Download presentation
1 -

FewShotAdversarialDomainAdaptation


SaeidMotiianQuinnJonesSeyedMehdiIranmaneshGianfrancoDorettoLaneDepartmentofComputerScienceandElectricalEngineeringWestVirginiaUniversitysamotiianqjones1seiranmaneshgidorettomixwvueduAbstractThisworkpr

mila/ milly's Recent Documents

ToAll EmployeesDateOctober 27 20
ToAll EmployeesDateOctober 27 20

FromJoel HawkinsSubjectEmployment Opportunity Seattle WA Marine Assurance CoordinatorUMMARYResponsible for providing support to Marine Assurance Group personnel en

published 0K
NCBI Handout Series  SRA Sequence Read Archive  Last Updated on Septem
NCBI Handout Series SRA Sequence Read Archive Last Updated on Septem

SRA Sequence Read Archive Collection of sequence data from next-generation sequencing technology for different organisms https//wwwncbinlmnihgov/sra/ https//wwwncbinlmnihgov/Traces/sra/ National Cen

published 0K
YearEthnicit
YearEthnicit

yinEnteringYearGenderGraduatedyears5years6years4years5

published 0K
IRC WATCHLIST 2021
IRC WATCHLIST 2021

19 18 2021 marks a decade of con30ict in Syria Despite its protracted nature the crisis continues to reach new lows as con30ict displacement and needs grow while humanitarians146 cross-border

published 0K
GETTING ORGANIZED
GETTING ORGANIZED

SCHOOL AGE1Getting OrganizedYour Child146s School-Age YearsBeing an advocate for your child is easier when you are organized and prepared to discuss your child146s needs and any concerns With organize

published 0K
Sun Mon Tue Wed Thu Fri Sat 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
Sun Mon Tue Wed Thu Fri Sat 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17

El CENTRO PUBLIC LIBRARY Library Board Meeting Learning Express Library Whatever your goal LearningEx-press Librarys resources will help you succeed 375 S 1st St El Centro CA 92243 wwwecliborg 9

published 0K
Islam is an Arabic word meaning  submission and in the
Islam is an Arabic word meaning submission and in the

Islamreligious context means submission to the will of God Islam is derived from the Arabic word salm which literally means peaceThe religion demonstrates peace and tolerance Muslimsthe followers of

published 0K
Nature Vol 276 7 December 1978 of crocodilians We are also told that s
Nature Vol 276 7 December 1978 of crocodilians We are also told that s

spend most of their time halfasleep and the only stimuli that can rouse them from their state of torpor are hunger-pangs due to a previous period of fasting due in turn to laziness and sex On the othe

published 0K
Download Section

Download - The PPT/PDF document "" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.






Document on Subject : "FewShotAdversarialDomainAdaptation"— Transcript:

1 Few-ShotAdversarialDomainAdaptation Saei
Few-ShotAdversarialDomainAdaptation SaeidMotiian,QuinnJones,SeyedMehdiIranmanesh,GianfrancoDorettoLaneDepartmentofComputerScienceandElectricalEngineeringWestVirginiaUniversity{samotiian,qjones1,seiranmanesh,gidoretto}@mix.wvu.eduAbstractThisworkprovidesaframeworkforaddressingtheproblemofsuperviseddomainadaptationwithdeepmodels.Themainideaistoexploitadversariallearningtolearnanembeddedsubspacethatsimultaneouslymaximizestheconfusionbetweentwodomainswhilesemanticallyaligningtheirembedding.Thesupervisedsettingbecomesattractiveespeciallywhenthereareonlyafewtargetdatasamplesthatneedtobelabeled.Inthisfew-shotlearningscenario,alignmentandseparationofsemanticprobabilitydistributionsisdifcultbecauseofthelackofdata.Wefoundthatbycarefullydesigningatrainingschemewherebythetypicalbinaryadversarialdiscriminatorisaugmentedtodistinguishbetweenfourdifferentclasses,itispossibletoeffectivelyaddressthesupervisedadaptationproblem.Inaddition,theapproachhasahigh“speed”ofadaptation,i.e.itrequiresanextremelylownumberoflabeledtargettrainingsamples,evenonepercategorycanbeeffective.Wethenextensivelycomparethisapproachtothestateoftheartindomainadaptationintwoexperiments:oneusingdatasetsforhandwrittendigitrecognition,andoneusingdatasetsforvisualobjectrecognition.1IntroductionAsdeeplearningapproacheshavegainedprominenceincomputervisionwehaveseentasksthathavelargeamountsofavailablelabeleddataourishwithimprovedresults.Therearestillmanyproblemsworthsolvingwherelabeleddataonanequallylargescaleistooexpensivetocollect,annotate,orboth,andbyextensionastraightforwarddeeplearningapproachwouldnotbefeasible.Typically,insuchascenario,practitionerswilltrainorreuseamodelfromacloselyrelateddatasetwithalargeamountofsamples,herecalledthesourcedomain,andthentrainwiththemuchsmallerdatasetofinterest,referredtoasthetargetdomain.Thisprocessiswell-knownunderthenamenetuning.Finetuning,whilesimpletoimplement,hasbeenfoundtobesub-optimalwhencomparedtolatertechniquessuchasdomainadaptation[5].DomainAdaptationcanbesupervised[58,27],unsupervised[15,34],orsemi-supervised[16,21,63],dependingonwhatdataisavailableinalabeledformatandhowmuchcanbecollected.Unsuperviseddomainadaptation(UDA)algorithmsdonotneedanytargetdatalabels,buttheyrequirelargeamountsoftargettrainingsamples,whichmaynotalwaysbeavailable.Conversely,superviseddomainadaptation(SDA)algorithmsdorequirelabeledtargetdata,andbecauselabelinginformationisavailable,forthesamequantityoftargetdata,SDAoutperformsUDA[38].Therefore,iftheavailabletargetdataisscarce,SDAbecomesattractive,evenifthelabelingprocessisexpensive,becauseonlyfewsamplesneedtobeprocessed.Mostdomainadaptationapproachestrytondafeaturespacesuchthattheconfusionbetweensourceandtargetdistributionsinthatspaceismaximum(domainconfusion).Becauseofthat,itishardtosaywhetherasampleinthefeaturespacehascomefromthesourcedistributionorthetargetdistribution.Recently,generativeadversarialnetworks[18]havebeenintroducedforimagegenerationwhichcanalsobeusedfordomainadaptation.In[18],thegoalistolearnadiscriminator31stConferenceonNeuralInformationProcessingSystems(NIPS2017),LongBeach,CA,USA. Figure1:ExamplesfromMNIST[32]andSVHN[40]ofgroupedsamplepairs.G1iscomposedofsamplesofthesameclassfromthesourcedatasetinthiscaseMNIST.G2iscomposedofsamplesofthesameclass,butoneisfromthesourcedatasetandtheotherisfromthetargetdataset.InG3thesamplesineachpairare

