NAGSC: NESTEROV’S ACCELERATED GRADIENT METHODS FOR SPARSE CODING

sparse coding

Abstract

This paper proposes efficient algorithms for Sparse Coding.  Firstly, Sparse Coding is divided into two sub-convex problems including L1 and L2 prob-lems.  Secondly, we transform the nonsmooth L1 problem into two smooth sub-problems, and alternatively optimize them by Nesterov’s Accelerated Gradient methods (NAG).  Thirdly, we apply NAG to optimize L2 problem.  Finally, L1 and L2 problems are iteratively solved until convergence.  Experiments show that our proposed algorithms are effective to optimize L1, L2 and learn over-complete bases.