Evaluation of the accuracy of artificial neural networks based on memristive devices based on the theory of experimental planning

Abstract

: To date, computer modeling is one of the most common approaches to assessing the accuracy of artificial neural networks based on memristive devices (ANNM), taking into account variations in their resistances. In the process of such modeling, models are usually used that describe the relationship between the applied voltage and current in a circuit with a memristor. These models reproduce the behavior of the memristor well, however, when evaluating the accuracy of the ANNM, they have a number of disadvantages related to the complexity and ambiguity in approaches to accounting and specifying resistance variations from cycle to cycle and from device to device, the resource intensity of the INSM modeling process, etc. In addition, with this approach, there is no connection with the parameters of the resistance setting signal, which adds additional difficulties in synthesizing specific values of the signal parameters after analyzing the accuracy. In this paper, a new approach to estimating the accuracy of the ANNM at the research design stage is proposed, which consists in the fact that the accuracy of a specific version of the ANNM should be evaluated in relation to the parameters of the resistance setting signal of a memristive device, based on the established relationship between these parameters and the values of resistances and weights actually obtained in the experiment. This approach involves the use of experimental planning theory to create a model of the dependence of the resistance of a memristive device on the parameters of its target signals and a model of the dependence of the weights of the INF synapses on the resistances and the weight formation scheme and their application in computer modeling when evaluating the accuracy of the network.