已阅读5页,还剩31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
Ch7.SparseKernelMachinesPatternRecognitionandMachineLearning,C.M.Bishop,2006.,SummarizedbyS.KimBiointelligenceLaboratory,SeoulNationalUniversityhttp:/bi.snu.ac.kr/,2,(C)2007,SNUBiointelligenceLab,http:/bi.snu.ac.kr/,Contents,MaximumMarginClassifiersOverlappingClassDistributionsRelationtoLogisticRegressionMulticlassSVMsSVMsforRegressionRelevanceVectorMachinesRVMforRegressionAnalysisofSparsityRVMsforClassification,3,(C)2007,SNUBiointelligenceLab,http:/bi.snu.ac.kr/,MaximumMarginClassifiers,ProblemsettingsTwo-classclassificationusinglinearmodelsAssumethattrainingdatasetislinearlyseparableSupportvectormachineapproachesThedecisionboundaryischosentobetheoneforwhichthemarginismaximized,supportvectors,4,(C)2007,SNUBiointelligenceLab,http:/bi.snu.ac.kr/,MaximumMarginSolution,Foralldatapoints,ThedistanceofapointtothedecisionsurfaceThemaximummarginsolution,5,(C)2007,SNUBiointelligenceLab,http:/bi.snu.ac.kr/,DualRepresentation,IntroducingLagrangemultipliers,Min.pointssatisfythederivativesofLw.r.t.wandbequal0Dualrepresentation,FindAppendixEformoredetails,6,(C)2007,SNUBiointelligenceLab,http:/bi.snu.ac.kr/,ClassifyingNewData,OptimizationsubjectstoFoundbysolvingaquadraticprogrammingproblem,Karush-Kuhn-Tucker(KKT)ConditionsAppendixE,:supportvectors,or,O(N3),7,(C)2007,SNUBiointelligenceLab,http:/bi.snu.ac.kr/,ExampleofSeparableDataClassification,Figure7.2,8,(C)2007,SNUBiointelligenceLab,http:/bi.snu.ac.kr/,OverlappingClassDistributions,AllowsomemisclassifiedexamplessoftmarginIntroduceslackvariables,9,(C)2007,SNUBiointelligenceLab,http:/bi.snu.ac.kr/,SoftMarginSolution,MinimizeKKTconditions:,:trade-offbetweenminimizingtrainingerrorsandcontrollingmodelcomplexity,:supportvectors,or,10,(C)2007,SNUBiointelligenceLab,http:/bi.snu.ac.kr/,DualRepresentation,DualrepresentationClassifyingnewdataandobtainingb(hardmarginclassifiers),11,(C)2007,SNUBiointelligenceLab,http:/bi.snu.ac.kr/,AlternativeFormulation,v-SVM(Schlkopfetal.,2000),-Upperboundonthefractionofmarginerrors-Lowerboundonthefractionofsupportvectors,12,(C)2007,SNUBiointelligenceLab,http:/bi.snu.ac.kr/,ExampleofNonseparableDataClassification(v-SVM),13,(C)2007,SNUBiointelligenceLab,http:/bi.snu.ac.kr/,SolutionsoftheQPProblem,Chunking(Vapnik,1982)Idea:thevalueofLagrangianisunchangedifweremovetherowsandcolumnsofthekernelmatrixcorrespondingtoLagrangemultipliersthathavevaluezeroProtectedconjugategradients(Burges,1998)Decompositionmethods(Osunaetal.,1996)Sequentialminimaloptimization(Platt,1999),14,(C)2007,SNUBiointelligenceLab,http:/bi.snu.ac.kr/,RelationtoLogisticRegression(Section4.3.2),Fordatapointsonthecorrectside,Fortheremainingpoints,:hingeerrorfunction,15,(C)2007,SNUBiointelligenceLab,http:/bi.snu.ac.kr/,RelationtoLogisticRegression(Contd),FrommaximumlikelihoodlogisticregressionErrorfunctionwithaquadraticregularizer,16,(C)2007,SNUBiointelligenceLab,http:/bi.snu.ac.kr/,ComparisonofErrorFunctions,Hingeerrorfunction,Errorfunctionforlogisticregression,Misclassificationerror,Squarederror,17,(C)2007,SNUBiointelligenceLab,http:/bi.snu.ac.kr/,MulticlassSVMs,One-versus-the-rest:KseparateSVMsCanleadinconsistentresults(Figure4.2)ImbalancedtrainingsetsPositiveclass:+1,negativeclass:-1/(K-1)(Leeetal.,2001)AnobjectivefunctionfortrainingallSVMssimultaneously(WestonandWatkins,1999)One-versus-one:K(K-