Abstract:
At present, in typical architectures of feedforward neural networks, the values of the weights of connections and thresholds of neurons are determined by adjusting the values of the weights performed by means of typical learning algorithms. Also known are the architectures of feedforward neural networks, implemented on the basis of metric recognition methods, for which the values of the weights of neurons are pre-calculated analytically. The analytical calculation of the values of the weights is carried out on the basis of metric expressions and allows to immediately obtain a workable neural network without training. In this case, the effectiveness of the obtained neural network will depend on the selected set and the number of samples, as well as on the selected dimension of the table of weights. It was also shown that such neural networks can be additionally trained with typical learning algorithms, which makes it possible to increase the efficiency of the neural network with the calculated weights by additional training of the neural network. In this case, the process of calculating the values of the weights and further training of the neural network is also faster than training the neural network in the traditional way, by means of a random initial generation of the weight values of the neural network. In this work, on the basis of these networks, the possibility of determining the values of weights and thresholds of a neural network using the parameters of the electrostatic field: tension, potential is considered. That is, it is proposed to use the values of the parameters of the electrostatic field as the values of the weights of the neural network. In other words, the possibility of creating a workable neural network without analytical calculations and without the use of learning algorithms is considered. This approach allows to make the process of determining the values of the neural network weights almost instantaneous. The technical possible implementations of this approach and the problematic aspects of using the parameters of the electrostatic field as the weights of the neural network, as well as possible approaches to resolving these difficulties are considered.
Keywords:neural networks, electric field, electric field strength, electric potential, learning algorithms, neurocomputer.