Technical reports

# Bayesian inference for latent variable models

**Ulrich Paquet**

July 2008, 137 pages

This technical report is based on a dissertation submitted March 2007 by the author for the degree of Doctor of Philosophy to the University of Cambridge, Wolfson College.

## Abstract

Bayes’ theorem is the cornerstone of statistical inference. It provides the tools for dealing with knowledge in an uncertain world, allowing us to explain observed phenomena through the refinement of belief in model parameters. At the heart of this elegant framework lie intractable integrals, whether in computing an average over some posterior distribution, or in determining the normalizing constant of a distribution. This thesis examines both deterministic and stochastic methods in which these integrals can be treated. Of particular interest shall be parametric models where the parameter space can be extended with additional latent variables to get distributions that are easier to handle algorithmically.

Deterministic methods approximate the posterior distribution with a simpler distribution over which the required integrals become tractable. We derive and examine a new generic α-divergence message passing scheme for a multivariate mixture of Gaussians, a particular modeling problem requiring latent variables. This algorithm minimizes local α-divergences over a chosen posterior factorization, and includes variational Bayes and expectation propagation as special cases.

Stochastic (or Monte Carlo) methods rely on a sample from the posterior to simplify the integration tasks, giving exact estimates in the limit of an infinite sample. Parallel tempering and thermodynamic integration are introduced as ‘gold standard’ methods to sample from multimodal posterior distributions and determine normalizing constants. A parallel tempered approach to sampling from a mixture of Gaussians posterior through Gibbs sampling is derived, and novel methods are introduced to improve the numerical stability of thermodynamic integration.

A full comparison with parallel tempering and thermodynamic integration shows variational Bayes, expectation propagation, and message passing with the Hellinger distance α = 1/2 to be perfectly suitable for model selection, and for approximating the predictive distribution with high accuracy.

Variational and stochastic methods are combined in a novel way to design Markov chain Monte Carlo (MCMC) transition densities, giving a variational transition kernel, which lower bounds an exact transition kernel. We highlight the general need to mix variational methods with other MCMC moves, by proving that the variational kernel does not necessarily give a geometrically ergodic chain.

## Full text

PDF (2.0 MB)

## BibTeX record

@TechReport{UCAM-CL-TR-724, author = {Paquet, Ulrich}, title = {{Bayesian inference for latent variable models}}, year = 2008, month = jul, url = {https://www.cl.cam.ac.uk/techreports/UCAM-CL-TR-724.pdf}, institution = {University of Cambridge, Computer Laboratory}, number = {UCAM-CL-TR-724} }