Latent Variable Models and Expectation Maximization

Size: px

Start display at page:

Download "Latent Variable Models and Expectation Maximization"

April Mosley
5 years ago
Views:

1 Latent Variable Models and Expectation Maximization Oliver Schulte - CMPT 726 Bishop PRML Ch. 9

2 small then only the white circle is seen, and there is no extrema in entropy. in entropy. There There is anis entropy an entropy extrema extrema when when the the µ, Σ) α f trema scale scale is slightly is slightly larger larger than the thanradius the radius of theofbright bright circle, circle, tative p relative to a ref- which which has has In practice In practice this method this method gives gives stable stable identification identification of fea- of fea- to a refnitydensity re arts assumed are assumed to totures turesaover variety a variety of sizes of sizes and copes and copes well with well with intra-class intra-class ckground model model asin(within a range a range r). r). ant toant scaling, to scaling, although although experimental experimental tests show tests show that this thatis this We as-variability. discussed variability. Theprobabilistic saliency saliency measure measure models is designed is designed attolength betoinvari- invari- is α f d p r f p, U p ) dp r f er. e finder. 1 p(d θ) (N, f) p(d θ) Learning and thereafter and Parameters thereafter entropy the entropy todecreases Probability theasscale scale increases. increases. Distributions detected tures detected using using The n M. second The second is is variable and the and last the last ll ible possible occlusion occlusion l. shape s the shape and ocpearancthe appearance and and compasses many many of of and oc- obabilistic istic way, this way, this constrained objects objects ll a small covariance) covariance) Given fully observed training data, setting parameters θ not entirely i not entirely the case thedue casetodue aliasing to aliasing and other and other effects. effects. Note, Note, for Bayes only monochrome only nets monochrome is straight-forward information information is used is used to detect to detect and represent features. sent infeatures. many settings not all variables are observed and repre- However, (labelled) in the training data: x i = (x i, h i ) 2.3. e.g Feature Speech Feature representation recognition: have speech signals, but not The phoneme feature The feature detector labels detector identifies identifies regions regions of interest of interest on each on each image. e.g. image. Object The coordinates Therecognition: coordinates of theof centre the have centre give object us give Xus and labels Xthe andsize (car, the size bicycle), of but theof not region thepart region gives labels gives S. Figure S. (wheel, Figure 2 illustrates 2door, illustrates this seat) on this two ontypical two typical images images from the frommotorbike the motorbike dataset. dataset. Unobserved variables are called latent variables figs from Fergus et al. Figure Figure 2: Output 2: Output of theof feature the feature detector detector

3 Latent Variables and Simplicity Latent variables can explain observed correlations with a simple model. Fewer parameters. Common in science: The heart, genes, energy, gravity, Smoking Diet Exercise Smoking Diet Exercise 54 HeartDisease Symptom 1 Symptom 2 Symptom Symptom 1 Symptom 2 Symptom 3 (a) (b) Fig. Russell and Norvig 2.1

4 Latent Variable Models: Pros Statistically powerful, often good predictions. Many applications: Learning with missing data. Clustering: missing cluster label for data points. Principal Component Analysis: data points are generated in linear fashion from a small set of unobserved components. (more later) Matrix Factorization, Recommender Systems: Assign users to unobserved user types, assign items to unobserved item types. Use similarity between user type, item type to predict preference of user for item. Winner of $1M Netflix challenge. If latent variables have an intuitive interpretation (e.g., action movies, factors ), discovers new features.

5 Latent Variable Models: Cons From a user s point of view, like a black box if latent variables don t have an intuitive interpretation. Statistically, hard to guarantee convergence to a correct model with more data (the identifiability problem). Harder computationally, usually no closed form for maximum likelihood estimates. However, the Expectation-Maximization algorithm provides a general-purpose local search algorithm for learning parameters in probabilistic models with latent variables.

