




已阅读5页,还剩26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
Semi-supervisedLearning,Introduction,Supervisedlearning:,=1E.g.:image,:classlabelsSemi-supervisedlearning:,=1,=+Asetofunlabeleddata,usuallyURTransductivelearning:unlabeleddataisthetestingdataInductivelearning:unlabeleddataisnotthetestingdataWhysemi-supervisedlearning?Collectingdataiseasy,butcollecting“labelled”dataisexpensiveWedosemi-supervisedlearninginourlives,Whysemi-supervisedlearninghelps?,Labelleddata,Unlabeleddata,cat,dog,(Imageofcatsanddogswithoutlabeling),Whysemi-supervisedlearninghelps?,Thedistributionoftheunlabeleddatatellussomething.,Usuallywithsomeassumptions,Whoknows?,Outline,Semi-supervisedLearningforGenerativeModel,SupervisedGenerativeModel,Givenlabelledtrainingexamples1,2lookingformostlikelypriorprobabilityP(Ci)andclass-dependentprobabilityP(x|Ci)P(x|Ci)isaGaussianparameterizedbyand,1,2,1|=|11|11+|22,With1,2,1,2,DecisionBoundary,Semi-supervisedGenerativeModel,Givenlabelledtrainingexamples1,2lookingformostlikelypriorprobabilityP(Ci)andclass-dependentprobabilityP(x|Ci)P(x|Ci)isaGaussianparameterizedbyand,DecisionBoundary,Theunlabeleddatahelpre-estimate1,2,1,2,1,2,Semi-supervisedGenerativeModel,Initialization:=1,2,1,2,Step1:computetheposteriorprobabilityofunlabeleddataStep2:updatemodel,1|,1=1+1|,1=111+11|1|,Backtostep1,Dependingonmodel,:totalnumberofexamples1:numberofexamplesbelongingtoC1,Thealgorithmconvergeseventually,buttheinitializationinfluencestheresults.,E,M,Why?,MaximumlikelihoodwithlabelleddataMaximumlikelihoodwithlabelled+unlabeleddata,=,=,+,=|11+|22,(cancomefromeitherC1andC2),Closed-formsolution,Solvediteratively,=1,2,1,2,Semi-supervisedLearningLow-densitySeparation,非黑即白,“Black-or-white”,Self-training,Given:labelleddataset=,=1,unlabeleddataset=+Repeat:TrainmodelfromlabelleddatasetApplytotheunlabeleddatasetObtain,=+Removeasetofdatafromunlabeleddataset,andaddthemintothelabeleddataset,Independenttothemodel,Howtochoosethedatasetremainsopen,Regression?,Pseudo-label,Youcanalsoprovideaweighttoeachdata.,Self-training,Similartosemi-supervisedlearningforgenerativemodelHardlabelv.s.Softlabel,Consideringusingneuralnetwork,Newtargetforis10,Class1,70%Class130%Class1,(networkparameter)fromlabelleddata,Newtargetforis0.70.3,Doesntwork,Itlookslikeclass1,thenitisclass1.,Hard,Soft,Entropy-basedRegularization,Distribution,Good!,Good!,Bad!,=15,Entropyof:Evaluatehowconcentratethedistributionis,=0,=0,=5,=15,Assmallaspossible,=,+,labelleddata,unlabeleddata,Outlook:Semi-supervisedSVM,Findaboundarythatcanprovidethelargestmarginandleasterror,Enumerateallpossiblelabelsfortheunlabeleddata,Semi-supervisedLearningSmoothnessAssumption,近朱者赤,近墨者黑,“Youareknownbythecompanyyoukeep”,SmoothnessAssumption,Assumption:“similar”hasthesameMoreprecisely:xisnotuniform.If1and2arecloseinahighdensityregion,1and2arethesame.,connectedbyahighdensitypath,Sourceofimage:/files/pinwheel.png,1,2,3,1and2havethesamelabel,2and3havedifferentlabels,SmoothnessAssumption,“indirectly”similarwithsteppingstones,(TheexampleisfromthetutorialslidesofXiaojinZhu.),similar?,Notsimilar?,Sourceofimage:,SmoothnessAssumption,Carticles,(TheexampleisfromthetutorialslidesofXiaojinZhu.),SmoothnessAssumption,Carticles,(TheexampleisfromthetutorialslidesofXiaojinZhu.),ClusterandthenLabel,Cluster1,Cluster2,Cluster3,Class1,Class2,Class2,Usingallthedatatolearnaclassifierasusual,Graph-basedApproach,Howtoknow1and2arecloseinahighdensityregion(connectedbyahighdensitypath),Representedthedatapointsasagraph,E.g.Hyperlinkofwebpages,citationofpapers,Graphrepresentationisnaturesometimes.,Sometimesyouhavetoconstructthegraphyourself.,Graph-basedApproach-GraphConstruction,Definethesimilarity,betweenandAddedge:KNearestNeighbore-NeighborhoodEdgeweightisproportionaltos,=2,GaussianRadialBasisFunction:,TheimageisfromthetutorialslidesofAmarnagSubramanyaandParthaPratimTalukdar,Graph-basedApproach,Class1,Class1,Propagatethroughthegraph,Thelabelleddatainfluencetheirneighbors.,x,Graph-basedApproach,Definethesmoothnessofthelabelsonthegraph,=12,2,Smallermeanssmoother,x1,x2,x3,x4,2,3,1,1,1=0,2=1,3=1,4=0,=0.5,=3,Foralldata(nomatterlabelledornot),Graph-basedApproach,Definethesmoothnessofthelabelsonthegraph,=,L:(R+U)x(R+U)matrix,GraphLaplacian,=,y:(R+U)-dimvector,=,=0220301031000110,D=5003000000005001,=12,2,Graph-basedApproach,Definethesmoothnessofthelabelsonthegraph,=,=12,2,Dependingonnetworkparameters,=,+,J.Weston,F.Ratle,andR.Collobert,“Deeplearningviasemi-supervisedembedding,”ICML,2008,Asaregularizationterm,smooth,smooth,smooth,Semi-supervisedLearningBetterRepresentation,去蕪存菁,化繁為簡,LookingforBetterRepresentation,Findthelate
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版高级女方离婚协议书撰写规范与样本释读
- 2025版苏州工业园区住宅租赁合同管理规范
- 2025版水泥行业人才培训合同样本
- 2025年燃料油运输安全责任保险合同范本
- 2025大理石大板石材工程安装、施工、监理与验收合同
- 2025年冰箱组件采购与集成服务合同模板
- 海南省文昌市2025年上半年公开招聘村务工作者试题含答案分析
- 2025年度企业社会责任报告编辑服务委托合同范本
- 2025年土方运输车租赁与新能源项目运输合同
- 2025版事业单位劳动违约赔偿与劳动合同续签赔偿协议
- 学生营养餐(中央厨房)集中配送项目计划书
- (新)精神卫生知识技能竞赛理论考试题库(含答案)
- 液碱卸车安全操作规程
- 建筑用砂石料采购 投标方案(技术方案)
- 中华护理学会成人肠内营养支持护理团标解读
- 医疗器械质量安全风险会商管理制度
- DLT 5175-2021 火力发电厂热工开关量和模拟量控制系统设计规程-PDF解密
- 电工仪表与测量(第六版)中职技工电工类专业全套教学课件
- 确保工期的资源保障措施
- 天津市二手房买卖通用版合同合集3篇
- 苏州团餐行业分析
评论
0/150
提交评论