Recognition and restoration of periodic patterns with recurrent neural network
Ryotaro Kamimura · 2002
Using the fully recurrent network with the temporal supervised learning algorithm developed by Williams and Zipser, the author performed several experiments aimed at recognizing and restoring periodic patterns. The results can be summarized as follows: the recurrent network could recognize complex and multiple patterns simultaneously, if appropriate number of hidden units are given; the network could regenerate infinitely the patterns with appropriate precision; the recurrent network could recognize and generate infinitely the patterns in spite of the existence of a large number of noises in the course of learning. >