Optimization of a cognitron type neural network

Optimization studies on a recognition neural network based on K. Fukushima´s cognitron (1975) are presented. The goal was to increase the selectivity and robustness of the network, which was used as the final stage identifier in an integrated vision network invariant to translation, rotation, and scaling. Unlike the original cognitron, different inhibitory parameters were introduced for differential layers so that selectivity of excitatory cells of different layers could be adjusted in a flexible manner. A supervised learning scheme was adopted in the last layer so that different learning samples could be related to the output elements in a desired order. Choosing relatively large values of the inhibitory parameter for the input layer and supervised learning parameter for the output layer improved the performance of the recognition system. The network used 64*64 binary M-transformed images as its input patterns. Computer simulation indicated that by adjusting the structure parameters of the network a tradeoff between selectivity and robustness could be made.