Probabilistic Graphical Models

Introduction

Welcome to the PGM Class

Daphne Koller

Probabilistic Graphical Models

Daphne Koller

Course Structure
10 weeks + final Videos + quizzes 9 problem sets
25% of score Multiple submissions

Daphne Koller

Course Structure
9 programming assignments
Genetically inherited diseases Optical character recognition Recognizing activities from Kinect sensor 9 x 7% = 63% of score

Final exam
12% of score

Daphne Koller

Background
Required
Basic probability theory Some programming Some algorithms and data structures

Recommended
Machine learning Simple optimization Matlab or Octave
Daphne Koller

Other Issues
Honor code Time management (10-15 hrs / week) Discussion forum & study groups

Daphne Koller

What youll learn

Fundamental methods Real-world applications How to use these methods in your work

M. Pradhan , G. Provan , B. Middleton , M.Henrion, UAI 94

Daphne Koller

Probabilis#c' Graphical' Models'

Introduc#on'

Mo#va#on' and'Overview'
Daphne Koller

predisposing factors symptoms test results diseases treatment outcomes

millions of pixels or thousands of superpixels each needs to be labeled {grass, sky, water, cow, horse, }

Daphne Koller

Probabilistic Graphical Models

Daphne Koller

domain expert

Models
Declarative representation

Data
Learning

elicitation

Model
Algorithm

Algorithm

Algorithm

Daphne Koller

Uncertainty
Partial knowledge of state of the world Noisy observations Phenomena not covered by our model Inherent stochasticity
Daphne Koller

Probability Theory
Declarative representation with clear semantics Powerful reasoning patterns Established learning methods
Daphne Koller

Complex Systems
predisposing factors symptoms test results diseases treatment outcomes class labels for thousands of superpixels

Random variables X1,, Xn Joint distribution P(X1,, Xn)

Daphne Koller

Graphical Models
Bayesian networks
Difficulty Intelligence Grade Letter SAT D C

Markov networks
A B

Daphne Koller

Graphical Models

M. Pradhan, G. Provan, B. Middleton, M. Henrion, UAI 94

Daphne Koller

Graphical Representation
Intuitive & compact data structure Efficient reasoning using general-purpose algorithms Sparse parameterization
feasible elicitation learning from data
Daphne Koller

Many Applications
Medical diagnosis Computer vision Image segmentation Fault diagnosis 3D reconstruction Natural language Holistic scene analysis processing Speech recognition Traffic analysis Social network models Robot localization & mapping Message decoding
Daphne Koller

Image Segmentation

Daphne Koller

Medical Diagnosis -

Thanks to: Eric Horvitz, Microsoft Research

Daphne Koller

Textual Information Extraction

Mrs. Green spoke today in New York. Green chairs the finance committee.

Daphne Koller

Multi-Sensor Integration: Traffic

Multiple views on traffic

Trained on historical data Learn to predict current & future road speed, including on unmeasured roads Dynamic route optimization

Weather

Learned Model

Incident reports

Thanks to: Eric Horvitz, Microsoft Research

I95 corridor experiment: accurate to 5 MPH in 85% of cases Fielded in 72 cities

Daphne Koller

This figure may be used for non-commercial and classroom purposes only. Any other uses require the prior written permission from AAAS

Biological Network Reconstruction

PKC PKA Plc Jnk PIP3 Erk PIP2 Akt Subsequently validated in wetlab
Daphne Koller

Phospho-Proteins Phospho-Lipids Perturbed in data

Raf P38 Mek

Known Supported Reversed Missed

15/17 2/17 1 3

Causal protein-signaling networks derived from multiparameter single-cell data Sachs et al., Science 2005

 Representation

Overview

Directed and undirected Temporal and plate models Exact and approximate Decision making Parameters and structure With and without complete data

Inference

Learning

Daphne Koller

Probabilis0c+ Graphical+ Models+

Introduc0on+

Preliminaries:+ Distribu0ons+
Daphne Koller

Joint Distribution
Intelligence (I) Difficulty (D) Grade (G)
i0 (low), i1 (high), d0 (easy), d1 (hard) g1 (A), g2 (B), g3 (C)

I i0 i0 i0 i0 i0 i0 i1 i1 i1 i1 i1 i1

D d0 d0 d0 d1 d1 d1 d0 d0 d0 d1 d1 d1

G g1 g2 g3 g1 g2 g3 g1 g2 g3 g1 g2 g3

Prob. 0.126 0.168 0.126 0.009 0.045 0.126 0.252 0.0224 0.0056 0.06 0.036 0.024
Daphne Koller

