Abstract by Spencer Newcomb
An Improved Merge-Split Sampler for Dirichlet Process Mixture Models
The Dirichlet process mixture (DPM) model is a popular Bayesian nonparametric model. Markov chain Monte Carlo (MCMC) algorithms are often used to sample from the posterior distribution. Several MCMC algorithms exist, including Gibbs samplers. In finite sampling, these algorithms often do not adequately explore valleys of low probability density, getting “stuck” in local modes – a problem that is accentuated in higher dimensions. In this presentation, we provide a simple overview of the DPM model, review existing sampling schemes, and introduce a new merge-split MCMC method that borrows ideas from Sequential Importance Sampling. We show that this new method offers better computational efficiency in terms of distributional convergence and scaling than other available split-merge algorithms.