PURPOSE: To shorten the setting time of a connection coefficient of a neural network.
CONSTITUTION: At least two neurons are included in a 1st intermediate layer other than the one that is used for correction of the threshold value, and the neurons used for correction of the threshold value and having the output always kept at 1 are included in the layers excluding an output layer. Then each input data is multiplied by a proper input coefficient so that the input total sum is calculated in common to each case. These input total sums are shown in the order of their value together with their corresponding teacher data. The outputs of neurons of the 1st intermediate layer which correspond to the maximum and minimum input total sums are designated with the changing directions of their large and small values set oriented and their numerical value also shifted from each other. The left side of a multi-variables simultaneous linear equations including a connection coefficient defined as an unknown number is constructed with the output of each neuron of the intermediate layer used as a coefficient. Meanwhile, the right side of the equation is defined as the input that is needed to acquire each output of an output layer from a sigmoid function. Then the multi-variables simultaneous linear equations is solved to find the connection coefficient. Thus, the connection coefficient can be set in a short period of time.
UCHIUMI TOMOAKI
ENDO FUMIMASA
YAMAGIWA TOKIO