An optical neural network reads the hidden geometry of light

Scientists at the University of the Witwatersrand have built an optical neural network that sorts information carried by structured light. The university described the work on 13 October: specially configured optical layers direct an incoming beam towards detection regions, allowing its relevant features to be read from where the light lands. The calculation takes place partly in the propagating light itself.[1]

Imagine a receiver asked to recognise a symbol after its appearance has changed. One approach is to recover every detail of the picture and then interpret it. Another is to find a property that identifies the symbol across several appearances. The attraction of the Wits approach is that it makes this second kind of question physical: what must the receiver actually measure?

The information is in the arrangement

Polarisation describes how a light wave’s electric field moves at a particular position. Across a structured beam, that behaviour can vary from point to point. Physicists can represent each local polarisation state on a sphere; the beam then becomes a map onto that sphere. For the optical skyrmions studied here, the skyrmion number counts how many times this mapping covers the sphere, including its orientation. It describes the organisation of the polarisation field across the beam.[4]

That is why two different-looking patterns can share a topological class. Smoothly changing a pattern need not change its number. Measuring the number, however, is delicate: a conventional reconstruction samples the field over an area, and noise or a truncated measurement region can spoil the calculation. Robust information and a robust way of reading it are separate engineering problems.[4]

A useful everyday comparison is sorting handwritten characters. Recording every contour gives more information than identifying the letter. The analogy concerns the choice of question, not the physics: light’s topological classes have mathematical definitions, and ordinary handwriting does not gain that protection.

How a learned pattern becomes a physical calculation

The broader idea of a diffractive neural network predates this experiment. In a 2018 Science paper, researchers demonstrated optical inference using layers designed through machine learning. Each layer changes the passing wave, and the wave spreads before reaching the next one. Contributions from different locations interfere, reinforcing some outputs and cancelling others. Training selects the layer settings that produce useful output patterns.[5]

This makes the learned settings part of the optical system. The light is not carrying out a software instruction one line at a time; the configured wave interaction produces the transformation as the beam passes through. Training and preparing the layers remain distinct from using a trained network.[5]

Eighteen ingredients, eighty-one combinations

The Wits detector uses two polarisation channels, each with five phase layers. Each channel is trained on nine modes with different orbital angular momentum and differently winding wavefronts: 18 basis modes altogether. Pairing the two sets produces 81 input configurations spanning 15 topological classes. The network identifies the component modes, then uses their known relationship to infer the skyrmion number. This is a deliberately structured classification problem, with the physics doing part of the interpretive work.[2]

In a detector-noise experiment, selecting the strongest output in each channel correctly labelled all 81 inputs even when added incoherent white light reached approximately 95% of the average signal intensity. Noise was added at the camera; atmospheric turbulence was not tested. The paper also demonstrates image reconstruction from topological symbols.[2][3]

Its optical power measurements do not provide a complete system energy comparison: the laser, camera, programmable optical hardware and control electronics also consume electricity. The published work is a laboratory detector demonstration, with the communications performance of an installed system still to be established.[2]

A receiver designed around the question

Wits researchers frame the work as part of a measurement toolbox for future classical and quantum communications. Their published account identifies single-photon information extraction as a further direction for investigation. The immediate contribution is a way to read selected information from complex light fields without rebuilding every feature of those fields first.[1]

That design choice is the interesting lesson. A camera image may be a convenient intermediate representation, but it is not automatically the best one for every task. When the question is precisely defined, some of the interpretation can be built into the measurement. Here, the reader and the information carrier begin to speak the same physical language.

Report an error or suggest a correction