Neuronal realizations based on memristive devices

Abstract

Development of hardware-based spiking neural networks calls for novel building blocks, such as artificial neurons. This could lead to the development of systems with reduced power consumption, fault tolerance, and biomimetic artificial intelligence. Memristors, which are devices with a signal history dependent resistance, are realized via various physical mechanisms such as phase-change phenomena, redox reactions, Ovonic switching, Mott insulator-to-metal transition, and magnetoresistance. These devices possess unique dynamics which could potentially replicate biological neuronal behaviors such as leaky integrate-and-fire function. Spiking networks enabled by such artificial neurons have demonstrated intrinsic bio-realistic unsupervised learning protocols, which promises a compact and energy-efficient hardware implementation of neuromorphic computing.