ECE 6504: Advanced Topics in Machine Learning Probabilistic Graphical Models and Large-Scale Learning

Transcription

1 ECE 6504: Advanced Topics in Machine Learning Probabilistic Graphical Models and Large-Scale Learning Topics Markov Random Fields: Representation Conditional Random Fields Log-Linear Models Readings: KF 4.1-3; Barber Dhruv Batra Virginia Tech

2 Administrativia No class Next week (Tue, Thu) Project Proposal Due: Mar 12, Mar 5, 11:59pm <=2pages, NIPS format HW2 Out later today Due: Mar 12, 11:59pm Implementation: Variable Elimination in BNs (C) Dhruv Batra 2

3 Recap of Last Time (C) Dhruv Batra 3

4 Markov Nets Set of random variables Undirected graph Encodes independence assumptions Unnormalized Factor Tables Joint distribution: Product of Factors (C) Dhruv Batra 4

5 Pairwise MRFs Pairwise Factors A function of 2 variables Often unary terms are also allowed (although strictly speaking unnecessary) On board (C) Dhruv Batra 5

6 Pairwise MRF: Example (C) Dhruv Batra 6

7 Computing probabilities in Markov networks vs BNs In a BN, can compute prob. of an instantiation by multiplying CPTs In an Markov networks, can only compute ratio of probabilities directly (C) Dhruv Batra Slide Credit: Carlos Guestrin 7

8 Normalization for computing probabilities To compute actual probabilities, must compute normalization constant (also called partition function) Computing partition function is hard! Must sum over all possible assignments (C) Dhruv Batra Slide Credit: Carlos Guestrin 8

9 Nearest-Neighbor Grids Low Level Vision Image denoising Stereo Optical flow Shape from shading Superresolution Segmentation unobserved or hidden variable local observation (C) Dhruv Batra Slide Credit: Erik Sudderth 9

10 General Gibbs Distribution Arbitrary Factors Induced MRF Graph (C) Dhruv Batra 10

11 Factorization in Markov networks Given an undirected graph H over variables X={X 1,...,X n } A distribution P factorizes over H if there exist subsets of variables D 1 X,, D m X, such that D i are fully connected in H non-negative potentials (or factors) φ 1 (D 1 ),, φ m (D m ) also known as clique potentials such that Also called Markov random field H, or Gibbs distribution over H (C) Dhruv Batra Slide Credit: Carlos Guestrin 11

12 MRFs Given a graph H, are factors unique? (C) Dhruv Batra 12

13 Active Trails and Separation A path X 1 X k is active when set of variables Z are observed if none of X i {X 1,,X k } are observed (are part of Z) Variables X are separated from Y given Z in graph If no active path between any X X and any Y Y given Z T 1 T 2 T 3 T 4 T 5 T 6 T (C) Dhruv Batra 7 T 8 T 9 13

14 Markov networks representation Theorem 1 If joint probability distribution P: Then H is an I-map for P If you can write distribution as a normalized product of factors Then Can read independencies from graph (C) Dhruv Batra Slide Credit: Carlos Guestrin 14

15 What about the other direction for Markov networks? If H is an I-map for P Then joint probability distribution P: Counter-example: X 1,,X 4 are binary, and only eight assignments have positive probability: For example, X 1 X 3 X 2,X 4 : E.g., P(X 1 =0 X 2 =0, X 4 =0) But distribution doesn t factorize!! (C) Dhruv Batra Slide Credit: Carlos Guestrin 15

16 Representation Theorem for Markov Networks Hammersley Clifford theorem If joint probability distribution P: Then H is an I-map for P If H is an I-map for P and P is a positive distribution Then joint probability distribution P: (C) Dhruv Batra Slide Credit: Carlos Guestrin 16

17 Markov Blanket = Markov Blanket of variable x 8 Parents, children and parents of children (C) Dhruv Batra Slide Credit: Simon J.D. Prince 17

18 Independence Assumptions in MNs Separation defines global independencies Pairwise Markov Independence: Pairs of non-adjacent variables A,B are independent given all others T 3 T 1 T 4 T 2 T 5 T 6 Markov Blanket: Variable A independent of rest given its neighbors T 7 T 8 T 9 (C) Dhruv Batra Slide Credit: Carlos Guestrin 18

19 P-map Perfect map G is a P-map for P if I(P) = I(G) Question: Does every distribution P have P-map? (C) Dhruv Batra 19

20 Structure in cliques Possible potentials for this graph: A B C

21 Factor graphs Bipartite graph: variable nodes (ovals) for X 1,,X n factor nodes (squares) for φ 1,,φ m edge X i φ j if X i ε Scope[φ j ] A C B Very useful for approximate inference Make factor dependency explicit

22 Types of Graphical Models Directed Factor Undirected (C) Dhruv Batra Slide Credit: Erik Sudderth 22

23 Factor Graphs show Fine-grained Factorization p(x) = 1 Z f F ψ f (x f ) (C) Dhruv Batra Slide Credit: Erik Sudderth 23

24 Plan for today Undirected Graphical Models: Representation Conditional Random Fields Log-Linear Models Undirected Graphical Models: Inference Variable Elimination (C) Dhruv Batra 24

25 Conditional Random Fields What s the difference between Naïve Bayes & Logistic Regression? (C) Dhruv Batra 25

26 Nearest-Neighbor Grids Low Level Vision Image denoising Stereo Optical flow Shape from shading Superresolution Segmentation unobserved or hidden variable local observation

27 (C) Dhruv Batra 27

28 Logarithmic representation Standard model: Log representation of potential (assuming positive potential): also called the energy function Log representation of Markov net:

29 Log-linear Markov network (most common representation) Feature (or Sufficient Statistic) is some function f [D] for some subset of variables D e.g., indicator function Log-linear model over a Markov network H: a set of features f 1 [D 1 ],, f k [D k ] each D i is a subset of a clique in H two f s can be over the same variables a set of weights w 1,,w k usually learned from data

30 CRFs Felzenszwalb, Huttenlocher, IJCV 04 (C) Dhruv Batra 30

31 CRFs Node Features Edge Features f i (Y i ) f ij (Y i,y j ) f j (Y j ) (C) Dhruv Batra 31

32 Node Feature -- Color Superpixel Feature Extraction Step 1 Rmean Gmean Bmean Hmean Smean Vmean 5-dim hist on H 3-dim hist on S Hoiem, Efros, Hebert, IJCV 2007

33 Node Feature Color Clustering (µ 1," 1 ),(µ 2," 2 ),,(µ N," N ) K-means/X-means, Pelleg, Moore Auton Lab implementation Dictionary Feature Space

34 Node Feature -- Color Superpixel Feature Extraction Step 1 Feature Extraction Step 2 Pr (Cluster feature) R mean G mean B mean H mean S mean V mean 5-dim hist on H 3-dim hist on S Pr(Cluster 1 " i )! Pr(Cluster i " i )! Pr(Cluster N " i ) Hoiem, Efros, Hebert, IJCV 2007

35 Conditional Random Fields (C) Dhruv Batra 35

36 Summary of types of Markov nets Pairwise Markov networks very common potentials over nodes and edges General MRFs Factor graphs explicit representation of factors you know exactly what factors you have very useful for approximate inference Log-linear models log representation of potentials linear coefficients learned from data most common for learning MNs