METHOD OF HIERARCHICAL LEARNING OF FEEDFORWARD NEURAL NETWORK

摘要

PURPOSE: A method of hierarchical learning of feedforward neural network is provided to reduce an error of output layer by defining new error function for an intermediate layer. CONSTITUTION: The method is comprising the steps of performing change of an output layer weighting value and setting a target value of neuron according to general hierarchical learning method, defining new error function for an intermediate layer, calculating a differential value for a weighting value of the intermediate layer in an error function defined, calculating the optimum learning rate of the weighting value of the intermediate layer, and changing the weighting value of the intermediate layer by using the differential value and the optimum learning rate. In the method, linear separating problems in the target value of the intermediate layer can be solved.