A neuron powered by light gives AI some of the brain’s most important traits

Training today’s largest AI models requires enormous data centres filled with processors that consume power around the clock. Some analyses predict that, in the coming years, AI could account for a significant share of the world’s growing electricity consumption.

Meanwhile, the human brain performs tasks that still challenge even the most advanced AI systems while consuming just 20 watts of power – roughly the same as an LED bulb.

“The transistors we use today are actually incredibly energy-efficient. The problem is that information has to be moved back and forth between them and the memory all the time. It is the transport of information that consumes energy. The brain works differently. Here, computation, memory and communication take place within the same networks of connections,” says Anders Mikkelsen, who led the new study together with colleagues from Lund University in Sweden and the University of Copenhagen in Denmark.

This is why researchers around the world are looking for new ways to build AI. If AI models continue to grow at the same pace as they do today, energy consumption could become one of the biggest constraints on how far the technology can develop.

In a new study, researchers from Lund University and the University of Copenhagen have developed an artificial neuron that brings together several of the functions researchers have long been seeking in photonic neurons.

Many photonic neurons have previously been able to demonstrate only a single biological function at a time. The new neuron combines several of them in a single component. If the technology can be scaled up to larger networks, it could eventually contribute to a new generation of energy-efficient AI hardware. But the researchers also see another possibility.

“You could almost think of it as an intelligent pixel that both senses and processes information,” says Joachim E. Sestoft, Assistant Professor at the Niels Bohr Institute of the University of Copenhagen and one of the study’s first authors.

For Joachim E. Sestoft, the neuron is also a building block.

“We have built the Lego brick. The next step is to put several of them together,” he says.

In time, such components may enable information to be analysed where the data is generated instead of first sending everything to a separate computer.

AI is built like a calculator – not a brain

The modern AI revolution is built on a technology that fundamentally works very differently from the human brain.

When a chatbot answers a question or an image generator creates a photograph, the calculations take place in enormous data centres filled with advanced computer chips. They are extremely fast, but they also consume enormous amounts of energy.

The problem is not necessarily the calculations themselves. A large share of the energy consumption goes into moving information back and forth between different parts of the computer.

This has led researchers to revisit an idea that has existed for decades: perhaps the AI of the future should not be built like traditional computers.

“What we are doing today is, in reality, trying to simulate a brain with a pocket calculator,” says Anders Mikkelsen.

In a conventional computer, memory and logic are separate, and data must constantly be moved between them. In the brain, by contrast, memory, signal processing and decision-making take place within the same neurons. Researchers are trying to learn from precisely this organisation in the field known as neuromorphic computing.

Light can do what electrons struggle to do

One particularly promising direction is photonic neural networks, in which light and not electrons carries information.

“Electrons often have to queue up. Light can pass through the same system simultaneously to a much greater extent. That opens up possibilities that conventional electronics struggles to match,” explains Anders Mikkelsen.

Yet although photonic neurons have been demonstrated before, they have often lacked some of the properties that make biological neurons so interesting.

A biological neuron must be able to receive both excitatory and inhibitory signals, integrate information from many sources, retain traces of previous activity and respond non-linearly to the world around it.

One function in particular has proved difficult to incorporate: inhibition.

The brain works as much by inhibition as by activation

The brain contains both signals that promote activity and signals that dampen it. The balance between the two helps to highlight important signals, suppress noise and control which neurons are allowed to respond. This is crucial for everything from learning to visual processing.

“From a biological perspective, inhibition is hugely important. But from a hardware perspective, it is also one of the hardest things to build,” explains Joachim E. Sestoft.

When we perceive the world, extensive signal processing is already taking place in the retina of the eye. Here, inhibitory signals are used to enhance contrasts, suppress background noise and sharpen edges long before the information reaches the brain.

“One of the most fascinating things about biological visual systems is that they do not just register the world. They start interpreting it straight away,” says Anders Mikkelsen.

This is why researchers have long been looking for a way to bring as many of the brain’s key functions as possible together in a single compact component.

That is precisely the problem the new study attempts to solve.

Three nanowires mimic several of the brain’s most important functions

To solve this problem, the researchers developed an artificial neuron consisting of just three nanowires.

Two of them act as light-sensitive receptors. The third acts as the decision-maker itself.

Together, they form a structure capable of mimicking several of the functions normally associated with biological neurons.

“We are not building a brain. We are building a building block. The goal is to show that a single component can behave like a biological nerve cell,” says Joachim E. Sestoft.

This includes the ability to receive both excitatory and inhibitory signals, combine information from multiple inputs, respond non-linearly and retain a short-term trace of previous activity.

This does not make the system an artificial brain. But it does bring it a step closer to the way biological neurons work.

The neuron passed the test on several fronts

When the researchers began testing their artificial neuron, the goal was not simply to show that it worked.

The goal was to investigate how many of the brain’s key properties could be brought together in a single component.

“We did not just want to demonstrate one property. We wanted to show that many of the functions you would want in an artificial neuron can exist within the same component,” says Joachim E. Sestoft.

