Graphical models are a unifying framework for describing the statistical relationships between large sets of variables. The course addresses the basic questions of computing the marginal distribution, partition function and feature expectation etc. The focus is on sparse graph structures and theoretical analysis with the goal of providing a unifying roadmap for navigating and understanding the broad array of approximate algorithms for marginalization and learning in graphical models. Topics include: message passing algorithms; belief propagation; survey propagation; sparse graph codes; random combinatorial optimization (random K-satisfiability). This course will show how a wide class of methods----including mean field theory, sum-product or belief propagation algorithms, expectation-propagation, and max-product algorithms----are all variational methods, meaning that they can be understood as algorithms for solving particular optimization problems on graphs. The perspective also forges connections to convex optimization, including linear programming and other type of conic relaxations.
Time: Tuesday/Thursday 6:00 pm -7:15 pm; Location: Russ 302
428, Russ Engineering Center Building
Office hours: Tuesday/Thursday 2:00PM-3:30PM
T. Richardson and R. Urbanke
Modern Coding Theory
Cambridge University Press , 2008.
Attendance and paper presentation 100%