搜索
433Vol.43,No.320173ACTAAUTOMATICASINICAMarch,2017GAN1;21;31;31;341;5GAN(Generativeadversarialnetworks).GAN,,..,GAN,.GAN,.GAN,GAN,GAN,,ACP(Arti¯cialsocieties,computationalexperiments,andparallelexecution).,,,,,ACP,,,,,.GAN.,2017,43(3):321¡332DOI10.16383/j.aas.2017.y000003GenerativeAdversarialNetworks:TheStateoftheArtandBeyondWANGKun-Feng1;2GOUChao1;3DUANYan-Jie1;3LINYi-Lun1;3ZHENGXin-Hu4WANGFei-Yue1;5AbstractGenerativeadversarialnetworks(GANs)havebecomeahotresearchtopicinarti¯cialintelligence.Inspiredbythetwo-playerzero-sumgame,GANiscomposedofageneratorandadiscriminator,bothtrainedwiththeadversariallearningmechanism.TheaimofGANistoestimatethepotentialdistributionofexistingdataandgeneratenewdatasamplesfromthesamedistribution.Sinceitsinitiation,GANhasbeenwidelystudiedduetoitsenormousprospectforapplications,includingimageandvisioncomputing,speechandlanguageprocessing,informationsecurity,andchessgame.InthispaperwesummarizethestateoftheartofGANandlookintoitsfuture.Firstofall,wesurveytheGAN0sbackground,theoreticandimplementationmodels,application¯elds,advantagesanddisadvantages,anddevelopmenttrends.Then,weinvestigatetherelationbetweenGANandparallelintelligencewiththeconclusionthatGANhasagreatpotentialinparallelsystemsespeciallyincomputationalexperiments,intermsofvirtual-realinteractionandintegration.Finally,weclarifythatGANcanprovidespeci¯candsubstantialalgorithmicsupportfortheACPtheory.KeywordsGenerativeadversarialnetworks,generativemodels,zero-sumgame,adversariallearning,parallelintelli-gence,ACPmethodologyCitationWangKun-Feng,GouChao,DuanYan-Jie,LinYi-Lun,ZhengXin-Hu,WangFei-Yue.Generativeadversarialnetworks:thestateoftheartandbeyond.ActaAutomaticaSinica,2017,43(3):321¡3322017-02-012017-03-01ManuscriptreceivedFebruary1,2017;acceptedMarch1,2017(61533019,71232006,91520301)SupportedbyNationalNaturalScienceFoundationofChina(61533019,71232006,91520301)RecommendedbyAssociateEditorLIUDe-Rong1.1001902.2660003.1000494.MN554145.4100731.TheStateKeyLaboratoryofManagementandControlforComplexSystems,InstituteofAutomation,ChineseAcademyofSciences,Beijing100190,China2.QingdaoAcademyofIntelligentIndustries,Qingdao266000,China3.UniversityofChineseAcademyofSciences,Beijing100049,China4.De-partmentofComputerScienceandEngineering,UniversityofGAN(Generativeadversarialnetworks)Goodfellow[1]2014.GAN(,),.,;,.