,
0.0865626335144043,
-0.058468885719776154,
-4.39372804805771e-8,
-0.04605238884687424,
-0.02502519078552723,
-0.02415984869003296,
0.06422465294599533,
0.06548866629600525,
-0.055075813084840775,
-0.03715109825134277,
0.09204818308353424,
-0.02349182218313217,
-0.03901803120970726,
0.062937892973423,
0.004001758527010679,
-0.012055907398462296,
-0.07841136306524277,
0.04006463661789894,
0.08242695033550262,
0.0325494222342968,
0.07844890654087067,
-0.012001234106719494,
-0.03472751006484032,
0.06659147888422012,
0.008884602226316929,
0.0009393233340233564,
-0.017595365643501282,
-0.023843063041567802,
-0.013399098068475723,
-0.030472399666905403,
0.05479460209608078,
0.02474026568233967,
-0.06095883995294571,
0.0913802981376648,
0.05919180437922478,
-0.03808862343430519,
-0.0686279907822609,
-0.009128873236477375,
0.07462950050830841,
0.007959207519888878,
-0.05216233804821968,
0.03206991031765938,
0.02447040006518364,
0.07514239847660065,
0.0777854323387146,
0.03804873675107956,
0.0881846472620964,
0.06217871606349945,
-0.017335588112473488,
0.01838136836886406,
-0.14466384053230286,
-0.0044838073663413525,
-0.0514146126806736,
-0.027553610503673553,
0.0430859811604023,
-0.02159862034022808,
0.05116058886051178,
0.019177403301000595,
0.03625249117612839,
0.1100420132279396,
-0.0016359997680410743,
-0.0022862101905047894,
0.07140684127807617,
0.11392819881439209,
0.05191988870501518,
-0.010363832116127014,
-0.02637697383761406
] |
Neuron Platonic Intrinsic Representation From Dynamics Using Contrastive Learning | https://openreview.net/forum?id=vFanHFE4Qv | [
"Wei Wu",
"Can Liao",
"zizhen Deng",
"Zhengrui Guo",
"Jinzhuo Wang"
] | Poster | The Platonic Representation Hypothesis posits that behind different modalities of data (what we sense or detect), there exists a universal, modality-independent representation of reality. Inspired by this, we treat each neuron as a system, where we can detect the neuron’s multi-segment activity data under different peripheral conditions. We believe that, similar to the Platonic idea, there exists a time-invariant representation behind the different segments of the same neuron, which reflects the intrinsic properties of the neuron’s system. Intrinsic properties include the molecular profiles, brain regions and morphological structure, etc. The optimization objective for obtaining the intrinsic representation of neurons should satisfy two criteria: (I) segments from the same neuron should have a higher similarity than segments from different neurons; (II) the representations should generalize well to out-of-domain data. To achieve this, we employ contrastive learning, treating different segments from the same neuron as positive pairs and segments from different neurons as negative pairs. During the implementation, we chose the VICReg, which uses only positive pairs for optimization but indirectly separates dissimilar samples via regularization terms. To validate the efficacy of our method, we first applied it to simulated neuron population dynamics data generated using the Izhikevich model. We successfully confirmed that our approach captures the type of each neuron as defined by preset hyperparameters. We then applied our method to two real-world neuron dynamics datasets, including spatial transcriptomics-derived neuron type annotations and the brain regions where each neuron is located. The learned representations from our model not only predict neuron type and location but also show robustness when tested on out-of-domain data (unseen animals). This demonstrates the potential of our approach in advancing the understanding of neuronal systems and offers valuable insights for future neuroscience research. | representation learning, biology, neuroscience, contrastive learning | Obtaining Intrinsic Representations of Single Neurons from Dynamics via Contrastive Learning | 191 | 2502.10425 | [
-0.06277109682559967,
-0.1069876179099083,
0.02946716919541359,
-0.04813815653324127,
-0.001937405439093709,
0.005568043794482946,
0.038387905806303024,
-0.03057721257209778,
0.06247299909591675,
-0.03774964436888695,
-0.06007097661495209,
-0.04340508207678795,
-0.061714209616184235,
0.15221251547336578,
-0.062431614845991135,
-0.028808601200580597,
0.050861701369285583,
0.1319108009338379,
-0.08902351558208466,
0.0038016417529433966,
0.07656285166740417,
0.03577810525894165,
-0.018234647810459137,
0.013092604465782642,
0.035156574100255966,
0.08886606246232986,
0.035564638674259186,
-0.02237485721707344,
0.06615205854177475,
-0.053312335163354874,
0.10505630820989609,
-0.0007656759698875248,
-0.025797830894589424,
-0.04094838723540306,
-0.03603053092956543,
0.06461673974990845,
-0.06841868162155151,
-0.003101672511547804,
-0.04660535231232643,
-0.05470627173781395,
-0.009452840313315392,
0.06853992491960526,
-0.03073284961283207,
0.023356016725301743,
0.060350146144628525,
0.060730285942554474,
0.05912439897656441,
-0.08426354080438614,
0.01860293745994568,
-0.0190826877951622,
0.01774577610194683,
-0.012739596888422966,
-0.049307681620121,
0.0407358855009079,
0.021921517327427864,
0.09351814538240433,
0.007501592859625816,
0.008526707999408245,
0.025101734325289726,
-0.0619332455098629,
0.03816572204232216,
0.006383243016898632,
0.007109344936907291,
-0.005006954539567232,
0.0027300661895424128,
0.05483970791101456,
0.03267974033951759,
-0.017088893800973892,
-0.056645240634679794,
0.029057422652840614,
-0.02058609575033188,
0.10563241690397263,
-0.04405338317155838,
-0.02814572863280773,
0.014794867485761642,
0.014358976855874062,
0.018416857346892357,
0.04404619336128235,
0.01944856159389019,
-0.10003764182329178,
0.04776103049516678,
0.09774293005466461,
0.07154449075460434,
-0.0029316875152289867,
0.09411462396383286,
0.09018221497535706,
-0.04903000593185425,
0.023158442229032516,
-0.03300802782177925,
0.017639316618442535,
-0.029865773394703865,
0.023275211453437805,
-0.04692792519927025,
-0.058072641491889954,
-0.0016967740375548601,
-0.07066342979669571,
0.032559316605329514,
-0.011776385828852654,
0.11650996655225754,
0.09492439776659012,
-0.02469499781727791,
-0.019818821921944618,
0.0586250014603138,
0.053937122225761414,
0.05326912924647331,
0.008936439640820026,
0.03731904551386833,
-0.02880433388054371,
0.031887833029031754,
-0.01657267101109028,
0.053037069737911224,
0.017341021448373795,
0.016418440267443657,
0.08838824927806854,
0.05593249946832657,
-0.017251139506697655,
-0.04116171598434448,
0.007911940105259418,
0.13894590735435486,
0.025172846391797066,
-0.04780879616737366,
-0.0451166070997715,
-0.022934582084417343,
0.10733752697706223,
0.046590790152549744,
-0.017360687255859375,
-0.058777496218681335,
1.8292749213908733e-33,
-0.026926493272185326,
0.023919077590107918,
0.045814692974090576,
-0.05859033763408661,
-0.013210851699113846,
-0.0184453334659338,
-0.016342394053936005,
-0.04262533038854599,
0.09482672065496445,
0.061696458607912064,
-0.11925913393497467,
0.03762412816286087,
0.011379536241292953,
0.11778484284877777,
0.008787629194557667,
0.04441481828689575,
-0.05945296958088875,
-0.043738871812820435,
-0.011454874649643898,
-0.08022553473711014,
0.015683621168136597,
0.009680530987679958,
-0.05769265815615654,
-0.017389537766575813,
-0.08180027455091476,
-0.05086664855480194,
-0.014076336286962032,
-0.009479810483753681,
0.006680965423583984,
-0.005162288900464773,
-0.06648906320333481,
0.031204640865325928,
-0.01090429536998272,
0.024901319295167923,
0.0482131652534008,
0.017846370115876198,
0.06125521659851074,
-0.030101468786597252,
0.003555544652044773,
-0.0659857764840126,
0.03988930210471153,
0.0559166818857193,
0.018200667575001717,
-0.0779750645160675,
-0.034732889384031296,
-0.006603173445910215,
0.052808985114097595,
0.058498840779066086,
-0.09977073222398758,
-0.18063120543956757,
-0.020990699529647827,
-0.004987988155335188,
0.029125504195690155,
-0.09818361699581146,
-0.026570633053779602,
-0.010132315568625927,
-0.04285530000925064,
0.03916134312748909,
0.04977107793092728,
0.09801202267408371,
0.013506833463907242,
-0.014789354987442493,
0.03983769193291664,
0.00923063512891531,
-0.038896430283784866,
0.09854210913181305,
-0.05400877445936203,
0.022723715752363205,
0.05451428145170212,
-0.04056118428707123,
-0.011831306852400303,
0.042665332555770874,
-0.009407215751707554,
-0.10447747260332108,
0.05109855905175209,
-0.04146192967891693,
0.061881035566329956,
-0.03639111667871475,
-0.11055264621973038,
0.026641907170414925,
-0.0910780131816864,
-0.02882307395339012,
-0.06072735786437988,
-0.05747578665614128,
-0.02732839621603489,
-0.03868592903017998,
0.1143629178404808,
-0.10506825149059296,
-0.0877247080206871,
0.018577158451080322,
-0.07350973784923553,
-0.0480722039937973,
-0.020404493436217308,
0.01087524276226759,
-0.01365366205573082,
-2.2850539378100067e-33,
-0.0533045195043087,
0.016138385981321335,
0.006542250048369169,
0.03649832308292389,
-0.0037509656976908445,
-0.0349825955927372,
-0.013936704024672508,
0.034066665917634964,
-0.05113588273525238,
-0.02861722931265831,
0.015456611290574074,
-0.030227413401007652,
-0.008110923692584038,
-0.08708321303129196,
-0.01238911971449852,
0.04115912690758705,
-0.06746840476989746,
0.030551819130778313,
0.042502839118242264,
0.029655855149030685,
0.0019031309057027102,
0.11179225146770477,
