Seminar by Frederik Eaton from University of Cambridge

Date and time

May 19th, 2011(Thu.), 13:30 - (about one hour)
Admission Free, No Booking Necessary

Speaker

Frederik Eaton (University of Cambridge)

Place

Seminar room 5 of ISM

Title

A conditional game for comparing approximations
(The talk is based on his recent work which has been awarded as Notable Paper Award in AISTATS 2011.)

Abstract

The task of approximate inference is to compute approximate marginal probabilities in a statistical model. Traditionally, researchers have evaluated the accuracy of approximate inference algorithms by running them on models which are small enough so that exact marginals can be computed for comparison. In this talk we will present a novel method for comparing two approximations in the absence of exact marginals. This makes it possible for the first time to directly compare the accuracy of our algorithms on large models. The method is based on a "conditional game" which is played between two approximate inference algorithms. We show how the game can be used to rank two approximations, and prove that the true marginal probabilities are an optimal strategy. We make an analogy between the conditional game and legal procedure. We present the results of some empirical investigations into the game's effectiveness.