20 Aug 10:15 Bin Yu: Lasso -- algorithm, theory and extension

HIIT seminar, Thursday Aug 20, 10:15 a.m. (coffee from 10), Exactum C222

Prof. Bin Yu
University of California, Berkeley
Department of Statistics &
Department of Electrical Engineering and Computer Science

Lasso -- algorithm, theory, and extension

Abstract:
Machine learning has been at the frontier of modern statistics because of its serious consideration of computation. Machine learning algorithms such as boosting and support vector machines have shown impressive successes in prediction for large data sets. Another important goal of statistics, in addition to prediction, is interpretation. Now much attention is paid on
L1 penalized empirical minimizations (Lasso or compressed sensing) because of the sparsity in the models induced by the L1 penalty and sparisty is a good proxy for interpretation.

In this talk, I would like to give an overview of recent results by my group on research related to L1 penalized minimization. In particular, an approximate Lasso algorithm, BLasso, is proposed and related to epsilon-L2Boosting; an irrepresentable condition is introduced for Lasso (and L1 penalized Gaussian graphical likelihood) to be model selection consistent and the consequence of this condition's relaxation is studied; finally, I will briefly cover a new penalization framework, CAP, for grouped and hierachical selection of predictors, and its fast implementation.

Bio:
Bin Yu is Chancellor's Professor in the departments of Statistics and of Electrical Engineering \& Computer Science at UC Berkeley.
She is currently the chair of department of Statistics, and a founding co-director of the Microsoft Lab on Statistics and Information Technology at Peking University, China.

She got her B.S. in mathematics from Peking University in 1984 and her M.A. and Ph.D. in statistics from UC Berkeley in 1987 and 1990, respectively.  She held faculty positions at University of Wisconsin-Madison and Yale, visiting faculty positions at ETH, Poincare Institute and Columbia Univ, and was a Memeber of Technical Staff in the math center at Lucent-Bell Labs.
Jointly with others, she holds two U.S. patents on information technology. Her current research interests include statistical machine learning for high dimensional data, information theory, and data problems from remote sensing, neuroscience, sensor networks, and finance.

She was a 2006 Guggenheim Fellow, and is a Fellow of the American Association for the Advancement of Science, IEEE, IMS (Institute of Mathematical Statistics) and ASA (American Statistical Association).

Homepage: www.stat.berkeley.edu/~binyu