Blind Source Separation with L1 Regularized Sparse Autoencoder

Abstract

Blind source separation of co-channel communication signals can be performed by structuring the problem with an over-complete dictionary of the channel and solving for the sparse coefficients, which represent the latent transmitted signals. $L_{1}$ regularized least squares is a common approach to imposing sparsity on the latent signal representation while minimizing the reconstruction error. In this paper we propose an unsupervised learning approach for blind source separation using an $L_{1}$ regularized sparse autoencoder with a softthreshold activation function at the hidden layer that is able to separate and fully recover multiple overlapping binary phase shift keying co-channel signals.