深度学习及其应用_ 课件 25-生成对抗网络基础.pptxVIP

机器视觉前沿领域介绍生成对抗网络模型基础

DeepgenerativemodelsSJTUDeepLearningLecture.2unsupervisedlearningsupervisedlearningclassificationregressionspeechrecognitionimagerecognitionmachineTranslationdimensionalityreduction(PCA),clusteringandmoregenerativemodelsVariationalAuto-Encoder(VAE)GenerativeAdversarialNets(GAN)…

GenerativeAdversarialNetworksSJTUDeepLearningLecture.3IanGoodfellowGameTheoryZero-SumTwo-PlayerGamesDatagenerationCompetitionandcooperationbetweenGANandVAE

SJTUDeepLearningLecture.4TrendofnumberofGANpaperspublishedatarXivNumberofpaper/nightrome/really-awesome-gan/kozistr/Awesome-GANs

ModelStructuresofGANSJTUDeepLearningLecture.5Generator:generatethefakedatafromrandomvariableDiscriminator:discriminatethefakedataandrealdataAlternatelytraintheDiscriminatorandGenerator(ksteps)(1step)Makethebothnetworksworkbetter???DiscriminatorDreal/fake

ModelStructuresofGANSJTUDeepLearningLecture.6????DiscriminatorDreal/fake

ModelStructuresofGANSJTUDeepLearningLecture.7

SJTUDeepLearningLecture.8????????????RVGTrainingDTrainingGNashequilibrium?

ThebasicmodelofGANSJTUDeepLearningLecture.9?????DiscriminatorDreal/fakemaxforDminforDmaxforG

DoubleLosses?SJTUDeepLearningLecture.10

AlgorithmSJTUDeepLearningLecture.11

NetworkArchitectureSJTUDeepLearningLecture.12OriginalGAN—Multi-LayerPerceptronforGandDDCGAN—CNNforGandD(detailsonpage25)Forgeneratorimagetranslation—U-Netimageenhancement—ResNettextgeneration—CNN+RNNGeneratorArchitectureofDCGAN

ShortcomingsofOriginalGANSJTUDeepLearningLecture.13VanishinggradientModecollapseOnlyforcontiunuousdata(image),butnotfordiscretedata(text)Evaluations(subjective)VanishinggradientModecollapse

ModecollapseSJTUDeepLearningLecture.14

EvaluationMetricsSJTUDeepLearningLecture.15InceptionScore（IS）?Drawbacks