A method for adjusting network parameters in a multi-layer perceptron device, and perceptron device provided with means for executing the method

摘要

The invention provides a perceptron device with an improved learning behaviour. The training is effected by alternating forward propagating steps wherein an input vector is presented and the processing result compared to an associated target vector; and backward propagating steps wherein the comparison difference, by means of a learning rate is used for updating th network parameters. Improvement is attained by two stratagems.a) the learning rate is eta_i = eta_o x M/KN, wherein N is the number of inputs to the processing element fed by the parameter value to be updated, K is the number of outputs from that processing element and M is the number of inputs to a processing element of the next layer;b) the learning is done in three steps, first forward propagation, then backward propagation, then again forward propagation. The improvement attained by the updating is compared to a discrimination level. If the improvement is bigger, the learning rate is decreased, if smaller, the learning rate is increased.