资源描述：
《On the Quantitative Analysis of Decoder-Based Generative Models 》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、UnderreviewasaconferencepaperatICLR2017ONTHEQUANTITATIVEANALYSISOFDECODER-BASEDGENERATIVEMODELSYuhuaiWuYuriBurdaRuslanSalakhutdinovDepartmentofComputerScienceOpenAISchoolofComputerScienceUniversityofTorontoyburda@openai.comCarnegieMellonUniversityywu@cs.tor
2、onto.edursalakhu@cs.cmu.eduRogerGrosseDepartmentofComputerScienceUniversityofTorontorgrosse@cs.toronto.eduABSTRACTThepastseveralyearshaveseenremarkableprogressingenerativemodelswhichproduceconvincingsamplesofimagesandothermodalities.Asharedcomponentofmanypo
3、werfulgenerativemodelsisadecodernetwork,aparametricdeepneuralnetthatdefinesagenerativedistribution.Examplesincludevariationalautoencoders,generativeadversarialnetworks,andgenerativemomentmatchingnetworks.Unfortunately,itcanbedifficulttoquantifytheperformanceo
4、fthesemodelsbecauseoftheintractabilityoflog-likelihoodestimation,andinspectingsamplescanbemisleading.WeproposetouseAnnealedImportanceSamplingforevaluatinglog-likelihoodsfordecoder-basedmodelsandvalidateitsaccuracyusingbidirectionalMonteCarlo.Usingthistechni
5、que,weanalyzetheperformanceofdecoder-basedmodels,theeffectivenessofexistinglog-likelihoodestimators,thedegreeofoverfitting,andthedegreetowhichthesemodelsmissimportantmodesofthedatadistribution.1INTRODUCTIONInrecentyears,deepgenerativemodelshavedramaticallypu
6、shedforwardthestate-of-the-artingenerativemodellingbygeneratingconvincingsamplesofimages(Radfordetal.,2016),achievingstate-of-the-artsemi-supervisedlearningresults(Salimansetal.,2016),andenablingautomaticimagemanipulation(Zhuetal.,2016).Manyofthemostsuccess
7、fulapproachesaredefinedintermsofaprocesswhichsampleslatentvariablesfromasimplefixeddistribution(suchasGaussianoruniform)andthenappliesalearneddeterministicmappingwhichwewillrefertoasadecodernetwork.Importantexamplesincludevariationalautoencoders(VAEs)(Kingma&
8、Welling,2014;arXiv:1611.04273v1[cs.LG]14Nov2016Rezendeetal.,2014),generativeadversarialnetworks(GANs)(Goodfellowetal.,2014),generativemomentmatchingnetworks(GMMNs)(Li&Swersky,2015;Dziugaiteetal.,2015),
