BNN Training Algorithm with Ternary Gradients and BNN based on MRAM Array

Internet of Things (IoT) devices have only limited computing resources, which means we need to reduce the scale of operation circuits and energy consumption to build a neural network (NN). The binarized neural network (BNN) and computing-in-memory (CiM) have been proposed to fulfill these requirements, and recently, magnetic random access memory (MRAM), next-generation memory for CiM-based architectures has attracted interest. In this study, we utilize a CiM architecture based on an MRAM array to build a BNN on the edge side. We also implement an XNOR gate on our MRAM array using voltage-controlled magnetic anisotropy (VCMA)-based magnetization switching to reduce the scale of the multiply-and-accumulate (MAC) operation circuits by half. Further, we propose a BNN training algorithm utilizing ternary gradients to enable both training and inference on the edge side using only binary weights and ternary gradients. Experiments on the MNIST dataset showed that our MRAM array can achieve an accuracy of around 80%.