Simulation Modeling of an Analog Impulse Neural Network Based on a Memristor Crossbar Using Parallel Computing Technologies

Abstract

The issues of simulation modeling of an analog impulse neural network based on memristive elements in the problem of pattern recognition are studied. Simulation modeling allows us to configure the network at the level of a mathematical model, and subsequently use the obtained parameters directly in the process of operation. The network model is given as a dynamic system, which can consist of tens or hundreds of thousands of ordinary differential equations. Naturally, there is a need for an efficient and parallel implementation of an appropriate simulation model. Open multiprocessing (OpenMP) is used as the technology for parallelizing calculations, since it allows us to easily create multithreaded applications in various programming languages. The efficiency of parallelization is evaluated on the problem of modeling the process of training the network to recognize a set of five images of a size of 128 by 128 pixels, which leads to the solution of about 80 000 differential equations. In this problem, the calculations are accelerated by a factor of over six. According to the experimental data, the operating character of memristors is stochastic, as shown by the scatter in the current-voltage characteristics (VACs) when switching between high-resistance and low-resistance states. To take this feature into account, a memristor model with interval parameters is used, which gives upper and lower limits on the values of interest, and encloses the experimental curves in corridors. When simulating the operation of the entire analog self-learning impulse neural network, in each epoch of training, the parameters of the memristors are set randomly from the selected intervals. This approach makes it possible to dispense with the use of a stochastic mathematical apparatus, thereby further reducing computational costs.