Trainable neural network consisting of FitzHugh-Nagumo systems

Abstract

ANNs have been evolving towards more powerful and biologically realistic models. In the last decade, spiking neural networks (SNNs) consisting of spiking neurons have been developed and improved. Information transmission in these neurons imitates information transmission in biological neurons, i.e., using the duration or sequence of pulses. To facilitate learning in such networks, new learning algorithms based on different degrees of biological plausibility have been recently developed. Here, neuronal models, namely FitzHugh-Nagumo (FHN) systems, will be used as elements of such networks. Such replacement will allow to approximate the model to real biological system and to estimate the peculiarities of implementation of these models in the network, which will allow to further refine the learning algorithms.
The peculiarities of behavior of ensembles and networks consisting of mathematical models of biological neurons is one of the subjects of nonlinear dynamics, which allows to study the behavior of connected neurons from the fundamental point of view. Using such models as neurons would bring the trained network closer to the biological one. This paper discusses how to implement a deep neural network with FitzHugh-Nagumo systems in a hidden layer. It is shown that the average accuracy of the resulting neural network is about 82.7% and 81.1%. on the training and validation sets of the MNIST database with handwritten digit images.
The work was supported by the Ministry of Science and Higher Education of the Russian Federation as part of the President Scholarship SP-749.2022.5.

Speaker

Nadezhda Semenova
Saratov State University
Russia

Discussion

Ask question