A Construction of Recurrent Neural Network IIR Filter by means of BPTT Learning using M Sequence

Abstract

We developed a new method to construct a digital filter with recurrent neural networks by using the back propagation through time (BPTT) learning. We presented a systematic method to create teacher signals by multiplying the Maximum length linear shift register sequence (M sequence) and desired dynamics in the frequency domain and by transforming it to the time domain. Digital filters were constructed from fully cross-coupled and partially cross-coupled recurrent neural networks. Partially cross-coupled recurrent neural networks modeled after Direct Type I and II of IIR digital filter acquired appropriate properties through the learning. Our method is featured to be construct any complicated filter with desired dynamics from a recurrent neural network.