A 'Light-Powered Brain' That Reads Messages Hidden in Light

On October 13, researchers at the University of the Witwatersrand in Johannesburg announced what they call a "light-powered brain": an optical neural network that recognizes hidden information inside complex patterns of light without taking a detailed picture of them first. The research, published in Nature Communications on October 7, was conceived and supervised by Dr Isaac Nape and Professor Andrew Forbes at the Wits Structured Light Laboratory. It was led by MSc student Hadrian Bezuidenhout, with contributions from Cade Peters and Ram Nandan Kumar. The team says the approach could lead to faster, more resilient optical communications, and eventually to new measurement tools for quantum technologies.

The problem it attacks sits at the heart of a growing field called structured light. Just as letters combine into words, light can be sculpted into specially shaped spatial patterns, and each distinct pattern can stand for a symbol in a communications alphabet. The larger the alphabet, the more information a single pulse of light can carry, and there is in principle no limit to how many patterns exist. But alphabets are only useful if someone can read them, and that has been the stubborn half of the equation.

These patterns do not necessarily keep their appearance as they travel. Atmospheric turbulence, stressed fiber, or even the physics of propagation itself can change how a structured beam looks while leaving its underlying identity intact. A receiver that works by matching the exact appearance of a pattern can therefore fail even when the message has survived the journey perfectly well. Conventional solutions take a detailed image and analyze thousands of individual points with heavy computation, which is slow and expensive.

Counting twists instead of creases

The Wits team's answer is to stop looking at the whole picture and read a hidden feature instead. Nape offers an analogy: imagine a ribbon that has been twisted, then bent and folded into many shapes. The number of twists stays the same no matter how you fold it. Light carries similar hidden features, which mathematicians and physicists call topology: properties that survive stretching and distortion.

Instead of imaging the beam, the new system passes light through specially designed layers that steer it toward a set of detectors. Where the light lands reveals the information it carries. In effect, the neural network is built into the optics themselves: the computation happens as light travels, rather than in a chip afterward. The team found that after training on just 15 examples, the system successfully recognized 81 different light structures from the family under study, including patterns it had never encountered before. It also kept working when the researchers deliberately added noise to the detector, a sign it could function under the imperfect conditions of real-world links.

"We don't need to map every crease in our fabric of light to count the twists," says Bezuidenhout, now an optics engineer at APSI. "Our network picks out the information that matters."

That robustness is the crux. Free-space optical links, such as laser connections between buildings or between ground stations and satellites, suffer exactly the kind of scrambling that topology ignores. Reporting on the result notes the authors' claim that a trained detector could eventually be built into fixed optics that draw no power, effectively a passive reader for structured light. The information sits in a feature of the pattern that survives distortion, so a receiver no longer needs the pristine image it was trained to expect.

Why it matters beyond the lab

For communications engineers, the prize is bandwidth without new spectrum: encoding more bits per pulse by expanding the alphabet of distinguishable symbols. Structured light has long promised that, and decoding has been the lagging edge. A reader that extracts information optically, in a single pass, without imaging or heavy digital processing, attacks the latency and energy cost of the receiver, two of the practical barriers to deployment.

The work also signals something broader: the merging of AI principles with photonics. Professor Benjamin Rosman, director of Wits's Machine Intelligence and Neural Discovery (MIND) Institute, described the research as an example of what becomes possible when ideas from artificial intelligence and fundamental physics meet, by embedding neural network principles directly into the behavior of light. Forbes, who directs the South African Quantum Technology Initiative, says the measurement toolbox for reading structured light is still taking shape, and that reaching beyond photonics is how the field builds it.

There are limits, and the team does not hide them. This is a laboratory demonstration on a specific family of light structures, not a product. Scaling to the giant alphabets and the long, turbulent links that would matter commercially will take more work, and quantum applications, where information must be extracted from individual particles of light, pose even steeper measurement challenges. Nape is explicit that this is just the beginning.

Still, the direction of travel is clear. Computing is increasingly moving into the medium itself, from optical neural networks to processing-in-memory chips, because moving data costs more than transforming it. A neural network that runs on light, trained on a handful of examples and unbothered by noise, is a small but genuine step in that direction, and it came out of a Johannesburg lab, not one of the usual coastal clusters. That geography matters too: world-class photonics research is not confined to the Northern Hemisphere's funding centers, and Nature Communications agreed.