Rylan Schaeffer

Logo About Me
Resume
Research
Learning
Blog
Talks
Teaching
Jokes
Kernel Papers


15 April 2021

A sequential algorithm for fast fitting of Dirichlet process mixture models

by Nott, Zhang, Yau and Jasra (Journal of Computational and Graphical Statistics 2014)

Background

Previously, Wang and Dunson (2011) proposed a fast inference algorithm for Dirichlet process mixture models that could be applied in an online setting, fancifully called Sequential Updating and Greedy Search (SUGS). The authors actually didn’t intend for their algorithm to function online, recommending that to remove the dependency on observation order, one should average over random orderings of the observations. The way SUGS made inference easy was by taking the MAP of the latent cluster’s posterior, deterministically placing the observation at a single cluster.

Overview

Conceptual Notes

tags: mixture-models - dirichlet-process - bayesian-nonparametrics - clustering