2 fromthesourcedatasetbutwithdifferingclas
fromthesourcedatasetbutwithdifferingclasslabels.Finally,pairsinG4arecomposedofsamplesfromthetargetandsourcedatasetswithdifferingclasslabels.todistinguishbetweenrealsamplesandgenerated(fake)samplesandthentolearnageneratorwhichbestconfusesthediscriminator.Domainadaptationcanalsobeseenasagenerativeadversarialnetworkwithonedifference,indomainadaptationthereisnoneedtogeneratesamples,instead,thegeneratornetworkisreplacedwithaninferencenetwork.Sincethediscriminatorcannotdetermineifasampleisfromthesourceorthetargetdistributiontheinferencebecomesoptimalintermsofcreatingajointlatentspace.Inthismanner,generativeadversariallearninghasbeensuccessfullymodiedforUDA[33,59,49]andprovidedverypromisingresults.Hereinstead,weareinterestedinadaptingadversariallearningforSDAwhichwearecallingfew-shotadversarialdomainadaptation(FADA)forcaseswhenthereareveryfewlabeledtargetsamplesavailableintraining.Inthisfew-shotlearningregime,ourSDAmethodhasprovencapableofincreasingamodel'sperformanceataveryhighratewithrespecttotheinclusionofadditionalsamples.Indeed,evenoneadditionalsamplecansignicantlyincreaseperformance.Ourrstcontributionistohandlethisscarcedatawhileprovidingeffectivetraining.Oursecondcontributionistoextendadversariallearning[18]toexploitthelabelinformationoftargetsamples.Weproposeanovelwayofcreatingpairsofsamplesusingsourceandtargetsamplestoaddresstherstchallenge.Weassignagrouplabeltoapairaccordingtothefollowingprocedure:0ifsamplesofapaircomefromthesourcedistributionandthesameclasslabel,1iftheycomefromthesourceandtargetdistributionsbutthesameclasslabel,2iftheycomefromthesourcedistributionbutdifferentclasslabels,and3iftheycomefromthesourceandtargetdistributionsandhavedifferentclasslabels.Thesecondchallengeisaddressedbyusingadversariallearning[18]totrainadeepinferencefunction,whichconfusesawell-traineddomain-classdiscriminator(DCD)whilemaintainingahighclassicationaccuracyforthesourcesamples.TheDCDisamulti-classclassierthattakespairsofsamplesasinputandclassiesthemintotheabovefourgroups.ConfusingtheDCDwillencouragedomainconfusion,aswellasthesemanticalignmentofclasses.OurthirdcontributionisanextensivevalidationofFADAagainstthestate-of-the-art.Althoughourmethodisgeneral,andcanbeusedforalldomainadaptationapplications,wefocusonvisualrecognition.2RelatedworkNaivelytrainingaclassierononedatasetfortestingonanotherisknowntoproducesub-optimalresults,becauseaneffectknownasdatasetbias[42,57,56]orcovariateshift[51]occursduetoadifferenceinthedistributionsoftheimagesbetweenthedatasets.Priorworkindomainadaptationhasminimizedthisshiftlargelyinthreeways.Sometrytondafunctionwhichcanmapfromthesourcedomaintothetargetdomain[47,28,19,16,11,55,52].Othersndasharedlatentspacethatbothdomainscanbemappedtobeforeclassication[35,2,39,13,14,41,37,38].Finally,someuseregularizationtoimprovethetonthetargetdomain[4,1,62,10,3,8].UDAcanleveragethersttwoapproacheswhileSDAusesthesecond,third,oracombinationofthetwoapproaches.Inadditiontothesemethods,[6,36,50]haveaddressedUDAwhenanauxiliarydataview[31,37],isavailableduringtraining,butthatisbeyondthescopeofthiswork.Forthisapproachwearefocusedonndingasharedsubspaceforboththesourceandtargetdistributions.Siamesenetworks[7]workwellforsubspacelearningandhaveworkedverywellwithdeepconvolutionalneuralnetworks[9,53,30,61].Siamesenetworkshavealsobeenusefulin2 G 2 1 G 3

