Applied Nonparametric Bayes
Michael I. Jordan · 2009
In this talk I will discuss some recent progress in Bayesian nonparametric modeling and inference. Focusing on the needs of applications, I will discuss some of the limitations of the popular Dirichlet process, in particular its lack of power-law marginals and its poor scaling in problems involving large numbers of clusters. I will present some alternatives that are based on the theory of completely random measures and on stick-breaking constructions. I will present applications to problems in computational vision and bioinformatics. [Joint work with Erik Sudderth and Romain Thibaux.]