Light-Powered Neural Network Reads Hidden Patterns in Shaped Beams

Researchers at the University of the Witwatersrand in Johannesburg have built what they describe as a light-powered brain, an optical system that can recognise hidden information embedded in complex patterns of light without ever taking a detailed picture of it. The work, published in Nature Communications, was conceived and supervised by Dr Isaac Nape and Professor Andrew Forbes at the Wits Structured Light Laboratory and led by MSc student Hadrian Bezuidenhout, with contributions from Cade Peters and Ram Nandan Kumar. Its central promise is deceptively simple: if engineers can shape light into elaborate patterns that carry messages, they need equally fast ways to read those messages back out, and the new system offers a route that bypasses the slow, computationally expensive step of imaging and digital analysis altogether.

The challenge the team set out to solve lies at the heart of a growing research effort in structured light. Light can be sculpted into specially shaped spatial patterns, much as letters combine to form words, and each distinct pattern can in principle encode a symbol in an optical communications alphabet. The larger the alphabet, the more information each pulse of light can carry. But there is a catch that has long frustrated researchers: these patterns do not necessarily keep their appearance as they travel. Distortions, turbulence and the physics of propagation can change how a structured beam looks while leaving its underlying identity intact. A receiver that relies on matching the exact appearance of a pattern can therefore fail even when the message itself has survived the journey perfectly well.

The Wits team’s insight was to stop reading the pattern’s appearance and start reading its topology instead. Nape offers an intuitive analogy: imagine a ribbon that has been twisted a certain number of times. If you bend and fold that ribbon into many different shapes, the number of twists remains unchanged no matter what the ribbon looks like. Light can carry analogous hidden features, robust quantities that persist even as the visible structure of the beam is transformed. These topological properties behave like a fingerprint that survives distortion, which makes them attractive carriers of information for any communication system that must operate in the imperfect, noisy real world rather than on a pristine laboratory bench.

Reading topology directly, however, is not straightforward. The conventional approach would be to capture a detailed image of the light pattern and analyse thousands of individual points to infer the hidden quantity, a process that demands heavy computation and introduces delays that undermine the speed advantage of using structured light in the first place. As Nape puts it, creating light patterns that form large alphabets for communications is only useful if they can be read as fast as they can be created. The new system addresses this by embedding the recognition task into the light itself, so that the measurement happens optically, in a single pass, rather than digitally, after the fact.

Technically, the light-powered brain works by passing the incoming structured beam through specially designed layers that transform and direct the light towards an array of detectors. The layers act like the weights of a neural network, but implemented physically in optics rather than in software running on silicon. Where the light lands on the detectors reveals the information it carries. Instead of mapping every crease in what the researchers call the fabric of light to count the twists, the network picks out the features that matter and identifies the topology quickly. Bezuidenhout, who is now an optics engineer at APSI, describes the approach as a simpler way to extract the needed information without first taking a detailed picture, allowing unusual light structures to serve as practical carriers of information for future communication systems.

The training procedure highlights how remarkably efficient the scheme is. The team trained the system on just 15 examples, after which it successfully recognised 81 different light structures within the family studied, including patterns it had never encountered before. That generalisation from a handful of training cases to a much larger set of unseen patterns is the kind of behaviour usually associated with machine learning systems, and here it emerges from the interplay between the physical design of the optical layers and the mathematical structure of the light fields themselves. The result is a classifier that runs at the speed of light, because the classification is performed by light propagating through the layers rather than by a processor crunching pixel data.

Robustness under real-world conditions was a second key test. When the researchers deliberately introduced noise into the detector, the system continued to work, correctly reading out the information despite the degraded measurements. This resilience matters enormously for practical deployment, because any future communication receiver will face imperfect detectors, background light and fluctuating channel conditions. A measurement scheme that only functions in idealised settings would be of limited use; one that tolerates noise brings structured-light communications a step closer to being engineered into working systems, whether for classical high-capacity links or for the quantum communication networks that many laboratories around the world are now trying to build.

The research is also a showcase for interdisciplinary science, combining Wits University’s strengths in artificial intelligence and quantum science, two fields in which the institution leads on the African continent. Professor Benjamin Rosman, Director of the Wits Machine Intelligence and Neural Discovery (MIND) Institute, notes that the work is a wonderful example of what becomes possible when ideas from artificial intelligence and fundamental physics are brought together. By embedding neural network principles directly into the behaviour of light, he says, the team is opening up new possibilities for how information is processed, and the result demonstrates the potential of interdisciplinary research, which is central to what the MIND Institute is building.

For Forbes, who directs the South African Quantum Technology Initiative (SA QuTI) and WitsQ, the breakthrough addresses a major bottleneck on the road to future communications technologies. At the Structured Light Laboratory, researchers are exploring how topology could support a future quantum and classical internet, but creating these light structures is only part of the challenge. Practical ways to read them are equally essential, and that measurement toolbox, as Forbes describes it, is still taking shape. The new work shows how reaching beyond photonics, in this case by borrowing ideas from machine intelligence, can help build that toolbox. It is a reminder that progress in optical communications often depends as much on better measurement as on better sources of structured light.

The team sees the light-powered brain as a beginning rather than an endpoint. Nape, who is the Optica Emerging Leader in Optics Chair and a MIND Institute Fellow as well as being affiliated with the laboratory, says the researchers hope to explore how optical intelligence can contribute to quantum technologies in South Africa. Extracting information from individual particles of light presents even greater measurement challenges than working with full beams, and the group is excited to investigate whether these tools can help tackle them. If they succeed, the same principle demonstrated here, that a physical system can be trained to recognise what matters in a signal without exhaustively measuring it, could shape how future quantum networks, sensors and communication systems extract meaning from light itself.

Subject of Research: Optical neural networks for reading topological information in structured light for communications

Article Title: Wits scientists create ‘light-powered brain’ that could transform future communications

Article References: Wits scientists create ‘light-powered brain’ that could transform future communications. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: structured light, optical neural network, topology, light-powered computing, quantum technology, communications, University of the Witwatersrand, Nature Communications, photonics, machine learning, information encoding, measurement