Conditioning
condition on g1
I i0 i0 i0 i0 i0 i0 i1 i1 i1 i1 i1 i1 D d0 d0 d0 d1 d1 d1 d0 d0 d0 d1 d1 d1 G g1 g2 g3 g1 g2 g3 g1 g2 g3 g1 g2 g3 Prob. 0.126 0.168 0.126 0.009 0.045 0.126 0.252 0.0224 0.0056 0.06 0.036 0.024
Daphne Koller

Conditioning: Reduction
I i0 D d0 G g1 Prob. 0.126 i0 d1 g1

0.009

i1

d0

g1

0.252

i1

d1

g1

0.06

Daphne Koller

Conditioning: Renormalization
I i0 i0 i1 i1 D d0 d1 d0 d1 G g1 g1 g1 g1 Prob. 0.126 0.009 0.252 0.06 I i0 i0 i1 i1 D d0 d1 d0 d1 Prob. 0.282 0.02 0.564 0.134

P(I, D, g1)

0.447

P(I, D | g1)

Daphne Koller

Marginalization
Marginalize I
I i0 i0 i1 i1 D d0 d1 d0 d1 Prob. 0.282 0.02 0.564 0.134

D d0 d1

Prob. 0.846 0.154

Daphne Koller

Probabilis1c+ Graphical+ Models+

Introduc1on+

Preliminaries:+ Factors+
Daphne Koller

Factors
A factor (X1,,Xk)
: Val(X1,,Xk) R

Scope = {X1,,Xk}

Daphne Koller

Joint Distribution
I i0 i0 i0 i0 i0 i0 i1 i1 i1 i1 i1 i1 D d0 d0 d0 d1 d1 d1 d0 d0 d0 d1 d1 d1 G g1 g2 g3 g1 g2 g3 g1 g2 g3 g1 g2 g3 Prob. 0.126 0.168 0.126 0.009 0.045 0.126 0.252 0.0224 0.0056 0.06 0.036 0.024
Daphne Koller

P(I,D,G)

Unnormalized measure P(I,D,g1)

P(I,D,g1)
I i0 i0 i1 i1 D d0 d1 d0 d1 G g1 g1 g1 g1 Prob. 0.126 0.009 0.252 0.06

Daphne Koller

Conditional Probability Distribution (CPD)

P(G | I,D)
i0,d0 i0,d1 i1,d0 i1,d1 g1 0.3 0.05 0.9 0.5 g2 0.4 0.25 0.08 0.3 g3 0.3 0.7 0.02 0.2

Daphne Koller

General factors
A a0 a0 a1 a1 B b0 b1 b0 b1

30 5 1 10

Daphne Koller

Factor Product
a1 a1 a1 a1 a1 a1 a2 a2 a3 a3 b1 b2 b1 b2 b1 b2 0.5 0.8 0.1 0 0.3 0.9 b1 b1 b2 b2 c1 c2 c1 c2 0.5 0.7 0.1 0.2 a2 a2 a2 a2 a3 a3 a3 a3 b1 b1 b2 b2 b1 b1 b2 b2 b1 b1 b2 b2 c1 c2 c1 c2 c1 c2 c1 c2 c1 c2 c1 c2 0.50.5 = 0.25 0.50.7 = 0.35 0.80.1 = 0.08 0.80.2 = 0.16 0.10.5 = 0.05 0.10.7 = 0.07 00.1 = 0 00.2 = 0 0.30.5 = 0.15 0.30.7 = 0.21 0.90.1 = 0.09 0.90.2 = 0.18
Daphne Koller

Factor Marginalization
a1 a1 a1 a1 a2 a2 a2 a2 a3 a3 a3 a3 b1 b1 b2 b2 b1 b1 b2 b2 b1 b1 b2 b2 c1 c2 c1 c2 c1 c2 c1 c2 c1 c2 c1 c2 0.25 0.35 0.08 0.16 0.05 0.07 0 0 0.15 0.21 0.09 0.18
Daphne Koller

a1 a1 a2 a2 a3 a3

c1 c2 c1 c2 c1 c2

0.33 0.51 0.05 0.07 0.24 0.39

Factor Reduction
a1 a1 a1 a1 a2 a2 a2 a2 a3 a3 a3 a3 b1 b1 b2 b2 b1 b1 b2 b2 b1 b1 b2 b2 c1 c2 c1 c2 c1 c2 c1 c2 c1 c2 c1 c2 0.25 0.35 0.08 0.16 0.05 0.07 0 0 0.15 0.21 0.09 0.18
Daphne Koller

a1 a1 a2 a2 a3 a3

b1 b2 b1 b2 b1 b2

c1 c1 c1 c1 c1 c1

0.25 0.08 0.05 0 0.15 0.09

Why factors?
Fundamental building block for defining distributions in high-dimensional spaces Set of basic operations for manipulating these probability distributions

Daphne Koller