PASCAL - Pattern Analysis, Statistical Modelling and Computational Learning

Gaussian quadrature based expectation propagation.
Onno Zoeter and Tom Heskes
In: AI & Statistics 2005, 6-8 Jan 2005, Barbados.

Abstract

We present a general approximation method for Bayesian inference problems. The method is based on Expectation Propagation (EP). Projection steps in the EP iteration that cannot be done analytically are done using Gaussian quadrature. By identifying a general form in the projections, the only quadrature rules that are required are for exponential family weight functions. The corresponding cumulant and moment generating functions can then be used to automatically derive the necessary quadrature rules. In this article the approach is restricted to approximating families that factorize to a product of one-dimensional families. The final algorithm has interesting similarities with particle filtering algorithms. We discuss these, and also discuss the relationship with variational Bayes and Laplace propagation. Experimental results are given for an interesting model from mathematical finance.

EPrint Type:Conference or Workshop Item (Invited Talk)
Project Keyword:Project Keyword UNSPECIFIED
Subjects:Theory & Algorithms
ID Code:562
Deposited By:Bert Kappen
Deposited On:26 December 2004