3 1 G 4 1 G 1 1 Figure2:Few-shotadversari
1 G 4 1 G 1 1 Figure2:Few-shotadversarialdomainadaptation.Forsimplicityweshowournetworksinthecaseofweightsharing(gs=gt=g).(a)Intherststep,weinitializedgandhusingthesourcesamplesDs.(b)WefreezegandtrainaDCD.ThepictureshowsapairfromthesecondgroupG2whenthesamplescomefromtwodifferentdistributionsbutthesameclasslabel.(c)WefreezetheDCDandupdategandh.domainadaptationrecently.In[58],whichisadeepSDAapproach,unlabeledandsparselylabeledtargetdomaindataareusedtooptimizefordomaininvariancetofacilitatedomaintransferwhileusingasoftlabeldistributionmatchingloss.In[54],whichisadeepUDAapproach,unlabeledtargetdataisusedtolearnanonlineartransformationthatalignscorrelationsoflayeractivationsindeepneuralnetworks.SomeapproacheswentbeyondSiameseweight-sharingandusedcouplenetworksforDA.[27]usestwoCNNstreams,forsourceandtarget,fusedattheclassierlevel.[45],whichisadeepUDAapproachandcanbeseenasanSDAafterne-tuning,alsousesatwo-streamsarchitecture,forsourceandtarget,withrelatedbutnotsharedweights.[38],whichisanSDAapproach,createspositiveandnegativepairsusingsourceandtargetdataandthenndsasharedfeaturespacebetweensourceandtargetbybringingtogetherthepositivepairsandpushingapartthenegativepairs.Recently,adversariallearning[18]hasshownpromisingresultsindomainadaptationandcanbeseenasexamplesofthesecondcategory.[33]introducedacoupledgenerativeadversarialnetwork(CoGAN)forlearningajointdistributionofmulti-domainimagesfordifferentapplicationsincludingUDA.[59]hasusedtheadversariallossfordiscriminativeUDA.[49]introducesanapproachthatleveragesunlabeleddatatobringthesourceandtargetdistributionscloserbyinducingasymbioticrelationshipbetweenthelearnedembeddingandagenerativeadversarialframework.Hereweuseadversariallearningtotraininferencenetworkssuchthatsamplesfromdifferentdistributionsarenotdistinguishable.Weconsiderthetaskwhereveryfewlabeledtargetdataareavailableintraining.Withthisassumption,itisnotpossibletousethestandardadversariallossusedin[33,59,49],becausethetrainingtargetdatawouldbeinsufcient.Weaddressthatproblembymodifyingtheusualpairingtechniqueusedinmanyapplicationssuchaslearningsimilaritymetrics[7,23,22].Ourpairingtechniqueencodesdomainlabelsaswellasclasslabelsofthetrainingdata(sourceandtargetsamples),producingfourgroupsofpairs.Wethenintroduceamulti-classdiscriminatorwithfouroutputsanddesignanadversariallearningstrategytondasharedfeaturespace.Ourmethodalsoencouragesthesemanticalignmentofclasses,whileotheradversarialUDAapproachesdonot.3Few-shotadversarialdomainadaptationInthissectionwedescribethemodelweproposetoaddresssuperviseddomainadaptation(SDA).WearegivenatrainingdatasetmadeofpairsDs=f(xsi;ysi)gNi=1.Thefeaturexsi2XisarealizationfromarandomvariableXs,andthelabelysi2YisarealizationfromarandomvariableYs.Inaddition,wearealsogiventhetrainingdataDt=f(xti;yti)gMi=1,wherexti2XisarealizationfromarandomvariableXt,andthelabelsyti2Y.Weassumethatthereisacovariateshift[51]betweenXsandXt,i.e.,thereisadifferencebetweentheprobabilitydistributionsp(Xs)andp(Xt).WesaythatXsrepresentsthesourcedomainandthatXtrepresentsthetargetdomain.Underthissettingsthegoalistolearnapredictionfunctionf:X!Ythatduringtestingisgoingtoperformwellondatafromthetargetdomain.Theproblemformulatedthusfaristypicallyreferredtoassuperviseddomainadaptation.Inthisworkweareespeciallyconcernedwiththeversionofthisproblemwhereonlyveryfewtargetlabeled3 Source g 1

4 h 1  1 h 1 Loss (1) (a) (b) (c) Lo
h 1  1 h 1 Loss (1) (a) (b) (c) Loss (3) Loss (1) Loss (4) DCD G 2 1 G 2 1 Loss (1) Algorithm1FADAalgorithm 1:TraingandhonDsusing(1).2:UniformlysampleG1,G3fromDsxDs.3:UniformlysampleG2,G4fromDsxDt.4:TrainDCDw.r.t.gt=gs=gusing(3).5:whilenotconvergentdo6:Updategandhbyminimizing(5).7:UpdateDCDbyminimizing(3).8:endwhile samplesperclassareavailable.Weaimathandlingcaseswherethereisonlyonetargetlabeledsample,andtherecanevenbesomeclasseswithnotargetsamplesatall.InabsenceofcovariateshiftavisualclassierfistrainedbyminimizingaclassicationlossLC(f)=E[`(f(Xs);Y)];(1)whereE[]denotesstatisticalexpectationand`couldbeanyappropriatelossfunction.WhenthedistributionsofXsandXtaredifferent,adeepmodelfstrainedwithDswillhavereducedperformanceonthetargetdomain.IncreasingitwouldbetrivialbysimplytraininganewmodelftwithdataDt.However,Dtissmallanddeepmodelsrequirelargeamountsoflabeleddata.Ingeneral,fcouldbemodeledbythecompositionoftwofunctions,i.e.,f=hg.Hereg:X!ZwouldbeaninferencefromtheinputspaceXtoafeatureorinferencespaceZ,andh:Z!Ywouldbeafunctionforpredictingfromthefeaturespace.Withthisnotationwewouldhavefs=hsgsandft=htgt,andtheSDAproblemwouldbeaboutndingthebestapproximationforgtandht,giventheconstraintsontheavailabledata.Ifgsandgtareabletoembedsourceandtargetsamples,respectively,toadomaininvariantspace,itissafetoassumefromthefeaturetothelabelspacethatht=hs=h.Therefore,domainadaptationparadigmsarelookingforsuchinferencefunctionssothattheycanusethepredictionfunctionhsfortargetsamples.TraditionalunsupervisedDA(UDA)paradigmstrytoalignthedistributionsofthefeaturesinthefeaturespace,mappedfromthesourceandthetargetdomainsusingametricbetweendistributions,MaximumMeanDiscrepancy[20]beingapopularoneandothermetricslikeKullbackLeiblerdivergence[29]andJensen–Shannon[18]divergencebecomingpopularwhenusingadversariallearning.Oncetheyarealigned,aclassierfunctionwouldnolongerbeabletotellwhetherasampleiscomingfromthesourceorthetargetdomain.RecentUDAparadigmstrytondinferencefunctionstosatisfythisimportantgoalusingadversariallearning.AdversarialtraininglooksforadomaindiscriminatorDthatisabletodistinguishbetweensamplesofsourceandtargetdistributions.InthiscaseDisabinaryclassiertrainedwiththestandardcross-entropylossLadv�D(Xs;Xt;gs;gt)=�E[log(D(gs(Xs)))]�E[log(1�D(gt(Xt)))]:(2)Oncethediscriminatorislearned,adversariallearningtriestoupdatethetargetinferencefunctiongtinordertoconfusethediscriminator.Inotherwords,theadversarialtrainingislookingforaninferencefunctiongtthatisabletomapatargetsampletoafeaturespacesuchthatthediscriminatorDwillnolongerdistinguishitfromasourcesample.Fromtheabovediscussionitisclearthatinordertoperformwell,UDAneedstoalignthedistributionseffectivelyinordertobesuccessful.Thiscanhappenonlyifdistributionsarerepresentedbyasufcientlylargedataset.Therefore,UDAapproachesareinapositionofweaknesswhenweassumeDttobesmall.Moreover,UDAapproacheshavealsoanotherintrinsiclimitation;evenwithperfectconfusionalignment,thereisnoguaranteethatsamplesfromdifferentdomainsbutwiththesameclasslabelwillmapnearbyinthefeaturespace.Thislackofsemanticalignmentisamajorsourceofperformancereduction.3.1HandlingScarceTargetDataWeareinterestedinthecasewhereveryfewlabeledtargetsamples(aslowas1sampleperclass)areavailable.Wearefacingtwochallengesinthissetting.First,sincethesizeofDtissmall,weneedto&#