-0.09454932063817978,
0.03553839400410652,
0.007313255220651627,
0.04271923378109932,
-0.06389720737934113,
0.0643155500292778,
-0.022101126611232758,
0.05444304272532463,
-0.0569608211517334,
0.010317926295101643,
-0.1136750876903534,
-0.015165786258876324,
-0.06624405831098557,
0.04249078780412674,
-0.05242282897233963,
-0.012683046981692314,
-0.00980142317712307,
0.014825322665274143,
0.03350025415420532,
0.005288904998451471,
-0.02581869438290596,
0.04185463488101959,
0.08304598182439804,
0.02246544137597084,
-0.05417386814951897,
-0.0031036539003252983,
-0.08238574862480164,
0.018117355182766914,
-0.042050860822200775,
0.038304559886455536,
-0.040985215455293655,
-0.038478948175907135,
-0.028034746646881104,
0.0013871790142729878,
0.05357085540890694,
-0.028164135292172432,
0.06516502797603607,
-0.006453410722315311,
-0.03647985681891441,
-0.10004410892724991,
-0.03294086456298828,
0.0736440122127533,
0.027839668095111847,
0.036009032279253006,
0.016552088782191277,
-0.0045896596275269985,
0.030868815258145332,
0.039737679064273834,
-0.022054333239793777,
0.013379869051277637,
0.045827608555555344,
0.007981776259839535,
-0.012887478806078434,
-0.053414586931467056,
0.01602732017636299,
-0.02724325656890869,
-0.010236241854727268,
-0.0055470094084739685,
-0.01425135787576437,
-0.000023980737751116976,
0.024034522473812103,
0.033976245671510696,
-0.007295704446732998,
0.09219140559434891,
0.0011666216887533665,
0.047176606953144073,
0.02987680211663246,
-0.015368232503533363,
0.012030662037432194,
0.012208876200020313,
-0.026055091992020607,
-0.01868337206542492,
0.020434878766536713,
-5.166946337453737e-8,
-0.020233729854226112,
-0.036234430968761444,
0.06459829211235046,
-0.03379570692777634,
-0.025082439184188843,
-0.04232948645949364,
0.03630014508962631,
0.03805378079414368,
-0.08798743039369583,
-0.038046833127737045,
0.08823546767234802,
-0.05704900994896889,
-0.07824402302503586,
-0.0436878576874733,
-0.016584299504756927,
0.09914358705282211,
-0.0038550556637346745,
0.01257425919175148,
-0.025068646296858788,
0.02195015735924244,
0.02510225400328636,
0.039975907653570175,
-0.0036924683954566717,
-0.014981756918132305,
0.06750430911779404,
-0.09833142906427383,
-0.01712113991379738,
-0.006246867123991251,
0.07880571484565735,
-0.001202924526296556,
0.01593850739300251,
0.07799316197633743,
0.03382083773612976,
-0.022724101319909096,
-0.04777286574244499,
0.07630148530006409,
-0.033988676965236664,
0.059197064489126205,
-0.09500224888324738,
-0.03934648633003235,
-0.032997824251651764,
0.03522954136133194,
0.005927001591771841,
-0.01895875856280327,
0.047159723937511444,
0.011990200728178024,
0.08138672262430191,
-0.03599175810813904,
0.055673014372587204,
0.02904990129172802,
0.024775689467787743,
0.04947294294834137,
0.03228102624416351,
0.03958352655172348,
0.0003807450120802969,
-0.04209475219249725,
-0.07145338505506516,
-0.008671848103404045,
0.01810366101562977,
-0.016696149483323097,
0.020241806283593178,
0.04260317236185074,
-0.09420676529407501,
-0.06898793578147888
] |
Where Am I and What Will I See: An Auto-Regressive Model for Spatial Localization and View Prediction | https://openreview.net/forum?id=NuHYh4YKNe | [
"Junyi Chen",
"Di Huang",
"Weicai Ye",
"Wanli Ouyang",
"Tong He"
] | Poster | Spatial intelligence is the ability of a machine to perceive, reason, and act in three dimensions within space and time.
Recent advancements in large-scale auto-regressive models have demonstrated remarkable capabilities across various reasoning tasks. However, these models often struggle with fundamental aspects of spatial reasoning, particularly in answering questions like "Where am I?" and "What will I see?". While some attempts have been done, existing approaches typically treat them as separate tasks, failing to capture their interconnected nature. In this paper, we present **G**enerative **S**patial **T**ransformer (GST), a novel auto-regressive framework that jointly addresses spatial localization and view prediction. Our model simultaneously estimates the camera pose from a single image and predicts the view from a new camera pose, effectively bridging the gap between spatial awareness and visual prediction. The proposed innovative camera tokenization method enables the model to learn the joint distribution of 2D projections and their corresponding spatial perspectives in an auto-regressive manner. This unified training paradigm demonstrates that joint optimization of pose estimation and novel view synthesis leads to improved performance in both tasks, for the first time, highlighting the inherent relationship between spatial awareness and visual prediction. | Generative Models, Novel View Synthesis, Camera Pose Estimation | null | 190 | 2410.18962 | [
-0.020759107545018196,
-0.10591376572847366,
0.02662278525531292,
-0.008283426985144615,
0.11285866051912308,
0.007943210192024708,
0.06892018765211105,
0.04395972937345505,
0.040230002254247665,
0.018398940563201904,
-0.07525759935379028,
-0.04059172421693802,
0.009137558750808239,
0.12313593178987503,
-0.0241702888160944,
-0.04331990331411362,
0.05011310800909996,
0.0582892931997776,
-0.07618305832147598,
-0.0056641302071511745,
-0.01750940829515457,
-0.0689181238412857,
0.04391355812549591,
0.05976679176092148,
0.08781968057155609,
0.01944245956838131,
0.07196545600891113,
-0.07015527039766312,
-0.04518260434269905,
0.03780198097229004,
0.028432779014110565,
-0.02123509719967842,
-0.0563463419675827,
0.027426602318882942,
-0.07098373770713806,
0.052293065935373306,
-0.025697674602270126,
0.06715425103902817,
0.012568684294819832,
-0.009494034573435783,
0.0015216195024549961,
-0.02075137384235859,
0.025208810344338417,
0.0003030684310942888,
0.05620379373431206,
0.07727989554405212,
0.07389660179615021,
-0.011342919431626797,
0.016113746911287308,
-0.04160916060209274,
-0.08767660707235336,
-0.02959585003554821,
-0.04091821610927582,
-0.06484846770763397,
-0.017744259908795357,
0.019755495712161064,
0.016034267842769623,
-0.011969116516411304,
0.07315770536661148,
0.027323808521032333,
0.034912168979644775,
-0.09682239592075348,
-0.022036414593458176,
0.025236565619707108,
-0.009744326584041119,
-0.041600801050662994,
0.011580889113247395,
-0.013488294556736946,
-0.030896198004484177,
0.02155333198606968,
-0.026373326778411865,
0.08936773240566254,
0.011331412009894848,
-0.04980044811964035,
0.021220291033387184,
0.05129438638687134,
0.005139859858900309,
-0.05714702606201172,
0.05947637930512428,
-0.08784619718790054,
0.1109711155295372,
0.012000142596662045,
-0.014736320823431015,
-0.004708945285528898,
0.07225187867879868,
0.003943440970033407,
-0.060208771377801895,
0.097665935754776,
0.02883492037653923,
-0.03556675463914871,
-0.030911685898900032,
-0.08278288692235947,
-0.04361071437597275,
-0.020968899130821228,
0.05660654976963997,
-0.020087042823433876,
0.007658923044800758,
-0.07930371910333633,
0.00244248122908175,
0.09440042078495026,
-0.017272360622882843,
0.09353715926408768,
0.05794902145862579,
0.0011227659415453672,
0.09049224108457565,
-0.034361936151981354,
-0.06608828902244568,
0.02943723276257515,
0.06830798089504242,
-0.08526522666215897,
0.01683148182928562,
-0.00038382766069844365,
0.02572394162416458,
0.03161380812525749,
-0.07695633918046951,
-0.015050237067043781,
0.006936743389815092,
-0.005452994257211685,
0.030127571895718575,
-0.06540201604366302,
0.023907579481601715,
-0.04732508957386017,
-0.05735987424850464,
0.04298824071884155,
0.07346684485673904,
-0.05403982475399971,
-0.015899360179901123,
-7.951333989129812e-34,
0.020319076254963875,
0.02779010869562626,
0.1257665902376175,
0.058319855481386185,
0.0011208908399567008,
0.015495538711547852,
-0.04820210486650467,
0.03242090344429016,
0.026408260688185692,
0.032909464091062546,
0.024092305451631546,
-0.05171450227499008,
-0.03168504685163498,
0.11812255531549454,
-0.014542736113071442,
0.02577264979481697,
-0.03599373623728752,
0.039989173412323,
-0.040796488523483276,
0.03996632620692253,
0.027762074023485184,
0.014241977594792843,
-0.024588260799646378,
-0.014628320001065731,
-0.026018349453806877,
-0.019666317850351334,
0.09951227158308029,
-0.026409713551402092,
-0.08312547951936722,
-0.03132171556353569,
-0.08387770503759384,
-0.00662205321714282,
-0.02632588893175125,
0.03209979832172394,
0.02520843781530857,
-0.08787461370229721,
-0.028862239792943,
-0.07564692199230194,
0.04855572432279587,
-0.02820551209151745,
-0.0023071563336998224,
0.07371477782726288,
-0.04339485242962837,
-0.00330626149661839,
0.0025388787034898996,
-0.01576334610581398,
-0.05370746925473213,
0.055627934634685516,
-0.07874833792448044,
0.0746995136141777,
0.05734052509069443,
0.015413261950016022,
-0.09717842936515808,
-0.05973641946911812,
-0.07644060999155045,
-0.0005917256930842996,
-0.026388229802250862,
-0.048210468143224716,
0.06233258917927742,
-0.008005926385521889,
-0.030370812863111496,
-0.0024804016575217247,
-0.03629240021109581,
0.03384462743997574,
0.009680865332484245,
0.0017090418841689825,
0.03849773481488228,
0.07249704003334045,
0.09615398943424225,
0.06934423744678497,
0.034499719738960266,
0.0011284344363957644,
0.025715133175253868,
-0.03310370072722435,
0.048279717564582825,
0.06758963316679001,
0.010096079669892788,
-0.030207717791199684,
