Yue Lu (Harvard)
Nov 4, 2016.
Title and Abstract
High-Dimensional Analysis of Stochastic Optimization
Algorithms: Limiting Dynamics and Phase Transitions
In this talk, we will present a framework for analyzing the exact dynamics of these methods in the high-dimensional limit. For concreteness, we consider two prototypical problems: regularized linear regression (e.g. LASSO) and sparse principal component analysis. For each case, we show that the time-varying estimates given by the algorithms will converge weakly to a deterministic “limiting process” in the high-dimensional (scaling and mean-field) limit. Moreover, this limiting process can be characterized as the unique solution of a nonlinear PDE, and it provides exact information regarding the asymptotic performance of the algorithms. For example, performance metrics such as the MSE, the cosine similarity and the misclassification rate in sparse support recovery can all be obtained by examining the deterministic limiting process. A steady-state analysis of the nonlinear PDE also reveals interesting phase transition phenomenons related to the performance of the algorithms. Although our analysis is asymptotic in nature, numerical simulations show that the theoretical predictions are accurate for moderate signal dimensions. What makes our analysis tractable is the notion of exchangeability, a fundamental property of symmetry that is inherent in many of the optimization problems encountered in signal processing and machine learning.
Yue M. Lu was born in Shanghai. After finishing undergraduate studies at Shanghai Jiao Tong University, he attended the University of Illinois at Urbana-Champaign, where he received the M.Sc. degree in mathematics and the Ph.D. degree in electrical engineering, both in 2007. He was a Research Assistant at the University of Illinois at Urbana-Champaign, and has worked for Microsoft Research Asia, Beijing, and Siemens Corporate Research, Princeton, NJ. Following his work as a postdoctoral researcher at the Audiovisual Communications Laboratory at Ecole Polytechnique Fédérale de Lausanne (EPFL), Switzerland, he joined Harvard University in 2010, where he is currently an Associate Professor of Electrical Engineering at the John A. Paulson School of Engineering and Applied Sciences.
He received the Most Innovative Paper Award (with Minh N. Do) of IEEE International Conference on Image Processing (ICIP) in 2006, the Best Student Paper Award of IEEE ICIP in 2007, and the Best Student Presentation Award at the 31st SIAM SEAS Conference in 2007. Student papers supervised and coauthored by him won the Best Student Paper Award (with Ivan Dokmanic and Martin Vetterli) of IEEE International Conference on Acoustics, Speech and Signal Processing in 2011 and the Best Student Paper Award (with Ameya Agaskar and Chuang Wang) of IEEE Global Conference on Signal and Information Processing (GlobalSIP) in 2014. He has been an Associate Editor of the IEEE Transactions on Image Processing since 2014, an Elected Member of the IEEE Image, Video, and Multidimensional Signal Processing Technical Committee since 2015, and an Elected Member of the IEEE Signal Processing Theory and Methods Technical Committee since 2016. He received the ECE Illinois Young Alumni Achievement Award in 2015