November 4, 2009 : Iain Murray

University of Toronto

Elliptical Slice Sampling

Many probabilistic models introduce strong dependencies between variables using a latent multivariate Gaussian distribution or a Gaussian process. Inference in such models with non-Gaussian observations is not possible in closed form. Approximate methods such as Expectation Propagation (EP) or Markov chain Monte Carlo (MCMC) are needed to marginalize over the latent variables.

We present a new Markov chain Monte Carlo operator for performing inference in models with multivariate Gaussian priors. Its key properties are: 1) it has simple, generic code applicable to any such model, 2) it has no free parameters, 3) it has worked well in some non-parametric and semi-parametric settings. These properties make our method ideal for use while model building, removing the need to spend time deriving and tuning updates for more complex algorithms.

This is very recent work with Ryan P. Adams and David J.C. MacKay.

seminars/seminaritems/2009-11-04.txt · Last modified: 2009/11/09 09:58 by koray