Neural Networks with Configurable Affinity Functions for Binary Object Classification

Abstract

The paper addresses computer systems for the recognition of discrete objects encoded by binary and bipolar feature vectors. It is shown that the well-known Hamming neural network, which uses the scalar product of bipolar vectors and the Hamming distance for recognition, does not allow the application of more refined similarity functions when object features are binary-encoded. Modifications of the network architecture are proposed, implementing computing systems for evaluating object similarity using the Jaccard, Russell-Rao, Sokal-Michener, Kulczynski, and Yule functions. Neuron and neural block structures are developed that ensure data transmission between the network's layers and perform diagnostics of systems for determining the operability of recognition algorithms through the parallel computation of match/mismatch variables for the features of compared binary vectors. A numerical example is given, confirming the correctness of the proposed computer tools for classification. It is shown that the developed approach extends the scope of application of neural-network-based computer systems for recognition and makes it possible to synthesize networks that determine several equally valid solutions simultaneously.