[Ml-stat-talks] Mathematics Colloquium: Sourav Chatterjee, Feb. 20 @ 4:30pm

Philippe Rigollet rigollet at Princeton.EDU
Fri Feb 15 15:39:08 EST 2013


This talk by Sourav Chatterjee on problems related to matrix completion looks very interesting.

Department Colloquium
Topic:  Hidden structures in large matrices
Presenter:  Sourav Chatterjee, New York University
Date:  February 20, 2013, Time:  4:30 p.m., Location:  Fine Hall 314
Abstract:  Consider the problem of estimating the entries of a large matrix, when most of the entries are either hidden from us or blurred by noise. Of course, one needs to assume that the matrix has some structure for this estimation to be possible. This problem has received widespread attention in statistics, computer science and information theory in recent times, especially after the pioneering works of Emmanuel Candes and collaborators. I will introduce a simple estimation procedure, called Universal Singular Value Thresholding (USVT), that automatically detects the inherent structure of the matrix whenever any structure is present, and uses this knowledge to estimate the unknown entries. Surprisingly, this  simple estimator is "minimax optimal" up to a constant factor. The method gives easy solutions to a wide range of difficult questions about large data matrices.






-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.cs.princeton.edu/pipermail/ml-stat-talks/attachments/20130215/c6f77ff7/attachment.htm>


More information about the Ml-stat-talks mailing list