1)/2SVMsBasedonerror-correctingoutputcodes(Allweinetal.,2000)Generalizationofthevotingschemeoftheone-versus-one,18,(C)2007,SNUBiointelligenceLab,http:/bi.snu.ac.kr/,SVMsforRegression,Simplelinearregression:minimize-insensitiveerrorfunction,quadraticerrorfunction,-insensitiveerrorfunction,19,(C)2007,SNUBiointelligenceLab,http:/bi.snu.ac.kr/,SVMsforRegression(Contd),Minimize,20,(C)2007,SNUBiointelligenceLab,http:/bi.snu.ac.kr/,DualProblem,21,(C)2007,SNUBiointelligenceLab,http:/bi.snu.ac.kr/,Predictions,KKTconditions:,(fromderivativesoftheLagrangian),22,(C)2007,SNUBiointelligenceLab,http:/bi.snu.ac.kr/,AlternativeFormulation,v-SVM(Schlkopfetal.,2000),fractionofpointslyingoutsidethetube,23,(C)2007,SNUBiointelligenceLab,http:/bi.snu.ac.kr/,Exampleofv-SVMRegression,24,(C)2007,SNUBiointelligenceLab,http:/bi.snu.ac.kr/,RelevanceVectorMachines,SVMOutputsaredecisionsratherthanposteriorprobabilitiesTheextensiontoK2classesisproblematicThereisacomplexityparameterCKernelfunctionsarecenteredontrainingdatapointsandrequiredtobepositivedefiniteRVMBayesiansparsekerneltechniqueMuchsparsermodelsFasterperformanceontestdata,25,(C)2007,SNUBiointelligenceLab,http:/bi.snu.ac.kr/,RVMforRegression,RVMisalinearforminChapter3withamodifiedprior,26,(C)2007,SNUBiointelligenceLab,http:/bi.snu.ac.kr/,RVMforRegression(Contd),andaredeterminedusingevidenceapproximation(type-2maximumlikelihood)(Section3.5),Fromtheresult(3.49)forlinearregressionmodels,Maximize,27,(C)2007,SNUBiointelligenceLab,http:/bi.snu.ac.kr/,RVMforRegression(Contd),TwoapproachesByderivativesofmarginallikelihoodEMalgorithmSection9.3.4Predictivedistribution,Section3.3.2,28,(C)2007,SNUBiointelligenceLab,http:/bi.snu.ac.kr/,ExampleofRVMRegression,MorecompactthanSVMParametersaredeterminedautomaticallyRequiremoretrainingtimethanSVM,RVMregression,v-SVMregression,29,(C)2007,SNUBiointelligenceLab,http:/bi.snu.ac.kr/,MechanismforSparsity,onlyisotropicnoise,=,afinitevalueof,30,(C)2007,SNUBiointelligenceLab,http:/bi.snu.ac.kr/,SparseSolution,Pulloutthecontributionfromiin,Using(C.7),(C.15)inAppendixC,31,(C)2007,SNUBiointelligenceLab,http:/bi.snu.ac.kr/,SparseSolution(Contd),ForlogmarginallikelihoodfunctionL,Stationarypointsofthemarginallikelihoodw.r.t.i,Sparsity:measurestheextenttowhichoverlapswiththeotherbasisvectors,Qualityof:representsameasureofthealignmentofthebasisvectorwiththeerrorbetweentandy-i,32,(C)2007,SNUBiointelligenceLab,http:/bi.snu.ac.kr/,SequentialSparseBayesianLearningAlgorithm,InitializeInitializeusing,with,withtheremainingEvaluateandforallbasisfunctionsSelectacandidateIf(isalreadyinthemodel),updateIf,addtothemodel,andevaluateIf,removefromthemodel,andsetUpdateGoto3untilconverged,33,(C)2007,SNUBiointelligenceLab,http:/bi.snu.ac.kr/,RVMforClassification,Probabilisticlinearclassificationmodel(Chapter4)withARDprior,-Initialize-BuildaGaussianapproximationtotheposterio
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
- 中级在线营销推广方案策划与实践
- 融资计划及资金筹措方案
- 社会体育指导员地掷球职业生涯规划指导手册
- 针对年轻客群的互联网保险产品开发策划
- 餐饮餐具产业升级路径-洞察及研究
- 电子销售工程师销售合同签订与风险控制
- 建筑施工企业三类人员B证资料管理实务
- 融资租赁租机合同范本
- 绿色印刷技术在实践中的应用案例
- 设计团队的质量控制与反馈机制建设
- 高三第三次联考动员班会课件
- 华为股权分配协议书
- 学宪法讲宪法主题班会
- 采购工程师转正述职报告
- 近视的成因和预防
- DB12T 1118-2021 地面沉降监测分层标施工技术规程
- 齐鲁名家 谈方论药知到智慧树章节测试课后答案2024年秋山东中医药大学
- 【MOOC】油气地质与勘探-中国石油大学(华东) 中国大学慕课MOOC答案
- 《晶体的缺陷》课件
- 人教版八年级上册数学期中复习课件
- 2024年全国营养师技能大赛备赛试题库(含答案)
评论
0/150
提交评论