Aadaptive signal de-noising based on feedback networks and counterpropagation network

摘要

The main purpose of this paper is to realize adaptive signal denoising simulation of some kind of feedback neural network models. The bidirectional associative memory (BAM) neural network, the discrete Hopfield feedback network (DHN), and the counterpropagation network (CPN) are discussed under the conditions of outside and within the maximal memory capacity. The experimental simulations of the three kind of networks are realized to data de-noise, the experimental results are compared and analyzed, show that both BAM network and discrete Hopfield network within the maximal memory capacity have all good de-noise effect, fewer iterations, less training time, and operation stability. The CPN is sensitive to initial weight values, good de-noising effect, but more iterations. When noise is increased and outside the maximal memory capacity of BAM network or DHN, we find that the CPN is of better de-noise performance than discrete Hopfield networks and Kosko's BAM net under the condition of overstepping the maximal memory capacity. Full CPN is of better de-noise performance than one-way CPN, but the former takes a longer training time.