Gaussian Belief Propagation -- a survey and some experimental results

by Prof. Erik Aurell (KTH)

Tuesday 19 Jan 2010, 13:30 → 14:30 Europe/Stockholm

122:028

Belief propagation is an algorithmic version of Bethe approximation i.e. a way to compute marginals of a probability distribution efficiently (but not exactly). The approach has generated considerable attention in information theory and artificial intelligence for several years, and underlies several industrially important techniques such as iterative decoding and inference / learning models from data. The performance analysis has also motivated quite some activity in the spin glass community cf. the recent monograph by M Mezard and A. Montanari. Belief propagation in Gaussian models has a special role since it is exact for the means of marginals although not for the variance, in a wide class of such models. I will describe these results, some modern applications of Gaussian Belief propagation in information technology, and numerical experiments on one version of the model. This is joint work with Rene Pfitzner published in Europhys. Lett. (2009) vol 87 68004.