-0.04790952801704407,
0.03797945752739906,
-0.04995338246226311,
0.005560936871916056,
-0.05659235268831253,
0.04374902695417404,
-0.00681193545460701,
-0.02038463205099106,
0.057162005454301834,
-0.1344641000032425,
-0.035670191049575806,
-0.0576879046857357,
-0.05674556642770767,
-0.08349630981683731,
-0.07661720365285873,
-0.005238511599600315,
-0.03727816417813301,
1.4841408260285742e-34,
0.0837724506855011,
-0.07156457751989365,
-0.01595045067369938,
-0.06699550896883011,
-0.044498708099126816,
-0.1018139123916626,
0.009849042631685734,
0.02899031899869442,
-0.03584691882133484,
-0.09601990878582001,
-0.08374238014221191,
-0.006843034643679857,
0.0842532217502594,
0.0683668851852417,
0.02848278544843197,
0.042001768946647644,
-0.015583565458655357,
0.02089061215519905,
-0.10105102509260178,
0.059232279658317566,
-0.009727450087666512,
0.0692429393529892,
-0.0022117120679467916,
0.025198081508278847,
-0.019537869840860367,
0.06513211131095886,
0.07066193222999573,
0.09526845067739487,
-0.07254040241241455,
0.017040850594639778,
0.00524570606648922,
-0.050126951187849045,
0.049889467656612396,
-0.03756224364042282,
-0.09401187300682068,
0.09100420773029327,
0.0269855298101902,
-0.07052309066057205,
-0.07138411700725555,
0.04307752102613449,
0.012843815609812737,
-0.0430632159113884,
0.000018318460206501186,
0.0013696840032935143,
-0.06210888922214508,
0.012843613512814045,
0.00795583613216877,
0.05086607113480568,
-0.019313054159283638,
0.03874697536230087,
-0.06588967144489288,
0.06863095611333847,
-0.08668825775384903,
-0.023037012666463852,
0.022932877764105797,
-0.035007741302251816,
-0.058489225804805756,
-0.008616050705313683,
0.05014543980360031,
-0.040321871638298035,
-0.012579022906720638,
-0.01013195887207985,
0.022334212437272072,
0.03449665755033493,
-0.02974270097911358,
0.008361661806702614,
-0.044293150305747986,
-0.029878951609134674,
0.02616288885474205,
0.06511308252811432,
0.01857910305261612,
-0.006206784397363663,
0.001716310391202569,
0.03917417675256729,
0.06642400473356247,
0.016584010794758797,
0.02089591510593891,
-0.006004353519529104,
0.027601297944784164,
-0.041213665157556534,
0.014889680780470371,
-0.11166709661483765,
0.054916199296712875,
0.11803239583969116,
0.01741825044155121,
0.033691391348838806,
0.04182799905538559,
0.07424911856651306,
0.07801742106676102,
-0.02071264013648033,
-0.060501787811517715,
0.05552133172750473,
-0.05855979770421982,
0.052702877670526505,
0.003327112179249525,
-4.1198557454436013e-8,
-0.12103065848350525,
0.074565090239048,
0.020939722657203674,
-0.0006099664024077356,
-0.07367610186338425,
-0.0563887320458889,
0.03384803980588913,
0.0515650138258934,
-0.06254847347736359,
-0.05142996460199356,
0.05798634514212608,
-0.07924969494342804,
-0.0021369829773902893,
-0.03183194622397423,
-0.01558552123606205,
0.033023983240127563,
0.03343901410698891,
-0.008247475139796734,
-0.044299568980932236,
0.07700587809085846,
-0.025654194876551628,
0.015571448020637035,
0.020715981721878052,
-0.020948510617017746,
0.04760370030999184,
-0.05853664129972458,
-0.055263832211494446,
0.08498773723840714,
0.06522572040557861,
-0.01994473487138748,
0.05977841466665268,
0.026820823550224304,
0.03991333767771721,
0.01090287696570158,
-0.0011907365405932069,
-0.031127460300922394,
-0.02975638397037983,
-0.040757082402706146,
0.03341234102845192,
-0.09591743350028992,
-0.0117042800411582,
0.0437844954431057,
-0.031649041920900345,
0.08669687062501907,
-0.00006844251038273796,
0.07483281195163727,
0.04400425776839256,
-0.06973371654748917,
-0.015427852049469948,
-0.03077596426010132,
-0.07151556015014648,
0.014322991482913494,
-0.004506811965256929,
0.08523158729076385,
0.04092038795351982,
0.004911703057587147,
-0.008276987820863724,
-0.004774543456733227,
-0.015293080359697342,
0.07785720378160477,
-0.039132874459028244,
0.05896584689617157,
-0.09668151289224625,
0.00786714069545269
] |
Achieving Dimension-Free Communication in Federated Learning via Zeroth-Order Optimization | https://openreview.net/forum?id=omrLHFzC37 | [
"Zhe Li",
"Bicheng Ying",
"Zidong Liu",
"Chaosheng Dong",
"Haibo Yang"
] | Poster | Federated Learning (FL) offers a promising framework for collaborative and privacy-preserving machine learning across distributed data sources.
However, the substantial communication costs associated with FL significantly challenge its efficiency.
Specifically, in each communication round, the communication costs scale linearly with the model's dimension, which presents a formidable obstacle, especially in large model scenarios.
Despite various communication-efficient strategies, the intrinsic dimension-dependent communication cost remains a major bottleneck for current FL implementations.
This paper proposes a novel dimension-free communication algorithm - DeComFL, which leverages the zeroth-order optimization techniques and reduces the communication cost from $\mathcal{O}(d)$ to $\mathcal{O}(1)$ by transmitting only a constant number of scalar values between clients and the server in each round, regardless of the dimension $d$ of the model parameters.
Theoretically, in non-convex functions, we prove that our algorithm achieves state-of-the-art rates, which show a linear speedup of the number of clients and local steps under standard assumptions. With additional low effective rank assumption, we can further show that the convergence rate is independent of the model dimension $d$ as well.
Empirical evaluations, encompassing both classic deep learning training and large language model fine-tuning, demonstrate significant reductions in communication overhead.
Notably, DeComFL achieves this by transmitting only around 1MB of data in total between the server and a client to fine-tune a model with billions of parameters.
The code is available at https://github.com/ZidongLiu/DeComFL. | Federated Learning, Zeroth-Order Optimization, Communication Efficiency, Low Rank, Large Language Models, Fine-Tuning | null | 183 | 2405.15861 | [
-0.04455010965466499,
-0.06355553865432739,
0.032095856964588165,
0.07106412947177887,
0.032943159341812134,
-0.05483047664165497,
-0.03850202634930611,
-0.05002085119485855,
0.04764644801616669,
0.021198226138949394,
-0.006769874133169651,
0.04606087878346443,
0.028904631733894348,
-0.01262222696095705,
-0.005734620150178671,
0.04707397520542145,
0.08711665123701096,
0.040052663534879684,
-0.06573595851659775,
-0.013313643634319305,
-0.03942215070128441,
-0.020392874255776405,
0.01261546928435564,
0.05506265163421631,
0.05389373004436493,
-0.05466269701719284,
0.0611807256937027,
0.0007826891960576177,
-0.029378755018115044,
0.019337927922606468,
0.07679582387208939,
0.01236279122531414,
-0.012079160660505295,
-0.04134821891784668,
-0.022800263017416,
0.038950711488723755,
0.016272246837615967,
0.04873808100819588,
-0.017282478511333466,
0.03081042319536209,
-0.0021093508694320917,
0.029176481068134308,
0.005345216020941734,
-0.0018527609063312411,
0.03220397233963013,
0.0020531693007797003,
0.030622530728578568,
-0.025991225615143776,
0.005057210102677345,
-0.023211626335978508,
0.014774180017411709,
0.004811775404959917,
-0.10509291291236877,
0.0983278751373291,
-0.031469620764255524,
-0.05207454413175583,
-0.05032015219330788,
-0.04211457818746567,
-0.09049949795007706,
0.06385275721549988,
-0.08616089075803757,
-0.09849826991558075,
-0.007338083349168301,
0.0000846454786369577,
-0.06451225280761719,
0.015153070911765099,
0.025527264922857285,
0.06447851657867432,
-0.0954102873802185,
0.03119630739092827,
-0.05099615827202797,
0.016815414652228355,
-0.02055484987795353,
-0.013038329780101776,
0.03235727548599243,
-0.0010493022855371237,
0.019060181453824043,
0.034038230776786804,
0.06420636922121048,
0.010871471837162971,
0.00031107786344364285,
0.051570337265729904,
-0.006290664430707693,
-0.010359670966863632,
0.060869596898555756,
-0.07162731140851974,
0.0322355180978775,
-0.034633371978998184,
0.03067178465425968,
0.031462591141462326,
-0.05450993776321411,
0.08959907293319702,
-0.05128680169582367,
0.014322291128337383,
-0.024783015251159668,
-0.01965050771832466,
0.02149135060608387,
-0.017635449767112732,
-0.035099659115076065,
0.0975538119673729,
-0.1022312119603157,
0.058619871735572815,
0.02455514296889305,
-0.030568039044737816,
0.023848729208111763,
0.022844966500997543,
-0.00437886081635952,
0.05201665684580803,
0.06276319921016693,
-0.05491063371300697,
0.03595595061779022,
0.04240225628018379,
0.010410676710307598,
0.06561888754367828,
-0.03535381332039833,
0.029638681560754776,
-0.020776238292455673,
-0.025214623659849167,
0.016675662249326706,
0.054313044995069504,
-0.05403098091483116,
-0.05186466500163078,
0.031018124893307686,
0.05947573110461235,
-0.053459156304597855,
0.0022884795907884836,
-0.050213202834129333,
1.6644783275872658e-33,
0.05110175535082817,
0.10762408375740051,
0.02990024723112583,
-0.07366763055324554,
0.044558439403772354,
0.007421863730996847,
0.09316001832485199,
0.016073979437351227,
-0.07149938493967056,
0.009403389878571033,
-0.00306884478777647,
-0.00632878765463829,
0.03904183581471443,
0.0192783921957016,
0.03442041575908661,
-0.007530849426984787,
0.04749472811818123,
0.07682908326387405,
0.054252978414297104,
-0.09453542530536652,
0.12323534488677979,
-0.05149873346090317,
-0.0424419641494751,