5 2;ndawaytoaugmentit.Second,weneedtosomeh
2;ndawaytoaugmentit.Second,weneedtosomehowusethelabelinformationofDt.Therefore,wecreatepairsofsamples.Inthisway,weareabletoalleviatethelackoftrainingtargetsamplesby4 pairingthemwitheachtrainingsourcesample.In[38],wehaveshownthatcreatingpositiveandnegativepairsusingsourceandtargetdataisveryeffectiveforSDA.Sincethemethodproposedin[38]doesnotencodethedomaininformationofthesamples,itcannotbeusedinadversariallearning.Hereweextend[38]bycreating4groupsofpairs(Gi;i=1;2;3;4)asfollows:webreakdownthepositivepairsintotwogroups(Groups1and2),wherepairsoftherstgroupconsistofsamplesfromthesourcedistributionwiththesameclasslabels,whilepairsofthesecondgroupalsohavethesameclasslabelbutcomefromdifferentdistributions(onefromthesourceandonefromthetargetdistribution).Thisisimportantbecausewecanencodebothlabelanddomaininformationoftrainingsamples.Similarly,webreakdownthenegativepairsintotwogroups(Groups3and4),wherepairsofthethirdgroupconsistofsamplesfromthesourcedistributionwithdifferentclasslabels,whilepairsoftheforthgroupcomefromdifferentclasslabelsanddifferentdistributions(onefromthesourceandonefromthetargetdistributions).SeeFigure1.InordertogiveeachgroupthesameamountofmembersweuseallpossiblepairsfromG2,asitisthesmallest,andthenuniformlysamplefromthepairsinG1,G3,andG4tomatchthesizeofG2.Anyreasonableamountofportionsbetweenthenumbersofthepairscanalsobeused.Inclassicaladversariallearningwewouldatthispointlearnadomaindiscriminator,butsincewehavesemanticinformationtoconsideraswell,weareinterestedinlearningamulti-classdiscriminator(wecallitdomain-classdiscriminator(DCD))inordertointroducesemanticalignmentofthesourceandtargetdomains.Byexpandingthebinaryclassiertoitsmulticlassequivalent,wecantrainaclassierthatwillevaluatewhichofthe4groupsagivensamplepairbelongsto.WemodeltheDCDwith2fullyconnectedlayerswithasoftmaxactivationinthelastlayerwhichwecantrainwiththestandardcategoricalcross-entropylossLFADA�D=�E[4Xi=1yGilog(D((Gi)))];(3)whereyGiisthelabelofGiandDistheDCDfunction.isasymbolicfunctionthattakesapairasinputandoutputstheconcatenationoftheresultsoftheappropriateinferencefunctions.TheoutputofispassedtotheDCD(Figure2).Inthesecondstep,weareinterestedinupdatinggtinordertoconfusetheDCDinsuchawaythattheDCDcannolongerdistinguishbetweengroups1and2,andalsobetweengroups3and4usingthelossLFADA�g=�E[yG1log(D((G2)))+yG3log(D((G4)))]:(4)(4)isinspiredbythenon-saturatinggame[17]andwillforcetheinferencefunctiongttoembedtargetsamplesinaspacethatDCDwillnolongerbeabletodistinguishbetweenthem.Connectionwithmulti-classdiscriminators:Consideranimagegenerationtaskwheretrainingsamplescomefromkclasses.Learningtheimagegeneratorcanbedonebyanystandardk-classclassierandaddinggeneratedsamplesasanewclass(generatedclass)andcorrespondinglyincreasingthedimensionoftheclassieroutputfromktok+1.Duringtheadversariallearning,onlythegeneratedclassisconfused.Thishasproveneffectiveforimagegeneration[48]andothertasks.However,thisisdifferentthantheproposedDCD,wheregroup1isconfusedwith2,andgroup3isconfusedwith4.Inspiredby[48],weareabletocreateak+4classiertoalsoguaranteeahighclassicationaccuracy.Therefore,wesuggestthat(4)needstobeminimizedtogetherwiththemainclassierlossLFADA�g=� E[yG1log(D(g(G2)))+yG3log(D(g(G4)))]+E[`(f(Xs);Y)]+E[`(f(Xt);Y)];(5)where strikesthebalancebetweenclassicationandconfus

