Improving dictionary learning using the Itakura-Saito divergence

Abstract

This paper presents an improved and efficient algorithm for overcomplete, nonnegative dictionary learning for nonnegative sparse representation (NNSR) of signals. We adopt the Itakura-Saito (IS) divergence as the error measure, which is quite different from the conventional dictionary learning methods using the Euclidean (EUC) distance as the error measure. In addition, for enforcing the sparseness of coefficient matrix, we impose ℓ1-norm minimization as the sparsity constraint. Numerical experiments on recovery of a dictionary show that the proposed dictionary learning algorithm performs better than other currently available algorithms which use Euclidean distance as the error measure.