Most image-processing pipelines throw away information before they ever start computing. A color photograph arrives with three channels — red, green, and blue — and a common first step collapses them into a single grayscale channel, discarding whatever the color itself was saying. For medical images, where a lesion’s hue can be diagnostic, that is an expensive habit. And in photonic computing, where the selling point is processing many channels of information in parallel through light, it is also an odd one: the medium is built for multichannel work, yet most optical neural networks have been confined to a single channel.
Two recent studies from Koç University in İstanbul, both led by Uğur Teğin, attack the problem from complementary directions. The first, published in Communications Engineering in August 2026, is a programmable photonic neural network that processes RGB images natively, keeping the color channels separate and letting them interact optically before any electronic classifier sees the result. The second, published in Communications Physics, takes a different and more radical step: it turns a multimode fiber laser itself into the processor, letting the laser cavity’s own dynamics transform image data into patterns that are easy to classify. Both systems still need cameras and electronic control hardware; neither is an all-optical computer. But both point at the same practical lesson: photonic hardware works best when it learns from its own imperfections rather than from an idealized simulation of itself.
Why color is a harder problem than it looks
Optical computing’s pitch rests on parallelism. A beam of light can carry many independent channels — different wavelengths, different spatial modes, different polarizations — and a single optical element can act on all of them at once. In practice, though, most photonic neural network demonstrations have processed one channel at a time, or have converted color images to grayscale before encoding them optically. That underuses the medium and, for tasks where color carries signal, throws away the very information that makes the task solvable.
The Koç team’s first system, described in “Self-optimizing multichannel optical computing” by Fatma Nur Kılınç and Uğur Teğin, is built around multi-plane light conversion: light passes through a sequence of programmable phase masks, and the interference between planes performs a transformation on the input. The key design choice is that the RGB channels are encoded and mixed optically from the start, so color, shape, and texture features are enhanced together before a camera records the resulting pattern for a simple classifier to read.
The numbers reported are substantial. Across medical imaging, natural image classification, and regression tasks, multichannel optical preprocessing improved accuracy by 25 to 45 percentage points over raw-pixel baselines, according to the paper’s abstract. RGB encoding consistently beat grayscale encoding by 6 to 11 percentage points — a direct measurement of what the discarded color channel was worth. And in a dermatological disease classification task highlighted in the university’s coverage of the work, accuracy rose from 68 percent using raw color pixel data to 98 percent after optical processing. That last figure comes from trade and university coverage rather than the paper’s abstract, so it should be read as the team’s headline demonstration rather than as a peer-reviewed benchmark claim — but it sits comfortably inside the range the paper itself reports.
Learning from the hardware, not from a model of it
The more interesting claim in the first paper is not the accuracy but how it was achieved. The system is described as self-optimizing, and it uses two complementary strategies that share one premise: the optical hardware should be tuned by measuring what it actually does, not by trusting a simulation of what it should do.
The first strategy is Bayesian optimization applied to channel mixing at the input level — in effect, a search procedure that tries different ways of combining the color channels and keeps what works, guided by measured outcomes rather than by gradients computed through a model. The second is a hardware-in-the-loop protocol based on self-organized criticality, used to navigate the high-dimensional phase landscapes of the phase masks. Neither requires a differentiable forward model of the optics, which is the usual prerequisite for training a photonic network by backpropagation.
This matters because real optical hardware is imperfect: lenses slightly misaligned, fabricated elements deviating from design, thermal drift moving things around. A network trained on a perfect simulation degrades when it meets the real device — the simulation-to-reality gap that has frustrated many photonic demonstrations. By learning from experimental measurements, the Koç system adapts to the particular device it runs on, misalignments and manufacturing variations included. The paper reports the self-optimization strategies added a further 7 to 12 percentage points on top of the preprocessing gains.
This is a pragmatic philosophy that generalizes beyond this one system. Photonic hardware will probably never be as precisely reproducible as digital electronics, at least not in the near term. Training methods that treat the physical device as the ground truth — measuring, adapting, and re-measuring — are therefore not a workaround but a requirement. The cost is speed: hardware-in-the-loop optimization is slower than simulation and must be repeated if the device drifts. The benefit is a system that actually works on the hardware it was built for.
A laser that computes
The second study, “Multimode fiber laser cavities as nonlinear optical processors” by Dilem Eşlik, Bahadır Utku Kesgin, Fatma Nur Kılınç, and Teğin, takes the physical-computing idea further. Instead of building a network out of separate optical components, it uses a continuous-wave multimode fiber laser cavity as the processor itself.
The setup is conceptually simple. Input images are encoded as phase patterns on a spatial light modulator and injected into the laser cavity. Inside the cavity, the light undergoes high-dimensional transformation through the interplay of multimode interference — many spatial modes mixing as they propagate — and gain saturation dynamics, the nonlinear process by which the laser’s amplification responds to the light passing through it. The output is a spatially stable intensity distribution that is, in the authors’ phrase, class-separable: images from different categories land in different, distinguishable patterns, so a simple linear classifier can read off the answer.
The results span a deliberately diverse set of tasks: recyclable waste sorting, weather and satellite imagery, retinal scans, and skin lesions. Across these benchmarks, the system achieved 85 to 99 percent accuracy using only that simple linear readout, with orders of magnitude fewer trainable parameters than a conventional deep neural network, according to the paper’s abstract.
