One Beam of Light, Four AI Jobs at Once

Diffractive optical neural networks have a versatility problem. A diffractive network is, in effect, a fixed piece of optics: light passes through engineered layers, interference does the computation, and the answer lands on a camera. That fixed character is the source of the speed — computation at the speed of light propagation, with no clock cycles — and also the source of the limitation. One set of optics has historically meant one job. Want the same platform to recognize handwritten digits and classify clothing images? Until now, the standard answers were to multiplex — give each task its own wavelength or polarization — or to reconfigure the hardware mechanically between tasks. Both add components, cost, and alignment fragility: exactly the things that keep diffractive systems in the laboratory.

A paper from a team led by China Information Communication Technologies Group Corporation (CICT), with Peng Cheng Laboratory, Wuhan University, and Nanyang Technological University, proposes a different answer: do not add hardware at all. Shape the beam. Published in Opto-Electronic Technology as “Neural holography enabled multi-beam steering for multi-task classifier,” the work builds an end-to-end Holographic Neural Network (HNN) around a single phase-only spatial light modulator (SLM) that both encodes the input data and acts as the diffraction layer, using a phase-multiplexing scheme to steer independent wavefronts — one per task — to different regions of the camera sensor. The team’s shorthand is “one brain for multiple uses.” The claim that matters is not the slogan but the mechanism: multi-task capability achieved through beam shaping rather than more hardware.

The deployment objection this attacks

In a diffractive optical neural network (DONN), computation is performed by light propagation itself: as a wavefront passes through engineered diffractive layers, each point of the wave interferes with every other point, performing what amounts to a massive parallel matrix multiplication in a single pass. There is no transistor switching, no memory fetch, no clock. The result is inference at propagation speed with, in principle, far lower energy than electronic operations.

The catch is that the optics are task-specific. A diffractive layer is a physical object whose phase profile encodes one trained transformation. In the conventional multi-task playbook, serving several tasks from one platform means either assigning each task a distinct physical channel — a wavelength band, a polarization state — which demands additional lasers, filters, and precise tuning, or physically reconfiguring the system between tasks, which means moving parts and realignment. Each route increases hardware cost and system complexity, and each makes the alignment procedures that already plague diffractive systems more delicate. The CICT paper attacks a practical problem this desk has circled from other directions — the Koç work on learning from measured hardware, and the Shanghai B-ONN work on training without a reciprocal backward path — by asking what happens when the customer wants one box to do more than one thing.

That is a deployment objection, not a benchmark objection, and it is worth naming as such. A photonic accelerator that must be rebuilt, retuned, or realigned every time the workload changes is not an accelerator; it is a demonstration rig. The question the CICT team set out to answer is whether a single fixed optical architecture can carry several inference jobs concurrently, with the task separation done by the light itself.

One modulator, two jobs

The architecture’s elegance lies in what it eliminates. Conventional DONNs use cascades of physical diffractive layers. The HNN uses one device: a single phase-only SLM that plays two roles at once. At the input layer, it encodes electronic information onto light by imposing phase patterns on an incoming wavefront. In its computing role, it acts as the diffraction-based modulation layer, sculpting the light as it propagates. There are no cascades of fabricated layers to align, no wavelength or polarization multiplexing to separate tasks, and no moving parts.

The multi-task capability rests on what the authors call phase multiplexing, inspired by the cross-stitch architecture from multi-task machine learning. In the cross-stitch design, separate task-specific networks share information through learned linear combinations of their activations — a controlled leak between tasks, tuned so related tasks help each other and unrelated tasks do not interfere. The holographic analogue superimposes task-specific phase patterns onto a single modulated wavefront, and a joint training algorithm learns how to allocate the optical resources among them. The wavefronts for different tasks travel the same optical path and are steered to different regions of the camera sensor, where the camera directly reads out the classification results with no intermediate electronic processing before the answer appears.

