Learning-based accelerated sparse signal recovery algorithms

Abstract

In this paper, we propose an accelerated sparse recovery algorithm based on inexact alternating direction of multipliers. We formulate a sparse recovery problem with a concave regularizer and solve it with the relaxed and accelerated alternating method of multipliers (R-A-ADMM). We introduce learnable parameters to optimize the algorithm with given data sets. The derived algorithm is an accelerated version of LISTA-AT that controls the threshold for each entry according to the previously recovered estimate. Numerical results show that the proposed Accel-LISTA-AT algorithm converges much faster and recovers the sparse signals with lower mean squared errors than the other learning-based sparse recovery algorithms.

Keywords

• Iterative soft thresholding
• Alternating direction method of multipliers
• Deep neural network