-0.01838834211230278,
0.03000110574066639,
-0.06112460792064667,
0.005403670482337475,
-0.025977643206715584,
-0.0023224649485200644,
0.007854431867599487,
-0.055397871881723404,
-0.018257323652505875,
0.024468565359711647,
0.05312228947877884,
0.0668296217918396,
-0.011697161011397839,
-0.03306810185313225,
-0.04215926676988602,
-0.01657600700855255,
-0.024621548131108284,
0.011896717362105846,
-0.05035315454006195,
0.035903945565223694,
-0.0582733228802681,
-0.18043971061706543,
-0.03831572085618973,
-0.05602962151169777,
-0.0018208352848887444,
-0.0026536034420132637,
-0.07358969002962112,
0.020043449476361275,
-0.019339382648468018,
-0.05985118821263313,
0.0264119990170002,
-0.0057080890983343124,
-0.011320078745484352,
0.012516040354967117,
0.018978262320160866,
0.01733451709151268,
0.11007579416036606,
-0.05000954493880272,
-0.013850590214133263,
-0.00289027183316648,
0.029744010418653488,
0.0164047684520483,
-0.015224623493850231,
-0.03162437677383423,
-0.03430645912885666,
0.12457658350467682,
-0.14416693150997162,
0.054037757217884064,
0.09567496180534363,
0.011984555050730705,
-0.045583758503198624,
0.13989616930484772,
-0.021350394934415817,
0.1054011881351471,
-0.07673554122447968,
0.012203645892441273,
0.02283318154513836,
-0.0007629034807905555,
0.054412417113780975,
-0.034932248294353485,
0.022471187636256218,
-0.04590195044875145,
-0.07124339789152145,
-0.03347019478678703,
0.04338018596172333,
-0.07615236192941666,
0.012590888887643814,
-0.1311529576778412,
-0.011573956348001957,
0.007058833260089159,
-0.022836830466985703,
0.013091190718114376,
-3.816068496473707e-35,
-0.06204993650317192,
0.023478787392377853,
-0.04790966957807541,
0.005388178862631321,
0.054855018854141235,
-0.051218025386333466,
-0.07924377918243408,
0.08524128794670105,
-0.025370009243488312,
-0.04789911210536957,
-0.09129868447780609,
-0.0418737456202507,
0.025445610284805298,
-0.005246893968433142,
0.048016373068094254,
0.06403249502182007,
-0.010717522352933884,
-0.014387305825948715,
-0.002541548339650035,
0.03077479638159275,
-0.03858126327395439,
0.09137839823961258,
-0.06964312493801117,
0.054868511855602264,
0.08436529338359833,
0.04084758460521698,
-0.04675814509391785,
0.1352134793996811,
0.022225234657526016,
0.03133866563439369,
-0.00910220853984356,
0.031053345650434494,
-0.0529501847922802,
-0.07224246859550476,
-0.07570715993642807,
0.05841052159667015,
0.007296064868569374,
0.0344548374414444,
-0.07973086833953857,
0.10315321385860443,
0.013078340329229832,
-0.08002439886331558,
-0.14446070790290833,
-0.004345577675849199,
0.07859418541193008,
-0.04356379434466362,
-0.03107275255024433,
-0.04519958794116974,
0.03285488486289978,
0.0018133786506950855,
-0.011883090250194073,
0.08880768716335297,
0.007155240047723055,
0.0185353122651577,
-0.03594962880015373,
-0.009342566132545471,
0.02746821567416191,
0.049346115440130234,
0.0977759063243866,
0.0037848479114472866,
0.07189839333295822,
-0.032259080559015274,
0.026434369385242462,
0.03774246945977211,
0.047624096274375916,
0.01327135693281889,
0.051118239760398865,
0.04769914969801903,
-0.03976812586188316,
0.05235806107521057,
-0.022480465471744537,
-0.002017751568928361,
-0.045342300087213516,
-0.0011057527735829353,
0.017186611890792847,
0.02288389764726162,
0.02234206348657608,
0.021366702392697334,
-0.06271694600582123,
0.09950126707553864,
0.07799771428108215,
0.0513497032225132,
-0.003048497950658202,
0.05147915706038475,
0.06821814924478531,
-0.020415380597114563,
0.15699872374534607,
0.02179555594921112,
-0.027846824377775192,
-0.04869328439235687,
0.02224166877567768,
-0.02160855196416378,
-0.017617447301745415,
0.015744589269161224,
0.004140337463468313,
-4.3958390705256534e-8,
-0.08660675585269928,
-0.019204681739211082,
-0.01511338260024786,
0.013395801186561584,
0.043322689831256866,
0.044327326118946075,
0.012179998680949211,
0.016393151134252548,
-0.03969963639974594,
-0.00950199831277132,
0.03772822394967079,
-0.0070318784564733505,
0.00020996255625505,
-0.03302058205008507,
-0.03923683986067772,
0.052917174994945526,
-0.027746498584747314,
-0.1446804702281952,
-0.017293384298682213,
0.020788663998246193,
-0.05037662014365196,
0.0066110300831496716,
-0.030994238331913948,
-0.05735171586275101,
0.04241054877638817,
-0.11694031953811646,
0.09417266398668289,
0.05230196565389633,
0.0904720276594162,
0.028008226305246353,
-0.10121706128120422,
-0.0011321742786094546,
0.044292837381362915,
0.01588485762476921,
-0.03736119717359543,
0.06779295951128006,
0.006689084693789482,
-0.05054105445742607,
-0.03779231011867523,
-0.018029917031526566,
0.02223920077085495,
0.061284877359867096,
-0.04511003941297531,
-0.020661620423197746,
0.06818336248397827,
0.003448322881013155,
-0.0013596651842817664,
-0.07129044830799103,
0.05229109898209572,
-0.0056926412507891655,
0.06645257771015167,
0.052287813276052475,
-0.02111802250146866,
0.06018495932221413,
0.06995079666376114,
-0.018647484481334686,
-0.031751226633787155,
-0.014833000488579273,
0.03977946564555168,
0.060353536158800125,
0.02283380925655365,
0.046753183007240295,
-0.058009471744298935,
-0.0008504270808771253
] |
MindSimulator: Exploring Brain Concept Localization via Synthetic fMRI | https://openreview.net/forum?id=vgt2rSf6al | [
"Guangyin Bao",
"Qi Zhang",
"Zixuan Gong",
"Zhuojia Wu",
"Duoqian Miao"
] | Poster | Concept-selective regions within the human cerebral cortex exhibit significant activation in response to specific visual stimuli associated with particular concepts. Precisely localizing these regions stands as a crucial long-term goal in neuroscience to grasp essential brain functions and mechanisms. Conventional experiment-driven approaches hinge on manually constructed visual stimulus collections and corresponding brain activity recordings, constraining the support and coverage of concept localization. Additionally, these stimuli often consist of concept objects in unnatural contexts and are potentially biased by subjective preferences, thus prompting concerns about the validity and generalizability of the identified regions. To address these limitations, we propose a data-driven exploration approach. By synthesizing extensive brain activity recordings, we statistically localize various concept-selective regions. Our proposed MindSimulator leverages advanced generative technologies to learn the probability distribution of brain activity conditioned on concept-oriented visual stimuli. This enables the creation of simulated brain recordings that reflect real neural response patterns. Using the synthetic recordings, we successfully localize several well-studied concept-selective regions and validate them against empirical findings, achieving promising prediction accuracy. The feasibility opens avenues for exploring novel concept-selective regions and provides prior hypotheses for future neuroscience research. | Neuroscience, fMRI encoding, Generative model, fMRI generation, fMRI functional localizer, Concept-selective voxel | null | 180 | 2503.02351 | [
0.00787521991878748,
-0.1274418979883194,
0.05056788772344589,
0.03472800552845001,
0.07052911818027496,
-0.027033349499106407,
0.03999324515461922,
-0.029807189479470253,
0.01893952116370201,
-0.010072503238916397,
-0.06700458377599716,
-0.14044059813022614,
-0.031541455537080765,
0.04731447249650955,
0.01122758537530899,
0.0010352267418056726,
0.10566074401140213,
0.06681320071220398,
0.03262165188789368,
0.07987341284751892,
0.036075714975595474,
-0.01700420305132866,
0.04949856549501419,
-0.019708193838596344,
0.05342637002468109,
0.08736689388751984,
0.0410802848637104,
0.025929514318704605,
0.013501589186489582,
0.022382456809282303,
0.06566202640533447,
0.04304924234747887,
-0.013363679870963097,
0.029876388609409332,
-0.04468486085534096,
0.007318865042179823,
-0.10651040077209473,
0.04932784661650658,
0.014820362441241741,
-0.03816397488117218,
0.018932202830910683,
-0.005689035169780254,
0.04907562583684921,
0.016195373609662056,
-0.02436813712120056,
0.06326287239789963,
0.029180044308304787,
-0.0694502592086792,
-0.009481145069003105,
-0.02478138357400894,
-0.05400389805436134,
-0.0042273555882275105,
-0.028945479542016983,
0.03609094023704529,
0.009850389324128628,
0.09390038996934891,
0.06152721494436264,
0.008111953735351562,
-0.041868846863508224,
0.09323704987764359,
-0.024721216410398483,
-0.03014542907476425,
0.008218458853662014,
-0.05370722711086273,
-0.01847098208963871,
-0.028410620987415314,
-0.0001487319532316178,
0.027475906535983086,
-0.015524256974458694,
-0.023830629885196686,
0.018981261178851128,
0.08867760002613068,
-0.032930292189121246,
-0.005054104141891003,
0.06324733048677444,
0.056415993720293045,
-0.055313240736722946,
-0.05582013726234436,
-0.006309594959020615,
-0.12928371131420135,
-0.015946755185723305,
0.03000074252486229,
0.018799837678670883,
0.022370003163814545,
0.11291626840829849,
0.08820730447769165,
0.018152978271245956,
0.06225846707820892,
0.012211859226226807,
0.04322114586830139,
-0.06963401287794113,
-0.06354906409978867,
-0.0977717861533165,
-0.04289324954152107,
0.06054644659161568,
0.032085683196783066,
-0.03357795253396034,
-0.1004888191819191,
0.04723907634615898,
-0.006922014523297548,
-0.0035765734501183033,
0.07913468033075333,
-0.019741173833608627,
-0.03231444209814072,
