A mimetic pattern generator for multiple periodic signals using recurrent neural networks

A living body has a mechanism to acquire various periodic signals in the environment. In our nerve system, this is realized by using of simple neural oscillators. The oscillators optimize their internal parameters to synchronize with the external signal only from error between their own output and external signal. Furthermore, a single neural oscillator is considered to generate multiple complex signals. In this study, a simple neural oscillator model which can acquire external periodic signals is described. The proposed model is based on the multi-layered recurrent neural network which has time delayed feedback connections. A task to acquire two waveforms of Japanese vowels was applied to examine the capability of this model. According to results of computer simulations, it is capable to acquire these wave forms successfully.