Power of active and compressive queries for subspace learning

Jun 12

Thursday, June 12, 2014

11:30 am - 12:30 pm
Gross Hall 318

Presenter

Aarti Singh, Machine Learning Department, Caregie Mellon University

lunch and talk The ability to learn large-scale matrices from few observed entries is important in myriad applications including imputing latencies between hosts in a communication network, expression levels of genes under various drugs, and user ratings for movies. This goal is feasible since the data generating system typically has limited degrees of freedom, and by leveraging the underlying subspace characterizing the data, we can hope to minimize the number of entries needed. Most recent work on this problem focuses on random queries. However, in many of these applications, one can employ active queries that use judicious feedback-driven choice of which entries to observe to minimize network traffic, experimental cost or user effort needed. Compressive queries offer another alternative, where only linear sketches of the data need to be observed and stored. I will present recent work from my group that demonstrates how active and compressive queries can enable improvements in sample, computational, memory and communication efficiency for subspace learning, as well as the ability to handle coherent rows or columns that might arise due to anomalous hosts, genes or users. B.E. in Electronics and Communication Engineering from the Univ of Delhi in 2001, and M.S. and Ph.D. degrees in Electrical and Computer Engineering from the UW-Madison in 2003 and 2008, respectively. I was a Postdoctoral Research Associate at the Program in Applied and Computation Math at Princeton.

Contact

Peterson, Kathy
613-7829
kathy.peterson@duke.edu