-0.026424136012792587,
-0.05238468572497368,
0.027679668739438057,
0.03727206587791443,
0.03595733642578125,
0.004071660339832306,
-0.002857193583622575,
0.0209520161151886,
0.00825879443436861,
-0.010155165567994118,
-0.04188821092247963,
-0.02091674506664276,
0.036705005913972855,
0.02987859398126602,
0.01840382069349289,
0.03667503222823143,
0.0006171472487039864,
-0.010148400440812111,
-0.0018789562163874507,
0.01121260691434145,
0.07347656041383743,
-0.0769210234284401,
-0.1217460185289383,
9.927372135285733e-34,
0.04578148573637009,
-0.06796860694885254,
0.0562111921608448,
0.08333395421504974,
0.02124028280377388,
-0.03808235004544258,
0.0056772478856146336,
-0.021503565832972527,
0.003241548780351877,
0.05849838629364967,
-0.00971716083586216,
0.09640859067440033,
-0.0029779826290905476,
0.11539952456951141,
0.03283967822790146,
-0.007626903709024191,
-0.05710412934422493,
0.052207279950380325,
-0.05300121754407883,
-0.07311597466468811,
0.012339706532657146,
0.0050233579240739346,
0.011476689018309116,
-0.08381270617246628,
-0.02262013405561447,
-0.008994623087346554,
-0.024185283109545708,
-0.005446358118206263,
-0.010501855984330177,
0.0023092629853636026,
-0.13078001141548157,
0.036600563675165176,
0.027888517826795578,
-0.01393386535346508,
-0.001965905539691448,
0.023533206433057785,
0.011415359564125538,
-0.03666101396083832,
-0.06227831169962883,
0.04477636516094208,
0.06453397125005722,
0.02490045316517353,
-0.02337818033993244,
0.008656329475343227,
-0.07729659229516983,
-0.00008126012835418805,
0.03755559027194977,
0.04975420609116554,
-0.012777739204466343,
-0.02087101899087429,
-0.024005547165870667,
-0.010270367376506329,
0.011672316119074821,
-0.0536072812974453,
-0.04490003362298012,
-0.02826637774705887,
-0.021814990788698196,
0.02121281623840332,
0.040047429502010345,
-0.046589262783527374,
-0.06272177398204803,
0.038774486631155014,
-0.028024373576045036,
0.028675492852926254,
-0.01426686067134142,
0.06510449200868607,
-0.0051449500024318695,
0.008092219941318035,
0.030902395024895668,
0.07246007025241852,
-0.004191989079117775,
0.0935317799448967,
-0.013025048188865185,
-0.06553889065980911,
0.029098402708768845,
0.007026130333542824,
0.03367795795202255,
-0.002554673468694091,
-0.005808162037283182,
-0.08442457765340805,
0.10245215892791748,
-0.1135750487446785,
-0.0750979632139206,
0.031947311013936996,
-0.008651116862893105,
-0.009340593591332436,
0.08165780454874039,
-0.07118990272283554,
-0.18056629598140717,
-0.09810764342546463,
-0.06353018432855606,
-0.04258503019809723,
0.0013466734671965241,
-0.04949593171477318,
-0.047877050936222076,
-2.623503005982729e-33,
0.017359253019094467,
-0.05566217377781868,
0.04173024371266365,
-0.018032750114798546,
-0.003249457338824868,
-0.058834973722696304,
-0.009841745719313622,
-0.026834961026906967,
-0.03669503703713417,
-0.04497946798801422,
-0.08623773604631424,
-0.08651790767908096,
0.055104102939367294,
0.003609149483963847,
-0.042215730994939804,
-0.03756147250533104,
-0.0414496585726738,
0.0503888800740242,
-0.013152885250747204,
0.03903297707438469,
-0.057213619351387024,
0.09600196778774261,
-0.08530460298061371,
0.01819949597120285,
-0.03678302466869354,
0.06168727949261665,
0.004050480201840401,
0.07158330082893372,
-0.04657590389251709,
-0.02892535924911499,
-0.09450262784957886,
0.012848802842199802,
-0.02086849883198738,
-0.005106787197291851,
-0.045758701860904694,
0.11221370846033096,
0.05486252158880234,
-0.031707968562841415,
-0.029470017179846764,
0.04996369779109955,
0.02406523935496807,
-0.026713615283370018,
-0.020629454404115677,
0.004186413250863552,
-0.020701611414551735,
-0.05555352941155434,
-0.02763381041586399,
0.06922180205583572,
-0.03551971912384033,
0.051915403455495834,
-0.04188599810004234,
0.07541433721780777,
-0.07273439317941666,
-0.029963484033942223,
-0.030023789033293724,
-0.00863609742373228,
-0.018647126853466034,
-0.0594543032348156,
0.11763787269592285,
-0.017827235162258148,
-0.06121046468615532,
-0.022296935319900513,
0.006235986016690731,
-0.041215330362319946,
-0.05523006245493889,
0.03564150631427765,
0.021321719512343407,
0.011781998910009861,
0.08604948222637177,
-0.009652980603277683,
-0.04169837385416031,
0.00785774551331997,
0.01800590753555298,
-0.059439949691295624,
0.02779061160981655,
0.03049553371965885,
0.060505833476781845,
-0.007962261326611042,
0.007978607900440693,
0.020368190482258797,
0.06182517483830452,
-0.03616982698440552,
-0.0014310225378721952,
0.030659601092338562,
0.10713820904493332,
0.0877380296587944,
-0.06727779656648636,
0.0877353698015213,
0.05237927660346031,
-0.06874009221792221,
-0.05412260442972183,
0.006112908013164997,
-0.028789089992642403,
0.04161962866783142,
-0.05890972167253494,
-3.808730753007694e-8,
-0.11847128719091415,
-0.060065723955631256,
0.07401970028877258,
0.09147133678197861,
0.03942978009581566,
0.033944666385650635,
-0.0034457009751349688,
0.06789763271808624,
-0.08143676072359085,
-0.08824899792671204,
0.10625807195901871,
-0.006521278060972691,
-0.004022207576781511,
-0.02248113416135311,
0.05179283767938614,
-0.011770045384764671,
-0.00534451100975275,
0.041164904832839966,
-0.043162375688552856,
0.01528880838304758,
0.01847953349351883,
0.006891399621963501,
0.10707136243581772,
-0.07467443495988846,
0.026057252660393715,
-0.05424297973513603,
-0.07559248805046082,
0.05252924561500549,
-0.020458275452256203,
-0.03506523743271828,
0.06309950351715088,
0.004517195746302605,
0.03809136152267456,
0.004112611524760723,
0.041000846773386,
0.021334325894713402,
-0.016812123358249664,
0.025324514135718346,
-0.0562744140625,
-0.0647629052400589,
0.012621582485735416,
0.0157665703445673,
-0.038313645869493484,
0.06533386558294296,
-0.012076608836650848,
-0.038574256002902985,
-0.021279174834489822,
-0.04682247340679169,
0.06104230135679245,
0.010720010846853256,
-0.04209639132022858,
0.07563790678977966,
0.007828719913959503,
0.1049027144908905,
0.06300925463438034,
0.0313364639878273,
-0.058236222714185715,
0.010550396516919136,
-0.04703594371676445,
0.04212317243218422,
0.01872829720377922,
0.09467960894107819,
-0.07058339565992355,
0.01967630721628666
] |
ImDy: Human Inverse Dynamics from Imitated Observations | https://openreview.net/forum?id=br8YB7KMug | [
"Xinpeng Liu",
"Junxuan Liang",
"Zili Lin",
"Haowen Hou",
"Yong-Lu Li",
"Cewu Lu"
] | Poster | Inverse dynamics (ID), which aims at reproducing the driven torques from human kinematic observations, has been a critical tool for gait analysis. However, it is hindered from wider application to general motion due to its limited scalability. Conventional optimization-based ID requires expensive laboratory setups, restricting its availability. To alleviate this problem, we propose to exploit the recently progressive human motion imitation algorithms to learn human inverse dynamics in a data-driven manner. The key insight is that the human ID knowledge is implicitly possessed by motion imitators, though not directly applicable. In light of this, we devise an efficient data collection pipeline with state-of-the-art motion imitation algorithms and physics simulators, resulting in a large-scale human inverse dynamics benchmark as Imitated Dynamics (ImDy). ImDy contains over 150 hours of motion with joint torque and full-body ground reaction force data. With ImDy, we train a data-driven human inverse dynamics solver ImDyS(olver) in a fully supervised manner, which conducts ID and ground reaction force estimation simultaneously. Experiments on ImDy and real-world data demonstrate the impressive competency of ImDyS in human inverse dynamics and ground reaction force estimation. Moreover, the potential of ImDy(-S) as a fundamental motion analysis tool is exhibited with downstream applications. The project page is https://foruck.github.io/ImDy. | Motion Understanding, Inverse Dynamics, Biomechanics | null | 179 | 2410.17610 | [
-0.12604749202728271,
-0.0814911425113678,
0.026971612125635147,
0.022226786240935326,
0.004451563581824303,
-0.09787625074386597,
0.027567729353904724,
-0.022374067455530167,
0.04840002581477165,
0.04370877519249916,
-0.004090634174644947,
-0.04369577020406723,
0.014936260879039764,
0.047706954181194305,
-0.047389667481184006,
0.0070693292655050755,
0.056895721703767776,
0.06333606690168381,
0.012098622508347034,
-0.024654656648635864,
-0.058295924216508865,
0.059362173080444336,
0.03315384313464165,
0.029621407389640808,
0.004618438892066479,
-0.033986616879701614,
-0.008734242059290409,
-0.029228592291474342,
-0.052775729447603226,
-0.029001958668231964,
0.05871986970305443,
-0.053045786917209625,
0.04727976396679878,
-0.08494989573955536,
-0.11833939701318741,
0.0788712128996849,
-0.011471412144601345,
-0.012929817661643028,
-0.042869679629802704,
-0.010766471736133099,
-0.06288939714431763,
-0.008398904465138912,
0.0503610298037529,
0.04274865239858627,
0.10096706449985504,
0.10478680580854416,
-0.01070348545908928,
-0.10754404217004776,
0.06936661899089813,
0.11385120451450348,
-0.07790819555521011,
0.05074505880475044,
0.025672826915979385,
0.012367120943963528,
0.070086769759655,
0.004020767752081156,
0.05565192550420761,
-0.012836131267249584,
0.050635650753974915,
0.02748122252523899,
0.0510653518140316,
-0.03998146578669548,
0.02128951996564865,
0.004277398809790611,
