Copy-Learning-Based Elman Network for Recognizing Overlapped Character Strings

Abstract

This study proposes a copy-learning model Elman network (CE-net) which can recognize overlapped character strings by only learning the standard character strings. CE-net learns standard character strings by a part of the 3-layer net, copies obtained weights and biases on an unused part of the net by a copy-learning rule, and recognizes overlapped character strings. We had confirmed the effectiveness of CE-net from several experiments.