Scalable inference for Bayesian non-parametrics

Abstract

Bayesian non-parametric models, despite their theoretical elegance, face a serious computational burden that prevents their use in serious "big data'' scenarios. Furthermore, we cannot expect the data in "big data'' to exist solely on one processor, so we must have parallel algorithms that are valid Bayesian inference samplers. However, inherent dependencies in Bayesian non-parametric models make this task very difficult. Instead, we must either construct good approximations or develop clever reformulations of our models so that we perform inference with provably accurate results. This thesis will discuss four methods developed to parallelize inference in the Bayesian and Bayesian non-parametric setting.