6 ion.Misclassifyingpairsfromgroup2asgroup
ion.Misclassifyingpairsfromgroup2asgroup1andlikewiseforgroups4and3,meansthattheDCDisnolongerabletodistinguishpositiveornegativepairsofdifferentdistributionsfrompositiveornegativepairsofthesourcedistribution,whiletheclassierisstillabletodiscriminatepositivepairsfromnegativepairs.ThissimultaneouslysatisesthetwomaingoalsofSDA,domainconfusionandclassseparabilityinthe5 Table1:MNIST-USPS-SVHNdatasets.ClassicationaccuracyfordomainadaptationovertheMNIST,USPS,andSVHNdatasets.M,U,andSstandforMNIST,USPS,andSVHNdomain.LBisourbasemodelwithoutadaptation.FTandFADAstandforne-tuningandourmethod,respectively. TraditionalUDAAdversarialUDALB[60][45][15][33][59][49]SDA1234567 M!U65.447.860.791.891.289.492.5FT82.384.985.786.587.288.488.6[38]85.089.090.191.492.493.092.9FADA89.191.391.993.393.494.094.4 U!M58.663.167.373.789.190.190.8FT72.678.281.983.183.483.684.0[38]78.482.285.886.188.889.689.4FADA81.184.287.589.991.191.291.5 S!M60.1--82.076.0-84.7FT65.568.670.773.374.574.675.4FADA72.881.882.685.186.186.887.2 M!S20.3--40.1--36.4FT29.731.236.136.738.138.339.1FADA37.740.542.946.346.146.847.0 S!U66.0------FT69.471.874.376.278.177.978.9FADA78.383.285.285.786.287.187.5 U!S15.3------FT19.922.222.824.625.425.425.6FADA27.529.834.536.037.941.342.9 featurespace.UDAonlylooksfordomainconfusionanddoesnotaddressclassseparability,becauseofthelackoflabeledtargetsamples.ConnectionwithconditionalGANs:ConcatenationofoutputsofdifferentinferenceshasbeendonebeforeinconditionalGANs.Forexample,[43,44,64]concatenatetheinputtexttothepenultimatelayersofthediscriminators.[25]concatenatespositiveandnegativepairsbeforepassingthemtothediscriminator.However,allofthemusethevanillabinarydiscriminator.Relationshipbetweengsandgt:Thereisnorestrictionforgsandgtandtheycanbeconstrainedorunconstrained.Anobviouschoiceofconstraintisequality(weight-sharing)whichmakestheinferencefunctionssymmetric.Thiscanbeseenasaregularizerandwillreduceovertting[38].Anotherapproachwouldbelearninganasymmetricinferencefunction[45].Sincewehaveaccesstoveryfewtargetsamples,weuseweight-sharing(gs=gt=g).Choiceofgs,gt,andh:Sinceweareinterestedinvisualrecognition,theinferencefunctionsgsandgtaremodeledbyaconvolutionalneuralnetwork(CNN)withsomeinitialconvolutionallayers,followedbysomefullyconnectedlayerswhicharedescribedspecicallyintheexperimentssection.Inaddition,thepredictionfunctionhismodeledbyfullyconnectedlayerswithasoftmaxactivationfunctionforthelastlayer.TrainingProcess:Herewediscussthetrainingprocessfortheweight-sharingregularizer(gs=gt=g).Oncetheinferencefunctionsgandthepredictionfunctionharechosen,FADAtakesthefollowingsteps:First,gandhareinitializedusingthesourcedatasetDs.Then,thementionedfourgroupsofpairsshouldbecreatedusingDsandDt.ThenextstepistrainingDCDusingthefourgroupsofpairs.Thisshouldbedonebyfreezingg.Inthenextstep,theinferencefunctiongandpredictionfunctionhshouldbeupdatedinordertoconfuseDCDandmaintainhighclassicationaccuracy.ThisshouldbedonebyfreezingDCD.SeeAlgorithm1andFigure2.Thetrainingprocessforthenonweight-sharingcasecanbederivedsimilarly.4ExperimentsWepresentresultsusingtheOfcedataset[47],theMNISTdataset[32],theUSPSdataset[24],andtheSVHNdataset[40].4.1MNIST-USPS-SVHNDatasetsTheMNIST(M),USPS(U),andSVHN(S)datasetshaverecentlybeenusedfordomainadap-tation[12,45,59].Theycontainimagesofdigitsfrom0to9invariousdifferentenvironmentsincludinginth

7 ewildinthecaseofSVHN[40].Weconsideredsix
ewildinthecaseofSVHN[40].Weconsideredsixcross-domaintasks.ThersttwotasksincludeM!U,U!M,andfollowedtheexperimentalsettingin[12,45,33,59,49],whichinvolvesrandomlyselecting2000imagesfromMNISTand1800imagesfromUSPS.Fortherestof6 Table2:Ofcedataset.Classicationaccuracyfordomainadaptationoverthe31categoriesoftheOfcedataset.A,W,andDstandforAmazon,Webcam,andDSLRdomain.LBisourbasemodelwithoutadaptation. UnsupervisedMethodsSupervisedMethodsLB[60][34][15][58][27][38]FADA A!W61.20.961.80.468.50.468.70.382.70.884.51.788.21.088.11.2A!D62.30.864.40.367.00.467.10.386.11.286.30.889.01.288.21.0W!A51.60.952.20.453.10.354.090.565.00.565.71.772.11.071.10.9W!D95.60.798.50.499.00.299.00.297.60.297.50.797.60.497.50.6D!A58.50.852.10.854.00.456.00.566.20.366.51.071.80.568.106D!W80.10.695.00.596.00.396.40.395.70.595.50.696.40.896.40.8 Average68.270.672.973.682.282.685.884.9 thecross-domaintasks,M!S,S!M,U!S,andS!U,weusedalltrainingsamplesofthesourcedomainfortrainingandalltestingsamplesofthetargetdomainfortesting.Since[12,45,33,59,49]introducedunsupervisedmethods,theyusedallsamplesofatargetdomainasunlabeleddataintraining.Hereinstead,werandomlyselectednlabeledsamplesperclassfromtargetdomaindataandusedthemintraining.Weevaluatedourapproachfornrangingfrom1to4andrepeatedeachexperiment10times(weonlyshowthemeanoftheaccuraciesforthisexperimentbecausestandarddeviationisverysmall).SincetheimagesoftheUSPSdatasethave1616pixels,weresizedtheimagesoftheMNISTandSVHNdatasetsto1616pixels.Weassumegsandgtshareweights(g=gs=gt)forthisexperiment.Similarto[32],weused2convolutionallayerswith6and16ltersof55kernelsfollowedbymax-poolinglayersand2fullyconnectedlayerswithsize120and84astheinferencefunctiong,andonefullyconnectedlayerwithsoftmaxactivationasthepredictionfunctionh.Also,weused2fullyconnectedlayerswithsize64and4asDCD(4groupsclassier).TrainingforeachstagewasdoneusingtheAdamOptimizer[26].Wecompareourmethodwith1SDAmethod,underthesamecondition,and6recentUDAmethods.UDAmethodsusealltargetsamplesintheirtrainingstage,whileweonlyuseveryfewlabeledtargetsamplespercategoryintraining.Table1showstheclassicationaccuraciesacrossarangeforthenumberoftargetsamplesavailableintraining(n=1;:::;7).FADAworkswellevenwhenonlyonetargetsamplepercategory(n=1)isavailableintraining.Wecangetcomparableaccuracieswiththestate-of-the-artusingonly10labeledtargetsamples(onesampleperclassn=1)insteadofusingmorethanthousandsofunlabeledtargetsamples.Wealsoreportthelowerbound(LB)ofourmodelwhichcorrespondstotrainingthebasemodelusingonlysourcesamples.Moreover,wereporttheaccuraciesobtainedbyne-tuning(FT)thebasemodelonavailabletargetdataandalsotherecentworkpresentedin[38].AlthoughTable1showsthatFTincreasestheaccuraciesoverLB,ithasreducedperformancecomparedtoSDAmethods.Figure3showshowmuchimprovementcanbeobtainedwithrespecttothebasemodel.ThebasemodelisthelowerboundLB.Thisissimplyobtainedbytraininggandhwithonlytheclassicationlossandsourcetrainingdata;so,noadaptationisperformed.Weight-Sharing.Aswediscussedearlier,weight-sharingcanbeseenasaregularizerthatpreventsthetargetnetworkgtfromovertting.Thisisimportantbecausegtcanbeeasilyoverttedsincetargetdataissc

