Анализ информации в нейроморфных информационных моделях нейронов

Abstract

The article discusses the systemic principle of constructing of detector artificial neural networks (DNN). This principle is based on the determination and detection of structural elements of recognizable patterns, as well as their non-derivative and derivative characteristics. Non-derivative structural elements, as well as their qualitative and quantitative characteristics, are determined empirically. These elements and their characteristics are detected by specific neurons-detectors of the DNN at the stage of sensory perception. The process of detecting of non-derivative structural elements is based on the discovery by David Hubel and Torsten Wiesel of the selective response of neurons in the primary visual cortex to certain stimuli. However, non-derivative structural elements and their characteristics are not enough to solve the problem of image classification. This is due to t he fact that in the process of training of a neuron-detector of a class of images, information is lost that does not contain stable classification features. This loss of information reflects the generalizing ability of the DNN and leads to a decrease in its resolution. To increase the resolution of the DNN , additional information is needed. This information can be obtained as a result of the formation of derived characteristics of the structural elements of a recognizable image. The formation of derived characteristics reflects the process of information analysis, which is carried out by neurons-analyzers of the DNN. According to the authors, these neurons-analyzers are information models of biological neurons-analyzers. Then the process of information synthesis is implemented by the single derivatives neurons-detectors of the DNN. These neurons-detectors respond to whole images. The construction of information models of neurons is based on the hypotheses of the neural code put forward by the authors, which explain the information essence of th e reactions of neurons.