One of the most important findings concerned inhibition.

In the experiments, the researchers showed that light directed at one nanowire increased the neuron’s activity, whereas light directed at the other reduced it.

The system was thus able to mimic the balance between excitation and inhibition that biological neurons maintain continuously.

“In principle, two of the nanowires function as tiny solar cells. One generates a positive signal, the other a negative one. In that way, we can both excite and inhibit the neuron,” explains Joachim E. Sestoft.

The neuron was also able to integrate several signals simultaneously.

The neuron could remember what happened a moment ago

When excitatory and inhibitory inputs were sent in at the same time, the outcome was determined by the balance between them. If the two signals cancelled each other out, the neuron returned to its resting state.

The next question was whether the component could retain a short-lived trace of previous activity.

After a brief pulse of light, the system did not immediately return to its initial state. Instead, it retained a temporary imprint that could last from around 0.1 seconds to almost a second. This means that the neuron not only responds to the latest input but is also influenced by what happened just beforehand.

“A neuron should not just react to what is happening right now. It should also carry a little information with it from what happened a moment ago,” says Anders Mikkelsen.

Different colours of light can carry different messages

In another experiment, the team sent light of different wavelengths through the system and showed that the neuron could respond selectively to specific colours.

This means that, in principle, different neurons can communicate through different optical channels without necessarily interfering with one another: a little like several radio stations broadcasting simultaneously without their signals becoming mixed together.

“That is one of the advantages of photonics. You can have many signals existing side by side without necessarily mixing them together,” explains Joachim E. Sestoft.

But the neuron was not only functional.

The active part of the neuron occupies just 30 to 90 square micrometres. According to the researchers, that makes it at most one hundredth as large as previous photonic neurons.

“The problem with many photonic neural networks is not the speed. The problem is that the components are enormous,” explains Joachim E. Sestoft. “If you want to build large networks, the chip quickly becomes gigantic.”

Small enough to scale, efficient enough to matter

The experiments also showed that the neuron can operate using optical effects at the picowatt level – trillionths of a watt. The researchers estimate the energy consumption at around 200 femtojoules per operation.

“When you get down to these energy levels, you start to approach the kind of efficiency that makes biological systems so fascinating,” says Anders Mikkelsen.

The results did not simply demonstrate a smaller photonic neuron.

They revealed a system that combines several of the functions researchers have long been seeking in photonic neurons while also being significantly smaller and more energy-efficient than previous solutions.

“The real breakthrough is not a single function. It is that we can combine several of the functions you would want in an artificial neuron within the same component,” says Joachim E. Sestoft.

Could AI move out of the data centre?

The researchers are cautious about talking in terms of revolutions. After all, the new component is still just a single neuron demonstrated in a laboratory.

“This is not the end goal. It is a building block,” says Anders Mikkelsen.

If the technology is to have any real significance, thousands, millions or perhaps billions of artificial neurons will need to be linked together in complex networks.

Anders Mikkelsen points out that today’s computers are largely flat structures, whereas the brain is three-dimensional. According to him, the possibility of building three-dimensional structures is precisely what makes light an interesting building material for future neural systems.

It is precisely here that the researchers believe their design may have an advantage.

The difficult part is not one neuron – but millions

The neuron’s extremely small size enables far more of them to be packed onto the same chip, and light-based communication may make it easier for many components to operate simultaneously.

“The biggest problem is not necessarily making one good neuron. It is making enough of them,” says Anders Mikkelsen.

There is still a long way to go. For example, the researchers need to develop methods for mass production, connect the neurons into large networks and demonstrate practical computations that can compete with existing AI hardware.

“A bee solves navigation tasks that are still difficult for our technology, but it does so with energy consumption that is many orders of magnitude lower,” says Anders Mikkelsen.

“If a neuron uses energy without contributing to the task, it is weeded out. Nature is extremely good at eliminating what is not needed.”

Perhaps it is as much a sensor as a computer

Perhaps the most interesting aspect of the study is that the technology may be useful for more than just AI hardware. It could also point towards a new way of building sensors.

Today, sensors, memory and processors typically function as separate components.

First, the sensor detects a signal. The data are then sent to a processor, which analyses the information.

Biological systems work differently.

A neuron detects information, stores it and processes it at the same time.

“The exciting thing is that the same component can function both as a sensor and as a computer,” explains Anders Mikkelsen.

Inspired by the eye, computation can begin before the data move on

The analysis may begin where the data are generated. The sensor not only detects information but also performs the first part of the computation.

When light hits the eye, for example, the information is not simply passed on to the brain in its raw form. Already in the retina, contrasts are enhanced, background noise is suppressed and relevant patterns are highlighted.

“It is actually both a computer and a sensor built into the same component,” says Joachim E. Sestoft.

If similar principles can be transferred to future sensors, some of the data processing that currently takes place in large data centres may be moved all the way out to where the data are generated.

“If we can learn some of the principles that make the brain so efficient, it could change the way we build the technology of the future,” says Anders Mikkelsen.