Neural network apparatus and learning method thereof

A neural network apparatus comprising:

• a neural network including at least two neuron layers each having a plurality of neurons and at least one synapse layer having a plurality of synapses each arranged between said neuron layers, each synapse storing a weight value between the neurons and multiplying the weight value with an output value from each of the neurons in a previous-stage neuron layer to output a product to a next-stage neuron layer;
• means for causing an error signal between an output from said neural network and a desired output to back-propagate from an output-side neuron layer to an input-side neuron layer of said neural network;
• learning control means for updating the weight value in said synapse on the basis of the error signal and the output value from said previous-stage neuron;
• and
• input means for inputting a control signal for selecting a synapse whose weight value is to be updated by said learning control means when said learning control means is to update the weight values of a predetermined number of the plurality of synapses in a predetermined order, wherein the predetermined number of the plurality of synapses updated per learning cycle is limited to n to reduce a magnitude of an updating amount weight vector, .DELTA.W, given by:
• ##EQU6## where .DELTA.W.sub.i is non-zero weight updating amount by which the weights of the plurality of synapses can be updated and n is less than a total number of the plurality of synapses.

A learning method of a neural network apparatus, for updating weight values of a plurality of synapses on the basis of an error signal proportional to a difference between a desired output propagating from an output-layer neuron to an input-layer neuron and an actual output of a network so as to set an output from said neural network close to the desired output, the method comprising:

• selecting, based on repeated learning, a subset of the plurality of synapses whose weight values are to be updated in a learning cycle, wherein the number of synapses in the subset is limited to n to reduce a magnitude of an updating amount weight vector, .DELTA.W, given by:
• ##EQU7## where .DELTA.W.sub.i is non-zero weight updating amount by which the weights of the plurality of synapses can be updated and n is less than a total number of the plurality of synapses;
• and
• updating in the learning cycle only the weight values of the subset of the plurality of synapses selected in the selecting step.