[Colloquium] REMINDER: Talks at TTIC: Srinadh Bhojanapalli, Univ of Texas at Austin

[Dawn Ellis]

When:     Thursday, May 14th at 11am

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

Who:       Srinadh Bhojanapalli, Univ of Texas at Austin

Title:       Sampling Matrices and Tensors

Abstract:

In this talk we first address a basic question: given a general rank-r
matrix, with no additional structure like incoherence, can one randomly
sample the entries so that it can be provably reconstructed only from the
samples ? Clearly, the sampling should be biased in a way that depends on
the matrix; we characterize precisely - via upper and lower bounds on the
sample complexity - what this dependence should be. We also show that from
samples with this dependence, one can provably recover the matrix via a
simple but non-convex weighted alternating least squares algorithm.

This allows us to then develop new methods for:
(a) approximate PCA
(b) directly computing low-rank approximations of matrix products
(c) approximate tensor factorization

Computationally, these are well-suited to out of core architectures for
large problem instances; statistically, they provide significantly improved
approximation over existing methods.

Host:  Nati Srebro, nati at ttic.edu


-- 
*Dawn Ellis*
Administrative Coordinator,
Bookkeeper
773-834-1757
dellis at ttic.edu

TTIC
6045 S. Kenwood Ave.
Chicago, IL. 60637
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20150513/cc097df2/attachment.htm