Natural Gradient Algorithm Based on a Class of Activation Functions and its Applications in BSS

Blind source separation has become a dominant domain of artificial neural network. It attempts to recover unknown independent sources from a given set of observed mixtures. The natural gradient algorithm is a very important approach for blind source separation (BSS). The selection of activation function is the key to the algorithm. The aim of this paper is to investigate the blind source separation of a linear mixture of independent communication signals by using the natural gradient algorithm. We compare various activation functions for the algorithm and propose a better one. Simulation results not only demonstrate the algorithm can effectively separate the two kinds of random mixing signals, but also show that the algorithm with proposed activation function converges faster than other activation functions