A learning method for a data processing system comprising an input layer and output layer which comprises a plurality of neurons each of which outputs an predetermined data after igniting according to the predetermined processing results performed onto an input data, and an middle layer, arranged between the input and output layer, the middle layer comprising a plurality of neurons each of which is connected to each neuron of the input and output layer; characterized in the following steps: the ignition patterns of the input layer and the output layer are determined artificially according to a plurality of inputs and outputs; weights of synapses of the middle layer and the output layer are increased so as to obtain the tendency that the ignition pattern of middle layer becomes the nearest approximation to the ignition patterns of input layer and output layer according to each input and output; the same processings as above are performed with respect to all inputs and outputs.