资源描述:
《DRAW_ A Recurrent Neural Network For Image Generation》由会员上传分享,免费在线阅读,更多相关内容在学术论文-天天文库。
1、DRAW:ARecurrentNeuralNetworkForImageGenerationKarolGregorKAROLG@GOOGLE.COMIvoDanihelkaDANIHELKA@GOOGLE.COMAlexGravesGRAVESA@GOOGLE.COMDaniloJimenezRezendeDANILOR@GOOGLE.COMDaanWierstraWIERSTRA@GOOGLE.COMGoogleDeepMindAbstractThispaperintroducestheDeepRecur
2、rentAtten-tiveWriter(DRAW)neuralnetworkarchitectureforimagegeneration.DRAWnetworkscombineanovelspatialattentionmechanismthatmimicsthefoveationofthehumaneye,withasequentialvariationalauto-encodingframeworkthatallowsfortheiterativeconstructionofcompleximages
3、.ThesystemsubstantiallyimprovesonthestateoftheartforgenerativemodelsonMNIST,and,whentrainedontheStreetViewHouseNumbersdataset,itgeneratesimagesthatcannotbedistin-guishedfromrealdatawiththenakedeye.1.IntroductionTimeApersonaskedtodraw,paintorotherwiserecrea
4、teavisualscenewillnaturallydosoinasequential,iterativefashion,Figure1.AtrainedDRAWnetworkgeneratingMNISTdig-reassessingtheirhandiworkaftereachmodification.Roughits.Eachrowshowssuccessivestagesinthegenerationofasin-outlinesaregraduallyreplacedbypreciseforms,
5、linesaregledigit.Notehowthelinescomposingthedigitsappeartobe“drawn”bythenetwork.Theredrectangledelimitstheareaat-sharpened,darkenedorerased,shapesarealtered,andthetendedtobythenetworkateachtime-step,withthefocalpreci-finalpictureemerges.Mostapproachestoauto
6、maticim-sionindicatedbythewidthoftherectangleborder.agegeneration,however,aimtogenerateentirescenesatonce.Inthecontextofgenerativeneuralnetworks,thistyp-arXiv:1502.04623v2[cs.CV]20May2015icallymeansthatallthepixelsareconditionedonasinglelatentdistribution(
7、Dayanetal.,1995;Hinton&Salakhut-ThecoreoftheDRAWarchitectureisapairofrecurrentdinov,2006;Larochelle&Murray,2011).Aswellaspre-neuralnetworks:anencodernetworkthatcompressesthecludingthepossibilityofiterativeself-correction,the“onerealimagespresentedduringtra
8、ining,andadecoderthatshot”approachisfundamentallydifficulttoscaletolargereconstitutesimagesafterreceivingcodes.Thecombinedimages.TheDeepRecurrentAttentiveWriter(DRAW)ar-systemistrainedend-to-endwithstochasticg