0310生成对抗网络进阶.pptxVIP

DeepLearningAdvancedGenerativeadversarialNets

Wassersteindistancev.s.JSdivergenceSJTUDeepLearningLecture.2Whyvanishinggradient?JSDorKLD—Allthedensitieshavethesamedistancestoeachother.UseWassersteindistance,theygrowlinearlyfromlefttorightOptimalTransportadistancebetweentwoprobabilities

SJTUDeepLearningLecture.3OptimalTransport—Origins:Monge’sProblem

OptimalTransport—Origins:Monge’sProblemSJTUDeepLearningLecture.4

SJTUDeepLearningLecture.5OptimalTransport—Origins:Monge’sProblem

SJTUDeepLearningLecture.6OptimalTransport—Origins:Monge’sProblem

OptimalTransport—Origins:Monge’sProblemSJTUDeepLearningLecture.7

SJTUDeepLearningLecture.8OptimalTransport—Origins:Monge’sProblem

SJTUDeepLearningLecture.9OptimalTransport—Origins:Monge’sProblem

SJTUDeepLearningLecture.10OptimalTransport—Origins:Monge’sProblem

SJTUDeepLearningLecture.11OptimalTransport—Origins:Monge’sProblem

SJTUDeepLearningLecture.12OptimalTransport—Origins:Monge’sProblem

SJTUDeepLearningLecture.13OptimalTransport—Origins:Monge’sProblemMonge’sProblem:FindanoptimalT(transportmap)

MongeProblem:ExistenceandUniquenessSJTUDeepLearningLecture.14MongeProblemifthesolutionexists,itmaynotbeuniqueThesolutiontoaMongeProblemmaynotexist

KantorovichRelaxationSJTUDeepLearningLecture.15KantorovichProblem(continuous)ThreeclassesofOTproblemsallowforonevs.many

WassersteinDistancesSJTUDeepLearningLecture.16???

WassersteindistanceandGANSJTUDeepLearningLecture.17?WassersteindistanceDualdistance???

WGANSJTUDeepLearningLecture.18???

WGANSJTUDeepLearningLecture.19

WGANSJTUDeepLearningLecture.20

WGAN_GPSJTUDeepLearningLecture.21WhysomeexperimentsfailforWGAN?WhathappensinWGAN?

WGAN_GPSJTUDeepLearningLecture.22???

WGAN_GPSJTUDeepLearningLecture.23

WGAN_GPfortextgenerationSJTUDeepLearningLecture.24

ComparisonamongOriginalGAN,WGANandWGAN_GPSJT