Interval Model of a Memristor Crossbar Network

Abstract

This article covers the modeling of memristive oxide‐based elements within the context of a complex simulation model of an analog self‐learning spiking neural network. According to the experimental data, the nature of the memristor operation is stochastic, as evidenced by the variation in the current–voltage characteristics for different resistive switching cycles. Often, the existing mathematical models of memristors do not fully reproduce the experimental results, which can adversely affect the accuracy of neuromorphic network simulation. It is proposed to use the interval approach to take into account the variations in the element characteristics. The idea is to introduce interval parameters into the mathematical model. Thus, the simulation results at each moment are interval estimates of phase variables. The values of interval parameters are calculated in such a way that the intervals completely contain the experimental data. When simulating the operation of the entire analog self‐learning spiking neural network, at each training epoch the parameters of the memristors are set randomly from the found intervals. The approbation of the proposed approach is carried out on the problem of pattern recognition by the crossbar architecture network with the LiNbO3‐based memristive elements.