资源描述:
《Inferring and Executing Programs for Visual Reasoning》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、InferringandExecutingProgramsforVisualReasoningJustinJohnson1BharathHariharan2LaurensvanderMaaten2JudyHoffman1LiFei-Fei1C.LawrenceZitnick2RossGirshick21StanfordUniversity2FacebookAIResearchAbstractExistingmethodsforvisualreasoningattempttodirectlymapinputstooutputsusingbla
2、ck-boxarchitectureswithoutexplicitlymodelingtheunderlyingreasoningprocesses.Asaresult,theseblack-boxmodelsoftenlearntoexploitbiasesinthedataratherthanlearningtoperformvisualreason-Howmanychairsareatthetable?Isthereapedestrianinmylane?ing.Inspiredbymodulenetworks,thispaperp
3、roposesamodelforvisualreasoningthatconsistsofaprogramgen-eratorthatconstructsanexplicitrepresentationoftherea-soningprocesstobeperformed,andanexecutionenginethatexecutestheresultingprogramtoproduceananswer.Boththeprogramgeneratorandtheexecutionengineareimplementedbyneuraln
4、etworks,andaretrainedusingacombinationofbackpropagationandREINFORCE.UsingIsthepersonwiththebluehatIsthereamattecubethathasthetheCLEVRbenchmarkforvisualreasoning,weshowthattouchingthebikeintheback?samesizeastheredmetalobject?ourmodelsignificantlyoutperformsstrongbaselinesand
5、Figure1.Compositionalreasoningisacriticalcomponentneededgeneralizesbetterinavarietyofsettings.forunderstandingthecomplexvisualscenesencounteredinap-plicationssuchasroboticnavigation,autonomousdriving,and1.Introductionsurveillance.Currentmodelsfailtodosuchreasoning[19].Inma
6、nyapplications,computer-visionsystemsneedtoanswersophisticatedqueriesbyreasoningaboutthevisualtolearndatasetbiasesbutnotreasoning[7,18,19].world(Figure1).TodealwithnovelobjectinteractionsorInthispaper,wearguethattosuccessfullyperformcom-object-attributecombinations,visualr
7、easoningneedstobeplexreasoningtasks,itmightbenecessarytoexplicitlyin-compositional:withouteverhavingseena“persontouchingcorporatecompositionalreasoninginthemodelstructure.abike”,themodelshouldbeabletounderstandthephraseSpecifically,weinvestigateanewmodelforvisualques-arXiv:
8、1705.03633v1[cs.CV]10May2017byputtingtogetheritsunderstandingof“person”,“bike”tionansweri