Further Parameters Estimation of Neocognitron Neural Network Modification with FFT Convolution

This paper presents further development of an improved version of the neocognitron algorithm introduced by Fukushima. Some comparisons of other symbol recognition methods based on the neocognitron neural network are also performed, which led to the proposal of several modifications — namely, layer dimension adjustment, threshold function and connection Gaussian kernel estimation. The width and height are taken into account independently in order to improve the recognition of patterns of slightly different dimensions. The learning and recognition calculations are performed as FFT convolutions in order to utilize external specialized computing system. Finally, more detailed results of the neocognitron performance evaluation are provided