PAC-Bayes理论及应用研究综述.pdfVIP

ISSN 1673．9418 CODEN JKYTA8 JournalofFrontiersofComputerScienceandTechnology E-mail：fcst@vip．163．tom 1673--9418／2015／09(01)-·0001--13 httrl：／／www．ceaj．org doi：10．3778~．issn．1673—9418．1410046 Tel：+86．10 PAC—Bayes理论及应用研究综述 汤 莉 ’’，宫秀军 ，何 丽 1．天津财经大学 理工学院 信息科学与技术系，天津 300222 2．天津大学 计算机科学与技术学院，天津 300072 3．天津市认知计算与应用重点实验室，天津 300072 SurveyonPAC-BayesTheoryandApplicationResearch TANGLi-．。GONGXiuiun，针．HELi 1．DepartmentofInformationScienceandTechnology,SchoolofScienceandTechnoloyg,TianjinUniversityofFinance andEconomics，Tianjin300222，China 2．SchoolofComputerScienceandTechnology,TianjinUniversiyt,Tianjin300072，China 3．TianjinKeyLaboratoryofCognitiveComputingandApplication，Tianjin300072，China +Correspondingauthor：E—mail：gong~@tju．edu．cn TANGLi，GONGXiujan，HELi．SurveyonPAC-Bayestheoryandapplicationresearch．JournalofFrontiers ofComputerScienceandTechnology,2015，9(1)：1-13． Abstract：PAC—Bayestheory integratingtheoriesofBayesianparadigm nadstructureriskminimizationforstochastic classifiershasbeenconsideredasaframeworkofrderivingsomeofthetightestgeneralizationbounds．Thispaper analyzestheresearchbackgroundandprofoundsignificanceofPAC—Bayestheory,andintroducestherfameworkof PAC—Bayeshteoryanditsapplicationtosupportvectormachine(SVM)．Then，thispaperdiscussesPAC—Bayesbound ofmnaymachinelearningalgorithms，nadspeciallynaalyzeshteboun d、)lrim thenon—IID data．Furthermore，thispaper elaboratesresearchstatusanddevelopmentofthePAC—Bayesboundapplicationrfom fourdirections，andcompares differentresearchmethodsandfeatures．Finally,thispaperdrawstheresearchprospectofthePAC—Bayesboun d． Keywords：PAC—Bayesbound；supF3rtvectormachine(SVM)；generalizationcapability；classifier 摘 要：PAC—Bayes理论融合 了贝叶斯定理和随机分类器的结构风险最小化原理，它作为一个理论框架，可得 到最紧的泛化风险边界。分析 了PAC—Bayes理论的研究背景和重要意义，介绍了PAC—Bayes理论框架及其在 TheNationalNaturalScienceFoundationofCh