Title: Nonparametric maximum likelihood estimation of mixture models: Recent developments
Authors: Ivan Mizera - University of Alberta (Canada) [presenting]
Sile Tao - University of Alberta (Canada)
Abstract: The purpose is to review primal and dual formulations for the nonparametric maximum likelihood estimation of the mixing distribution in mixture models, in the context of the empirical Bayes methodology known as the Kiefer-Wolfowitz, or also Robbins method. While the original, primal formulation, is an infinite-dimensional convex optimization problem, its dual has finite-dimensional objective function and infinite-dimensional constraint. This opens room for possible alternative strategies, with an eye of their potential scalability to high-dimensional problems, which are subjected to some theoretical analysis, and also tried on practical examples.