[Colloquium] Research at TTIC: Dahua Lin

[Liv Leader]

REMINDER:

When:     Friday, February 15th at Noon

Where:    TTIC, 6045 S Kenwood Avenue, 5th Floor, Room #526

Who:      Dahua Lin

Title:       New Efforts on Bayesian Inference

In this talk, I will present my latest efforts on developing new tools for
Bayesian inference, which can potentially be useful a variety of
application domains. This talk consists of two parts. In the first part, I
will present a general construction of Nonparametric Bayesian mixture
model, and a new algorithm to estimate the model. In the second part, I
will present a project that aims to provide an expressive language to
describe Bayesian graphical models and the related inference procedure.

Dirichlet process mixture models (DPMMs) have become an important tool to
describe complex data in the past decade. However, learning multiple DPMMs
over related sets of observations remains an open question. A popular
approach to this problem is the Hierarchical Dirichlet Process (HDP), which
is limited in that models are required to be organized as a tree. In this
paper, we present a generic framework to construct dependent DP mixtures,
drawing on a new formulation that we proposed recently. This framework
breaks the limitations of HDP, allowing each mixture model to inherit atoms
from multiple sources with covariate-dependent probabilities. We show,
through experiments on real data, that the proposed framework allows one to
devise models that capture the dependency between different data sets more
accurately.

Whereas graphical models have become widely used in practice, Implementing
and testing such models remain a time consuming and error prone process.
Researchers have to spend an enormous amount of time on implementation and
debugging instead of solving the original problems. Though codes for
certain algorithms are sometimes available from the paper authors, such
codes are often designed in a way that only works under a specific setting,
making it extremely difficult to generalize and extend the algorithms to
other application domain. In a recent project, I am working with MIT to
develop a new framework, which is based on a very young language called
Julia. With this framework, you may express a graphical model and the
inference procedure in an incredibly simple way -- while the run-time
performance is near the same as the codes tailored to a specific model.

******************************
*********
Research at TTIC Seminar Series

TTIC is hosting a weekly seminar series presenting the research currently
underway at the Institute. Every week a different TTIC faculty member will
present their research.  The lectures are intended both for students
seeking research topics and adviser, and for the general TTIC and
University of Chicago communities interested in hearing what their
colleagues are up to.

To receive announcements about the seminar series, please subscribe to the
mailing list: https://groups.google.com/a/ttic.edu/group/talks/subscribe

Speaker details can be found at: http://www.ttic.edu/tticseminar.php.

For additional questions, please contact Nati Srebro at nati at ttic.edu.


-- 
Liv Leader
Director of Human Resources and International Affairs

Toyota Technological Institute Chicago
6045 S Kenwood Ave
Chicago, IL 60637
Phone- (773) 702-5033
Fax-     (773) 834-9881
Email-  lleader at ttic.edu
Web-   www.ttic.edu
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20130214/f322efce/attachment.htm