Neural network training method and neural network classifying device

PURPOSE: To perform a learning process not based on the constitution of point- sets which are centers of gravity, but by searching the boundary between the point-sets, namely, the areas where the input vectors of the sets do not exist. CONSTITUTION: Test vectors are supplied to a neural network in which arithmetic parameters (synapse coefficients W1 -WN and a neuron threshold S) have random values in the initial stages. The test vectors are classified in a probabilistic classifying procedure in which the output actuation of the network is mutually related and, at the same time, a weight factor which adjusts the probability at the time of randomly selecting a class is generated at every test vector. When the test vectors are classified in such a way, the arithmetic parameter of the network is modified so as to emphasize the difference between output actuating patterns to all test vectors.