Speaker: Abhishek Kumar (IBM TJ Watson Research Center)

Title: Topics in Nonnegative Matrix Factorization: Separable NMF and 

Date and time: Jan 4, 2016 (3:00-4:30pm)
Venue: RM Building, Room 101

Abstract: The goal in nonnegative matrix factorization (NMF) is to 
express, exactly or approximately, a given matrix as a product of two 
nonnegative matrices of smaller inner dimension. NMF and its variants 
have been widely used for extracting interpretable features and patterns 
in various applications, including text, vision, and speech. Computing 
NMF has been shown to be NP-hard. In the first part, I will talk about 
fast conical hull algorithms for NMF under the so-called separability 
assumption which makes the NMF problem tractable. I will show 
applications of separable NMF to the problem of Video 
foreground-background separation, comparing it with Robust PCA which is 
a widely used method for this problem. In the second part, I will talk 
about a related problem of Semi-nonnegative matrix factorization where 
only one of the factors is constrained to be nonnegative. I will talk 
about conditions for tractability, and exact and heuristic algorithms 
for computing Semi-NMF. This talk is based on joint works with Vikas 
Sindhwani and Nicolas Gillis.

Speaker bio: Abhishek Kumar is a Research Staff Member at IBM TJ Watson 
Research Center with research interests broadly lying in the area of 
machine learning, with work ranging from learning from multiple views of 
data, multitask learning, and efficient non-negative matrix 
factorization algorithms 
(http://www.umiacs.umd.edu/~abhishek/index.html). Before joining IBM, he 
graduated with a Ph.D. from Department of Computer Science at the 
University of Maryland in 2013.