Statistical research and multilayered neural networks

Discusses two possible trends in research into three-layered neural networks synthesized from the paradigm of variable-structure neural networks and used for pattern recognition. These trends are: (i) minimizing the dimension of the space of attributes, and (ii) evaluating the permissible spread in weighting coefficients in order to determine the class of accuracy of electrical parameters of the circuitry used to simulate a neural network. The authors describe the algorithm and technique used to minimize the space of attributes and evaluate the class of accuracy. The proposed algorithms were used in experiments conducted on five neural networks obtained for the solution of practical problems