文档介绍
GANforText:HypeorRealDeal?XiangZhangPart1:GenerativeAdversarialNetsin5MinutesGanforText:HypeorRealDeal?GenerativeAdversarialNets鈼ProposedbyIanJ.Goodfellowetal.in2014RandomzGeneratorgg(z)d(g(z))DiscriminatordSampleyd(y)鈼Learningprocess:鈼Optimized:d(y)鈫d(g(z))鈫鈼Optimizeg:d(g(z))鈫GenerativeAdversarialNets鈼ProposedbyIanJ.Goodfellowetal.in2014RandomzGeneratorgg(z)d(g(z))DiscriminatordSampleyd(y)鈼Itpartiallysolvesthemultimodal(asinstatistics)trainingproblem鈼Generatorgisafunctionmapping.Theonlywayforittogeneratedifferentoutputsistousedifferentinputrandomvaluesz.鈼Usingthediscriminatord,wenolongerneedexplicit(z,y)pairsGenerativeAdversarialNets鈼ProposedbyIanJ.Goodfellowetal.in2014RandomzGeneratorgg(z)d(g(z))DiscriminatordSampleyd(y)鈼Butitdoesnotguaranteediversity鈼Itdoesnotencouragegtooutputdifferent