Probabilistic Graphical Models

CVFX
2015.04.21

1

representation
inference
learning

2

probabilistic graphical models?

? ?

3

Probabilistic Graphical Models: Principles and
Techniques
Daphne Koller and Nir Friedman
http://pgm.stanford.edu/
MOOC course on Coursera
Graphical Models in a Nutshell
http://ai.stanford.edu/~koller/Papers/Koller+al:SR
L07.pdf

4
Graphs

nodes and links

directed undirected

5

random variables

joint probability

Independence

marginal probability

conditional probability

6
 +
a tool for modeling uncertainty

a general-purpose modeling language for

exploiting the independence properties in the
distribution

7
uncertainty:
probabilities

logical structure:
independence constraints

8
uncertainty:
1. observations are partial
2. observations are noisy
3. innate nondeterministic

9
uncertainty:
1. observations are partial
2. observations are noisy
3. innate nondeterministic

10
uncertainty:
1. observations are partial
2. observations are noisy
3. innate nondeterministic

11
uncertainty:
1. observations are partial
2. observations are noisy
3. innate nondeterministic

12
uncertainty:
1. observations are partial
2. observations are noisy
3. innate nondeterministic

13
uncertainty:
1. observations are partial
2. observations are noisy
3. innate nondeterministic

structure:
1. joint probability distribution P(A,B)
2. posterior distribution P(A|B = b)
3. conditional independence and
factorization

14
 PRML, Chris Bishop
: image de-noising

15
: image labeling
Efficient Inference in Fully Connected CRFs with Gaussian Edge
Potentials, Krahenbuhl and Koltun

16
Graphical models

nodes: random variables

links: probabilistic constraints between variables

Bayesian network Markov network

17
 Bayesian networks
directed acyclic graphs (DAGs)
conditional probability distributions (CPDs)
decompose the distribution as a product of CPDs

Markov networks
undirected graphs
cliques (complete subgraphs) and factors
non-negativity: the only constraint on the
parameters in the factor

18
Estimating joint distributions?

For me, estimating joint distributions is a bit like

playing God.
You cant do everything!
Vladimir Vapnik

quote from Graphical Models for Machine Learning and Digital Communication,
Brendan J. Frey
19
Estimating joint distributions

Season 4-valued

Flu Hayfever
binary-valued binary-valued

Muscle-pain Congestion
binary-valued binary-valued

modeling P(S, F, H, C, M)
4 2 2 2 2 = 64 configurations
20


joint

21
Table representations

joint marginal

22
Conditional probabilities

23
 0
Evidence (| = )

24
Factorization
chain rule

large factor large table ()

conditional independence factors

25
Factorization and graphs

directed graphs undirected graphs

Bayesian networks Markov networks
d-separation blanket
parent-child neighbors
causality clique

26
Conditional independence

Season

Flu Hayfever

Muscle-pain Congestion

27
Bayesian networks
directed acyclic graphs (DAGs)
joint distribution factorization of
conditional probability distributions (CPDs)

28
Bayesian networks

Season

Flu Hayfever

Muscle-pain Congestion

29
Conditional independence
assumptions in Bayesian networks

Season

Flu Hayfever

Muscle-pain Congestion

30
Flow of influence
Consider a simple three-node path X Z Y. If
influence can flow from X to Y via Z, we say that
the path X Z Y is active.
causal path evidential path
X Z Y X Z Y
common causal common effect
Z X Y X Y
X Y Z Z
W
31
Flow of influence

32
Active paths

33
Directed separation (d-separation)

34
Independence and factorization in BN
Coherence

Difficult Intelligence

Grade
SAT

Letter
example from PGM,
Happy Job Koller and Friedman

35
 PRML, Chris Bishop
Markov networks
Markov random fields (MRF)
undirected graphs
Nodes: variables
Links: connect a pair of nodes

specify a factorization and a set of conditional

independence relations for the joint
distribution of the random variables

36
Conditional independence properties
in MRF
consider all possible paths that connect nodes
in set A to nodes in set B
if all such paths pass through one or more nodes
in set C, then all such paths are 'blocked' and so
the conditional independence property holds

C B

A
37
Factorization properties
expressing the joint distribution as a product
of functions defined over sets of variables that
are local to the graph

38
Clique
a complete subgraph
A subset of the nodes in a graph such that there
exists a link between every pair of nodes in the
subset
a maximal clique is a clique such that it is not
possible to include any other nodes from the
graph in the set without it ceasing to be a
clique

39
Maximal cliques

40
Potential functions
the joint distribution can be written as a
product of potential functions over the
maximal cliques of the graph

41
Computational limitation

42
Strictly positive potential functions
express the potential functions as exponentials

the joint distribution is defined as the product of

potentials, and so the total energy is obtained by
adding the energies of each of the maximal
cliques

43
 PRML, Chris Bishop
MRF modes as binary pixels

cliques?

44
 PRML, Chris Bishop
: image de-noising

noise model
E.g., flipping the sign of the pixels with probability
10%

45
Joint probability

state state-state image-state

noisy image compatibility compatibility
function function
neighboring local
state nodes observations

46
Energy functions
we need to choose energy functions for the
cliques
a suitable energy function should express the
relations among the nodes of a cliques
E.g.,

minimizing energy = maximizing probability

47
How to minimize the energy function?
Iterated conditional modes (ICM)
Coordinate-wise gradient descent
Not guaranteed to find the global minimum

inference ()

48
Iterated conditional modes (ICM)

49
ICM example

-1
+1

50
Factorization and graphs

directed graphs undirected graphs

Bayesian networks Markov networks
d-separation blanket
parent-child neighbors
causality clique

51
: Bayesian network
SamIam

52
Summary
graphical models ?

graphical models ?

53
Markov networks A
D B
undirected graphs C
cliques (complete subgraphs)

54
55
Maximal cliques

56
Factor and energy function

57
Independencies in Markov networks

58
 PRML, Chris Bishop
Image de-noising

iterated conditional modes graph-cut

59