seminar: Graphical models: Variational methods and message-passing
speaker: Martin Wainwright (UC Berkeley, Departments of EECS and Statistics)abstract: Graphical models provide a unified framework for modeling and computation of large collections of random variables, and are widely used in various fields, including communication theory, control theory, signal processing, and statistics. Use of these models in practice requires efficient algorithms for solving various basic problems, including computing marginal distributions and maximum a posteriori (MAP) assignments for a known model, as well as the learning-theoretic problem of estimating unknown model parameters and structure from data. All of these problems can be understood within a common variational framework, based on convex duality and the geometry of exponential families from statistics. In Part I, we introduce this variational framework, and how various known message-passing algorithms (sum-product, max-product, Kikuchi methods, Gauss-Seidel, Kalman filtering) can be derived as distributed algorithms for solving specific instances (or approximations) of this general variational principle. We also discuss how various learning algorithms (iterative proportional fitting, expectation-maximization) can be understood within the same framework.
timetable:
Tue 4 Dec, 16:30 - 18:00, Aula Dini
<< Go back