CIS 620 - Advanced Topics in AI - Spring 07
Probabilistic Graphical Models

Instructors: Ben Taskar and Koby Crammer
Lectures: Levine 307, Tuesday and Thursday, 1:30-3pm
Office hours: GRW 464, Thursday, 3-4pm (Ben); GRW 170, Tuesday, 3-4pm (Koby) or by appointment


A list of presentations can be found here.

Project guidelines have been posted. Please talk to Ben or Koby before submitting your proposal on Thursday, Feb 15.

Older Announcements

Please fill out the questionnaire if you haven't already and bring to class on Thursday, Jan 11.

Course description

Effective modeling of uncertainty and complexity of interactions is a fundamental problem in understanding and designing complex systems. By combining ideas from statistics and graph theory, probabilistic graphical models provide a general representation and an algorithmic framework for reasoning about statistical dependencies. The graphical model family includes very well-known members: Bayesian networks (BNs), Markov random fields (MRFs), hidden Markov models (HMMs), Kalman filters, Ising models, mixture models. These models underly many statistical approaches in computer vision, robotics, natural language understanding, signal processing, computational biology and other fields. The course will cover the following topics:
Part I: Representation
Markov properties
Directed graphical models
Undirected graphical models
Exponential family, generalized linear models
Part II: Inference
Elimination algorithm
Junction tree algorithm
Variational methods, mean field, belief propagation
Sampling methods, Gibbs, MCMC
Part III: Learning
Parameter estimation
Model selection
Incomplete data, expectation maximization (EM)
Mixture models, factor analysis, Dirichlet process
Part IV: Student presentations


There is no published textbook for this course. Photocopies of selected chapters from the following books will be handed out in class.


CIS 520 or equivalent. Knowledge of basic probabilty and statisics, linear algebra, dynamic programming. Exposure to graph theory and algorithms, information theory, optimization will be helpful.