Laser-Based Artificial Neuron Replicates Nerve Cell Functions at High Speed

In the human body, there are different types of nerve cells. Graded neurons, for instance, encode information through continuous changes in their membrane potential, allowing for nuanced and precise signal processing. On the other hand, biological spiking neurons communicate using all-or-none action potentials, which is a more binary form of communication.

“Our laser graded neuron surpasses the speed limitations of existing photonic spiking neurons and has the potential for even faster operations,” explained Chaoran Huang, the lead researcher from the Chinese University of Hong Kong. “By utilizing its neuron-like nonlinear dynamics and rapid processing, we built a reservoir computing system that excels in AI tasks such as pattern recognition and sequence prediction.”

Published in Optica, a journal by Optica Publishing Group for high-impact research, the researchers detail that their chip-based quantum-dot laser graded neuron reaches a signal processing speed of 10 GBaud. This speed enabled them to process data from 100 million heartbeats or 34.7 million handwritten digital images in just one second.

“Our technology can speed up AI decision-making in urgent applications while maintaining high accuracy,” said Huang. “We aim to integrate our technology into edge computing devices—those processing data near its source—to develop faster and smarter AI systems for real-world applications, with reduced energy consumption in the future.”

Faster Laser Neurons

Laser-based artificial neurons, which can react to input signals similarly to biological neurons, are being explored to significantly enhance computing due to their ultrafast data processing speeds and low energy consumption. However, most of these developed so far are photonic spiking neurons, which have limited response speeds, potential information loss, and require extra laser sources and modulators.

The speed limitations of photonic spiking neurons arise because they generally work by injecting input pulses into the laser’s gain section, causing a delay that restricts their response speed. For the laser graded neuron, researchers used a different strategy by injecting radio frequency signals into the quantum dot laser’s saturable absorption section, avoiding this delay. They also designed high-speed radio frequency pads for the saturable absorption section, resulting in a faster, simpler, and more energy-efficient system.

“With strong memory effects and excellent information processing capabilities, a single laser graded neuron can function like a small neural network,” said Huang. “Thus, even a single laser graded neuron without extra complex connections can efficiently perform machine learning tasks.”

High-Speed Reservoir Computing

To further demonstrate the potential of their laser graded neuron, the researchers used it to create a reservoir computing system. This computing method employs a specific type of network, known as a reservoir, to process time-dependent data, such as that used in speech recognition and weather prediction. The neuron-like nonlinear dynamics and fast processing speed of the laser graded neuron make it ideally suited for high-speed reservoir computing.

In tests, the reservoir computing system showed excellent pattern recognition and sequence prediction capabilities, especially for long-term predictions, across various AI applications with high processing speed. For instance, it processed 100 million heartbeats per second and identified arrhythmic patterns with an average accuracy of 98.4%.

“In this work, we used a single laser graded neuron, but we believe that combining multiple laser graded neurons could further unlock their potential, much like the brain’s billions of neurons working together in networks,” said Huang. “We are working to improve the processing speed of our laser graded neuron while also developing a deep reservoir computing architecture that incorporates cascaded laser graded neurons.”