[Ml-stat-talks] Kurt Miller, Monday 11/8

David Blei blei at CS.Princeton.EDU
Thu Nov 4 11:32:26 EDT 2010


hi ml-stat-talks and ml-reading

kurt miller (uc berkeley, go bears) will be giving a talk on monday
11/8 at 3PM in computer science room 402.  his work will be
interesting to enthusiasts of graphical models, bayesian nonparametric
methods and matrix factorization.

[reading group: note the change in room.]

best,
dave

---

Kurt Miller, U.C. Berkeley
Non-Exchangeable Bayesian Nonparametric Latent Feature Models
Monday, November 8, 3:00PM
Computer Science Room 402

Bayesian nonparametric methods are based on prior distributions
expressed as general stochastic processes.  Due to the need to
integrate over these priors at inference time, strong constraints such
as exchangeability are often placed on the kinds of models that can be
considered.  Over the past five years, two infinitely exchangeable
nonparametric priors for latent feature models have been
introduced--the Indian Buffet Process for which the De Finetti mixing
distribution is Hjort's beta process, and the Infinite Gamma-Poisson
Feature Model for which the De Finetti mixing distribution is the
gamma process.  These are priors over infinite binary matrices and
infinite non-negative integer valued matrices, respectively, that
allow us to perform nonparametric latent feature inference.

Following on these developments, we aim to extend the range of
Bayesian nonparametric latent feature modeling by presenting two
non-exchangeable generalizations for each of these two models in which
efficient posterior inference is possible.  Rather than having an
exchangeable model in which the features for each of the objects are
conditionally independent given the De Finetti mixing distribution, we
draw the features for all objects jointly using a stochastic process
that allows us to utilize prior knowledge about objects to infer
better features.  Our models are applicable to the general settings in
which the dependencies between objects can be captured using either a
rooted tree expressing how closely related objects are or a chain
which expresses a linear relationship amongst the objects.

This is joint work with Michael I. Jordan, and Thomas L. Griffiths.


More information about the Ml-stat-talks mailing list