11School of Computer ScienceLearning generalized linear models and tabular CPT of structured full BNProbabilistic Graphical Models (10Probabilistic Graphical Models (10--708)708)Lecture 9, Oct 15, 2007Eric XingEric XingReceptor AKinase CTF FGene GGene HKinaseEKinase DReceptor BX1X2X3X4X5X6X7X8Receptor AKinase CTF FGene GGene HKinaseEKinase DReceptor BX1X2X3X4X5X6X7X8X1X2X3X4X5X6X7X8Reading: J-Chap. 7,8. Eric Xing 2Linear Regressionz Let us assume that the target variable and the inputs are related by the equation:where ε is an error term of unmodeled effects or random noisez Now assume that ε follows a Gaussian N(0,σ), then we have:iiTiyεθ+= x⎟⎟⎠⎞⎜⎜⎝⎛−−=22221σθσπθ)(exp);|(iTiiiyxypx2Eric Xing 3Logistic Regression (sigmoid classifier)z The condition distribution: a Bernoulliwhere µis a logistic functionz We can used the brute-force gradient method as in LRz But we can also apply generic laws by observing the p(y|x) is an exponential family function, more specifically, a generalized linear modelyyxxxyp−−=11 ))(()()|(µµxTexθµ−+=11)(Eric Xing 4Exponential familyz For a numeric random variable Xis an exponential family distribution with natural (canonical) parameter ηz Function T(x) is a sufficient statistic.z Function A(η) = log Z(η) is the log normalizer.z Examples: Bernoulli, multinomial, Gaussian, Poisson, gamma,...{}{})(exp)()()()(exp)()|(xTxhZAxTxhxpTTηηηηη1=−=XnN3Eric Xing 5Multivariate Gaussian Distributionz For a continuous vector random variable X∈Rk:z Exponential family representationz Note: a k-dimensional Gaussian is a (d+d2)-parameter distribution with a (d+d2)-element vector of sufficient statistics (but because of symmetry and positivity, parameters are constrained and have lower degree of freedom)()()(){}Σ−Σ−Σ+Σ−=⎭⎬⎫⎩⎨⎧−Σ−−Σ=Σ−−−−logtrexp)()(exp),(///µµµπµµπµ121112121212212121TTTkTkxxxxxxp()[]()[]()[]()()22211122112112121121121122/)(log)(trlog)(vec;)( and ,vec,vec;kTTTxhAxxxxT−−−−−−=−−−=Σ+Σ==Σ−=Σ==Σ−Σ=πηηηηµµηηµηηηµηMoment parameterNatural parameterEric Xing 6Multinomial distributionz For a binary vector random variable z Exponential family representation),|(multi~πxx⎭⎬⎫⎩⎨⎧==∑kkkxKxxxxpKπππππlnexp)( L2121[]110111=⎟⎠⎞⎜⎝⎛=⎟⎠⎞⎜⎝⎛−−==⎥⎦⎤⎢⎣⎡⎟⎠⎞⎜⎝⎛=∑∑=−=)(lnln)()(;lnxheAxxTKkKkkKkkηπηππη⎭⎬⎫⎩⎨⎧⎟⎠⎞⎜⎝⎛−+⎟⎟⎠⎞⎜⎜⎝⎛∑−=⎭⎬⎫⎩⎨⎧⎟⎠⎞⎜⎝⎛−⎟⎠⎞⎜⎝⎛−+=∑∑∑∑∑−=−=−=−=−=−=1111111111111111KkkKkkKkkkKkkKkKKkkkxxxπππππlnlnexplnlnexp4Eric Xing 7Why exponential family?z Moment generating property{}{}[])()()(exp)()()(exp)()()()()(logxTEdxZxTxhxTdxxTxhddZZddZZddddATT=====∫∫ηηηηηηηηηηη11{}{}[][][])()()()()()()(exp)()()()(exp)()(xTVarxTExTEZddZdxZxTxhxTdxZxTxhxTdAd2TT=−=−=∫∫22221ηηηηηηηηEric Xing 8Moment estimationz We can easily compute moments of any exponential family distribution by taking the derivatives of the log normalizerA(η).z The qthderivative gives the qthcentered moment.z When the sufficient statistic is a stacked vector, partial derivatives need to be considered.Lvariance)(mean)(==22ηηηηdAdddA5Eric Xing 9Moment vs canonical parametersz The moment parameter µ can be derived from the natural (canonical) parameterz A(h) is convex sincez Hence we can invert the relationship and infer the canonical parameter from the moment parameter (1-to-1):z A distribution in the exponential family can be parameterized not only by η −the canonical parameterization, but also by µ −the moment parameterization.