Metal Oxide Resistive Memory (OxRAM) and Phase Change Memory (PCM) as Artificial Synapses in Spiking Neural Networks

Abstract

Brain-inspired computing systems are attracting considerable attention due to their high energy efficiency with respect to conventional computing systems when applied to problems in artificial intelligence, sensing and robotics. This is attributed to the spike-based computational mechanism and architectural co-localization of processing and memory. To realize this neuron circuits must be integrated with further circuits modelling synapses. Synapses require to exhibit plasticity, that is modulation in their efficacy, to support online learning algorithms, manifesting in changes in their conductivity. Metal Oxide Resistive Memory (OxRAM) and Phase Change Memory (PCM) can be used as synaptic elements thanks to tunable conductivity, compatibility with advanced CMOS fabrication process and scalability. In this work, we present recent advances in implementing synaptic plasticity using both OxRAM and PCM arrays. We demonstrate that certain properties of these devices, usually considered non-idealities (i.e. conductance variability and non-linear change of conductance upon identical programming pulses), can improve the performance of Spiking Neural Networks (SNNs) trained with a bio-inspired, on-line and spike based unsupervised learning algorithm called Spike-Timing Dependent Plasticity (STDP).