[Ml-stat-talks] Wed: Ryan Adams on Bayesian non-parametrics

David Mimno mimno at CS.Princeton.EDU
Fri Mar 30 16:21:46 EDT 2012


For this week's Machine Learning lunch we have Ryan Adams from
Harvard. Ryan will be around all day -- please let me know if you
would like to meet with him.

Ryan Adams (Harvard, Computer Science)
Wed April 4, 12:30, CS402

Title: Tree-Structured Stick Breaking Processes for Hierarchical Data

Abstract:
The Dirichlet process and related distributions over
infinite-dimensional random measures have had a large impact on
Bayesian nonparametric modeling, but they can be limited by the lack
of structure they place over their random partitions. In this work, we
use a recursive stick breaking process to construct a Bayesian
nonparametric prior for inferring hierarchies in data, yielding an
exchangeable urn scheme on trees of unbounded width and depth. We use
a novel slice sampling approach based on an auxiliary variable model
to update the structure of the tree, as well as Gibbs moves that take
advantage of invariance under size-biased permutation. We use this
MCMC sampling scheme to perform hierarchical clustering on a set of
50,000 color images, and also to discover structure from natural
language in a corpus of scientific documents.

This is joint work with Zoubin Ghahramani (Cambridge) and Michael
Jordan (Berkeley).


More information about the Ml-stat-talks mailing list