8 arce.WerepeatedtheexperimentfortheU!Mwit
arce.WerepeatedtheexperimentfortheU!Mwithn=5withoutsharingweights.Thisprovidesanaverageaccuracyof84:1over10repetitions,whichislessthantheweight-sharingcase.4.2OfceDatasetTheofcedatasetisastandardbenchmarkdatasetforvisualdomainadaptation.Itcontains31objectclassesforthreedomains:Amazon,Webcam,andDSLR,indicatedasA,W,andD,foratotalof4,652images.TherstdomainA,consistsofimagesdownloadedfromonlinemerchants,thesecondW,consistsoflowresolutionimagesacquiredbywebcams,thethirdD,consistsofhighresolutionimagescollectedwithdigitalSLRs.Weconsiderfourdomainshiftsusingthethreedomains(A!W,A!D,W!A,andD!A).SincethereisnotaconsiderabledomainshiftbetweenWandD,weexcludeW!DandD!W.7 Figure3:MNIST-USPS-SVHNsummary.ThelowerbarofeachcolumnrepresentstheLBasreportedinTable1forthecorrespondingdomainpair.Themiddlebaristheimprovementofne-tuningFTthebasemodelusingtheavailabletargetdatareportedinTable1.ThetopbaristheimprovementofFADAoverFT,alsoreportedinTable1.Wefollowedthesettingdescribedin[58].Allclassesoftheofcedatasetand5train-testsplitsareconsidered.Forthesourcedomain,20examplespercategoryfortheAmazondomain,and8examplespercategoryfortheDSLRandWebcamdomainsarerandomlyselectedfortrainingforeachsplit.Also,3labeledexamplesarerandomlyselectedforeachcategoryinthetargetdomainfortrainingforeachsplit.Therestofthetargetsamplesareusedfortesting.Notethatweusedthesamesplitsgeneratedby[58].InadditiontotheSDAalgorithms,wereporttheresultsofsomerecentUDAalgorithms.TheyfollowadifferentexperimentalprotocolcomparedtotheSDAalgorithms,anduseallsamplesofthetargetdomainintrainingasunlabeleddatatogetherwithallsamplesofthesourcedomain.So,wecannotmakeanexactcomparisonbetweenresults.However,sinceUDAalgorithmsuseallsamplesofthetargetdomainintrainingandweuseonlyveryfewofthem(3perclass),wethinkitisstillworthlookingathowtheydiffer.Hereweareinterestedinthecasewheregsandgtshareweights(gs=gt=g).Fortheinferencefunctiong,weusedtheconvolutionallayersoftheVGG-16architecture[53]followedby2fullyconnectedlayerswithoutputsizeof1024and128,respectively.Forthepredictionfunctionh,weusedafullyconnectedlayerwithsoftmaxactivation.Similarto[58],weusedtheweightspre-trainedontheImageNetdataset[46]fortheconvolutionallayers,andinitializedthefullyconnectedlayersusingallthesourcedomaindata.WemodeltheDCDwith2fullyconnectedlayerswithasoftmaxactivationinthelastlayer.Table2reportstheclassicationaccuracyover31classesfortheOfcedatasetandshowsthatFADAhasperformancecomparabletothestate-of-the-art.5ConclusionsWehaveintroducedadeepmodelcombiningaclassicationandanadversariallosstoaddressSDAinfew-shotlearningregime.WehaveshownthatadversariallearningcanbeaugmentedtoaddressSDA.Theapproachisgeneralinthesensethatthearchitecturesub-componentscanbechanged.Wefoundthataddressingthesemanticdistributionalignmentswithpoint-wisesurrogatesofdistributiondistancesandsimilaritiesforSDAworksveryeffectively,evenwhenlabeledtargetsamplesareveryfew.Inaddition,wefoundtheSDAaccuracytoconvergeveryquicklyasmorelabeledtargetsamplespercategoryareavailable.Theapproachshowsclearpromiseasitsetsnewstate-of-the-artperformanceintheexperiments.8 (d) U!M (b) M!S (e) S!U (c) U!S (a) M!U (f) S!M References[1]Y.AytarandA.Zisserman.Tabularasa:Modeltransferforobjectcategorydetection.InComputerVision(ICCV),2011IEEEInternationalConferenceon,pages2252–2259.IEEE,2011.[2]M.Baktashmotlagh,M.T.Harandi,B.C.

