The growth of the corrective ability of neural network structures with redundancy due to the replacement of binary neurons in them with ternary neurons

Abstract

Background. Obtaining a numerical estimate of the transition effect from ordinary binary neurons with two output states “0”, “1” to neurons with three output states “– 1”, “0”, “1”. Materials and methods. As an example, we consider a neural network that generalizes three classical statistical criteria for testing the hypothesis of independence of samples of 100 experiments. The Pearson-Edgeworth-Edleton test (1890–1900), the Kenuya test (1965) and the modified Nelson test (1983) were used. For these criteria, artificial neurons equivalent to them with binary and ternary quantizers have been constructed. As a result, we get a binary and ternary output code with a threefold code redundancy. The folding of these codes makes it possible to correct the errors present in them. Results. The ternary self-correcting output code of the neural network in terms of its corrective ability turned out to be one and a half times more powerful in comparison with its binary counterpart. The latter is explained by the increase in the amount of information available for analysis and the greater information content of data on error syndromes. Conclusions. It has been suggested that the effect of increasing the growth of the corrective ability of ternary neurons compared to binary neurons will increase as the number of artificial neurons that combine the currently known statistical criteria for joint use increases.

Keywords

• artificial neurons
• binary neurons
• ternary neurons
• self-correcting code