New neural networks for determining proximity and distance functions when comparing binary objects

Abstract

Artificial neural networks are effectively used to solve various problems (recognition, clustering, classification, etc.) in conditions where information about objects is given by vectors with binary components. The Hamming neural network is an effective tool for solving these problems when the initial information is given in the form of bipolar vectors. However, when comparing binary objects with qualitative features, more than 70 different distances (proximity measures) and proximity functions of binary objects are used. Synthesizing dozens of neural networks for matching binary objects is too laborious. Therefore, a universal approach to the synthesis of such neural networks is proposed. In addition, the Hamming network can select only one object from its memory that is closest to the input. It cannot function normally if there are two or more such objects. The proposed neural network does not have this shortcoming. And last but not least, a neural network with this architecture allows you to calculate various distances and similarity measures of the input data and data stored in the network memory