9 Lovell,andM.Salzmann.Unsuperviseddomaina
Lovell,andM.Salzmann.Unsuperviseddomainadaptationbydomaininvariantprojection.InIEEEICCV,pages769–776,2013.[3]C.J.Becker,C.M.Christoudias,andP.Fua.Non-lineardomainadaptationwithboosting.InAdvancesinNeuralInformationProcessingSystems,pages485–493,2013.[4]A.BergamoandL.Torresani.Exploitingweakly-labeledwebimagestoimproveobjectclassication:adomainadaptationapproach.InAdvancesinNeuralInformationProcessingSystems,pages181–189,2010.[5]J.Blitzer,R.McDonald,andF.Pereira.Domainadaptationwithstructuralcorrespondencelearning.InProceedingsofthe2006conferenceonempiricalmethodsinnaturallanguageprocessing,pages120–128.AssociationforComputationalLinguistics,2006.[6]L.Chen,W.Li,andD.Xu.RecognizingRGBimagesbylearningfromRGB-Ddata.InCVPR,pages1418–1425,June2014.[7]S.Chopra,R.Hadsell,andY.LeCun.Learningasimilaritymetricdiscriminatively,withapplicationtofaceverication.InComputerVisionandPatternRecognition,2005.CVPR2005.IEEEComputerSocietyConferenceon,volume1,pages539–546.IEEE,2005.[8]H.DaumeIIIandD.Marcu.Domainadaptationforstatisticalclassiers.JournalofArticialIntelligenceResearch,26:101–126,2006.[9]J.Donahue,Y.Jia,O.Vinyals,J.Hoffman,N.Zhang,E.Tzeng,andT.Darrell.DeCAF:adeepconvolutionalactivationfeatureforgenericvisualrecognition.InarXiv:1310.1531,2013.[10]L.Duan,I.W.Tsang,D.Xu,andS.J.Maybank.Domaintransfersvmforvideoconceptdetection.InComputerVisionandPatternRecognition,2009.CVPR2009.IEEEConferenceon,pages1375–1381.IEEE,2009.[11]B.Fernando,A.Habrard,M.Sebban,andT.Tuytelaars.Unsupervisedvisualdomainadaptationusingsubspacealignment.InIEEEICCV,pages2960–2967,2013.[12]B.Fernando,T.Tommasi,andT.Tuytelaarsc.Jointcross-domainclassicationandsubspacelearningforunsupervisedadaptation.PatternRecogitionLetters,2015.[13]Y.GaninandV.Lempitsky.Unsuperviseddomainadaptationbybackpropagation.arXivpreprintarXiv:1409.7495,2014.[14]Y.Ganin,E.Ustinova,H.Ajakan,P.Germain,H.Larochelle,F.Laviolette,M.Marchand,andV.Lempitsky.Domain-adversarialtrainingofneuralnetworks.JournalofMachineLearningResearch,17(59):1–35,2016.[15]M.Ghifary,W.B.Kleijn,M.Zhang,D.Balduzzi,andW.Li.Deepreconstruction-classicationnetworksforunsuperviseddomainadaptation.InEuropeanConferenceonComputerVision,pages597–613.Springer,2016.[16]B.Gong,Y.Shi,F.Sha,andK.Grauman.Geodesicowkernelforunsuperviseddomainadaptation.InComputerVisionandPatternRecognition(CVPR),2012IEEEConferenceon,pages2066–2073.IEEE,2012.[17]I.Goodfellow.Nips2016tutorial:Generativeadversarialnetworks.arXivpreprintarXiv:1701.00160,2016.[18]I.Goodfellow,J.Pouget-Abadie,M.Mirza,B.Xu,D.Warde-Farley,S.Ozair,A.Courville,andY.Bengio.Generativeadversarialnets.InZ.Ghahramani,M.Welling,C.Cortes,N.D.Lawrence,andK.Q.Weinberger,editors,AdvancesinNeuralInformationProcessingSystems27,pages2672–2680.CurranAssociates,Inc.,2014.[19]R.Gopalan,R.Li,andR.Chellappa.Domainadaptationforobjectrecognition:Anunsupervisedapproach.InIEEEICCV,pages999–1006,2011.[20]A.Gretton,K.M.Borgwardt,M.Rasch,B.Schölkopf,andA.J.Smola.Akernelmethodforthetwo-sample-problem.InNIPS,2006.[21]Y.GuoandM.Xiao.Crosslanguagetextclassicationviasubspaceco-regularizedmulti-viewlearning.InProceedingsofthe29thInternationalConferenceonMachineLearning,ICML2012,Edinburgh,Scotland,UK,June26-July1,2012,2012.[22]E.HofferandN.Ailon.Deepmetriclearningusingtripletnetwork.InIn

10 ternationalWorkshoponSimilarity-BasedPat
ternationalWorkshoponSimilarity-BasedPatternRecognition,pages84–92.Springer,2015.[23]J.Hu,J.Lu,andY.-P.Tan.Discriminativedeepmetriclearningforfacevericationinthewild.InComputerVisionandPatternRecognition(CVPR),2014IEEEConferenceon,pages1875–1882,June2014.[24]J.J.Hull.Adatabaseforhandwrittentextrecognitionresearch.IEEETransactionsonpatternanalysisandmachineintelligence,16(5):550–554,1994.[25]P.Isola,J.-Y.Zhu,T.Zhou,andA.A.Efros.Image-to-imagetranslationwithconditionaladversarialnetworks.arXivpreprintarXiv:1611.07004,2016.9 [26]D.P.KingmaandJ.Ba.Adam:Amethodforstochasticoptimization.CoRR,abs/1412.6980,2014.[27]P.Koniusz,Y.Tas,andF.Porikli.Domainadaptationbymixtureofalignmentsofsecond-orhigher-orderscattertensors.arXivpreprintarXiv:1611.08195,2016.[28]B.Kulis,K.Saenko,andT.Darrell.Whatyousawisnotwhatyouget:Domainadaptationusingasymmetrickerneltransforms.InComputerVisionandPatternRecognition(CVPR),2011IEEEConferenceon,pages1785–1792.IEEE,2011.[29]S.KullbackandR.A.Leibler.Oninformationandsufciency.Theannalsofmathematicalstatistics,22(1):79–86,1951.[30]B.Kumar,G.Carneiro,I.Reid,etal.Learninglocalimagedescriptorswithdeepsiameseandtripletconvolutionalnetworksbyminimisinggloballossfunctions.InProceedingsoftheIEEEConferenceonComputerVisionandPatternRecognition,pages5385–5394,2016.[31]M.Lapin,M.Hein,andB.Schiele.Learningusingprivilegedinformation:SVM+andweightedSVM.NeuralNetworks,53:95–108,2014.[32]Y.LeCun,L.Bottou,Y.Bengio,andP.Haffner.Gradient-basedlearningappliedtodocumentrecognition.ProceedingsoftheIEEE,86(11):2278–2324,1998.[33]M.-Y.LiuandO.Tuzel.Coupledgenerativeadversarialnetworks.InAdvancesinNeuralInformationProcessingSystems,pages469–477,2016.[34]M.Long,Y.Cao,J.Wang,andM.I.Jordan.Learningtransferablefeatureswithdeepadaptationnetworks.InICML,pages97–105,2015.[35]M.Long,G.Ding,J.Wang,J.Sun,Y.Guo,andP.S.Yu.Transfersparsecodingforrobustimagerepresentation.InProceedingsoftheIEEEconferenceoncomputervisionandpatternrecognition,pages407–414,2013.[36]S.MotiianandG.Doretto.Informationbottleneckdomainadaptationwithprivilegedinformationforvisualrecognition.InEuropeanConferenceonComputerVision,pages630–647.Springer,2016.[37]S.Motiian,M.Piccirilli,D.A.Adjeroh,andG.Doretto.Informationbottlenecklearningusingprivilegedinformationforvisualrecognition.InProceedingsoftheIEEEConferenceonComputerVisionandPatternRecognition,pages1496–1505,2016.[38]S.Motiian,M.Piccirilli,D.A.Adjeroh,andG.Doretto.Unieddeepsuperviseddomainadaptationandgeneralization.InTheIEEEInternationalConferenceonComputerVision(ICCV),Oct2017.[39]K.Muandet,D.Balduzzi,andB.Schölkopf.Domaingeneralizationviainvariantfeaturerepresentation.InICML(1),pages10–18,2013.[40]Y.Netzer,T.Wang,A.Coates,A.Bissacco,B.Wu,andA.Y.Ng.Readingdigitsinnaturalimageswithunsupervisedfeaturelearning.InNIPSworkshopondeeplearningandunsupervisedfeaturelearning,2011.[41]S.J.Pan,I.W.Tsang,J.T.Kwok,andQ.Yang.Domainadaptationviatransfercomponentanalysis.IEEETNN,22(2):199–210,2011.[42]J.Ponce,T.L.Berg,M.Everingham,D.A.Forsyth,M.Hebert,S.Lazebnik,M.Marszalek,C.Schmid,B.C.Russell,A.Torralba,etal.Datasetissuesinobjectrecognition.InTowardcategory-levelobjectrecognition,pages29–48.Springer,2006.[43]S.Reed,Z.Akata,X.Yan,L.Logeswaran,B.Schiele,andH.Lee.Generativeadversarialtexttoimagesynthesis.InProceedings

