Natural Gradient Learning Algorithms for Decorrelation

摘要

In this paper, we derive a new adaptive learning algorithm for spatio-temporal decorrelation of signals using a dynamic re urrent neural network and natural gradient approach. First, we derivje rigorously a on-line learning algorithm for blind source separation (separation of instandtaneous mixture), from an information-theoretic approach, by using a fully-connected memoryless recurrent neural network. Next. we extend this concept for a dynamic recurrent neural network with synaptic weights of the form FIR (finite impulse response) filter. Comparison was made with temporal anti-Hebbian learning, Extensive computer simulation results confirmed the validity and high performance of the developed algorithm.