Researchers at the University of the Witwatersrand have developed an innovative optical neural network capable of reading information embedded in complex light patterns without capturing traditional, detailed images. Hailing from the Wits Structured Light Laboratory, the breakthrough introduces a streamlined method for processing high-capacity data streams and improving the resilience of optical communications against environmental interference.
Read: ASUS ROG Zephyrus Duo Review: Two-screen titan
Modern high-speed communications increasingly look toward structured light, where laser beams are manipulated into intricate spatial patterns acting much like an expanded alphabetic system. Because individual patterns can represent distinct symbols, a single pulse of light can theoretically carry vastly more data than conventional telecommunication signals. However, a major engineering bottleneck has persisted: receivers could only decode these patterns by capturing high-resolution images and analysing thousands of individual data points, a process too slow to match generation speeds.
The Wits research team bypassed this limitation by creating a hardware-level optical neural network. Instead of mapping every micro-feature of a light beam, the system utilizes a pair of channels containing five diffractive layers, optical patterns that bend light deterministically to steer it toward specific detector spots. The physical landing point immediately reveals the underlying data code. By designing the network with a compact parameter set rather than pixel-by-pixel processing, the team reduced training overhead by a factor of roughly one thousand.
The system focuses on reading “optical skyrmions,” topological light structures defined by an unchanging whole-number property called a skyrmion number. Much like a twisted ribbon retains its core twist count regardless of how it is bent or folded, skyrmion numbers remain intact even when light beams undergo severe distortion. This property makes the approach exceptionally well-suited for open-air laser links, such as inter-building connections or satellite communications, where atmospheric turbulence typically scrambles traditional signals.
Laboratory testing demonstrated high reliability under adverse conditions. The neural network successfully classified 81 different structured patterns spanning 15 values with approximately 93 percent accuracy. Furthermore, when subjected to intense background interference by mixing incoherent white light into the signal, the system maintained complete accuracy until noise levels reached 95 percent of the signal strength, a threshold far exceeding the limits of conventional Stokes polarimetry measurement techniques. The researchers also successfully utilized the framework to transmit and recover visual imagery encoded via a 15-letter topological alphabet.
Led by MSc student Hadrian Bezuidenhout under the supervision of Isaac Nape and Andrew Forbes, the study was published on 7 October in Nature Communications. The project reflects extensive cross-disciplinary collaboration, drawing on principles from artificial intelligence and fundamental physics through affiliations with the Wits Machine Intelligence and Neural Discovery (Mind) Institute and the South African Quantum Technology Initiative (SA QuTI).
While the current setup remains a laboratory proof-of-concept operating at an optical throughput of roughly 15 percent due to component losses, university leadership emphasizes its broader strategic importance. As academic spin-out ventures like Button Optics begin commercializing quantum-inspired toolkits, the team aims to extend the light-powered neural network architecture into quantum computing and secure communication domains, tackling the measurement challenges inherent in processing information from individual photons.