11 ofthe33rdInternationalConferenceonIntern
ofthe33rdInternationalConferenceonInternationalConferenceonMachineLearning-Volume48,pages1060–1069.JMLR.org,2016.[44]S.E.Reed,Z.Akata,S.Mohan,S.Tenka,B.Schiele,andH.Lee.Learningwhatandwheretodraw.InAdvancesinNeuralInformationProcessingSystems,pages217–225,2016.[45]A.Rozantsev,M.Salzmann,andP.Fua.Beyondsharingweightsfordeepdomainadaptation.arXivpreprintarXiv:1603.06432,2016.[46]O.Russakovsky,J.Deng,H.Su,J.Krause,S.Satheesh,S.Ma,Z.Huang,A.Karpathy,A.Khosla,M.Bernstein,A.C.Berg,andL.Fei-Fei.ImageNetLargeScaleVisualRecognitionChallenge.IJCV,2015.[47]K.Saenko,B.Kulis,M.Fritz,andT.Darrell.Adaptingvisualcategorymodelstonewdomains.InECCV,pages213–226,2010.[48]T.Salimans,I.Goodfellow,W.Zaremba,V.Cheung,A.Radford,andX.Chen.Improvedtechniquesfortraininggans.InAdvancesinNeuralInformationProcessingSystems,pages2234–2242,2016.[49]S.Sankaranarayanan,Y.Balaji,C.D.Castillo,andR.Chellappa.Generatetoadapt:Aligningdomainsusinggenerativeadversarialnetworks.arXivpreprintarXiv:1704.01705,2017.[50]N.Saraanos,M.Vrigkas,andI.A.Kakadiaris.Adaptivesvm+:Learningwithprivilegedinformationfordomainadaptation.arXivpreprintarXiv:1708.09083,2017.[51]H.Shimodaira.Improvingpredictiveinferenceundercovariateshiftbyweightingthelog-likelihoodfunction.JournalofStatisticalPlanningandInference,90(2):227–244,2000.[52]A.Shrivastava,T.Pster,O.Tuzel,J.Susskind,W.Wang,andR.Webb.Learningfromsimulatedandunsupervisedimagesthroughadversarialtraining.InTheIEEEConferenceonComputerVisionandPatternRecognition(CVPR),July2017.10 [53]K.SimonyanandA.Zisserman.Verydeepconvolutionalnetworksforlarge-scaleimagerecognition.CoRR,abs/1409.1556,2014.[54]B.SunandK.Saenko.Deepcoral:Correlationalignmentfordeepdomainadaptation.InComputerVision–ECCV2016Workshops,pages443–450.Springer,2016.[55]T.Tommasi,M.Lanzi,P.Russo,andB.Caputo.Learningtherootsofvisualdomainshift.InComputerVision–ECCV2016Workshops,pages475–482.Springer,2016.[56]T.Tommasi,N.Patricia,B.Caputo,andT.Tuytelaars.Adeeperlookatdatasetbias.InGermanConferenceonPatternRecognition,pages504–516.Springer,2015.[57]A.TorralbaandA.A.Efros.Unbiasedlookatdatasetbias.InComputerVisionandPatternRecognition(CVPR),2011IEEEConferenceon,pages1521–1528,2011.[58]E.Tzeng,J.Hoffman,T.Darrell,andK.Saenko.Simultaneousdeeptransferacrossdomainsandtasks.InICCV,2015.[59]E.Tzeng,J.Hoffman,K.Saenko,andT.Darrell.Adversarialdiscriminativedomainadaptation.InTheIEEEConferenceonComputerVisionandPatternRecognition(CVPR),July2017.[60]E.Tzeng,J.Hoffman,N.Zhang,K.Saenko,andT.Darrell.Deepdomainconfusion:Maximizingfordomaininvariance.arXivpreprintarXiv:1412.3474,2014.[61]R.R.Varior,B.Shuai,J.Lu,D.Xu,andG.Wang.Asiameselongshort-termmemoryarchitectureforhumanre-identication.InEuropeanConferenceonComputerVision,pages135–153.Springer,2016.[62]J.Yang,R.Yan,andA.G.Hauptmann.Adaptingsvmclassierstodatawithshifteddistributions.InDataMiningWorkshops,2007.ICDMWorkshops2007.SeventhIEEEInternationalConferenceon,pages69–76.IEEE,2007.[63]T.Yao,Y.Pan,C.-W.Ngo,H.Li,andT.Mei.Semi-superviseddomainadaptationwithsubspacelearningforvisualrecognition.InTheIEEEConferenceonComputerVisionandPatternRecognition(CVPR),June2015.[64]H.Zhang,T.Xu,H.Li,S.Zhang,X.Huang,X.Wang,andD.Metaxas.Stackgan:Texttophoto-realisticimagesynthesiswithstackedgenerativeadversarialnetworks.arXivpreprintarXiv:1612.03242,20