已阅读5页,还剩45页未读, 继续免费阅读
版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领
文档简介
UnsupervisedLearning:Generation,Creation,GenerativeModels:,WhatIcannotcreate,Idonotunderstand.,RichardFeynman,CreationImageProcessing,Now,Inthefuture,v.s.,Machinedrawsacat,GenerativeModels,PixelRNN,Tocreateanimage,generatingapixeleachtime,Ref:AaronvandenOord,NalKalchbrenner,KorayKavukcuoglu,PixelRecurrentNeuralNetworks,arXivpreprint,2016,E.g.3x3images,NN,NN,NN,Canbetrainedjustwithalargecollectionofimageswithoutanyannotation,PixelRNN,RealWorld,Ref:AaronvandenOord,NalKalchbrenner,KorayKavukcuoglu,PixelRecurrentNeuralNetworks,arXivpreprint,2016,Morethanimages,Audio:AaronvandenOord,SanderDieleman,HeigaZen,KarenSimonyan,OriolVinyals,AlexGraves,NalKalchbrenner,AndrewSenior,KorayKavukcuoglu,WaveNet:AGenerativeModelforRawAudio,arXivpreprint,2016,Video:NalKalchbrenner,AaronvandenOord,KarenSimonyan,IvoDanihelka,OriolVinyals,AlexGraves,KorayKavukcuoglu,VideoPixelNetworks,arXivpreprint,2016,PracticingGenerationModels:PokmonCreation,Smallimagesof792PokmonsCanmachinelearntocreatenewPokmons?Sourceofimage:,Dontcatchthem!Createthem!,Originalimageis40 x40,Makingtheminto20 x20,PracticingGenerationModels:PokmonCreation,Tips(?),R=50,G=150,B=100,Eachpixelisrepresentedby3numbers(correspondingtoRGB),Eachpixelisrepresentedbya1-of-Nencodingfeature,Clusteringthesimilarcolor,167colorsintotal,PracticingGenerationModels:PokmonCreation,Originalimage(40 x40):.tw/tlkagk/courses/ML_2016/Pokemon_creation/image.rarPixels(20 x20):.tw/tlkagk/courses/ML_2016/Pokemon_creation/pixel_color.txtEachlinecorrespondstoanimage,andeachnumbercorrespondstoapixel.tw/tlkagk/courses/ML_2016/Pokemon_creation/colormap.txtFollowingexperiment:1-layerLSTM,512cells,0,1,2,RealPokmon,Cover50%,Cover75%,Itisdifficulttoevaluategeneration.,Neverseenbymachine!,PokmonCreation,Drawingfromscratch,Needsomerandomness,GenerativeModels,DiederikPKingma,MaxWelling,Auto-EncodingVariationalBayes,arXivpreprint,2013,Auto-encoder,Ascloseaspossible,NNEncoder,NNDecoder,code,NNDecoder,code,Randomlygenerateavectorascode,Image?,NNEncoder,NNDecoder,code,input,output,Auto-encoder,VAE,NNEncoder,input,NNDecoder,output,2,3,Fromanormaldistribution,3,1,2,X,+,Minimizereconstructionerror,=131+2,exp,=+,Minimize,Cifar-10,Sourceofimage:/pdf/1606.04934v1.pdf,NNEncoder,input,NNDecoder,output,2,3,3,1,2,X,+,exp,PokmonCreation,Training,NNDecoder,3,1,2,10-dim,10-dim,Picktwodim,andfixtheresteight,?,Ref:http:/www.wired.co.uk/article/google-artificial-intelligence-poetrySamuelR.Bowman,LukeVilnis,OriolVinyals,AndrewM.Dai,RafalJozefowicz,SamyBengio,GeneratingSentencesfromaContinuousSpace,arXivprepring,2015,WritingPoetry,NNEncoder,sentence,NNDecoder,sentence,code,CodeSpace,iwenttothestoretobuysomegroceries.,comewithme,shesaid.,istoretobuysomegroceries.,iweretobuyanygroceries.,talktome,shesaid.,dontworryaboutit,shesaid.,WhyVAE?,?,encode,decode,code,IntuitiveReason,noise,noise,WhyVAE?,IntuitiveReason,NNEncoder,input,NNDecoder,output,2,3,3,1,2,X,+,exp,OriginalCode,Codewithnoise,Thevarianceofnoiseisautomaticallylearned,Whatwillhappenifweonlyminimizereconstructionerror?,=131+2,Minimize,WhyVAE?,IntuitiveReason,NNEncoder,input,NNDecoder,output,2,3,3,1,2,X,+,exp,OriginalCode,Codewithnoisy,Thevarianceofnoiseisautomaticallylearned,Whatwillhappenifweonlyminimizereconstructionerror?,=131+2,Minimize,Wewantcloseto0(variancecloseto1),L2regularization,WhyVAE?,Backtowhatwewanttodo,Estimatetheprobabilitydistribution,EachPokmonisapointxinthespace,P(x),P(x),=|,P(m),1,2,3,4,5,.,GaussianMixtureModel,Howtosample?,|,Eachxyougenerateisfromamixture,Distributedrepresentationisbetterthancluster.,(multinomial),misaninteger,P(x),z,VAE,0,|,zisavectorfromnormaldistribution,Eachdimensionofzrepresentsanattribute,InfiniteGaussian,Eventhoughzisfrom0,P(x)canbeverycomplex,=|,P(z)isnormaldistribution,MaximizingLikelihood,isgoingtobeestimated,TuningtheparameterstomaximizelikelihoodL,Weneedanotherdistributionq(z|x),Decoder,Encoder,NN,=|,|,|,=,Maximizingthelikelihoodoftheobservedx,=|,q(z|x)canbeanydistribution,=|,|,=|,|,=|,|+|,|,|,MaximizingLikelihood,0,=,P(z)isnormaldistribution,Maximizingthelikelihoodoftheobservedx,isgoingtobeestimated,=|,|,MaximizingLikelihood,=+|,=|,Maximizeby|,Find|and|maximizingLb,|willbeanapproximationof|intheend,=|,|,=|,=|,+|,|,MaximizingLikelihood,NN,P(z)isnormaldistribution,Maximizingthelikelihoodoftheobservedx,isgoingtobeestimated,=|,|,=,|,ConnectionwithNetwork,|,Minimizing,=131+2,Minimize,|,(RefertotheAppendixBoftheoriginalVAEpaper),=|,Maximizing,NN,NN,NN,close,Thisistheauto-encoder,ConditionalVAE,/pdf/1406.5298v2.pdf,Tolearnmore,CarlDoersch,TutorialonVariationalAutoencodersDiederikP.Kingma,DaniloJ.Rezende,ShakirMohamed,MaxWelling,“Semi-supervisedlearningwithdeepgenerativemodels.”NIPS,2014.Sohn,Kihyuk,HonglakLee,andXinchenYan,“LearningStructuredOutputRepresentationusingDeepConditionalGenerativeModels.”NIPS,2015.XinchenYan,JimeiYang,KihyukSohn,HonglakLee,“Attribute2Image:ConditionalImageGenerationfromVisualAttributes”,ECCV,2016Cooldemo:http:/vdumoulin.github.io/morphing_faces/http:/fvae.ail.tokyo/,ProblemsofVAE,Itdoesnotreallytrytosimulaterealimages,NNDecoder,code,Output,Onepixeldifferencefromthetarget,Onepixeldifferencefromthetarget,Realistic,Fake,VAEmayjustmemorizetheexistingimages,insteadofgeneratingnewimages,GenerativeModels,IanJ.Goodfellow,JeanPouget-Abadie,MehdiMirza,BingXu,DavidWarde-Farley,SherjilOzair,AaronCourville,YoshuaBengio,GenerativeAdversarialNetworks,arXivpreprint2014,YannLeCunscomment,YannLeCunscomment,擬態的演化,棕色,葉脈,蝴蝶不是棕色,蝴蝶沒有葉脈,.,Theevolutionofgeneration,NNGeneratorv1,Discri-minatorv1,Realimages:,NNGeneratorv2,Discri-minatorv2,NNGeneratorv3,Discri-minatorv3,GAN-Discriminator,NNGeneratorv1,Realimages:,Discri-minatorv1,image,1/0,(realorfake),DecoderinVAE,Vectorsfromadistribution,1,1,1,1,0,0,0,0,GAN-Generator,Discri-minatorv1,NNGeneratorv1,Randomlysampleavector,0.87,“Tuning”theparametersofgenerator,Theoutputbeclassifiedas“real”(ascloseto1aspossible),Generator+Discriminator=anetwork,Usinggradientdescenttofindtheparametersofgenerator,Fixthediscriminator,1.0,GANToyExample,Demo:/people/karpathy/gan/,Discri-minator,NNGenerator,x,z,1/0,Greendistribution,Bluecurve,Realdata(blackpoints),Cifar-10,Whichoneismachine-generated?,Ref:,Movingonthecodespace,AlecRadford,LukeMetz,SoumithChintala,UnsupervisedRepresentationLearningwithDeepConvolutionalGenerativeAdversarialNetworks,ICLR,2016,畫漫畫,Ref:,畫漫畫,Ref:,Webdemo:http:/mattya.github.io/chainer-DCGAN/,Inpractical,GANsaredifficulttooptimize
温馨提示
- 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
- 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
- 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
- 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
- 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
- 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
- 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。
最新文档
- 2025年超星尔雅学习通《员工团队合作与领导力发展培训方案》考试备考题库及答案解析
- 2025年超星尔雅学习通《认知科学与学习心理学》考试备考题库及答案解析
- 2025年超星尔雅学习通《网络传播学》考试备考题库及答案解析
- 2025年超星尔雅学习通《地球历史:《冰川时代探秘》》考试备考题库及答案解析
- 2025年超星尔雅学习通《工业 0 技术应用》考试备考题库及答案解析
- 2025年超星尔雅学习通《哲学思辨与智慧生活》考试备考题库及答案解析
- 2025年超星尔雅学习通《文化艺术品市场管理》考试备考题库及答案解析
- 2025年教学管理员人员岗位招聘面试参考试题及参考答案
- 2025年超星尔雅学习通《中国传统文化名著名篇名著名家作品解读》考试备考题库及答案解析
- 2025年超星尔雅学习通《信息系统运维与安全》考试备考题库及答案解析
- 成人脓毒症相关心肌损伤和(或)心功能障碍急诊专家共识2025
- 2025杭州西湖云创产业服务有限公司公开招聘工作人员6人考试笔试参考题库附答案解析
- 2025年下半年河北张家口市张北县硕博人才引进99人重点基础提升(共500题)附带答案详解
- 2025入团积极分子结业考试题库(含答案)
- 国网安全考试试卷题库及答案解析
- 空调卫生知识培训记录课件
- DB64 819-2024 煤质活性炭工业大气污染物排放标准
- 2025GCP(药物临床试验质量管理规范)相关知识考试试卷及答案
- 2025年部编新版道德与法治三年级上册期末复习计划
- 山东省日照市2025年中考英语真题附同步解析答案
- 生成式人工智能应用实战课件 第6章 AIGC视频创作
评论
0/150
提交评论