Blind source separation by sensor-signal identity mapping by auto-encoder with hidden-layer pruning

A new non-information-theoretic approach is described for the blind source separation (BSS) problem. It is based on an auto-encoder neural network which incorporates a pruning algorithm. Hidden units are nonlinear, and ones that survive the pruning become the source extractors. As such, no assumption is needed for the number of sources. Simulation results show that the auto-encoder can make BSS for a broad class of source-signal mixtures without changing the nonlinear activation function of the hidden units.