R202, Astronomy-Mathematics Building, NTU
(台灣大學天文數學館 202室)
A Dynamic Approach to Sparse Recovery
Yuan Yao (Peking University)
In this talk we aim to solve an open problem raised by Jianqing Fan et al. in 2001, where convex l1-regularization (LASSO) causes bias in linear regression and nonconvex regularization is thus introduced for debias which however suffers from NP-hardness in finding global optimizers. We show a novel approach utilizing a technique from dynamics. Instead of optimizing a potential (objective) function, we evolve ODEs formed by gradient descent in dual space of $l_1$-norm. Equipped with early stopping regularization, it simultaneously achieves variable selection consistency and unbiased estimator, which is thus better than LASSO estimator. The dynamics leads to a simple discretization as linearized Bregman iteration algorithm, which has been widely used in image processing, matrix completion, as well as robust ranking, etc.