lecture8 advanced machine learningaai高级机器学习.pptx

下载文档

0
0
约2.59千字
约 27页
2024-05-11 发布于北京
举报
版权申诉
保障服务

lecture8 advanced machine learningaai高级机器学习.pptx

1、本文档共27页，可阅读全部内容。
2、原创力文档（book118）网站文档一经付费（服务费），不意味着购买了该文档的版权，仅供个人/单位学习、研究之用，不得用于商业用途，未经授权，严禁复制、发行、汇编、翻译或者网络传播等，侵权必究。
3、本站所有内容均由合作方或网友上传，本站不对文档的完整性、权威性及其观点立场正确性做任何保证或承诺！文档内容仅供研究参考，付费前请自行鉴别。如您付费，意味着您自己接受本站规则且自行承担风险，本站不退款、不进行额外附加服务；查看《如何避免下载的几个坑》。如果您已付费下载过本站文档，您可以点击这里二次下载。
4、如文档侵犯商业秘密、侵犯著作权、侵犯人身权等，请点击“版权申诉”（推荐），也可以打举报电话：400-050-0827(电话支持时间：9:00-18:30)。

AdvancedArtificialIntelligence

Lecture8:Advancemachinelearning

OutlineClusteringK-MeansEMSpectralClusteringDimensionalityReduction2

Theunsupervisedlearningproblem3Manydatapoints,nolabels

K-Means4Manydatapoints,nolabels

K-MeansChooseafixednumberofclustersChooseclustercentersandpoint-clusterallocationstominimizeerrorcan’tdothisbyexhaustivesearch,becausetherearetoomanypossibleallocations.Algorithmfixclustercenters;allocatepointstoclosestclusterfixallocation;computebestclustercentersxcouldbeanysetoffeaturesforwhichwecancomputeadistance(carefulaboutscaling)

K-Means

K-Means*FromMarcPollefeysCOMP2562003

K-MeansIsanapproximationtoEMModel(hypothesisspace):MixtureofNGaussiansLatentvariables:CorrespondenceofdataandGaussiansWenotice:Giventhemixturemodel,it’seasytocalculatethecorrespondenceGiventhecorrespondenceit’seasytoestimatethemixturemodels

ExpectationMaximzation:IdeaDatageneratedfrommixtureofGaussiansLatentvariables:CorrespondencebetweenDataItemsandGaussians

GeneralizedK-Means(EM)

Gaussians11

MLFittingGaussians12

LearningaGaussianMixture

(withknowncovariance)M-StepE-Step

ExpectationMaximizationConverges!Proof[Neal/Hinton,McLachlan/Krishnan]:E/MstepdoesnotdecreasedatalikelihoodConvergesatlocalminimumorsaddlepointButsubjecttolocalminima

PracticalEMNumberofClustersunknownSuffers(badly)fromlocalminimaAlgorithm:Startnewclustercenterifmanypoints“unexplained”Killclustercenterthatdoesn’tcontribute(UseAIC/BICcriterionforallthis,ifyouwanttobeformal)15

SpectralClustering16

SpectralClustering17

SpectralClustering:OverviewDataSimilaritiesBlock-Detection

EigenvectorsandBlocksBlockmatriceshaveblockeigenvectors:Near-blockmatriceshavenear-blockeigenvectors:[Ngetal.,NIPS02]1100110000110011eigensolver.71.710000.71.71l1=2l2=2l3=0l4=011.20110-.2.20110-.211eigensolver.71.69.1400-.14.69.7