Probabilistic graphical models (PGMs) combine ideas from statistics and computer science into a unifying framework for modeling complex real-world phenomena. PGMs are now widespread in language and speech processing. PGMs are well suited to handle the inherent challenges of linguistic problems: complex and structured relationships, a large number of relevant attributes, and large volumes of data. This short course will provide students with advanced training in several specific applications of graphical models that are important in our field. After reviewing the essentials of directed and undirected graphical models, we will discuss complex CRFs, approximate inference including variational and MCMC methods, Bayesian models and non-parametric Bayesian models including Chinese Restaurant Processes. Students will also gain practical experience by solving problems using existing PGM software.

About the course

This is a one-credit short course (course number 600.405) that will comprise 12 lectures of 50 minutes each. The course will meet weekly on Fridays from 11:00am-12:00pm from February 4, 2011 to May 6, 2011 (no meeting on March 25 due to spring break).


Hackerman 209 

Office Hours

Tuesdays, 10:30am-11:50am, Hackerman 226E (Ves & Shane), Hackerman 226F (Alex), or by appointment

Grading policy

Marks will be assigned as follows:
  • 20% for class attendance and participation
  • 40% for written reviews/critiques of two PGM papers (we will mark your best 2 reviews out of 3)
  • 40% for presentation of a relevant research paper  *OR*  completion of a small PGM project

Course Materials

Students will read key research papers that illustrate the use of each of the main PGMs discussed in the class.  Students may also refer to suggested review material in “Daphne Koller and Nir Friedman, Probabilistic Graphical Models: Principles and Techniques.” [KF09]

Course Feedback

We welcome regular feedback on the course and our teaching, by e-mail or in person!  In addition, we have created a form to allow you to provide anonymous feedback.  Please take a moment to fill in the following: