资源描述：
《Contextual RNN-GANs for Abstract Reasoning Diagram Generation 》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。
1、ContextualRNN-GANsforAbstractReasoningDiagramGenerationArnabGhosh1,VivekaKulharia1,AmitabhaMukerjee1,VinayNamboodiri1,MohitBansal21IITKanpur2UNCChapelHillfarnabghosh93,vivekakulhariag@gmail.com,famit,vinaypng@iitk.ac.in,mbansal@cs.unc.eduAbstractUnderstan
2、ding,predicting,andgeneratingobjectmotionsandtransformationsisacoreprobleminartificialintelli-gence.Modelingsequencesofevolvingimagesmaypro-videbetterrepresentationsandmodelsofmotionandmayultimatelybeusedforforecasting,simulation,orvideogen-Figure1:Exampleab
3、stractreasoningproblem,whereourmodeleration.DiagrammaticAbstractReasoningisanavenueinwasabletogenerateanimageveryclosetothecorrectanswer.whichdiagramsevolveincomplexpatternsandoneneedstoinfertheunderlyingpatternsequenceandgeneratethenextFig.1showsanexamplep
4、roblemfromourDAT-DARimageinthesequence.Forthis,wedevelopanovelCon-datasetandhighlightstheintricaciesofthereasoningin-textualGenerativeAdversarialNetworkbasedonRecurrentNeuralNetworks(Context-RNN-GANs),whereboththegen-volvedininferringthecorrectanswer(i.e.,t
5、henextimageeratorandthediscriminatormodulesarebasedoncontex-inthesequence).Differentpatterncomponentsonboththetualhistory(modeledasRNNs)andtheadversarialdiscrim-sidesandboththecornersarechangingindifferentandmul-inatorguidesthegeneratortoproducerealisticima
6、gesfortipleways,makingitaninterestingchallengetocorrectlytheparticulartimestepintheimagesequence.Weevaluategeneratethenextimageinthesequence.1Accurategenera-theContext-RNN-GANmodel(anditsvariants)onanoveltionmodelsdevelopedforsuchareasoningtaskcanbeuseddata
7、setofDiagrammaticAbstractReasoning,whereitper-forgeneralAIapplicationssuchasforecastingandsimula-formscompetitivelywith10th-gradehumanperformancebuttiongeneration.Thesemodelswillalsobeusefulforgen-thereisstillscopeforinterestingimprovementsascomparederation
8、ofreal-worldimagesandvideos,arecentresearchtocollege-gradehumanperformance.Wealsoevaluateourdirectionincomputervisionanddeeplearning(Goodfel-modelonastandardvideonext-framepredictiontask,achiev-ingimpr
