1992 年 16 巻 11 号 p. 43-48
In this paper we describes the inevitable problems in enlarging the scale of neural networks for character recognition and approaches to them. At first we propose an initial setting rule of network coefficients, a method of avoiding inactive states in the initial learning period and a reduction algriosm of hidden units. And then we simulate neural network operations for recognizing drawn Chinese characters and alphabets by using above techniques. The simulation results indicate expected performances.