Variability-Aware Memristive Crossbars With ImageSplit Neural Architecture

Abstract

The errors in the memristive crossbar arrays due to device variations will impact the overall accuracy of neural networks or in-memory systems developed. For ensuring reliable use of memristive crossbar arrays, variability compensation techniques are essential to be part of the neural network design. In this paper, we present an input regulated variability compensation technique for memristive crossbar arrays. In the proposed method, the input image is split into non-overlapping blocks to be processed individually by small sized neural network blocks, which is referred to as imageSplit architecture. The memristive crossbar based Artificial Neural Network (ANN) blocks are used for building the proposed imageSplit. Circuit level analysis and integration is carried out to validate the proposed architecture. We test this approach on different datasets using various deep neural network architectures. The paper considers various device variations including $R_{OFF}/R_{ON}$ variations and aging using imageSplit. Along with hardware compensation techniques, algorithmic modifications like pruning and dropouts are also considered for analysis. The results show that splitting the input and independently training the smaller neural networks performs better in terms of output probabilistic values even with the presence of the significant amount of hardware variability.