The crucial control knob is the task correlation factor, introduced during error backpropagation. It governs how strongly the learning of one task influences another: set it one way, and related tasks share representations and reinforce each other; set it another, and unrelated tasks are kept apart so their computations do not contaminate each other. Crosstalk — unwanted interference when multiple computations share the same optical channel — is the failure mode of any multiplexed scheme, and the correlation factor is the paper’s answer to it. The team also adopted an end-to-end optoelectronic joint training framework in which the input encoding and the holographic diffraction modulation are optimized together, and the paper identifies the nonlinear characteristics of the electrical-to-optical conversion at the input layer as important to overall performance — a detail worth noting, because the nonlinearity gap is one of photonic computing’s persistent problems, and here an existing nonlinearity is exploited rather than engineered around.

A second practical consequence follows from the architecture’s simplicity. Because the optical system is low in complexity — one SLM, one propagation path, one camera — offline simulation training aligns closely with experimental inference results. That matters: the simulation-to-reality gap has frustrated many photonic demonstrations, and a system whose simulated performance transfers to hardware with minimal degradation is one that can plausibly be trained digitally and deployed optically — which is how any real product would have to work.

What the results actually show

The experimental results, as reported in the journal announcement and the trade coverage, fall into three claims of increasing ambition.

First, synergy. When two related tasks are trained together — the coverage describes two different sets of handwritten digits — the multi-task network surpasses the accuracy of comparable single-task networks. Shared optical computation is not merely tolerated; it is genuinely useful, because related tasks learn shared representations. This is the multi-task learning promise in its classic form, and it is notable that it survives translation into optics, where the shared representation is a physical wavefront rather than a layer of artificial neurons.

Second, suppression. For unrelated task types — handwritten digit recognition alongside fashion item recognition — optimizing the task correlation factor suppresses the mutual interference that would otherwise degrade performance. The coverage gives no per-task accuracy breakdown for this pairing, so the claim as reported is directional: tuning the correlation factor controls crosstalk well enough for unlike tasks to coexist.

Third, scale. A single holographic neural network simultaneously carried out four distinct recognition tasks while maintaining a classification accuracy of approximately 90 percent, all within the same fixed optical architecture. That is the headline figure, and it should be read with discipline. Approximately 90 percent across four concurrent tasks, on recognition benchmarks, is a strong result for a single fixed optical system — but the coverage does not specify which four tasks, does not break out per-task accuracy, and does not report the datasets by name. The honest statement is that four concurrent recognition tasks ran at roughly 90 percent; anything more specific would be inventing detail the sources do not contain.

It is also worth stating what is not in the record. The coverage carries no power consumption, latency, or energy-per-inference figures, and no system-level energy comparison against a digital baseline. The journal announcement describes “ultra-high-speed and low-energy information processing” in general terms, but general terms are not measurements. Until a full-pipeline figure exists — laser, modulator, camera, control electronics, and readout, against a digital baseline on the same tasks — the energy case for this system is a direction of travel, not a result.

Not the B-ONN story

Readers of this desk may notice a surface resemblance to the backpropagation-free photonic chip covered yesterday, and the distinction is worth drawing, because the two papers solve different problems. B-ONN, from the University of Shanghai for Science and Technology, is a training algorithm: it replaces the backward pass with local, optically computed targets so a photonic network can learn without a precise reciprocal optical path. The CICT paper is an inference architecture: it assumes a trained network and asks how one fixed optical system can serve several tasks at once. B-ONN’s innovation is in how the network learns; the HNN’s innovation is in how the light is shaped after learning. They are complementary, not competing.

There is also a useful contrast in where each paper’s risk sits. B-ONN’s risk is algorithmic: whether local learning scales beyond small benchmarks and shallow printed chips. The HNN’s risk is architectural and operational: whether beam-steered task separation holds up as tasks multiply, as tasks grow more complex than image recognition, and as the system leaves the bench for environments with vibration, thermal drift, and dust. The team’s own framing points that way: based in the Optoelectronic Devices Research Department of OCTN and NOEIC, they describe the classifier as foundational for next-generation intelligent beam control — platform groundwork, not a finished product.

The honest ledger

Several limitations deserve equal weight, because they define what this paper does not establish.

First, the evidence base available to this article is thinner than for the papers covered earlier this month. The numbers here come from the journal’s announcement and trade coverage of it, not from the paper’s full text, which this desk could not retrieve at writing time. The DOI is real and the journal is peer-reviewed, but a reader who wants per-task accuracies, dataset names, sample counts, or hardware specifications will need the paper itself.

