[Paper Review] A Laser Spiking Neuron in a Photonic Integrated Circuit

Mitchell A. Nahmias, Hsuan-Tung Peng+5 more|arXiv (Cornell University)|Dec 15, 2020

Neural Networks and Reservoir Computing4 citations

TL;DR

This paper presents a photonic integrated circuit (PIC)-based laser spiking neuron that emulates biological spiking behavior using excitable laser dynamics. By leveraging balanced photodetectors and a semiconductor optical amplifier, the device performs nonlinear operations such as weighted summation, excitation, inhibition, and cascadability across multiple wavelengths, demonstrating compatibility with wavelength-division multiplexing for scalable photonic neural networks.

ABSTRACT

There has been a recent surge of interest in the implementation of linear operations such as matrix multipications using photonic integrated circuit technology. However, these approaches require an efficient and flexible way to perform nonlinear operations in the photonic domain. We have fabricated an optoelectronic nonlinear device--a laser neuron--that uses excitable laser dynamics to achieve biologically-inspired spiking behavior. We demonstrate functionality with simultaneous excitation, inhibition, and summation across multiple wavelengths. We also demonstrate cascadability and compatibility with a wavelength multiplexing protocol, both essential for larger scale system integration. Laser neurons represent an important class of optoelectronic nonlinear processors that can complement both the enormous bandwidth density and energy efficiency of photonic computing operations.

Motivation & Objective

• To address the lack of efficient, flexible, and scalable nonlinear optical processing in photonic integrated circuits for artificial intelligence workloads.
• To implement a biologically inspired spiking neuron using laser dynamics in a PIC-compatible platform.
• To demonstrate simultaneous excitation, inhibition, and summation across multiple wavelengths in a single device.
• To validate cascadability and compatibility with wavelength-division multiplexing for system-level integration.
• To enable energy-efficient, high-bandwidth photonic computing by reducing reliance on electronic nonlinear conversion.

Proposed method

• The laser neuron uses a pair of balanced photodetectors (BPDs) to perform current summation of multiple input signals across distinct wavelengths.
• The BPD outputs drive a distributed feedback (DFB) laser with a semiconductor optical amplifier (SOA), creating excitable dynamics that generate spiking behavior.
• The system employs wavelength-division multiplexing (WDM), assigning each laser neuron a unique wavelength to enable parallel processing.
• Time-division multiplexing via a PPG (patterned pulse generator) with programmable delays is used to simulate Poisson-distributed inputs across channels.
• Fiber delay lines (90–720 ns) are used to introduce controlled time delays per wavelength, which are digitally compensated to align inputs in time.
• Power calibration is performed using continuous-wave reference traces to normalize input and output power measurements across all channels.

Experimental results

Research questions

• RQ1 Can a photonic integrated circuit implement a nonlinear spiking neuron using laser dynamics that mimic biological neurons?
• RQ2 Can the device perform simultaneous excitation and inhibition across multiple wavelength channels in a single node?
• RQ3 Is the laser neuron compatible with wavelength-division multiplexing and capable of cascading for larger network integration?
• RQ4 Can the system achieve accurate spiking behavior with programmable, time-multiplexed inputs such as Poisson-distributed bit patterns?
• RQ5 Does the photonic spiking neuron reduce reliance on energy-intensive electronic-to-optical conversion while maintaining high-speed, low-energy operation?

Key findings

• The laser neuron successfully demonstrated spiking behavior in response to both excitatory and inhibitory inputs across eight distinct wavelength channels.
• The device achieved simultaneous excitation and inhibition with precise temporal control, validated through time-domain traces and spectral measurements.
• Cascadability was demonstrated by using the output of one laser neuron as input to another, confirming compatibility with multi-layer network architectures.
• The system maintained signal integrity across multiple wavelengths, with input patterns correctly reconstructed after multiplexing and demultiplexing using AWGs.
• Power calibration enabled accurate normalization of input and output traces, with measured input powers consistent across channels and within expected loss margins.
• The device operated with a 5 GHz clock rate and 88 ns time window, supporting high-speed, low-jitter processing suitable for neural network inference.