PURPOSE: To attain the high speed of learning convergence by learning a learn ing pattern in which a teacher pattern is the subset of an input pattern as a first step, and in a second step, learning the learning pattern to be learnt originally by making the learnt result of the first step be an initial value.
CONSTITUTION: In the first step, an input pattern set A is inputted to the input of the neural network and the subset A0 of the input pattern set A is given as the teacher pattern, and learning is executed. In the second step. first of all, the subset A0 of the input pattern set A is given to the input of the neural network, and the subset B0 of a teacher pattern set B corresponding to the subset A0 of the input pattern set A is given as the teacher pattern. The learning part of the neural network learns a conversion rule from A to B on the basis of these input data. Subsequently, the learning is executed by giving successively the rest of the set A0 and the rest of the set B0. Since network structure after the learning of the first step is preserved and made to be the initial value of the learning of the second step, the number of times of the learning of the second step is reduced.