Abstract
Resistive switching memory (RRAM) devices have been proposed to boost the density and the bio-realistic plasticity in neural networks. One of the main limitations to the development of neuromorphic systems with RRAM devices is the lack of compact models for the simulation of spiking neural networks, including neuron spike processing, synaptic plasticity, and stochastic learning. Here, we present a predictive model for neuromorphic networks with unsupervised spike timing-dependent plasticity (STDP) in HfO2 RRAM devices. Our compact model can predict the learning behavior of experimental networks and can speed up the simulation of unsupervised learning compared to Monte Carlo (MC) approaches. The model can be used to optimize the classification accuracy of data sets, such as MNIST, and to estimate the time of learning and the energy consumption.
Highlights
T HE spiking neural network (SNN) is one of the most suitable ways to perform brain-inspired computing [1].Learning in the brain takes place via synaptic plasticity, whereManuscript received March 21, 2020; accepted April 28, 2020
We demonstrate that the model is robust and flexible versus device nonidealities, and it can describe other types of memory devices used as synaptic elements, like phase change memory (PCM)
Simulation results from the compact model and the Monte Carlo (MC) model were compared to experimental data for a perceptrontype SNN with 16 PRE fully connected to a single POST [20]
Summary
T HE spiking neural network (SNN) is one of the most suitable ways to perform brain-inspired computing [1]. STDP was shown to be fundamental to achieve energy-efficient continual learning in artificial neural networks (ANNs) [18], [19]. Despite these achievements, unsupervised learning via synaptic plasticity at the network level is still an open challenge. The design of neuromorphic circuits for STDP-based unsupervised learning is usually addressed by Monte Carlo (MC) simulations, which handle the stochastic spiking activities of the RRAM synapses [12] but suffer from excessive computational time. Based on rate equations of the synaptic weights in the network, the model allows to predict the learning speed and energy efficiency as a function of the learning parameters, such as the pattern density or the noise rate of submission. We demonstrate that the model is robust and flexible versus device nonidealities, and it can describe other types of memory devices used as synaptic elements, like PCMs
Sign-in/Register to view highlights & summary 
Published Version (
Free)
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call