中科院機(jī)器學(xué)習(xí)題庫(kù)

本文格式為Word版，下載可任意編輯——中科院機(jī)器學(xué)習(xí)題庫(kù)機(jī)器學(xué)習(xí)題庫(kù)

一、極大似然

1、MLestimationofexponentialmodel(10)

AGaussiandistributionisoftenusedtomodeldataontherealline,butissometimesinappropriatewhenthedataareoftenclosetozerobutconstrainedtobenonnegative.Insuchcasesonecanfitanexponentialdistribution,whoseprobabilitydensityfunctionisgivenby

1?xp?x??eb

bGivenNobservationsxidrawnfromsuchadistribution:

(a)Writedownthelikelihoodasafunctionofthescaleparameterb.(b)Writedownthederivativeoftheloglikelihood.(c)GiveasimpleexpressionfortheMLestimateforb.

x??2、換成Poisson分布：p?x|????e,y?0,1,2,...

x!l?????log?p?xi|?????xilog????log?x!?

i?1i?1N?????xi?log??N???log?xi!?i?1?i?1?NNN3、二、貝葉斯

假設(shè)在考試的多項(xiàng)選擇中，考生知道正確答案的概率為p，猜測(cè)答案的概率為1-p，并且假

設(shè)考生知道正確答案答對(duì)題的概率為1，猜中正確答案的概率為1m，其中m為多項(xiàng)選擇項(xiàng)的數(shù)目。那么已知考生答對(duì)題目，求他知道正確答案的概率。1、

p?known|correct??p?known,correct?p?known??pp??1?p?1mConjugatepriors

Thereadingsforthisweekincludediscussionofconjugatepriors.Givenalikelihoodp?x|??foraclassmodelswithparametersθ,aconjugatepriorisadistributionp??|??withhyperparametersγ,suchthattheposteriordistribution

p??|X,????p?X|??p??|???p??|???

與先驗(yàn)的分布族一致

(a)Supposethatthelikelihoodisgivenbytheexponentialdistributionwithrateparameterλ:p?x|????e??x

Showthatthegammadistribution

Gamma??|?,???????1???_

?e????isaconjugatepriorfortheexponential.Derivetheparameterupdategivenobservationsx1,?,xNandthepredictiondistributionp?xN?1|x1,?,xN?.

(b)Showthatthebetadistributionisaconjugatepriorforthegeometricdistribution

p?x?k|????1???k?1?

whichdescribesthenumberoftimeacoinistosseduntilthefirstheadsappears,whentheprobabilityofheadsoneachtossisθ.Derivetheparameterupdateruleandpredictiondistribution.

(c)Supposep??|??isaconjugatepriorforthelikelihoodp?x|??;showthatthemixtureprior

p??|?1,...,?M???wmp??|?m?

m?1Misalsoconjugateforthesamelikelihood,assumingthemixtureweightswmsumto1.

(d)Repeatpart(c)forthecasewherethepriorisasingledistributionandthelikelihoodisamixture,andthepriorisconjugateforeachmixturecomponentofthelikelihood.

somepriorscanbeconjugateforseveraldifferentlikelihoods;forexample,thebetaisconjugatefortheBernoulliandthegeometricdistributionsandthegammaisconjugatefortheexponentialandforthegammawithfixedα

(e)(Extracredit,20)Explorethecasewherethelikelihoodisamixturewithfixedcomponentsandunknownweights;i.e.,theweightsaretheparameterstobelearned.

