What would artificial neurons, a billion times faster than the human brain, look like?

Today, this sci-fi scenario has become a reality.

Scientists from the Chinese University of Hong Kong, the Institute of Physics of the Chinese Academy of Sciences and other institutions have successfully developed an artificial neuron based on "laser".

The latest research is published in the journal Optica.

Address: https://opg.optica.org/optica/fulltext.cfm?uri=optica-11-12-1690&id=565919

This chip-based quantum dot laser can not only fully mimic the function of real nerve cells, but also achieve amazing speed

That's 10GBaud signal processing speed, which means it's a billion times faster than biological neurons.

Based on a chip-point laser, biological cascade neurons can be simulated while achieving a signal processing speed of 10 GBaud

How to understand how fast this is?

Capable of processing 100 million heartbeats in 1 second; Able to analyze 34.79 million handwritten digital images in 1 second.

This breakthrough could revolutionize the field of AI and advanced computing, improving the ability to recognize patterns and predict sequences.

AI mimics biological neurons and surges by a factor of 1 billion

Why is this groundbreaking discovery so important?

In our bodies, there are different types of nerve cells.

graded neurons encode information by continuously changing the membrane potential for sophisticated signal processing.

spiking neurons, in contrast, use all-and-none action neurons to transmit information, creating a more binary communication.

The key technological breakthrough in the latest research is the innovative design approach.

Conventional photon-pulsed neurons typically work by injecting input pulses into the gain region of the laser, which causes a delay that limits the neuron's response rate.

As shown in the figure below, the input and output of pulsed neurons and stepped neurons are compared.

Laser neurons, fast speed and low power consumption

Laser artificial neurons, which are able to respond to input signals in a way that mimics the behavior of biological neurons, are being explored as a way to significantly enhance computing due to their ultra-fast data processing speed and low energy consumption.

However, most of those developed so far have been photon-pulsed neurons.

These artificial neurons have a limited response speed, can suffer from information loss, and require additional laser sources and modulators.

The speed limit for photons to pulse neurons has been broken in the latest study.

The research team took a different approach, choosing to inject the RF signal into the saturable absorption region of the quantum dot laser, and cleverly circumvented this limitation.

They also designed high-speed RF panels for the saturable absorption region, resulting in a faster, simpler, and energy efficient system.

Chaoran Huang, head of the Hong Kong Chinese research team, said, "The laser cascade neurons break the speed limit of the current photonic pulse neurons, and a reservoir computing system we built has demonstrated excellent performance in AI tasks such as pattern recognition and sequence prediction."

Laser cascade neurons demonstrate superior pattern recognition and sequence prediction in AI tasks such as arrhythmia detection and image classification

He also said that with powerful memory effects and excellent information processing capabilities, a single laser gradient neuron can behave like a small neural network.

So even a single laser cascade neuron without additional complex connections can efficiently perform machine learning tasks.

High-speed reservoir computing, processing 100 million heartbeats per second

To further demonstrate the power of laser cascade neurons, the research team used them to build a reservoir computing system.

This is a computational method that uses a specific network (called a repository) to process time-related data, often used in areas such as speech recognition and weather prediction.

The neuron-like nonlinear dynamics of laser cascade neurons, as well as their express processing speed, make them ideal for supporting high-speed reservoir computation.

The diagram below shows the architecture of reservoir calculation (RC).

Derived from recurrent neural networks, RC is a powerful and cost-effective computing framework that is well suited for processing time/sequence information.

It is mainly composed of input layer, storage layer and read layer. In the storage layer, the interconnection between nonlinear nodes is random and the weights are fixed, thus avoiding the training of the storage layer.

Here, only the readout layer needs to be trained, which can be done by simple and computationally efficient methods such as linear regression.

In the latest study, the authors chose to use laser cascade neurons as laser reservoirs to perform reservoir calculations. In the input layer, the input signal is encoded as an electrical pulse injected into the laser reservoir.

In specific experiments, the system demonstrated impressive performance.

For example, it can process 100 million heartbeats per second and detect arrhythmia patterns with an average accuracy of 98.4%.

Specifically, the researchers used the processed MIT-BIH arrhythmia dataset to turn on the baseline task of abnormal heartbeat detection.

The database contains extracts of 48 1/2 hours of electrocardiogram recordings obtained from 47 subjects and is the first test material that can be widely used to evaluate arrhythmia monitors.

In the two post-processing MIT-BIH arrhythmia datasets, the original ECG waveform was resamped and divided into a single heartbeat, each consisting of 50 time steps.

As shown in Figure a below, these heartbeats were classified into two groups - the healthy group and the arrhythmia group, labeled 0 and 1, respectively.

In addition, it has demonstrated excellent pattern recognition and sequence prediction capabilities in various AI applications, especially for long-term prediction tasks.

In the MNIST handwritten dataset task, the researchers evaluated the classification performance of the laser repository. The MNIST dataset consists of handwritten digital images composed of 28×28 grayscale pixels.

As shown in the figure below, the average accuracy calculated using the six-fold cross-validation method reached 92.3% in the four classes of MNIST handwritten digit classification tasks.

"Hearing about this breakthrough makes me more convinced that we are steadily moving forward on the exponential growth curve.

It's moving so fast now that I have no way of predicting where we'll be in six months or even a year. Maybe I'm getting ahead of myself, but I do feel strongly the acceleration of technological progress these days.

So what does the discovery of laser neurons mean?

It speeds up the decision-making process for AI in time-critical applications, significantly increasing processing speed while maintaining high accuracy.

If one day it is integrated into edge computing devices, enabling faster and smarter AI systems, it will significantly reduce energy consumption.

According to the researchers, the next step in the team's efforts will be to increase the processing speed of laser gradient neurons, while developing a computational architecture that includes a deep reservoir of cascade laser gradient neurons.