Artificial Neuron Using Lasers Mimics Nerve Cell Functions at Lightning Speed

Researchers have developed a laser-based artificial neuron that can fully simulate the function, dynamics, and information processing of a biological neuron. With a signal processing speed of 10 GBaud – one billion times faster than biological neurons – this new laser sorting neuron could lead to breakthroughs in fields such as artificial intelligence and other advanced computing.

The body contains many types of neurons, including hierarchical neurons that encode information through continuous changes in membrane potential, allowing for sophisticated and precise signal processing. In contrast, biological neurons transmit information using all-or-none action potentials, creating a more binary form of communication.

“Our laser-sorted neurons overcome the speed limitations of current photon versions of spiny neurons and have the potential to operate even faster,” said lead researcher Chaolan Huang of the Chinese University of Hong Kong. “By leveraging neuron-like nonlinear dynamics and high-speed processing, we have built a storage computing system that can outperform AI tasks such as pattern recognition and sequence prediction.”  

 In Optica , a high-impact research journal from Optica Publishing Group , the researchers report that their chip-based quantum dot laser-sorted neurons can achieve signal processing speeds of 10 GBaud. They used this speed to process data from 100 million heartbeats or 34.7 million handwritten digital images in just one second. 

“Our technology can accelerate AI decision-making in time-critical applications while maintaining high accuracy,” said Huang . “By integrating our technology into edge computing devices and processing data closer to the source, we hope to enable faster and smarter AI systems in the future while consuming less energy, making them more suitable for real-world applications.”  

Faster Laser Neurons

Laser-based artificial neurons can mimic the behavior of biological neurons to react to input signals, and due to their ultrafast data processing speed and low energy consumption, they are being investigated as a way to dramatically increase computing power. However, most neurons developed so far are photon-emitting neurons. These artificial neurons have limited response speed, can suffer from information loss, and require additional laser sources and modulators.

The speed limitation of optically firing neurons comes from the fact that neurons typically operate by pumping input pulses into the amplifying section of a laser. This introduces a delay that limits how fast neurons can respond. For laser-sorted neurons, the researchers took a different approach by injecting a radio frequency signal into the saturable absorption section of a quantum dot laser to circumvent this delay. They also designed a high-speed radio frequency pad for the saturable absorption section to achieve a faster, simpler, and more energy-efficient system.

“Laser-sorting neurons have strong memory effects and excellent information processing capabilities, and can function as small neural networks,” Huang said. “Therefore, even laser-sorting neurons without additional complex connections can perform machine learning tasks with high performance.”  

Fast Reservoir Computer

To further demonstrate the capabilities of laser sorting neurons, the researchers used them to create a reservoir computing system. This computational approach uses a specific type of network, called a reservoir, to process time-dependent data, such as that used for speech recognition or weather forecasting. Laser sorting neurons have neuron-like nonlinear dynamics and fast processing speeds, making them ideally suited to support high-speed reservoir computing.

In experiments, the resulting reservoir computing system demonstrated superior pattern recognition and sequence prediction capabilities across a range of AI applications, especially long-term prediction, at high processing speeds, for example processing 100 million heartbeats per second and detecting arrhythmia patterns with an average accuracy of 98.4%.

“In this study, we used a single laser-sorted neuron, but we believe that its potential can be further realized by cascading multiple laser-sorted neurons, just as the brain networks billions of neurons together, ” Huang said. “We are working to improve the processing speed of laser-sorted neurons and develop a deep storage computing architecture that combines cascaded laser-sorted neurons.”  

Để lại một bình luận

Email của bạn sẽ không được hiển thị công khai. Các trường bắt buộc được đánh dấu *