A neural network with multiple large‐scale subnetworks and its application to recognition of handwritten characters
Yoshihiro Hagihara, Hidefumi Kobatake · Systems and Computers in Japan · 2003
Abstract We propose a large‐scale network system and learning method that prevent increases in the amount of calculation, increases in the necessary memory capacity, and decreases of accuracy. It is shown that the system is effective for a task with a huge number of categories. The proposed system consists of multiple large‐scale neural networks (LSNNs) and each LSNN consists of multiple small‐scale neural networks (SSNNs). Each SSNN supports only part of the categories in recognition and learning, and each LSNN supports categories that overlap with other LSNNs. The combination of categories that each SSNN in each LSNN supports is different from that of the other LSNNs, because it is determined randomly. The output values of the LSNNs for each category are averaged and the maximum gives the final decision. Since the learning of the SSNNs and LSNNs is performed independently, realization of the proposed system is easy. The effectiveness of this system is shown in the recognition of handwritten characters using a multilayer perceptron and backpropagation. © 2003 Wiley Periodicals, Inc. Syst Comp Jpn, 34(6): 91–99, 2003; Published online in Wiley InterScience ( www.interscience.wiley.com ). DOI 10.1002/scj.1208