6 Outline K-Means EM Example: Gaussian Mixture Models The Expectation Maximization Algorithm

7 Outline K-Means EM Example: Gaussian Mixture Models The Expectation Maximization Algorithm

8 Unsupervised Learning 2 (a) 2 We will start with an unsupervised learning (clustering) problem: Given a dataset {x 1,..., x N }, each x i R D, partition the dataset into K clusters Intuitively, a cluster is a group of points, which are close together and far from others

9 Distortion Measure 2 (a) 2 2 (i) 2 Formally, introduce prototypes (or cluster centers) µ k R D Use binary r nk, 1 if point n is in cluster k, otherwise (1-of-K coding scheme again) Find {µ k }, {r nk } to minimize distortion measure: J = N n=1 k=1 K r nk x n µ k 2

10 Minimizing Distortion Measure Minimizing J directly is hard N K J = r nk x n µ k 2 n=1 k=1 However, two things are easy If we know µ k, minimizing J wrt r nk If we know r nk, minimizing J wrt µ k This chicken-and-egg scenario suggests an iterative procedure Start with initial guess for µ k Iteration of two steps: Minimize J wrt r nk Minimize J wrt µ k Rinse and repeat until convergence

11 Minimizing Distortion Measure Minimizing J directly is hard N K J = r nk x n µ k 2 n=1 k=1 However, two things are easy If we know µ k, minimizing J wrt r nk If we know r nk, minimizing J wrt µ k This chicken-and-egg scenario suggests an iterative procedure Start with initial guess for µ k Iteration of two steps: Minimize J wrt r nk Minimize J wrt µ k Rinse and repeat until convergence

12 Minimizing Distortion Measure Minimizing J directly is hard N K J = r nk x n µ k 2 n=1 k=1 However, two things are easy If we know µ k, minimizing J wrt r nk If we know r nk, minimizing J wrt µ k This chicken-and-egg scenario suggests an iterative procedure Start with initial guess for µ k Iteration of two steps: Minimize J wrt r nk Minimize J wrt µ k Rinse and repeat until convergence

13 2 (a) Determining Membership Variables Step 1 in an iteration of K-means is to minimize distortion measure J wrt cluster membership variables r nk 2 2 (b) 2 J = N n=1 k=1 K r nk x n µ k 2 Terms for different data points x n are independent, for each data point set r nk to minimize: K r nk x n µ k 2 How? k=1 Simply set r nk = 1 for the cluster center µ with smallest distance

14 2 (a) Determining Membership Variables Step 1 in an iteration of K-means is to minimize distortion measure J wrt cluster membership variables r nk 2 2 (b) 2 J = N n=1 k=1 K r nk x n µ k 2 Terms for different data points x n are independent, for each data point set r nk to minimize: K r nk x n µ k 2 How? k=1 Simply set r nk = 1 for the cluster center µ with smallest distance

15 2 (a) Determining Membership Variables Step 1 in an iteration of K-means is to minimize distortion measure J wrt cluster membership variables r nk 2 2 (b) 2 J = N n=1 k=1 K r nk x n µ k 2 Terms for different data points x n are independent, for each data point set r nk to minimize: K r nk x n µ k 2 How? k=1 Simply set r nk = 1 for the cluster center µ with smallest distance

16 Determining Cluster Centers Step 2: fix r nk, minimize J wrt the cluster centers µ k 2 (b) K N J = r nk x n µ k 2 switch order of sums k=1 n=1 2 2 (c) 2 Minimize wrt each µ k separately (how?) Take derivative, set to zero: 2 N r nk (x n µ k ) = n=1 µ k = n r nkx n n r nk i.e. mean of datapoints x n assigned to cluster k

17 Determining Cluster Centers Step 2: fix r nk, minimize J wrt the cluster centers µ k 2 (b) K N J = r nk x n µ k 2 switch order of sums k=1 n=1 2 2 (c) 2 Minimize wrt each µ k separately (how?) Take derivative, set to zero: 2 N r nk (x n µ k ) = n=1 µ k = n r nkx n n r nk i.e. mean of datapoints x n assigned to cluster k

18 K-means Algorithm Start with initial guess for µ k Iteration of two steps: Minimize J wrt r nk Assign points to nearest cluster center Minimize J wrt µ k Set cluster center as average of points in cluster Rinse and repeat until convergence

19 K-means example 2 (a) 2

20 K-means example 2 (b) 2

21 K-means example 2 (c) 2

22 K-means example 2 (d) 2

23 K-means example 2 (e) 2

24 K-means example 2 (f) 2

25 K-means example 2 (g) 2

26 K-means example 2 (h) 2

27 K-means example 2 (i) 2 Next step doesn t change membership stop

28 K-means Convergence Repeat steps until no change in cluster assignments For each step, value of J either goes down, or we stop Finite number of possible assignments of data points to clusters, so we are guaranteed to converge eventually Note it may be a local maximum rather than a global maximum to which we converge

