Method for training neural network

PURPOSE: To improve accuracy for recognizing characters or symbols by impressing a specified training sequence to a neural network. CONSTITUTION: Concerning a method for training the neural network for recognizing the symbol expressed by an input array P (K), the intermediate layer of the neural network is provided and when the input array P (K) expresses the symbol to be recognized by the neural network, a weight function W1 (K) related with an intermediate node is initialized to a value having extremely deep correlation with the input array P (K). Then, various steps are provided for impressing the exercise (training) sequence to the neural network. This weight function W1 (K) is defined by an expression. Thus, accuracy for recognizing characters or symbols can be improved.