training conditional random fields using virtual evidence boosting

《training conditional random fields using virtual evidence boosting》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、TrainingConditionalRandomFieldsusingVirtualEvidenceBoostingLinLiaoTanzeemChoudhury†DieterFoxHenryKautzUniversityofWashington†IntelResearchDepartmentofComputerScience&Engineering1100NE45thSt.Seattle,WA98195Seattle,WA98105Abstracteraldomains.However,nogeneralguidancehasbeengivenonwhenMPL

2、canbesafelyused,andindeedMPLhasbeenWhileconditionalrandomfields(CRFs)havebeenobservedtoover-estimatethedependencyparametersinsomeappliedsuccessfullyinavarietyofdomains,theirexperiments[GeyerandThompson,1992].trainingremainsachallengingtask.Inthispaper,Inaddition,neitherMLnorMPLperformsf

3、eatureselec-weintroduceanoveltrainingmethodforCRFs,tionexplicitly,andneitherofthemisabletoadequatelyhan-calledvirtualevidenceboosting,whichsimulta-dlecontinuousobservations.Theselimitationsmakethemneouslyperformsfeatureselectionandparameterunsuitableforsometasks,suchasactivityrecogniti

4、onbasedestimation.Toachievethis,weextendstandardonrealsensordataandidentifyingthesetoffeaturesthatboostingtohandlevirtualevidence,whereanob-aremostusefulforclassification.Alternatively,boostinghasservationcanbespecifiedasadistributionratherbeensuccessfullyusedforfeatureselectioninthecont

5、extofthanasinglenumber.Thisextensionallowsustoclassificationproblems[ViolaandJones,2002].However,itsdevelopaunifiedframeworkforlearningbothlocalapplicationtorelationaldataremainsanunsolvedproblemandcompatibilityfeaturesinCRFs.Inexperimentssinceitassumestheindependenceofhiddenlabels.onsyn

6、theticdataaswellasrealactivityclassifi-Inthispaper,weshowhowtoseamlesslyintegrateboost-cationproblems,ournewtrainingalgorithmout-ingandCRFtraining,therebycombiningthecapabilitiesofperformsothertrainingapproachesincludingmax-bothparadigms.Theintegrationisachievedbycuttingaimumlikelihood,

7、maximumpseudo-likelihood,andCRFintoindividualpatches,asdoneinMPL,andusingthesethemostrecentboostedrandomfields.patchesastraininginstancesforboosting.ThekeydifferencetoMPL,however,isthatinourframeworktheneighborlabels1Introductionarenottreatedasobserved,butasvirtualevidencesorbeliefs.T