Binarized Neural Network Comprising Quasi-Nonvolatile Memory Devices for Neuromorphic Computing

Abstract

This study presents a binarized neural network (BNN) comprising quasi-nonvolatile memory (QNVM) devices that operate in a positive feedback loop mechanism and exhibit an extremely low subthreshold swing (≤ 5 mV dec⁻¹) and a high on/off ratio (≥ 10⁷). A pair of QNVM devices are used for a single synaptic cell in a cell array, in which its memory state represents the synaptic weight, and the voltages applied to the pair act as input in a complementary fashion. The array of synaptic cells performs matrix multiply-accumulate (MAC) operations between the weight matrix and input vector using XNOR and current summation. All the results of the MAC operations and vector-matrix multiplications are equivalent. Moreover, the BNN features a high accuracy of 93.32% in the MNIST image recognition simulation owing to high device uniformity (1.35%), which demonstrates the feasibility of compact and high-performance neuromorphic computing.