Analysis of pulse frequency in a simple neural network (Q1842414)

One of the basic problems in the development of artificial intelligence systems is how to parallelize information processing with the aid of neuron-like structures so as to achieve substantial gains of speed, reliability, information capacity, and other indicators. Hopfield has engendered a lasting interest in mathematical models of neural networks. He solved the problem of pattern recognition by a network consisting of simple deterministic binary neurons. Various authors have since considered stochastic versions of this model. Another direction of research focuses on constructing and analysis of networks of more complex neurons, which instead of a qualitative ``on-off'' variable are characterized by a quantitative variable, such as degree of excitation or pulse frequency. The ``input-output'' characteristic of a single neuron is usually assumed a priori known. The principle of frequency encoding of information is closely linked with neurophysiological studies. The main purpose of this study is to elucidate the principle of frequency encoding of information and to construct the ``input-output'' dependence for an elementary converter (a neuron). Algorithms and programs have been developed for exact computation of the distribution of interpulse intervals in the output stream, which can be used to test various asymptotics, in particular approximate relationships. All the theoretical results have been tested using a simple simulation model.