This is, in the taxonomy this desk has used before, a reservoir computing system: a fixed, complex physical system transforms the input into a high-dimensional representation, and only a simple readout layer is trained. The reservoir here is not a passive medium but an active laser, and its nonlinearity comes from gain saturation rather than from a separate nonlinear element. That distinction matters for the nonlinearity gap in photonic AI — the problem, described in earlier coverage here, that photonic systems handle linear transformations elegantly but struggle with the nonlinear activation functions that neural networks need between layers. In the fiber laser, the nonlinearity is intrinsic to the device doing the computation. There is no separate activation stage to engineer, and no optoelectronic conversion in the middle of the processing.
The hybrid reality, stated plainly
Neither system is a standalone optical computer, and the researchers do not claim otherwise. Both still rely on electronic components: cameras to record the optical output, spatial light modulators and control electronics to set the input patterns and phase masks, and a digital classifier to produce the final answer. The optical processing replaces a portion of the computation — the feature transformation that would otherwise be done by the early layers of a digital neural network — not the whole pipeline.
That hybrid character is the honest frame for nearly all photonic computing work today, and it is the frame readers should keep when evaluating claims about energy efficiency. Optical preprocessing improved accuracy while shrinking the trainable model to a simple linear readout, which plausibly reduces electronic computation. But neither paper, as reported, provides a system-level energy measurement comparing the full pipeline — laser pump, modulator, camera, control electronics — against a digital baseline. Until such measurements exist, “more energy-efficient” should be read as a direction of travel, not a measured result.
There are also limits to what benchmark accuracy proves. The dermatological, retinal, waste-sorting, and satellite tasks are classification benchmarks on image datasets, not clinical or operational deployments. A system that classifies skin-lesion images at high accuracy in a laboratory is not a diagnostic device, and the step from benchmark to bedside involves validation, regulation, and robustness to data the lab never saw. The waste-sorting and satellite results are more immediately plausible as near-term applications, since they involve controlled imaging conditions and well-defined categories — but they, too, would need to survive real-world variation in lighting, orientation, and sensor quality.
What is genuinely new here
Set against those caveats, two things in this pair of papers deserve attention.
The first is the multichannel point. If photonic computing’s advantage is parallelism, then processing color natively rather than collapsing to grayscale is not a minor feature — it is the architecture using the medium as intended. The measured 6-to-11-point advantage of RGB over grayscale gives that intuition a number, and it suggests that future optical systems should be designed multichannel from the start rather than retrofitted.
The second is the self-optimizing, hardware-in-the-loop methodology. The photonic computing field has produced many demonstrations that work beautifully in simulation and degrade on real hardware. A system that learns from its own measurements, and that improves by 7 to 12 points by doing so, is a methodological contribution as much as a device contribution. It also lowers a barrier to deployment: if the system can calibrate itself to its own imperfections, it does not need the manufacturing precision that has kept photonic prototypes in the lab.
The fiber laser paper adds a third point: the processor is a commodity-like device. A multimode fiber laser is not an exotic nanofabricated chip; it is fiber-optic hardware with an established manufacturing base. If laser cavities can serve as compact, low-power physical processors, the path from laboratory demonstration to deployable device may be shorter than for platforms requiring custom silicon photonics fabrication — though that claim still needs system-level verification on energy and long-term stability.
What to watch
Three questions will determine whether this work matters beyond the laboratory.
First, stability and calibration over time. A system that adapts to its own imperfections must keep adapting as the device ages, warms up, or is moved. How often must the self-optimization run, and how much does performance degrade between calibrations? The papers establish that hardware-in-the-loop learning works; they do not yet establish a maintenance schedule.
Second, system-level energy and latency. The field needs a full-pipeline measurement — laser, modulator, camera, control electronics, and digital readout — compared against a digital baseline on the same task. Until that exists, efficiency claims remain plausible but unproven, and readers should treat them that way.
Third, scale and integration. Both demonstrations are free-space and fiber laboratory setups. The interesting question is whether the multichannel architecture and the laser-cavity processor can be miniaturized into chip-integrated optics, which the researchers themselves identify as the direction that could shift more computation onto photonic hardware. Faster light-control devices — modulators that can update patterns at higher rates — are part of that story too, since the speed of the whole system is limited by its slowest electronic interface, not by the speed of light in the cavity.
For readers following photonic AI, the Koç results are best understood as proofs of concept strengthening the same argument from two sides: light can handle richer inputs than the field has been giving it, and physical devices can do useful computation when allowed to learn from their own behavior. Neither paper claims to replace electronic AI hardware. Together, they make the hybrid case — optics for the transformations it does well, electronics for control and readout — a little more concrete, and a little harder to dismiss.
Sources and further reading
This article is based on two peer-reviewed papers and university and trade coverage of them. Kılınç, F.N., and Teğin, U., “Self-optimizing multichannel optical computing,” Communications Engineering (2026), published 21 August 2026, DOI 10.1038/s44172-026-00762-w. Eşlik, D., Kesgin, B.U., Kılınç, F.N., and Teğin, U., “Multimode fiber laser cavities as nonlinear optical processors,” Communications Physics (2026), DOI 10.1038/s42005-026-02823-0. Both papers acknowledge support from the Scientific and Technological Research Council of Türkiye (TÜBİTAK) under grant 122C150. Coverage: Koç University via AlphaGalileo, “A New Approach to Sustainable AI From Assistant Professor Uğur Teğin,” and Tech Xplore / ENGtechnica, “Light Takes On AI’s Growing Computing Burden,” 13 October 2026. The 68-to-98-percent dermatological figure and the task list for the laser system are as reported in that coverage; the percentage-point ranges (25–45, 6–11, 7–12) and the 85–99-percent range are from the papers’ abstracts. For background on the nonlinearity gap in photonic AI accelerators and on reservoir computing, see the earlier articles in this section, “Photonic AI Accelerators: Where Excitonic Devices Fit” and “Neuromorphic Photonics: Spiking With Light.”