Experiment for creating a neural network with weights determined by the potential of a simulated electrostatic field

In this work, based on the architecture of the neural network that implements the metric recognition method, experiments are carried out to determine the values of the weights and thresholds of the neural network using the parameter of the electrostatic field – potential, without using additional analytical calculations and without using learning algorithms. The simulation of the electrostatic field is implemented in the Builder C ++ software environment, which calculates the value of the total potential of the electrostatic field at the points of the proposed model where the potentiometer sensors are located. The same software module implements the possibility of creating a neural network based on metric recognition methods, for which the values of the weights of the first layer neurons are determined based on the values of the obtained potentials of the simulated electrostatic field. The effectiveness of the resulting neural network is checked against the MNIST (Modified National Institute of Standards and Technology) control base of numbers. Thus, 4 experiments were carried out with different initial conditions of the proposed simulated model for reading the potential of the electrostatic field: the values of charges from the photosensors, the distance between the sensors, the number of samples. The results of all the experiments carried out showed the efficiency of the proposed model and the possibility of determining the values of the weights of the neural network by the parameter of the electrostatic field – the potential. Checking the obtained neural networks with weights – the potentials of the electrostatic field on the MNIST control base (the performance of neural networks), showed the possibility of increasing the effectiveness of the neural networks obtained in this way. In particular, it was shown that the efficiency of the obtained neural networks increases with an increase in the number of samples used, as well as by changing the parameters of the proposed model for reading the parameters of the electrostatic field, such as the values of charges from the photosensors, the distance between the photosensors and the distance between charged panels.