[2].GANMinnesota,Minneapolis,MN55414,USA5.ResearchCen-terforComputationalExperimentsandParallelSystemsTech-nology,NationalUniversityofDefenseTechnology,Changsha410073,China32243(Minimaxgame),[3],.,GAN,.GAN,LeCun\.,GAN,,,,,,[4].,GAN[5¡6][7][8].GAN,.1GAN.2GAN,GNN.3GAN.4GAN,GAN,.,5.1GANGAN,GAN.1.1,,,[2;9].:.,,,,.,,.,,1.,;,,.,.,,,.,,Feynman\WhatIcannotcreate,Idonotunderstand.(,.).GAN,..,GAN.1Fig.1Thelevelsofarti¯cialintelligence1.2,...,.:.,,,.[10¡11][12¡14].,.,,,.,.,.,,[15],.3:GAN323,.,.GAN,,,.GAN,,.,.1.310,[16¡17],.,,,.,[18¡19],[20],[21]..,,;,,;,,.,,.1.4,..[22],,,.AlphaGo[23],AlphaGo,,.,[24],,.[25¡26]:,;,.[27¡28],,.,.,GAN.2GAN2.1GANGAN.(Generator)(Discriminator),,;,,,.GAN2.GAN,,DG,xz.G(z)Gpdata.,1.G(z),0.D:(x)(G(z)),GG(z)DD(G(z))xDD(x),D2GANFig.2ComputationprocedureandstructureofGAN32443G,D,,G.2.2GANGAN.,G,D.Sigmoid,D,:ObjD(µD;µG)=¡12Ex»pdata(x)[logD(x)]¡12Ez»pz(z)[log(1¡D(g(z)))](1),xpdata(x),zpz(z)(),E(¢).,pdata(x)(1)pg(x)(0).G,(1),,(1):ObjD(µD;µG)=¡12Zxpdata(x)log(D(x))dx¡12Zzpz(z)log(1¡D(g(z)))dz=¡12Zx[pdata(x)log(D(x))+pg(x)log(1¡D(x))]dx(2)mn,y2[0;1],¡mlog(y)¡nlog(1¡y)(3)mm+n.,G,(2)D¤G(x)=pdata(x)pdata(x)+pg(x)(4),.(4),GAN,.,D(x)x.x,DD(x)1,G(z),D,D(G(z))0,G1.GD,GObjG(µG)=¡ObjD(µD;µG).GAN|,GAN:minGmaxDff(D;G)=Ex»pdata(x)[logD(x)]+Ez»pz(z)[log(1¡D(G(z)))]g(5),GAN,DG(z),,Glog(1¡D(G(z))).:G,D,D;D,G,D.pdata=pg.GAN,,DkG1.2.3GANGoodfellow[1]2014GAN,GAN,.3.GAN,,Jensen-Shannon,.,Arjovsky[29]Wasser-steinGAN(W-GAN).W-GANEarth-MoverJensen-Shannon,fGAN,fLipschitz.,GAND,,D,.,Qi[30]Loss-sensitiveGAN(LS-GAN),Lipschitz,.,W-GANLS-GANGAN,.GAN(),.Odena[31]3:GAN3253GAN((a)GAN[1],W-GAN[29],LS-GAN[30];(b)Semi-GAN[31];(c)C-GAN[32];(d)Bi-GAN[33];(e)Info-GAN[34];(f)AC-GAN[35];(g)Seq-GAN[6])Fig.3ComputationproceduresandstructuresofGAN-derivedmodelsSemi-GAN,D.,ConditionalGAN(CGAN)[32]yGD,y.GAN,Donahue[33]Bidi-rectionalGANs(BiGANs),.GAN,BiGANsQx,32643minG;QmaxDf(D;Q;G).InfoGAN[34]GAN.GAN,z,In-foGAN.Gzc,zGAN,c,.GANpG(x)=pG(xjc),cG.G(z;c),[34]I(c;G(z;c)),minGmaxDffI(D;G)=f(D;G)¡¸I(c;G(z;c))g(6),p(cjx),.Odena[35]AuxiliaryClassi¯erGAN(AC-GAN),.,,,,AC-GAN,,.GAN,Yu[6]Seq-GAN.RNNG,CNND,DG.D,,D.3GAN\,GAN,,.GAN,.GAN,.GAN.3.1GAN.Twitter,Ledig[36]GAN.VGG[37],[19],4,GAN.4GAN[36]Fig.4IllustrationofGAN-generatedimage[36]GAN.Santana[38]GAN,RNN,5.GAN,GAN.Gou[39¡40],.Shrivastava[41]GAN(SimGAN),,.,,.3.2GAN.Li[5]GAN,.Zhang[42]GAN,CNN,LSTM,;,,CNN3:GAN3275GAN(,)[38]Fig.5AnotherillustrationofGAN-generatedimages(Oddcolumnsshowthegeneratedimages,andevencolumnsshowthetargetimages)[38].SeqGAN[6]G,,SeqGAN.Reed[43]GAN,,,,,.3.3GAN,GAN,SeqGAN[6].GAN[44¡45]GANActor-critic[46].Hu[7]MalGAN,GAN,GAN.Childambaram[8]GAN,,.4GAN4.1GANGAN.GAN,.,,.,.GAN,,,,.GAN,,.,.GAN,.,32843,.GAN,.GAN.GAN,GAN.,G