Image Classification by Neural Network on Crossbars with Memristor Defects

Memristor devices are promising in terms of speeding up and improving the power efficiency of deep learning systems. Crossbar architectures built using memristors can be used to efficiently implement various in-memory computational operations such as multiply-accumulate and convolution, which are widely used in deep and convolutional neural networks. The modeling of image processing by neural networks built on the basis of memristor crossbars is performed. The influence of memristor crossbar defects on the quality of image recognition is studied. The proportions of defective memristors for which memristor neural networks are able to correctly classify images are determined. Dependences of the performance of memristor neural networks on their architectures are revealed: the more layers in the network, the stronger the influence of defects in memristors on the network performance. The time of software simulation of a memristor crossbar operation is estimated. The high resistance of memristor crossbars to defects with a large size of crossbars is noted. It has been established that various models of memristors, both linear and non-linear, behave almost identically within convolutional neural networks.