Photonic spiking neural networks and CMOS-compatible graphene-on-silicon spiking neurons

TL;DR

This paper surveys photonic spiking neural networks, introduces a novel graphene-on-silicon spiking neuron, compares existing devices, and discusses training methods and potential applications for high-speed, noise-robust photonic neural processors.

Contribution

It presents a new CMOS-compatible graphene-on-silicon spiking neuron and compares it with existing photonic spiking devices, advancing scalable photonic neural hardware.

Findings

Graphene-on-silicon microring cavity enables efficient spiking neurons

Comparison shows advantages of the proposed device over existing ones

Discussion highlights training challenges and application potentials

Abstract

Spiking neural networks are known to be superior over artificial neural networks for their computational power efficiency and noise robustness. The benefits of spiking coupled with the high-bandwidth and low-latency of photonics can enable highly-efficient, noise-robust, high-speed neural processors. The landscape of photonic spiking neurons consists of an overwhelming majority of excitable lasers and a few demonstrations on nonlinear optical cavities. The silicon platform is best poised to host a scalable photonic technology given its CMOS-compatibility and low optical loss. Here, we present a survey of existing photonic spiking neurons, and propose a novel spiking neuron based on a hybrid graphene-on-silicon microring cavity. A comparison among a representative sample of photonic spiking devices is also presented. Finally, we discuss methods employed in training spiking neural networks, their challenges as well as the application domain that can be enabled by photonic spiking neural hardware.