Data-Driven MCMC for Learning and Inference in面向学习和推理的数据驱动MCMC 切换线性动态系统

《Data-Driven MCMC for Learning and Inference in面向学习和推理的数据驱动MCMC 切换线性动态系统》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、Data-DrivenMCMCforLearningandInferenceinSwitchingLinearDynamicSystemsSangMinOhJamesM.RehgTuckerBalchFrankDellaertCollegeofComputing,GeorgiaInstituteofTechnology801AtlanticDriveAtlanta,GA30332-0280{sangmin,tucker,rehg,dellaert}@cc.gatech.eduAbstractInferenceinanSLDSmodelinvolvescomputing

2、theposteriordistributionofthehiddenstates,whichconsistSwitchingLinearDynamicSystem(SLDS)modelsareofthe(discrete)switchingstateandthe(continuous)dy-apopulartechniqueformodelingcomplexnonlineardy-namicstate.Inthebehaviorrecognitionapplicationwhichnamicsystems.AnSLDShassignificantlymoredesc

3、riptivemotivatesthiswork,thediscretestaterepresentsdistinctpowerthananHMM,butinferenceinSLDSmodelsishoneybeebehaviorswhilethedynamicstaterepresentscomputationallyintractable.ThispaperdescribesanovelinferencealgorithmforSLDSmodelsbasedontheData-thebee’struemotion.Givenvideo-basedmeasurem

4、entsDrivenMCMCparadigm.Wedescribeanewproposalofthepositionandorientationofthebeeovertime,distributionwhichsubstantiallyincreasestheconvergenceSLDSinferencecanbeusedtoobtainaMAPestimatespeed.Comparisonstostandarddeterministicapproximationofthebehaviorandmotionofthebee.Inadditiontomethods

5、demonstratetheimprovedaccuracyofournewitscentralroleinapplicationssuchasMAPestimation,approach.WeapplyourapproachtotheproblemoflearninginferenceisalsothecrucialstepinparameterlearningviaanSLDSmodelofthebeedance.Honeybeescommuni-theEMalgorithm(Pavlovic,Rehg,&MacCormick2000).´catethelocat

6、ionanddistancetofoodsourcesthroughaApproximateinferencetechniqueshavebeendevelopedtodancethattakesplacewithinthehive.WelearnSLDSaddressthecomputationallimitationsoftheexactapproach.modelparametersfromtrackingdatawhichisautomaticallyextractedfromvideo.WethendemonstratetheabilitytoPreviou

7、sworkonapproximateinferenceinSLDSmod-successfullysegmentnovelbeedancesintotheirconstituentelshasfocusedprimarilyontwoclassesoftechniques:parts,effectivelydecodingthedanceofthebees.stage-wisemethodssuchasapproximateViterbiorGPB2whichmaintainaconstantrepresentationalsizeforeachti