Performance of Pulse-Programmed memristive crossbar array with bimodally distributed stochastic synaptic weights

• Simulation of artificial neural networks in the form of memristive crossbar arrays. • Pulse-programmed memristive devices show bimodal distribution for synaptic weight. • Efficient simulation of classification result by noise-based variability approach. • Influence of variability and stepping of weights on classification is explored. • Constant weight stepping with low variability is beneficial. In this paper, we present a method of implementing memristive crossbar arrays with bimodally distributed weights. The bimodal distribution is a result of pulse-based programming. The memristive devices are used for implementing synaptic weights and can only have an ON (logical “1″) or an OFF (logical ”0″) state. The state of the memristive device after programming is determined by the bimodal distribution. The highly efficient noise-based variability approach is used to simulate this stochasticity. The memristive crossbar array is used to classify the MNIST data set and comprises more than 15,000 weights. The interpretation of these weights is investigated. In addition, the influence of the stochasticity of the weights and the accuracy of the weights on the classification results is considered and various programming settings are examined.