资源描述：
《Adversarially Learned Inference》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、PublishedasaconferencepaperatICLR2017ADVERSARIALLYLEARNEDINFERENCEVincentDumoulin1,IshmaelBelghazi1,BenPoole2OlivierMastropietro1,AlexLamb1,MartinArjovsky3AaronCourville1y1MILA,UniversitédeMontréal,firstname.lastname@umontreal.ca.2NeuralDynamicsandCompu
2、tationLab,Stanford,poole@cs.stanford.edu.3NewYorkUniversity,martinarjovsky@gmail.com.yCIFARFellow.ABSTRACTWeintroducetheadversariallylearnedinference(ALI)model,whichjointlylearnsagenerationnetworkandaninferencenetworkusinganadversarialprocess.Thegenerat
3、ionnetworkmapssamplesfromstochasticlatentvariablestothedataspacewhiletheinferencenetworkmapstrainingexamplesindataspacetothespaceoflatentvariables.Anadversarialgameiscastbetweenthesetwonetworksandadiscriminativenetworkistrainedtodistinguishbetweenjointl
4、atent/data-spacesamplesfromthegenerativenetworkandjointsamplesfromtheinferencenetwork.Weillustratetheabilityofthemodeltolearnmutuallycoherentinferenceandgen-erationnetworksthroughtheinspectionsofmodelsamplesandreconstructionsandconfirmtheusefulnessofthel
5、earnedrepresentationsbyobtainingaperformancecompetitivewithstate-of-the-artonthesemi-supervisedSVHNandCIFAR10tasks.1INTRODUCTIONDeepdirectedgenerativemodelhasemergedasapowerfulframeworkformodelingcomplexhigh-dimensionaldatasets.Thesemodelspermitfastance
6、stralsampling,butareoftenchallengingtolearnduetothecomplexitiesofinference.Recently,threeclassesofalgorithmshaveemergedaseffectiveforlearningdeepdirectedgenerativemodels:1)techniquesbasedontheVariationalAutoencoder(VAE)thataimtoimprovethequalityandeffici
7、encyofinferencebylearninganinferencemachine(Kingma&Welling,2013;Rezendeetal.,2014),2)techniquesbasedonGenerativeAdversarialNetworks(GANs)thatbypassinferencealtogether(Goodfellowetal.,2014)and3)autoregressiveapproaches(vandenOordetal.,2016b;c;a)thatforeg
8、olatentrepresentationsandinsteadmodeltherelationshipbetweeninputvariablesdirectly.Whilealltechniquesareprovablyconsistentgiveninfinitecapacityanddata,inpracticetheylearnverydifferentkindsofgenerativemodelsontypicaldatasets.arXiv:1
