[Lisa_teatalk] Talk by Sungjin Ahn on 1 Apr at 14.00

Kyung Hyun Cho cho.k.hyun at gmail.com
Sat Mar 28 14:52:56 EDT 2015


Dear all,

Sungjin Ahn of UC Irvine (http://www.ics.uci.edu/~sungjia/) will visit us
this Wednesday (1 Apr). He is an expert in applying Bayesian approaches to
large-scale machine learning problems (stochastic MCMC). He will tell us
about "Scalable MCMC in the era of Big Data" at 14.00 in the usual place
(AA3195).

If you're interested in talking to Sungjin, please, email me and Sungjin.

Hope to see many of you on Wednesday!
- Cho

===
Speaker: Sungjin Ahn, UC Irvine
Date/Time: 1 April @14.00
Place: AA3195

Title: Scalable MCMC in the era of Big Data

Abstract:
Scaling up Markov chain Monte Carlo (MCMC) methods for the web-scale data
is one of the most important issues in probabilistic inference in this era.
Whereas the most efficient traditional MCMC algorithms, e.g. HMC, have
failed to adapt to this scale since it requires to evaluate all data points
per iteration, the stochastic gradient methods in optimization have
successfully shown its efficiency. In this talk, I introduce a family of
MCMC algorithms based the Stochastic Gradient Langevin Dynamics, and its
extension for distributed computation. The algorithms bring down the
computation of MCMC to the same level as the stochastic gradient
optimization by making it possible for the transition kernel to be based
only on mini-batches and Gaussian noises. Also, several successful results
in application to the most interesting large-scale problems such as topic
modeling, matrix factorization, and neural networks, are presented.

Bio:
Sungjin Ahn is a Ph.D. candidate in the department of Computer Science at
the University of California, Irvine. Prior to UC Irvine, he received his
M.Sc. at the Korea Advanced Institute of Science and Technology (KAIST) and
then worked at the Agency for Defense Development as a research scientist.
During his PhD, he has worked on scalable Bayesian inference methods, in
particular focusing on the stochastic gradient MCMC methods. Along with
Bayesian inference methods for machine learning, his research interest also
includes developing efficient computational models and learning procedures
for vision and natural language information.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://webmail.iro.umontreal.ca/pipermail/lisa_teatalk/attachments/20150328/e470a792/attachment.html 


More information about the Lisa_teatalk mailing list