An Improved RRAM-Based Binarized Neural Network With High Variation-Tolerated Forward/Backward Propagation Module

Binarized neural network (BNN) enables resistive switching random access memory (RRAM) with high nonlinearity and nonsymmetry to realize online training, using an RRAM comparator structure. In this work, a new hardware implementation approach is proposed to improve the efficiency of BNN. In the approach, an 1T1R array-based propagation module is introduced and designed to realize the computing acceleration of fully parallel vector-matrix multiplication (VMM) in both forward and backward propagations. Using the 1T1R-based propagation module, high computing efficiency is achieved in both training and inference tasks, improving by $50\times $ and $177\times $ , respectively. To solve the computation error caused by device variation, a novel operation scheme with low gate voltage is proposed. With the operation scheme, the RRAM variation is dramatically suppressed by 74.8 for cycle-to-cycle and 59.9 for device-to-device. It enables high-accuracy VMM calculation and, therefore, achieves 94.7 accuracy with a typical BNN, showing only 0.7 degradation from the ideal variation-free case.