Designing and Learning Algorithm of Neural Networks for Pattern Recognition

Abstract

There are many studies on designing neural networks. These researches are classified into two kinds of approaches. One is direct encoding method and the other is grammatical encoding method. The direct encoding method has some restrictions of neural network structures because the network connectivities are directly encoded into a matrix. The grammatical encoding method is more flexible than the direct method. The method enables the generation of any kind of neural network structures, but the connections have only three kinds of connection weights. Therefore, it is difficult to generate networks for complex pattern recognition. In the case of pattern recognition using neural networks, it is very difficult for researchers or users to design them. This chapter discusses a method of learning and designing feedforward neural networks. The chapter describes the designing and training neural networks which perform handwritten KATAKANA recognitions. The chapter also discusses the efficiency of the proposed method.