OnthePartitionFunctionandRandomMaximumAPosteriori在分区的作用随机最大后验

OnthePartitionFunctionandRandomMaximumAPosteriori在分区的作用随机最大后验

ID:36612105

大小:420.01 KB

页数:8页

时间:2019-05-12

OnthePartitionFunctionandRandomMaximumAPosteriori在分区的作用随机最大后验_第1页
OnthePartitionFunctionandRandomMaximumAPosteriori在分区的作用随机最大后验_第2页
OnthePartitionFunctionandRandomMaximumAPosteriori在分区的作用随机最大后验_第3页
OnthePartitionFunctionandRandomMaximumAPosteriori在分区的作用随机最大后验_第4页
OnthePartitionFunctionandRandomMaximumAPosteriori在分区的作用随机最大后验_第5页
资源描述:

《OnthePartitionFunctionandRandomMaximumAPosteriori在分区的作用随机最大后验》由会员上传分享,免费在线阅读,更多相关内容在学术论文-天天文库

1、OnthePartitionFunctionandRandomMaximumA-PosterioriPerturbationsTamirHazantamir@ttic.eduTommiJaakkolatommi@csail.mit.eduAbstractcoveryofthe3Djointpositionsfrom2Dimagesisof-teninherentlyambiguous.Similarly,inparsingthereInthispaperwerelatethepartitionfunctio

2、nmightbeequallylikelyparsetreesforthesamesen-tothemax-statisticsofrandomvariables.tence.Inafullyprobabilistictreatment,allpossibleInparticular,weprovideanovelframeworkalternativeassignmentsareconsidered.Thisrequiresforapproximatingandboundingtheparti-summi

3、ngovertheassignmentswiththeirrespectivetionfunctionusingMAPinferenceonran-weights{evaluatingthepartitionfunction{whichisdomlyperturbedmodels.Asaresult,wecanconsiderablyharder.Infact,anyalgorithmforcom-useecientMAPsolverssuchasgraph-cutsputingthepartitionf

4、unctionwouldbeabletoapprox-toevaluatethecorrespondingpartitionfunc-imatetheMAPvaluearbitrarilywellviaatempera-tion.Weshowthatourmethodexcelsinthetureargument(Landau&Lifshitz,1980).Incontrast,typicalhighsignal-highcoupling"regimeMAPinference(maximization)c

5、anbetractableeventhatresultsinraggedenergylandscapesdif-whentheproblemofevaluatingthepartitionfunction cultforalternativeapproaches.(weightedcounting)isnot.ThemainsurprisingresultofourworkisthatMAPin-1.Introductionferencecanbeusedtoapproximatethepartitionf

6、unc-tion.Inotherwords,givenanalgorithmforcomputingLearningandinferenceincomplexmodelsdrivesmuchtheMAPvalue,wecanapproximateorboundthepar-oftheresearchinmachinelearningapplications,fromtitionfunction.TheMAPvaluesinourcasearisefromcomputervision,naturallangu

7、ageprocessing,tocom-randomlyperturbedmodels.Whilemodelsbasedonputationalbiology.Examplesincludeobjectdetec-randomperturbationshavebeenconsideredrecentlytion(Felzenszwalbetal.,2009),stereovision(Szeliski(Papandreou&Yuille,2011;Keshetetal.,2011;Tar-etal.,200

8、7),parsing(Kooetal.,2010),orproteinlowetal.,2012),theirrelationtothepartitionfunctiondesign(Sontagetal.,2008).Theinferenceproblemhasnot.Speci cally,werelatethepartitionfunctioninsuchcasesinvolvesassessingthel

当前文档最多预览五页,下载文档查看全文

此文档下载收益归作者所有

当前文档最多预览五页,下载文档查看全文
温馨提示:
1. 部分包含数学公式或PPT动画的文件,查看预览时可能会显示错乱或异常,文件下载后无此问题,请放心下载。
2. 本文档由用户上传,版权归属用户,天天文库负责整理代发布。如果您对本文档版权有争议请及时联系客服。
3. 下载前请仔细阅读文档内容,确认文档内容符合您的需求后进行下载,若出现内容与标题不符可向本站投诉处理。
4. 下载文档时可能由于网络波动等原因无法下载或下载错误,付费完成后未能成功下载的用户请联系客服处理。