METHOD FOR TRAINING TOP-DOWN SELECTIVE ATTENTION IN ARTIFICIAL NEURAL NETWORKS

摘要

The present invention is an engineering implementation of the ability of selective attention occurring in the human brain, and applying it to a recognizer to arbitrarily increase precision, a plurality of synapses corresponding to the strength of each synaptic connection acquired through learning of an artificial neural network based on a multilayer perceptron network. The weight of is fixed based on a preset weight value, and a training pattern is presented to an input layer composed of a plurality of neurons to perform an operation in an artificial neural network corresponding to the training pattern, and the input vector corresponding to the data for each domain After calculating the output, by comparing the recognition rate of policy-based data for each domain through the calculated output, a weight is assigned to the output of the data with the highest recognition rate, and a top-down selective attention-based training for each synapse is performed. By defining the degree of attention to the candidate class as a new recognition scale, it is possible to output superior recognition results compared to the existing recognition system for one candidate class, and to perform calculations without lowering the maximum speed that can be implemented through weighted summation. The precision can be arbitrarily increased, and this top-down selective attention multilayer perceptron models a biologically selective attention mechanism, and at the same time, it is possible to implement a large-capacity general-purpose neural network computer and integrate it into a small semiconductor. Thus, it is intended to provide technologies applicable to various artificial neural network applications.