Summary

2011

FH Eaton. Combining Approximations for Inference (PhD thesis)

This is my PhD thesis, which explores the approximate statistical inference problem from an external structural perspective, asking what are the ways of "combining" two approximations and how these could be used to build new algorithms. It is the first work to apply a form of simulated evolution, analogous to Genetic Algorithms in the framework of optimisation, directly to the framework of inference.

FH Eaton, Z Ghahramani. Generality of pairwise, binary, and planar factor graphs for probabilistic inference. (in submission, Neural Computation)

This paper is an extended version of section 2.4.3 of my PhD thesis. It shows that it is impossible to reduce general factor graphs (containing zeroes) to binary-pairwise form. It also establishes that such reductions are possible when the graph is strictly positive, and are also possible for general graphs in a "limit" sense. A limiting reduction of general graphs to planar binary pairwise form is also described.

FH Eaton. A conditional game for comparing approximations. In Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics (AISTATS 2011). Notable Paper Award

Defines a game which can be used to compare the accuracy of two approximations to the marginal probabilities of a statistical model. Simple as it is, apparently this is the first method that anyone has proposed for making such comparisons!

2009

FH Eaton, Z Ghahramani. Choosing a Variable to Clamp: Approximate Inference Using Conditioned Belief Propagation. In Proceedings of the Twelfth International Conference on Artificial Intelligence and Statistics (AISTATS 2009)

Proposes an algorithm for applying Belief Propagation to a model using divide-and-conquer, by recursively conditioning on specific variables. This algorithm can be seen as an approximate version of cutset conditioning. The paper also introduces "BBP" or "Back Belief Propagation", an application of back-propagation to belief propagation. BBP is used here for choosing condition variables, but has many other applications, for instance to parameter learning tasks in computer vision (Domke, "Parameter Learning with Truncated Message-Passing". CVPR 2011) and to empirical risk minimisation in general Machine Learning (Stoyanov et al, "Empirical Risk Minimization of Graphical Model Parameters Given Approximate Inference, Decoding, and Model Structure". AISTATS 2011).