Determining Neural Network Weights Using an Electrostatic Field

Abstract

This study explores an approach to determining the weights and thresholds of a neural network based on the potential of an electrostatic field, without the need for additional analytical computations or traditional training algorithms. The neural network follows a metric-based recognition method, and the electrostatic field simulation is implemented in the Builder C++ programming environment. The software computes the total electrostatic potential at designated points in the proposed model (corresponding to sensor locations–potentiometers). The same software module also enables the creation of a neural network based on metric recognition methods, where the weights of the first-layer neurons are assigned based on the computed potentials of the simulated electrostatic field. The effectiveness of the resulting neural network is evaluated using the MNIST dataset for the task of handwritten digit classification. Additionally, the possibility of applying this approach to semantic text understanding tasks is considered.