Reduced-conflict learning for similar pattern recognition using backpropagation neural networks

Summary form only given, as follows. A problem of similar pattern recognition using backpropagation neural networks (BPNNs) was investigated. It was shown that a conflict emerges when similar patterns are input into a BPNN, and it was trained in an all-or-nothing fashion. Secondly, three kinds of learning techniques for reducing the conflict were proposed: similarity learning (SML), similarity relearning (SRL), and conflict-free learning (CFL). The effectiveness of SML, SRL, and CFL were confirmed by applying them to handwritten-digit recognition.<<ETX>>