Probabilis1c

Graphical Models

Representa1on Local Structure

Log-Linear Models
Daphne Koller

Log-Linear Representation

Each feature fj has a scope Dj Different features can have same scope
Daphne Koller

Representing Table Factors

(X1, X2) = a00 a10 a01 a11

Daphne Koller

Features for Language

Features: word capitalized, word in atlas or name list, previous word is Mrs, next word is Times,

Daphne Koller

Ising Model

Daphne Koller

Metric MRFs
All Xi take values in label space V
Xi

Xj

Distance function : V V R

want Xi and Xj to take similar values

Reflexivity: (v,v) = 0 for all v Symmetry: (v1,v2) = (v2,v1) for all v1, v2 Triangle inequality: (v1,v2) (v1,v3) + (v3, v2) for all v1, v2, v3
Daphne Koller

Metric MRFs
All Xi take values in label space V
Xi

Xj

Distance function : V V R

want Xi and Xj to take similar values

values of Xi and Xj far in

lower probability
Daphne Koller

Metric MRF Examples

0

(vk,vl) =

vk=vl otherwise

0 1 1 1

1 0 1 1

1 1 0 1

1 1 1 0

(vk,vl)

vk-vl

(vk,vl)

vk-vl
Daphne Koller

Metric MRF: Segmentation

0

(vk,vl) =

vk=vl otherwise

0 1 1 1

1 0 1 1

1 1 0 1

1 1 1 0

Daphne Koller

original

Metric MRF: Denoising

Gaussian noise stdev 20

denoised

Gaussian noise stdev 50

denoised

(vk,vl) = |vk-vl|

(vk,vl) = min(|vk-vl|,d)

vk-vl Similar idea for stereo reconstruction

vk-vl
Daphne Koller