PURPOSE: To speed up learning by considering the average and dispersion of an input pattern and the average and dispersion of the weight between the membrane potential of a neural element and the neural element and supplying the initial value of the weight between neural elements after clarifying their relation.
CONSTITUTION: All example input patterns are read, and the average value and dispersion are calculated based on this pattern. In this case, the value to be the dispersion of the weight of the coupling of the first layer is calculated as well as the value to be the dispersion of the weight of the coupling of the 0th layer. Further, the random number is generated using the dispersion of the first layer to decide the weight of the coupling of the first layer. Similarly, the random number is generated using the dispersion of the 0th layer to decide the weight of the coupling of the 0th layer. Thus, the dispersion of the weight of the coupling is made large when the dispersion of the input pattern is small and the dispersion of the weight of the coupling is made small when the dispersion of the input pattern is large to perform the learning of the multilayered neural circuit network at a high speed.