[]µηηdef)()(== xTEddA[]022>= )()(xTVardAdηη)(defµψη=48-2-101248-2-1012Aηη∗Eric Xing 10MLE for Exponential Familyz For iid data, the log-likelihood isz Take derivatives and set to zero:z This amounts to moment matching.z We can infer the canonical parameters using{}∑∑∏−⎟⎠⎞⎜⎝⎛+=−=nnnTnnnTnNAxTxhAxTxhD)()()(log)()(exp)(log);(ηηηηηl )()()( )()(∑∑∑==∂∂⇒=∂∂−=∂∂nnMLEnnnnxTNxTNAANxT110µηηηηη)l)(MLEMLEµψη))=6Eric Xing 11Sufficiencyz For p(x|θ), T(x) is sufficient for θif there is no information in Xregarding θyeyond that in T(x).z We can throw away Xfor the purpose pf inference w.r.t. θ. z Bayesian viewz Frequentist viewz The Neyman factorization theoremzT(x) is sufficient for θif T(x)θX))(|()),(|( xTpxxTpθθ=T(x)θX))(|()),(|( xTxpxTxp=θT(x)θX))(,()),(()),(,( xTxxTxTxp21ψθψθ=))(,()),(()|( xTxhxTgxpθθ=⇒Eric Xing 12Examplesz Gaussian:z Multinomial:z Poisson:()[]()[]()22112112112/)(log)(vec;)(vec;kTTxhAxxxxT−−−−=Σ+Σ==Σ−Σ=πµµηµη∑∑==⇒nnnnMLExNxTN111)(µ[]110111=⎟⎠⎞⎜⎝⎛=⎟⎠⎞⎜⎝⎛−−==⎥⎦⎤⎢⎣⎡⎟⎠⎞⎜⎝⎛=∑∑=−=)(lnln)()(;lnxheAxxTKkKkkKkkηπηππη∑=⇒nnMLExN1µ!)()()(logxxheAxxT1=====ηληλη∑=⇒nnMLExN1µ7Eric Xing 13Generalized Linear Models (GLIMs)z The graphical modelz Linear regressionz Discriminative linear classificationz Commonality: model Ep(Y)=µ=f(θTX)z What is p()? the cond. dist. of Y.z What is f()? the response function.z GLIMz The observed input xis assumed to enter into the model via a linear combination of its elementsz The conditional mean µis represented as a function f(ξ) of ξ, where f is known as the response functionz The observed output yis assumed to be characterized by an exponential family distribution with conditional mean µ. XnYnNxTθξ=Eric Xing 14GLIM, cont.z The choice of exp family is constrained by the nature of the data Yz Example: y is a continuous vector Æ multivariate Gaussiany is a class label Æ Bernoulli or multinomial z The choice of the response functionz Following some mild constrains, e.g., [0,1]. Positivity …z Canonical response function: z In this case θTxdirectly corresponds to canonical parameter η.ηψfθxµξ(){})()(exp)()|(ηηηφAyxyhypT−=⇒1)(⋅=−1ψf{})()(exp)()|(ηηηAyxyhypT−=8Eric Xing 15MLE for GLIMs with natural responsez Log-likelihoodz Derivative of Log-likelihoodz Online learning for canonical GLIMsz Stochastic gradient ascent = least mean squares (LMS) algorithm:()∑∑−+=nnnnnTnAyxyh )()(logηθl())()(µµθηηηθ−=−=⎟⎟⎠⎞⎜⎜⎝⎛−=∑∑yXxyddddAyxddTnnnnnnnnnnlThis is a fixed point function because µis a function of θ()ntnnttxyµρθθ−+=+1()size step
View Full Document