-0.018295736983418465,
-0.06271723657846451,
-0.010556422173976898,
-0.0089042903855443,
0.030829278752207756,
0.0035221711732447147,
-0.029768705368041992,
-0.009971478953957558,
-0.012252070009708405,
-0.0025486245285719633,
-0.0004989898880012333,
0.013611966744065285,
-0.003853488713502884,
-0.0009139692992903292,
0.05135905742645264,
-0.058020398020744324,
-0.024211624637246132,
-0.06667863577604294,
0.05519741773605347,
0.010105207562446594,
0.03907306492328644,
0.026378769427537918,
0.011407089419662952,
0.13358546793460846,
0.029001016169786453,
0.010521408170461655,
-0.06473025679588318,
-0.020598145201802254,
0.016961731016635895,
-0.03028038702905178,
0.08355801552534103,
0.030850104987621307,
-0.01637597754597664,
-0.03164862096309662,
-0.008869349025189877,
0.07677505165338516,
-0.054762933403253555,
0.07590898126363754,
-0.020047353580594063,
0.07123742997646332,
-0.04990973323583603,
-0.04536224156618118,
0.05722111091017723,
0.0004171493055764586,
0.02325240708887577,
0.0019634522031992674,
0.05110817775130272,
-0.051871512085199356,
0.037548232823610306,
0.041090238839387894,
0.09824579954147339,
-0.03038710728287697,
-0.026466093957424164,
-0.004670802969485521,
-0.018634386360645294,
0.01763378083705902,
-0.014442643150687218,
-0.032566387206315994,
-0.010308683849871159,
0.05824565887451172,
0.09024222195148468,
-0.005088505335152149,
-0.003572935936972499,
-4.353431389655841e-34,
0.0010868782410398126,
0.04272286593914032,
0.1417943835258484,
0.040338389575481415,
-0.04274535924196243,
-0.0681125819683075,
0.003972521051764488,
-0.026494409888982773,
-0.039984021335840225,
0.060583099722862244,
0.019045962020754814,
0.04182934761047363,
-0.045953623950481415,
0.07734561711549759,
-0.044325366616249084,
0.037394050508737564,
-0.029805701225996017,
-0.0020871737506240606,
-0.07783722877502441,
0.011133741587400436,
0.12404634803533554,
-0.04639953002333641,
-0.060995664447546005,
-0.0561426617205143,
0.016649046912789345,
0.012770635075867176,
0.0047299922443926334,
0.04130425676703453,
0.027213983237743378,
-0.020010152831673622,
-0.07100927829742432,
-0.032985836267471313,
-0.06480071693658829,
0.06777728348970413,
-0.04564768821001053,
0.0028305414598435163,
0.06819203495979309,
-0.07433845102787018,
-0.04660182073712349,
0.006661741062998772,
0.026044826954603195,
0.02780856564640999,
-0.013531409204006195,
-0.033897604793310165,
-0.11779659241437912,
-0.03532674163579941,
0.09100989252328873,
0.055963434278964996,
-0.0035882662050426006,
0.0007425346411764622,
-0.006594151258468628,
0.02141611836850643,
0.0066931238397955894,
-0.10428012162446976,
-0.04593371972441673,
-0.021015586331486702,
-0.0011733255814760923,
-0.0013522851513698697,
0.024013910442590714,
0.0723244696855545,
-0.020821763202548027,
-0.01339747104793787,
-0.0042593772523105145,
-0.006504868622869253,
-0.004651286173611879,
0.04138670116662979,
-0.03127382695674896,
0.07445041835308075,
0.04657767340540886,
0.08865931630134583,
0.017693040892481804,
-0.03923563286662102,
-0.01431724801659584,
-0.05159929022192955,
0.10319750010967255,
-0.004942247644066811,
0.05529167130589485,
-0.06534456461668015,
-0.04376157745718956,
-0.00510898744687438,
-0.023424455896019936,
-0.022604310885071754,
-0.04671109467744827,
0.004053655546158552,
0.04319893196225166,
-0.06551363319158554,
0.04981623589992523,
-0.057335883378982544,
-0.03872428461909294,
-0.0363614521920681,
-0.05909666046500206,
-0.06822782754898071,
-0.07121354341506958,
0.06300493329763412,
-0.008760754950344563,
-5.818816422715329e-35,
0.006186923943459988,
0.010117917321622372,
-0.022096557542681694,
0.06172439828515053,
0.007291474379599094,
-0.060089390724897385,
0.04388286918401718,
-0.019178448244929314,
-0.0004906289977952838,
-0.07303943485021591,
-0.018036257475614548,
-0.08299464732408524,
0.07357121258974075,
0.05088469013571739,
0.047004424035549164,
-0.007707721553742886,
-0.03585768863558769,
-0.0023350543342530727,
0.04204564541578293,
0.03537420928478241,
-0.006350757088512182,
0.09007003903388977,
-0.01610858179628849,
-0.008832315914332867,
-0.07116695493459702,
0.06705114245414734,
0.001862070057541132,
0.053951796144247055,
-0.03557850420475006,
0.054204944521188736,
-0.04256328195333481,
-0.0261378213763237,
-0.061491891741752625,
-0.01929466612637043,
-0.09634430706501007,
0.08947385847568512,
0.01846132054924965,
0.04517146199941635,
0.0028493155259639025,
-0.013242742046713829,
0.0530201755464077,
-0.01691177859902382,
-0.021759945899248123,
0.021579386666417122,
0.03148569166660309,
-0.0530991293489933,
-0.08327134698629379,
0.08214334398508072,
-0.03061717003583908,
0.008720873855054379,
-0.07186928391456604,
0.062020398676395416,
-0.07615409791469574,
-0.06067925691604614,
-0.03480111435055733,
-0.0412968248128891,
-0.02527475729584694,
0.016978735104203224,
0.03827798739075661,
-0.013553346507251263,
-0.031016049906611443,
0.00892889779061079,
0.004507217090576887,
0.04999789595603943,
-0.054861411452293396,
0.005478053353726864,
-0.0014427226269617677,
0.049161918461322784,
-0.06186710298061371,
-0.0045255753211677074,
-0.0072931223548948765,
-0.024317817762494087,
0.05174138769507408,
-0.005601812619715929,
0.10879947245121002,
-0.0927233174443245,
0.03176705166697502,
-0.003650984261184931,
-0.005383623298257589,
-0.035900264978408813,
0.08229182660579681,
-0.013583676889538765,
0.10206232964992523,
0.06050259619951248,
-0.04089564085006714,
0.1141095831990242,
-0.06313499063253403,
0.10656925290822983,
-0.0047136410139501095,
0.03835628926753998,
-0.04433326795697212,
0.06249452009797096,
0.013357870280742645,
0.028643064200878143,
-0.0021778044756501913,
-3.680596449839868e-8,
-0.12841379642486572,
0.06867602467536926,
0.17339307069778442,
-0.016786884516477585,
-0.024594131857156754,
0.07222414016723633,
-0.012733595445752144,
0.0046844808384776115,
-0.011282652616500854,
-0.057788632810115814,
0.06855763494968414,
-0.0025886090006679296,
0.07378660142421722,
-0.024314241483807564,
-0.0500350184738636,
0.11452294886112213,
-0.015360827557742596,
-0.02148122526705265,
-0.03427049145102501,
0.031273551285266876,
0.018273623660206795,
-0.0003724727430380881,
-0.019541017711162567,
-0.04695696383714676,
0.06991536915302277,
-0.060773592442274094,
-0.0954253226518631,
-0.016161996871232986,
-0.0012950379168614745,
-0.012379338033497334,
-0.02084849402308464,
-0.007782276254147291,
0.0274440236389637,
-0.030071774497628212,
0.0029560651164501905,
-0.01019864808768034,
0.05854735150933266,
0.00584174832329154,
-0.07301241159439087,
-0.06981150805950165,
-0.06318341195583344,
0.05003742128610611,
-0.04690719023346901,
-0.02467990107834339,
0.008646310307085514,
0.053013578057289124,
-0.09592492878437042,
-0.12366641312837601,
0.021536514163017273,
0.009603412821888924,
0.05244378000497818,
-0.11533725261688232,
0.032597485929727554,
0.08535176515579224,
0.10816264152526855,
0.031370311975479126,
-0.04173726961016655,
-0.01489226520061493,
0.07201957702636719,
0.02993856742978096,
0.01072059664875269,
-0.023277325555682182,
-0.002570680109784007,
-0.06107907369732857
] |
PIORF: Physics-Informed Ollivier-Ricci Flow for Long–Range Interactions in Mesh Graph Neural Networks | https://openreview.net/forum?id=qkBBHixPow | [
"Youn-Yeol Yu",
"Jeongwhan Choi",
"Jaehyeon Park",
"Kookjin Lee",
"Noseong Park"
] | Poster | Recently, data-driven simulators based on graph neural networks have gained attention in modeling physical systems on unstructured meshes. However, they struggle with long-range dependencies in fluid flows, particularly in refined mesh regions. This challenge, known as the 'over-squashing' problem, hinders information propagation. While existing graph rewiring methods address this issue to some extent, they only consider graph topology, overlooking the underlying physical phenomena. We propose Physics-Informed Ollivier--Ricci Flow (PIORF), a novel rewiring method that combines physical correlations with graph topology. PIORF uses Ollivier--Ricci curvature (ORC) to identify bottleneck regions and connects these areas with nodes in high-velocity gradient nodes, enabling long-range interactions and mitigating over-squashing. Our approach is computationally efficient in rewiring edges and can scale to larger simulations. Experimental results on 3 fluid dynamics benchmark datasets show that PIORF consistently outperforms baseline models and existing rewiring methods, achieving up to 26.2\% improvement. | graph neural network, fluid dynamics, simulation, mesh, physics, over-squashing, rewiring | PIORF is a physics-informed graph rewiring method that integrates Ollivier–Ricci curvature with velocity gradients to improve long-range interactions in mesh-based GNNs, significantly mitigating over-squashing in fluid dynamics simulations. | 173 | null | [
-0.041587695479393005,
-0.03604716807603836,
0.05442850664258003,
0.09579744189977646,
0.03003539703786373,
-0.018383529037237167,
-0.1320551037788391,
-0.016060713678598404,
0.01574757695198059,
-0.03809443116188049,
-0.06843654066324234,
0.01529955305159092,
-0.037598755210638046,
0.039380818605422974,
-0.010294831357896328,
0.00806584395468235,
-0.01602344773709774,
0.11595465987920761,
-0.07628920674324036,
-0.03097252920269966,
0.04130689054727554,
0.006874374113976955,
-0.028801802545785904,
