中科院机器学习题库new概要

《中科院机器学习题库new概要》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、机器学习题库一、极大似然1、MLestimationofexponentialmodel(10)AGaussiandistributionisoftenusedtomodeldataontherealline,butissometimesinappropriatewhenthedataareoftenclosetozerobutconstrainedtobenonnegative.Insuchcasesonecanfitanexponentialdistribution,whoseprobabilitydensityfu

2、nctionisgivenbyGivenNobservationsxidrawnfromsuchadistribution:(a)Writedownthelikelihoodasafunctionofthescaleparameterb.(b)Writedownthederivativeoftheloglikelihood.(c)GiveasimpleexpressionfortheMLestimateforb.2、换成Poisson分布：3、二、贝叶斯假设在考试的多项选择中，考生知道正确答案的概率为p，猜测答案的概率为

3、1-p，并且假设考生知道正确答案答对题的概率为1，猜中正确答案的概率为，其中m为多选项的数目。那么已知考生答对题目，求他知道正确答案的概率。1、ConjugatepriorsThereadingsforthisweekincludediscussionofconjugatepriors.Givenalikelihoodforaclassmodelswithparametersθ,aconjugatepriorisadistributionwithhyperparametersγ,suchthattheposteriord

4、istribution与先验的分布族相同(a)Supposethatthelikelihoodisgivenbytheexponentialdistributionwithrateparameterλ:Showthatthegammadistribution_isaconjugatepriorfortheexponential.Derivetheparameterupdategivenobservationsandthepredictiondistribution.(b)Showthatthebetadistributi

5、onisaconjugatepriorforthegeometricdistributionwhichdescribesthenumberoftimeacoinistosseduntilthefirstheadsappears,whentheprobabilityofheadsoneachtossisθ.Derivetheparameterupdateruleandpredictiondistribution.(a)Supposeisaconjugatepriorforthelikelihood;showthatthem

6、ixturepriorisalsoconjugateforthesamelikelihood,assumingthemixtureweightswmsumto1.(d)Repeatpart(c)forthecasewherethepriorisasingledistributionandthelikelihoodisamixture,andthepriorisconjugateforeachmixturecomponentofthelikelihood.somepriorscanbeconjugateforseveral

7、differentlikelihoods;forexample,thebetaisconjugatefortheBernoulliandthegeometricdistributionsandthegammaisconjugatefortheexponentialandforthegammawithfixedα(e)(Extracredit,20)Explorethecasewherethelikelihoodisamixturewithfixedcomponentsandunknownweights;i.e.,thew

8、eightsaretheparameterstobelearned.三、判断题（1）给定n个数据点，如果其中一半用于训练，另一半用于测试，则训练误差和测试误差之间的差别会随着n的增加而减小。（2）极大似然估计是无偏估计且在所有的无偏估计中方差最小，所以极大似然估计的风险最小。（３）回归函数A和B，如果A比B更简单，则