Title: Using particle Gibbs methods with Bayesian nonparametric models
Authors: Jim Griffin - University of Kent (United Kingdom) [presenting]
Abstract: Bayesian nonparametric models are characterized by an infinite number of parameters. In many nonparametric models, only a finite number of these parameters are used for a finite sample but the exact number is unknown. For example, with a finite sample, a Dirichlet process mixture model will only have a finite (but unknown) number of clusters or a Indian buffet process factor model will only have a finite (but unknown) number of factors. Defining effective Markov chain Monte Carlo (MCMC) methods for posterior inference in these models is challenging and often uses one-at-a-time updates which can lead to slow mixing. Particle MCMC methods are attractive as they can avoid one-at-a-time updates but defining effective particle filters for these models is also very challenging. We will consider versions of particle Gibbs methods which are useful with these models and certain time-dependent extensions (where particle filtering is a natural method for inference). The methods will be illustrated on a range of Bayesian nonparametric models and data settings.