GenerativeAdversarialNetwork(GAN)RestrictedBoltzmannMachine:://.tw/~tlkagk/courses/MLDS_2015_2/Lecture/RBM%20(v2).ecm.mp4/index.htmlGibbsSampling:://.tw/~tlkagk/courses/MLDS_2015_2/Lecture/MRF%20(v2).ecm.mp4/index.htmlOutlook:NIPS2016Tutorial:GenerativeAdversarialNetworksAuthor:IanGoodfellowPaper:s:///abs/1701.00160Video:YoucanfindtipsfortrainingGANhere:s://github/soumith/ganhacksReviewGenerationDrawing?WritingPoems?Review:Auto-encoderAscloseaspossibleNNEncoderNNDecodercodeNNDecodercodeRandomlygenerateavectorascodeImage?Review:Auto-encoderNNDecodercode2D-1.51.5

NNDecoder

NNDecoderReview:Auto-encoder-1.51.5NNEncoderNNDecodercodeinputoutputAuto-encoderVAENNEncoderinputNNDecoderoutputm1m2m3

Fromanormaldistribution

X+Minimizereconstructionerror

exp

MinimizeAuto-EncodingVariationalBayes,s:///abs/1312.6114ProblemsofVAEItdoesnotreallytrytosimulaterealimagesNNDecodercodeOutputAscloseaspossibleOnepixeldifferencefromthetargetOnepixeldifferencefromthetargetRealisticFakeTheevolutionofgenerationNNGeneratorv1Discri-minatorv1Realimages:NNGeneratorv2Discri-minatorv2NNGeneratorv3Discri-minatorv3BinaryClassifierTheevolutionofgenerationNNGeneratorv1Discri-minatorv1Realimages:NNGeneratorv2Discri-minatorv2NNGeneratorv3Discri-minatorv3GAN-DiscriminatorNNGeneratorv1Realimages:Discri-minatorv1image1/0(realorfake)SomethinglikeDecoderinVAERandomlysampleavector11110000GAN-GeneratorDiscri-minatorv1NNGeneratorv1Randomlysampleavector0.13UpdatingtheparametersofgeneratorTheoutputbeclassifiedas“real”(ascloseto1aspossible)Generator+Discriminator=anetworkUsinggradientdescenttoupdatetheparametersinthegenerator,butfixthediscriminator1.0v2GAN

–二次元人物頭像鍊成DCGAN:s://github/carpedm20/DCGAN-tensorflowGAN

–二次元人物頭像鍊成100roundsGAN

–二次元人物頭像鍊成1000roundsGAN

–二次元人物頭像鍊成2000roundsGAN

–二次元人物頭像鍊成5000roundsGAN

–二次元人物頭像鍊成10,000roundsGAN

–二次元人物頭像鍊成20,000roundsGAN

–二次元人物頭像鍊成50,000roundsBasicIdeaofGANMaximumLikelihoodEstimation

Likelihoodofgeneratingthesamples

MaximumLikelihoodEstimation

Itisdifficulttocomputethelikelihood.

BasicIdeaofGANGeneratorGGisafunction,inputz,outputxGivenapriordistributionPprior(z),aprobabilitydistributionPG(x)isdefinedbyfunctionGDiscriminatorDDisafunction,inputx,outputscalarEvaluatethe“difference”betweenPG(x)andPdata(x)ThereisafunctionV(G,D).

HardtolearnbymaximumlikelihoodBasicIdea

GivenG,whatistheoptimalD*maximizingGivenx,theoptimalD*maximizing

AssumethatD(x)canhaveanyvaluehere

Givenx,theoptimalD*maximizingFindD*maximizing:

aDbD0<<1

22

Jensen-Shannondivergence

Intheend……

0<<log2

Algorithm

Algorithm

DecreaseJS

divergence(?)DecreaseJS

divergence(?)Algorithm

DecreaseJS

divergence(?)

smaller

……

Don’tupdateGtoomuchInpractice…

Maximize

MinimizeCross-entropyBinaryClassifierOutputisD(x)Minimize–logD(x)IfxisapositiveexampleIfxisanegativeexampleMinimize–log(1-D(x))

PositiveexamplesNegativeexamples

MaximizeMinimize

MinimizeCross-entropyBinaryClassifierOutputisf(x)Minimize–logf(x)IfxisapositiveexampleIfxisanegativeexampleMinimize–log(1-f(x))

AlgorithmRepeatktimesLearningDLearningG

CanonlyfindlowerfoundofOnlyOnceObjectiveFunctionforGenerator

inRealImplementation

Realimplementation:labelxfromPGaspositive

SlowatthebeginningDemoThecodeusedindemofrom:s://github/osh/KerasGAN/blob/master/MNIST_CNN_GAN_v2.ipynbIssueaboutEvaluatingtheDivergenceEvaluatingJSdivergenceMartinArjovsky,

LéonBottou,TowardsPrincipledMethodsforTrainingGenerativeAdversarialNetworks,

2017,arXivpreprintEvaluatingJSdivergenceJSdivergenceestimatedbydiscriminatortellinglittleinformations:///abs/1701.07875WeakGeneratorStrongGeneratorDiscriminator

Reason1.Approximatebysampling

10=0

log2Weakenyourdiscriminator?CanweakdiscriminatorcomputeJSdivergence?Discriminator

Reason2.thenatureofdata

10=0

log2

UsuallytheydonothaveanyoverlapEvaluationBetterEvaluation

Better…………Notreallybetter……AddNoiseAddsomeartificialnoisetotheinputsofdiscriminatorMakethelabelsnoisyforthediscriminator

DiscriminatorcannotperfectlyseparaterealandgenerateddataNoisesdecayovertimeModeCollapseModeCollapseDataDistributionGeneratedDistributionModeCollapse

Whatwewant…Inreality…FlawinOptimization?

ModifiedfromIanGoodfellow’stutorial

Thismaynotbethereason(basedonIanGoodfellow’stutorial)SomanyGANs……ModifyingtheOptimizationofGANfGANWGANLeast-squareGANLossSensitiveGANEnergy-basedGANBoundary-seekingGANUnrollGAN……DifferentStructurefromtheOriginalGANConditionalGANSemi-supervisedGANInfoGANBiGANCycleGANDiscoGANVAE-GAN……ConditionalGANMotivationGeneratorScottReed,ZeynepAkata,XinchenYan,LajanugenLogeswaran,BerntSchiele,HonglakLee,“GenerativeAdversarialText-to-ImageSynthesis”,ICML2016TextImageScottReed,

ZeynepAkata,

SantoshMohan,

SamuelTenka,

BerntSchiele,

HonglakLee,“LearningWhatandWheretoDraw”,NIPS2016HanZhang,

TaoXu,

HongshengLi,

ShaotingZhang,

XiaoleiHuang,

XiaogangWang,

DimitrisMetaxas,“StackGAN:TexttoPhoto-realisticImageSynthesiswithStackedGenerativeAdversarialNetworks”,arXivprepring,2016MotivationChallengeNNTextImage(apoint,notadistribution)Text:“train”NN

output

ConditionalGANG

conditionPriordistributionLearntoapproximateP(x|c)D(