29 K-means Example - Image Segmentation Original image K-means clustering on pixel colour values Pixels in a cluster are coloured by cluster mean Represent each pixel (e.g. 24-bit colour value) by a cluster number (e.g. 4 bits for K = 1), compressed version

30 K-means Generalized: the set-up Let s generalize the idea. Suppose we have the following set-up. x denotes all observed variables (e.g., data points). Z denotes all latent (hidden, unobserved) variables (e.g., cluster means). J(x, Z θ) where J measures the goodness of an assignment of latent variable models given the data points and parameters θ. e.g., J = -dispersion measure. parameters = assignment of points to clusters. It s easy to maximize J(x, Z θ) wrt θ for fixed Z. It s easy to maximize J(x, Z θ) wrt Z for fixed θ.

31 Outline K-Means EM Example: Gaussian Mixture Models The Expectation Maximization Algorithm

32 Hard Assignment vs. Soft Assignment 2 (i) In the K-means algorithm, a hard 2 assignment of points to clusters is made However, for points near the decision boundary, this may not be a good idea Instead, we could think about making a soft assignment of points to clusters

Gaussian Mixture Model 1 (a) 1 (b).5.2.5.5.3.5 1.5 1 The Gaussian mixture model (or mixture of Gaussians MoG) models the data as a combination of Gaussians.

33 Gaussian Mixture Model 1 (a) 1 (b) The Gaussian mixture model (or mixture of Gaussians MoG) models the data as a combination of Gaussians. a: constant density contours. b: marginal probability p(x). c: surface plot. Widely used general approximation for multi-modal distributions. See text Sec for kernel version

34 Gaussian Mixture Model 1 (b) 1 (a) Above shows a dataset generated by drawing samples from three different Gaussians

35 ag=1) θ Generative Model ag C 1 (c).5 pper Hole X.5 1 (a) (b) The mixture of Gaussians is a generative model To generate a datapoint x, we first generate a value C = i for the component/cluster C {1, 2, 3} We then generate a vector x using P(x C = i) N (x µ i, Σ i ) using the Gaussian parameters for component i

36 B) P(Bag=1) θ MoG Marginal over Observed Variables Bag C vor Wrapper Hole X (a) (b) The marginal distribution p(x θ) for this model with parameter values θ is: p(x) = = 3 3 p(x, C = i) = p(c = i)p(x C = i) i=1 3 w i N (x µ i, Σ i ) i=1 i=1 A mixture of Gaussians

37 P(Bag=1) θ Learning Gaussian Mixture Models B) Bag C vor Wrapper Hole X The (a) EM algorithm can be used (b) to learn Bayes net parameters with unobserved variables 1. Initialize parameters randomly (or using k-means) 2. Repeat the following Expectation E-step Find probability of cluster assignments Maximization M-step Update parameter values

38 =1) E-step: Probability of Cluster Assignments g C 1 (b).5 per Hole X.5 1 ) (b) How to find p ij = P(C = i x j )? Use Bayes Theorem p ij P(x j C = i)p(c = i) Bayes net inference!

39 =1) E-step: Probability of Cluster Assignments g C 1 (b).5 per Hole X.5 1 ) (b) How to find p ij = P(C = i x j )? Use Bayes Theorem p ij P(x j C = i)p(c = i) Bayes net inference!

40 EM for Gaussian Mixtures: Update Parameters Update parameters wrt soft cluster assignments p ij n i j p ij w i P(C = i) := n i /N µ i := j p ij x j /n i Σ i = j p ij (x j µ i )(x j µ i ) T /n i n i is the effective (estimated) number of data points from component i

41 MoG EM - Example 2 (a) 2 Same initialization as with K-means before Often, K-means is actually used to initialize EM

42 MoG EM - Example 2 (b) 2 Calculate cluster probabilities p ij Shown using mixtures of colours.

43 MoG EM - Example 2 (c) 2 Calculate model parameters {P(C = i), µ i, Σ i } using the cluster probabilities

44 MoG EM - Example 2 (d) 2 Iteration 2

45 MoG EM - Example 2 (e) 2 Iteration 5

46 MoG EM - Example 2 Iteration 2 - converged (f) 2

47 Log-likelihood L EM and Maximum Likelihood 1. The EM procedure is guaranteed to increase at each step, the data log-likelihood ln p(d θ). 2. Therefore converges to local log-likelihood maximum. 3. Data log-likelihood of true model is around 65 in figure Iteration number Fig. Russell and Norvig 2.12