0.03855990618467331,
0.06170295923948288,
0.046664461493492126,
0.028898993507027626,
-0.0075730206444859505,
0.011411795392632484,
-0.005108844954520464,
0.12443778663873672,
-0.03710394352674484,
-0.1451495736837387,
-0.02419779635965824,
-0.07101879268884659,
0.019712403416633606,
-0.02471337467432022,
0.06148649379611015,
-0.06265156716108322,
-0.016608571633696556,
0.032991304993629456,
0.043870195746421814,
0.06966707855463028,
0.029047679156064987,
0.16380997002124786,
0.002079182304441929,
-0.04128919541835785,
-0.002902888460084796,
0.01132692489773035,
0.025643546134233475,
-0.09566633403301239,
-0.0709468349814415,
0.013527915813028812,
-0.0010873200371861458,
0.066881462931633,
0.0065003749914467335,
0.05042736232280731,
-0.05474241077899933,
0.03305383399128914,
0.032608628273010254,
0.1030745729804039,
-0.05792969837784767,
-0.014256330206990242,
0.027407346293330193,
-0.09391677379608154,
-0.0058220745995640755,
-0.009888894855976105,
0.07441550493240356,
0.033724598586559296,
0.07940750569105148,
0.08250615000724792,
0.06021703779697418,
-0.07046261429786682,
0.00526333786547184,
0.10075431317090988,
0.014080514200031757,
0.04053341969847679,
0.05181919410824776,
-0.034780800342559814,
0.016313428059220314,
0.042361289262771606,
0.003548207227140665,
0.026582082733511925,
-0.06069651246070862,
-0.00615126546472311,
0.017547138035297394,
0.01596670225262642,
-0.0053574698977172375,
-0.009870635345578194,
0.01155474130064249,
-0.020554939284920692,
0.0456455759704113,
-0.030155880376696587,
-0.06626255810260773,
0.05768955871462822,
-0.003388557117432356,
-0.031085293740034103,
-0.030796781182289124,
-0.02911747619509697,
0.06860983371734619,
-0.03416883945465088,
-0.043848853558301926,
-0.0179778840392828,
0.07853541523218155,
0.0015261761145666242,
0.03842673823237419,
0.06206667423248291,
0.08047071099281311,
0.012759560719132423,
-0.029883334413170815,
0.02805553376674652,
-0.010462157428264618,
-0.002468516118824482,
0.029384929686784744,
-0.0265754796564579,
0.020386582240462303,
-0.014606155455112457,
-0.04178337752819061,
0.009094065055251122,
0.1278706192970276,
-0.06963524222373962,
0.0046373107470571995,
-0.07581296563148499,
0.020831532776355743,
0.03435823321342468,
0.017061850056052208,
-0.07707847654819489,
1.902383323172327e-33,
0.018974697217345238,
0.013796854764223099,
0.0306764617562294,
0.043811649084091187,
0.06841014325618744,
0.03524060919880867,
-0.027552686631679535,
-0.06574226915836334,
0.026283184066414833,
-0.011653190478682518,
-0.11694280803203583,
0.09052947908639908,
-0.09701937437057495,
0.05153176560997963,
-0.006992821116000414,
-0.07413002848625183,
-0.0368136391043663,
-0.02836322970688343,
-0.06997448951005936,
-0.03515256941318512,
0.06126674637198448,
-0.10193386673927307,
-0.05225696787238121,
-0.022815890610218048,
-0.015293070115149021,
0.07555100321769714,
0.02973724529147148,
0.0019438127055764198,
0.009469742886722088,
0.028294790536165237,
-0.06978558748960495,
0.0642530545592308,
-0.02110995724797249,
0.06748974323272705,
-0.009180787019431591,
0.009911864995956421,
0.03376402705907822,
-0.061412788927555084,
-0.05425409600138664,
0.03027014434337616,
-0.017538411542773247,
0.0034222316462546587,
-0.03031826578080654,
-0.022106191143393517,
-0.037157025188207626,
-0.05301138758659363,
0.0111924447119236,
0.07652684301137924,
-0.02464228868484497,
-0.05707824230194092,
-0.013495274819433689,
0.063224196434021,
0.011902482248842716,
0.04413742572069168,
0.005409585312008858,
-0.08260910958051682,
0.10658609122037888,
-0.012916404753923416,
0.010132916271686554,
0.08206388354301453,
-0.0077896262519061565,
0.007515846285969019,
-0.019204547628760338,
-0.008426361717283726,
0.0026738732121884823,
0.08711685985326767,
-0.04009680822491646,
0.08706981688737869,
0.08081714063882828,
-0.015722783282399178,
-0.05469149723649025,
0.08926958590745926,
-0.04244963452219963,
-0.002263428643345833,
-0.0019278728868812323,
0.011016761884093285,
0.02796091139316559,
-0.02609865739941597,
-0.011426158249378204,
0.06522391736507416,
-0.0517546683549881,
0.02667853981256485,
0.01013264898210764,
0.054359834641218185,
-0.04835960268974304,
-0.06686250120401382,
0.0331055223941803,
-0.01189412735402584,
-0.0402863547205925,
-0.07945366203784943,
-0.05478992685675621,
-0.034205805510282516,
0.10465201735496521,
-0.01883506216108799,
0.014873272739350796,
-2.1289918180459858e-33,
-0.009603412821888924,
0.06262962520122528,
0.030288085341453552,
0.07799258828163147,
0.023995745927095413,
-0.02862134389579296,
0.009687008336186409,
0.039780519902706146,
0.03844866529107094,
-0.05699065700173378,
-0.014404556713998318,
-0.13453778624534607,
0.009843824431300163,
-0.0267841387540102,
0.13290338218212128,
-0.021764880046248436,
-0.07479705661535263,
-0.06847645342350006,
0.023296356201171875,
-0.051975566893815994,
-0.0708656758069992,
0.03913469985127449,
-0.05533585324883461,
-0.05126509070396423,
0.03783004730939865,
0.02259463258087635,
-0.017204836010932922,
-0.0412842333316803,
-0.005675423890352249,
0.02576850727200508,
-0.08811687678098679,
-0.03186173364520073,
-0.006222943775355816,
0.0277003962546587,
0.01567886397242546,
0.19502979516983032,
0.05430834740400314,
0.017078714445233345,
0.03026917204260826,
-0.007789051625877619,
0.007893430069088936,
-0.03084179013967514,
-0.13088931143283844,
0.00539013184607029,
-0.009594938717782497,
0.05106740444898605,
-0.03627771884202957,
0.02114962227642536,
-0.07889550924301147,
0.006103279069066048,
-0.02119590900838375,
-0.005190046038478613,
-0.005475060548633337,
-0.03345656767487526,
-0.07653596252202988,
-0.07795114815235138,
0.04997331276535988,
0.07105652987957001,
-0.01946769841015339,
-0.025054944679141045,
-0.056155964732170105,
-0.05043286085128784,
-0.013696226291358471,
0.035139359533786774,
-0.013848105445504189,
-0.007565935142338276,
-0.0029971685726195574,
-0.03019321896135807,
0.01800105720758438,
0.004280097782611847,
-0.045940808951854706,
-0.04029710218310356,
0.021729178726673126,
0.007828420028090477,
0.101494699716568,
-0.012830152176320553,
-0.03246675431728363,
0.002888427348807454,
-0.0244743749499321,
0.03703651204705238,
-0.0038356578443199396,
0.045487284660339355,
0.04458852484822273,
0.00888342596590519,
0.11705728620290756,
0.02443797141313553,
-0.04493129998445511,
0.02576579712331295,
0.07424914836883545,
0.10378997027873993,
0.006634509190917015,
-0.015050393529236317,
-0.04111035540699959,
0.02960989810526371,
-0.028718866407871246,
-3.903607392885533e-8,
-0.1046123132109642,
-0.012010496109724045,
0.07764139026403427,
0.00036627735244110227,
0.045953068882226944,
0.030438333749771118,
0.0950622409582138,
0.07874777913093567,
-0.020725596696138382,
0.02660595439374447,
0.03817534074187279,
-0.04847794771194458,
-0.050470612943172455,
-0.08625670522451401,
-0.0358448289334774,
0.048600804060697556,
-0.013328601606190205,
-0.01141491811722517,
0.028333226218819618,
-0.003252954687923193,
-0.07486952841281891,
-0.0037918798625469208,
0.0034401044249534607,
0.006202655378729105,
-0.005974354222416878,
-0.09628909826278687,
-0.06795457750558853,
0.04663684219121933,
0.047507144510746,
-0.04945872724056244,
-0.07824326306581497,
-0.055490944534540176,
0.03166452422738075,
0.0625532865524292,
-0.027561500668525696,
0.08709679543972015,
0.05423654988408089,
0.05942099168896675,
-0.09823830425739288,
-0.006715623661875725,
-0.06096937134861946,
0.051666565239429474,
-0.0025257912930101156,
-0.02304103597998619,
0.029816048219799995,
0.04621565341949463,
0.004310294985771179,
-0.009632116183638573,
0.02157355658710003,
-0.0012116756988689303,
-0.05285947397351265,
0.04000171273946762,
-0.014848429709672928,
0.01735091209411621,
0.003263791324570775,
0.02426973357796669,
-0.08978506922721863,
-0.033898744732141495,
-0.022780951112508774,
-0.021842097863554955,
-0.0866486132144928,
-0.015025222674012184,
-0.001720159430988133,
-0.022572383284568787
] |
ControlAR: Controllable Image Generation with Autoregressive Models | https://openreview.net/forum?id=BWuBDdXVnH | [
"Zongming Li",
"Tianheng Cheng",
"Shoufa Chen",
"Peize Sun",
"Haocheng Shen",
"Longjin Ran",
"Xiaoxin Chen",
"Wenyu Liu",
"Xinggang Wang"
] | Poster | Autoregressive (AR) models have reformulated image generation as next-token prediction, demonstrating remarkable potential and emerging as strong competitors to diffusion models. However, control-to-image generation, akin to ControlNet, remains largely unexplored within AR models. Although a natural approach, inspired by advancements in Large Language Models, is to tokenize control images into tokens and prefill them into the autoregressive model before decoding image tokens, it still falls short in generation quality compared to ControlNet and suffers from inefficiency. To this end, we introduce ControlAR, an efficient and effective framework for integrating spatial controls into autoregressive image generation models. Firstly, we explore control encoding for AR models and propose a lightweight control encoder to transform spatial inputs (e.g., canny edges or depth maps) into control tokens. Then ControlAR exploits the conditional decoding method to generate the next image token conditioned on the per-token fusion between control and image tokens, similar to positional encodings. Compared to prefilling tokens, using conditional decoding significantly strengthens the control capability of AR models but also maintains the model efficiency. Furthermore, the proposed ControlAR surprisingly empowers AR models with arbitrary-resolution image generation via conditional decoding and specific controls. Extensive experiments can demonstrate the controllability of the proposed ControlAR for the autoregressive control-to-image generation across diverse inputs, including edges, depths, and segmentation masks. Furthermore, both quantitative and qualitative results indicate that ControlAR surpasses previous state-of-the-art
