A digitalized RRAM-based Spiking Neuron Network system with 3-bit weight and unsupervised online learning scheme

Abstract

Resistive-switching Random Access Memory (RRAM) has emerged as a promising candidate for the artificial synaptic in neuromorphic computation circuits due to its similar electronic characteristics with the synaptic and features such as high integration density, non-volatile retention and supporting matrix-vector multiplication. In this paper, a digitalized RRAM-based fully-connected Spiking Neuron Network (SNN) system with 3-bit weight and unsupervised online learning scheme is proposed. It consists of 64 pre-neurons and 10 post-neurons, all the neurons are realized by digital circuits for low area overhead, low power consumption and high accuracy. An unsupervised online learning scheme based on binary STDP protocol is applied to train the synaptic weights. Experiments show that the system can be used to recognize the learned ten handwritten digits efficiently.