A Hardware and Energy-Efficient Online Learning Neural Network With an RRAM Crossbar Array and Stochastic Neurons

Abstract

This article presents a hardware- and power-efficient RRAM-based neural network capable of online learning. The network is modularized in consideration of scalability and consists of 11 modules. Each module comprises two 25 x 25 RRAM crossbar arrays, four analog multiplexers, and one neuron chip. A stochastic neuron that performs both quantization and activation of the column current flowing through the RRAM array is used to perform online learning without using the high-resolution analog-to-digital converters and digital-to-analog converters. The impact of the initialization of the RRAM device on the network performance is analyzed and a fast yet effective method of initialization is also presented. The network consists of three fully connected layers with 100 input and 10 output neurons. The two hidden layers are made up of 50 and 25 neurons, respectively, and a total of 13 x 103 synapses are implemented by combining 21 RRAM arrays. The neuron chip, which includes 32 neurons, is fabricated through a 0.18-mu m standard CMOS process. The network achieves a classification accuracy of 93.4% for the reduced Modified National Institute of Standards and Technology database (MNIST) dataset and the energy efficiency of 38.1 teraoperations per second perWatt (TOPS/W), the highest efficiency among recently reported state-of-the-art RRAM-based neural networks.