48 Outline K-Means EM Example: Gaussian Mixture Models The Expectation Maximization Algorithm

49 Hard EM Algorithm We assume a probabilistic model, specifically the complete-data log likelihood function P(x, Z = z θ). Goodness of the model is the log-likelihood L(x, Z = z θ) ln P(x, Z = z θ). Guess the value for latent variables that is the expected value given current parameter settings: E[Z] where the distribution over latent variables is given by P(Z x, θ (i) ). Given latent variable values, the next parameter values θ i+1 are evaluated by maximizing the likelihood L(x Z = E[Z], θ). Example: fill in missing values, single variable, Gaussian model.

50 Generalized EM Algorithm In Bayesian fashion, do not guess a best value for the latent variables Z. Instead, average over the distribution P(Z x, θ (i) ) given the current hypothesis. Given a latent variable distribution, the next parameter values θ (i+1) are evaluated by taking the expected likelihood L(x, Z = z θ (i+1) ) over all possible latent variable settings. In one equation: θ (i+1) = argmax θ P(Z x, θ (i) )L(x, Z = z θ) z

51 EM Algorithm: The procedure 1. Guess an initial parameter setting θ (i). 2. Repeat until convergence: 2.1 The E-step: Evaluate P(Z = z x, θ (i) ). 2.2 The M-step: Evaluate the function Q(θ, θ (i) ) z P(Z x, θ(i) )L(x, Z = z θ) Maximize Q(θ, θ (i) ) wrt θ. Update θ (i+1) to be the maximum.

52 Bag P(F=cherry B) 1 θ F1 2 θ F2 P(Bag=1) EM θfor Gaussian Mixtures: E-step (I) Bag C Flavor Wrapper Hole X (a) (b) Want to evaluate Q(θ, θ (i) ) z P(Z x, θ(i) )L(x, Z = z θ) Let Z ij = 1 denote the event that data point j is assigned to cluster i. Then L(x, Z = z θ) = ln P(x j C = i, θ) z ij P(C = i θ) z ij i j = z ij [ln P(x j C = i, θ) z ij + ln P(C = i θ)] i j z ij l ij i j

53 P(Bag=1) EM for Gaussian Mixtures: E-step (II) θ Bag P(F=cherry B) 1 θ F1 2 θ F2 Bag C Flavor Wrapper Hole (a) (b) Want to evaluate Q(θ, θ (i) ) z P(Z x, θ(i) ) i j z ijl ij This is the expectation of a sum = sum of expectations (see assignment). Therefore P(Z x, θ (i) ) z ij l ij = z i j = P(Z ij = 1 x, θ (i) )l ij = i j p ijl ij i j X E Zij x,θ (i)[z ijl ij ] i j

54 EM for Gaussian Mixtures: E-step (III) P(Bag=1) θ Bag P(F=cherry B) Bag C 1 θ F1 2 θ F2 Flavor Wrapper Hole X (a) Want to evaluate Q(θ, θ (i) ) z P(Z x, θ(i) )L(x, Z = z θ) We showed that this is equal to i j p ijl(x j, z ij θ) Deriving the maximization solution is more or less straightforward and leads to the updates shown before. (b) Variations on this theme work for many models where the likelihood function factors (think Bayes nets).

55 EM - Summary EM finds local maximum to (incomplete) data likelihood P(x θ) = Z p(x, Z θ) Iterates two steps: E step calculates the distribution of the missing variables Z (Hard EM fills in the variables). M step maximizes expected complete log likelihood (expectation wrt E step distribution) Often the updates can be obtained in closed form E.g., for discrete Bayesian networks (see Ch.2.3.2)

56 Conclusion K-means clustering Gaussian mixture model What about K? Model selection: either cross-validation or Bayesian version (average over all values for K) Expectation-maximization, a general method for learning parameters of models when not all variables are observed

Introduction to Machine Learning CMU-10701

Introduction to Machine Learning CMU-10701 Clustering and EM Barnabás Póczos & Aarti Singh Contents Clustering K-means Mixture of Gaussians Expectation Maximization Variational Methods 2 Clustering 3 K-