多模态相似性学习课件.ppt

《多模态相似性学习课件.ppt》由会员上传分享，免费在线阅读，更多相关内容在教育资源-天天文库。

1、LearningMulti-modalSimilarity1AbstractBodyIntroductionAGraphicalViewofSimilarityPartialOrderEmbeddingMultipleKernelEmbeddingExperimentsHardnessofDimensionalityReductionConclusionAcknowledgmentsAppendixAppendixA.EmbeddingPartialOrdersAppendixB.SolverAppendixC.RelationshiptoAUCReferences2

2、Aim:integratheterogeneousdataintoasingle,unifiedsimilarityspaceSolution:multiplekernellearningtechniqueKeywords:multiplekernellearning,metriclearning,similarityAbstract31IntroductionNoobviouschoiceofsimilaritymeasureinmorecomplex,multi-mediadomainsUseside-informationtooptimizeasimilarit

3、yfunctionMetriclearningmethods度量学习Multidimensionalscaling(MDS)techniques多维标度分析Learnasuitablemetricandthensimilaritytonew,unseendatacanbecomputedeitherdirectlyorviaout-of-sampleextensionsUsingsimilaritymeasurementsasside-informationtoguidetheconstructionofasimilarityspaceformulti-modalda

4、ta41IntroductionCollectsimilarityinformationintheformoftriadicorrelativecomparisonsReasons:SimilaritymaybeahighlysubjectiveconceptandvaryfromonelabelertothenextAsinglelabelermaynotbeabletoconsistentlydecideifortowhatextenttwoobjectsaresimilarButasinglelabelerisabletoreliablyproducearank

5、-orderingofsimilarityoverpairsExamples:Answerquestionsoftheform:“Isxmoresimilartoyorz?”51IntroductionDevelopaframeworkforintegratingmulti-modaldataThreeguidingprinciples:Robustagainstsubjectivityandinter-labelerdisagreement.Beabletointegratemulti-modaldatainanoptimalwayBepossibletocompu

6、tedistancestonew,unseendataFormulateasalearningproblem:GivenadatasetandacollectionofrelativecomparisonsbetweenpairsLearnarepresentationofthedatathatoptimallyreproducesthesimilaritymeasurements61.1ContributionsTwo-fold:developthepartialorderembedding(POE)frameworkformulateanovelmultiplek

7、ernellearning(MKL)algorithmFigure1:Anoverviewofproposedframeworkformulti-modalfeatureintegration71.2PreliminariesPartialordersatisfies:Irreflexivity,Transitivity,Anti-symmetryForadirectedgraphG,G∞denoteitstransitiveclosure;GmindenotetransitivereductionX={x1,x2,...,xn}denotethet