controllable diffusion models, e.g., ControlNet++. | controllable image generation, autoregressive models, autoregressive image generation, diffusion models, image generation | null | 170 | 2410.02705 | [
-0.07692482322454453,
-0.10253576189279556,
0.038530971854925156,
0.08960946649312973,
0.04838912934064865,
0.03997046500444412,
-0.04920656979084015,
-0.06686395406723022,
0.0701821967959404,
0.004586579743772745,
-0.03235046938061714,
-0.05303390324115753,
0.03971192613244057,
0.04875754937529564,
-0.014892338775098324,
-0.0010121195809915662,
0.038130540400743484,
0.05930288881063461,
-0.11373428255319595,
-0.051237888634204865,
0.012252459302544594,
-0.038352906703948975,
-0.02480345033109188,
-0.025009239092469215,
0.07119417935609818,
0.05333222076296806,
-0.007540019694715738,
-0.08462399989366531,
0.03511238470673561,
-0.0582280270755291,
0.04883701726794243,
-0.05196574702858925,
-0.013989598490297794,
0.028561707586050034,
-0.07185529172420502,
0.013372414745390415,
-0.03653339669108391,
0.025774264708161354,
0.024211429059505463,
-0.03608770668506622,
-0.020816367119550705,
0.016413316130638123,
0.03458705171942711,
0.0005563238519243896,
0.054355744272470474,
0.050249308347702026,
0.033918291330337524,
-0.026576893404126167,
-0.0035454262979328632,
0.027170708402991295,
-0.015241158194839954,
-0.014660345390439034,
-0.013924135826528072,
-0.027900073677301407,
0.013730654492974281,
-0.006692462135106325,
0.05406123027205467,
-0.009934812784194946,
0.04164302721619606,
-0.009256552904844284,
-0.049884721636772156,
-0.06537948548793793,
-0.03267457336187363,
-0.025997810065746307,
-0.041999366134405136,
-0.048245519399642944,
0.027540013194084167,
0.06052091345191002,
0.009679586626589298,
-0.012008234858512878,
-0.029489347711205482,
0.04108399152755737,
-0.0322754941880703,
-0.004099169280380011,
0.02357189729809761,
0.0680568590760231,
0.04699939489364624,
-0.03812029957771301,
0.04187093675136566,
-0.09789919853210449,
0.11735760420560837,
0.012979642488062382,
0.04810609295964241,
0.022308582440018654,
0.03429185971617699,
0.007048237603157759,
0.011569839902222157,
0.06710390746593475,
0.07789924740791321,
0.002781599760055542,
-0.0509619265794754,
-0.05086411535739899,
0.0597984716296196,
-0.004009196534752846,
0.04874913766980171,
0.007599058095365763,
-0.04585077986121178,
-0.12137514352798462,
-0.04261994734406471,
0.06155911087989807,
0.054013997316360474,
-0.02289487235248089,
-0.012434462085366249,
-0.02799869142472744,
0.12918785214424133,
-0.05358656495809555,
0.07617621123790741,
0.04529806599020958,
0.017742477357387543,
0.01795881986618042,
0.014673734083771706,
0.02575409598648548,
-0.019350431859493256,
0.007029039319604635,
0.03380659595131874,
-0.019868221133947372,
-0.017875930294394493,
-0.027091873809695244,
0.06862693279981613,
0.01106664352118969,
-0.042962446808815,
-0.03834623843431473,
-0.0908009335398674,
0.0559159554541111,
0.017514668405056,
-0.06496632099151611,
-0.02517084777355194,
2.7087318565471143e-33,
-0.045012690126895905,
0.05035719275474548,
0.03666919842362404,
0.013006173074245453,
0.04991193488240242,
0.009131176397204399,
-0.00408791471272707,
-0.03040480427443981,
-0.010353131219744682,
-0.04755565896630287,
-0.022743958979845047,
-0.07160867750644684,
-0.09250643849372864,
0.11755235493183136,
0.06092304363846779,
-0.025652863085269928,
-0.022174546495079994,
0.04172215610742569,
-0.010972462594509125,
0.034270018339157104,
-0.0385090671479702,
0.02578848972916603,
-0.04357759281992912,
-0.033446986228227615,
-0.0544729009270668,
-0.0019786290358752012,
0.05847948044538498,
-0.04380227252840996,
-0.0031328846234828234,
-0.011045695282518864,
-0.0988500788807869,
-0.020573755726218224,
0.026424361392855644,
0.04389527812600136,
-0.023052925243973732,
-0.0234933253377676,
-0.043084196746349335,
-0.0235943291336298,
-0.013746360316872597,
-0.008208788000047207,
-0.020356202498078346,
0.05741298943758011,
-0.08981884270906448,
-0.002789780730381608,
-0.06964882463216782,
-0.08272188156843185,
-0.0038643141742795706,
0.09053097665309906,
-0.1581978052854538,
0.005018743220716715,
0.09050368517637253,
0.07406292110681534,
-0.0501994788646698,
-0.055136412382125854,
-0.013079970143735409,
-0.08601780235767365,
0.019893640652298927,
-0.01251167245209217,
0.04771948233246803,
0.0076345414854586124,
0.0324154756963253,
-0.0008963800501078367,
0.04537133127450943,
0.05841401219367981,
0.051843900233507156,
-0.05931898206472397,
0.06173070892691612,
0.041387442499399185,
-0.003129680408164859,
0.04016220569610596,
-0.04568963497877121,
0.03856603056192398,
-0.06653589755296707,
-0.08747627586126328,
0.014406492002308369,
-0.008933519013226032,
0.04207706078886986,
-0.005150096025317907,
-0.04467430338263512,
0.08025011420249939,
-0.08434531837701797,
0.005475263111293316,
0.030584031715989113,
-0.021778788417577744,
0.017734814435243607,
-0.07982069253921509,
0.04326855391263962,
-0.05226782336831093,
-0.0388362854719162,
-0.07845745980739594,
-0.007750236429274082,
-0.06676516681909561,
-0.032959990203380585,
0.06193976476788521,
-0.005719405133277178,
-2.649427064522176e-33,
0.08165162801742554,
0.06422790884971619,
0.025859462097287178,
-0.008078621700406075,
-0.11161811649799347,
-0.07431096583604813,
0.04172774776816368,
0.015006527304649353,
-0.039977096021175385,
-0.08868181705474854,
-0.09095488488674164,
0.00491253612563014,
0.029261749237775803,
-0.036833811551332474,
0.00946907140314579,
-0.10296210646629333,
0.048691827803850174,
-0.04067395254969597,
-0.018901566043496132,
0.02008490450680256,
-0.010448180139064789,
0.07770760357379913,
0.006019046064466238,
0.07394257187843323,
-0.13145871460437775,
0.042240798473358154,
-0.00030411622719839215,
0.10841123014688492,
-0.045724768191576004,
0.022804219275712967,
0.009096193127334118,
-0.010137070901691914,
0.06980247795581818,
-0.007452181074768305,
-0.08823754638433456,
0.012971190735697746,
0.08448098599910736,
0.002992714988067746,
-0.08956900238990784,
0.08121725171804428,
0.04487556219100952,
-0.03594902530312538,
-0.027936313301324844,
0.07747522741556168,
-0.05985376983880997,
0.00464728893712163,
0.022974295541644096,
0.0027316855266690254,
0.02563713677227497,
0.018191838636994362,
-0.017709331586956978,
0.020349320024251938,
-0.13695573806762695,
0.04103582352399826,
-0.016329899430274963,
-0.08796168118715286,
-0.08389558643102646,
0.003591069020330906,
0.07895349711179733,
-0.011133464984595776,
-0.05317264050245285,
-0.07162829488515854,
-0.0006957739824429154,
-0.0521453432738781,
-0.011292573064565659,
-0.0027850959450006485,
-0.024773715063929558,
-0.0184183269739151,
0.02312314137816429,
0.08745116740465164,
0.07634737342596054,
-0.019937686622142792,
0.03443330153822899,
0.10946362465620041,
0.054916199296712875,
-0.0350952111184597,
0.06952396780252457,
-0.002772310981526971,
0.034114398062229156,
-0.06697600334882736,
-0.04479866847395897,
-0.03227638825774193,
0.054337307810783386,
0.09013465791940689,
0.027475187554955482,
0.05880133807659149,
-0.02904524840414524,
-0.012876597233116627,
0.08435305207967758,
0.051783207803964615,
-0.02572365291416645,
0.030048668384552002,
0.009944296441972256,
0.06966452300548553,
0.013702881522476673,
-4.059655367427695e-8,
-0.0382499061524868,
0.055374857038259506,
0.06557511538267136,
-0.011161279864609241,
-0.06349200010299683,
-0.0368923619389534,
0.05763111263513565,
-0.0038333022966980934,
0.0255119726061821,
-0.0006245715194381773,
0.04482583701610565,
-0.027914388105273247,
-0.04736761748790741,
-0.075218066573143,
-0.038390785455703735,
0.04271704703569412,
0.06546609103679657,
0.036053404211997986,
0.06002194434404373,
-0.0018634791485965252,
-0.05345612019300461,
0.03518248721957207,
0.018938597291707993,
-0.05241568014025688,
0.06106578931212425,
-0.031185433268547058,
-0.02907709591090679,
0.033769622445106506,
0.10476112365722656,
-0.018992925062775612,
0.08054748177528381,
0.024838970974087715,
0.0528721809387207,
-0.0100757060572505,
-0.03713229298591614,
0.05596693232655525,
-0.00588231161236763,
-0.04602217674255371,
-0.006153313908725977,
-0.059311144053936005,
0.011506840586662292,
0.07303713262081146,
-0.06323684751987457,
-0.020855911076068878,
-0.041577789932489395,
0.0419510155916214,
0.12729033827781677,
-0.10443425178527832,
0.03309469297528267,
-0.05699103698134422,
-0.0006731090252287686,
-0.005436426028609276,
0.003087269840762019,
0.07030357420444489,
0.07230887562036514,
0.006483899895101786,
0.044892046600580215,
-0.0251625943928957,
0.06326513737440109,
0.09785040467977524,
-0.0010441634804010391,
0.03216177225112915,
-0.09679201990365982,
-0.010175582952797413
] |
Subsets and Splits