Second, the demonstrations are recognition benchmarks. Image classification is the right first proving ground for a multi-task optical classifier — it is where the task-separation question can be asked cleanly — but it is not evidence about language, reasoning, or any workload beyond classification. The jump from “four recognition tasks at roughly 90 percent” to “a general-purpose optical accelerator” is not supported by anything in the record.

Third, the scaling question is open in both directions. Four tasks at roughly 90 percent is the reported point; the record does not say what happens at eight tasks, or whether accuracy degrades gracefully or collapses as the sensor is subdivided among more wavefronts. Sensor real estate is finite: steering each task to its own region of the camera means each task gets a smaller patch, and there must be a point where subdivision costs more than multiplexing gains.

Fourth, the energy and latency case is unmeasured in the available record, as noted above. A system with no moving parts and no laser retuning is plausibly cheaper to operate than a wavelength-multiplexed one, but plausibility is not a measurement.

What to watch

Three developments would indicate this work is becoming more than a laboratory result.

First, the paper itself, with per-task accuracies, dataset specifications, and hardware details — laser wavelength, SLM resolution, camera specifications, sample counts — so the four-task figure can be checked rather than quoted. The DOI is 10.29026/oet.2026.260019; read the full text before citing any number from this article in a technical context.

Second, a system-level energy and latency measurement: the full pipeline, laser to camera to readout, against a digital baseline on the same four tasks. That is the number that would let a data-center operator evaluate the claim that optical processors could sit alongside electronic accelerators on high-throughput inference workloads. Until it exists, the deployment case rests on architectural simplicity — real, but incomplete.

Third, task count and task diversity beyond recognition. If the phase-multiplexing scheme holds accuracy as tasks are added, and if it extends beyond image classification, then beam shaping becomes a genuine alternative to multiplexing as the standard answer to the multi-task problem. If accuracy falls off quickly past four tasks, it is a useful niche technique for small task bundles rather than a platform shift. Either outcome would be informative; the difference between them is the difference between a better diffractive classifier and a new default architecture.

The idea at the center of this paper is a small one, stated plainly: when the hardware cannot change, change the light. Diffractive optics has spent years adding components to solve the multi-task problem — more wavelengths, more polarizations, more moving parts — and each addition made the system harder to build and align. This team subtracted instead, down to a single modulator, and asked the beam to do the work of separating tasks. Whether that subtraction survives contact with larger task sets and real operating environments is unproven. But it reframes the problem in a way the field should take seriously: the versatility of an optical computer may come not from more optics, but from better control of the one beam it already has.

Sources and further reading

This article is based on the peer-reviewed paper: Wu, L., Wu, M., Li, Z., Qiu, Y., Liu, Z., Zeng, Y., Liang, G., Li, H., Zheng, G., He, Z., Xiao, X., Wang, Q. J., Tao, J., and Yu, S., “Neural holography enabled multi-beam steering for multi-task classifier,” Opto-Electronic Technology 2, 260019 (2026), DOI 10.29026/oet.2026.260019. The team is led by China Information Communication Technologies Group Corporation (CICT) — specifically the Optoelectronic Devices Research Department of the State Key Laboratory of Optical Communication Technologies and Networks (OCTN) and the National Information Optoelectronics Innovation Center (NOEIC) — with Peng Cheng Laboratory, Wuhan University, and Nanyang Technological University. Corresponding authors Jin Tao and Shaohua Yu. The full paper text could not be retrieved at writing time, so all figures in this article are as reported in the journal announcement (AlphaGalileo, 13 October 2026) and trade coverage (bioengineer.org, 14 October 2026): the single phase-only SLM dual role, the cross-stitch-inspired phase-multiplexing scheme with a tunable task correlation factor, related-task synergy exceeding single-task accuracy, crosstalk suppression for unlike tasks, and four concurrent recognition tasks at approximately 90 percent accuracy. No power, latency, or energy-per-inference figures appear in the available coverage, and none are claimed here. For the contrasting training-side approach, see this desk’s coverage of the Shanghai team’s backpropagation-free B-ONN chip; for the hardware-in-the-loop response to the simulation-to-reality gap, see the Koç University multichannel and fiber-laser coverage.