三、判斷題

（1）給定n個(gè)數(shù)據(jù)點(diǎn)，假使其中一半用于訓(xùn)練，另一半用于測(cè)試，則訓(xùn)練誤差和測(cè)試誤差之間的區(qū)別會(huì)隨著n的增加而減小。

（2）極大似然估計(jì)是無偏估計(jì)且在所有的無偏估計(jì)中方差最小，所以極大似然估計(jì)的風(fēng)險(xiǎn)最小。（３）回歸函數(shù)A和B，假使A比B更簡(jiǎn)單，則A幾乎一定會(huì)比B在測(cè)試集上表現(xiàn)更好。（４）全局線性回歸需要利用全部樣本點(diǎn)來預(yù)計(jì)新輸入的對(duì)應(yīng)輸出值，而局部線性回歸只需利用查詢點(diǎn)附近的樣本來預(yù)計(jì)輸出值。所以全局線性回歸比局部線性回歸計(jì)算代價(jià)更高。

（５）Boosting和Bagging都是組合多個(gè)分類器投票的方法，二者都是根據(jù)單個(gè)分類器的正確率決定其權(quán)重。

(６)Intheboostingiterations,thetrainingerrorofeachnewdecisionstumpandthetrainingerrorofthecombinedclassifiervaryroughlyinconcert（F）

Whilethetrainingerrorofthecombinedclassifiertypicallydecreasesasafunctionofboostingiterations,theerroroftheindividualdecisionstumpstypicallyincreasessincetheexampleweightsbecomeconcentratedatthemostdifficultexamples.

(７)OneadvantageofBoostingisthatitdoesnotoverfit.（F）

(８)Supportvectormachinesareresistanttooutliers,i.e.,verynoisyexamplesdrawnfromadifferentdistribution.（Ｆ）

（9）在回歸分析中，最正確子集選擇可以做特征選擇，當(dāng)特征數(shù)目較多時(shí)計(jì)算量大；嶺回歸和Lasso模型計(jì)算量小，且Lasso也可以實(shí)現(xiàn)特征選擇。

（10）當(dāng)訓(xùn)練數(shù)據(jù)較少時(shí)更簡(jiǎn)單發(fā)生過擬合。

（11）梯度下降有時(shí)會(huì)陷于局部微小值，但EM算法不會(huì)。

（12）在核回歸中，最影響回歸的過擬合性和欠擬合之間平衡的參數(shù)為核函數(shù)的寬度。(13)IntheAdaBoostalgorithm,theweightsonallthemisclassifiedpointswillgoupbythesamemultiplicativefactor.（T）

(14)True/False:Inaleast-squareslinearregressionproblem,addinganL2regularizationpenaltycannotdecreasetheL2errorofthesolutionw?onthetrainingdata.（F）

(15)True/False:Inaleast-squareslinearregressionproblem,addinganL2regularizationpenaltyalwaysdecreasestheexpectedL2errorofthesolutionw?onunseentestdata（F）.(16)除了EM算法，梯度下降也可求混合高斯模型的參數(shù)。(T)

(20)Anydecisionboundarythatwegetfromagenerativemodelwith

class-conditionalGaussiandistributionscouldinprinciplebereproducedwithanSVMandapolynomialkernel.

True!Infact,sinceclass-conditionalGaussiansalwaysyieldquadraticdecisionboundaries,theycanbereproducedwithanSVMwithkernelofdegreelessthanorequaltotwo.

(21)AdaBoostwilleventuallyreachzerotrainingerror,regardlessofthetypeofweak

classifierituses,providedenoughweakclassifiershavebeencombined.

False!Ifthedataisnotseparablebyalinearcombinationoftheweakclassifiers,AdaBoostcan’tachievezerotrainingerror.

(22)TheL2penaltyinaridgeregressionisequivalenttoaLaplacepriorontheweights.（F）

(23)Thelog-likelihoodofthedatawillalwaysincreasethroughsuccessiveiterationsoftheexpectationmaximationalgorithm.(F)

4、midterm2023problem4

6、Considertwoclassifiers:1)anSVMwithaquadratic(secondorderpolynomial)kernelfunctionand2)anunconstrainedmixtureoftwoGaussiansmodel,oneGaussianperclasslabel.TheseclassifierstrytomapexamplesinR2tobinarylabels.Weassumethattheproblemisseparable,noslackpenaltiesareaddedtotheSVMclassifier,andthatwehavesufficientlymanytrainingexamplestoestimatethecovariancematricesofthetwoGaussianc