Koç University Photonic AI Offloads Image Processing to Light, but the Energy Case Is Still Open

Koç University photonic AI researchers have presented two optical systems that offload image processing to light, despite relying on electronics around the optical core. One preserved color information while reaching 98% accuracy on a skin-image benchmark. The other used a fiber laser to obtain accuracies between 85% and 99% across several data sets.

The results give sustainable AI researchers a more concrete alternative to simply building larger electronic accelerators. Yet they do not establish that either system consumes less total energy than a modern GPU deployment. Cameras, light modulators, lasers, control electronics, and digital classifiers remain part of the workflow.

That gap defines the real story. Koç University has shown that physical optical systems can perform useful transformations instead of merely transmitting data. The next contest is not light against electricity in isolation. It is specialized photonic processing against complete, highly optimized electronic computing systems.

Koç University Photonic AI Moves Two Workloads Into Light

The research shifts a meaningful processing stage from software into physical optical hardware, but it does not replace the entire AI pipeline.

The reported research comes from assistant professor Uğur Teğin and students at Koç University in Istanbul. It combines two separate studies published in Communications Engineering and Communications Physics.

The first system addresses multichannel data, especially color images. Many optical machine-learning experiments simplify their input by converting color images into grayscale. That reduces the information passing through the optical system before classification begins.

Teğin and master’s student Fatma Nur Kılınç took a different route. Their programmable photonic neural network keeps the red, green, and blue channels and processes them together.

A photonic neural network is hardware that uses controlled light propagation to perform mathematical transformations associated with machine learning. It does not necessarily reproduce every layer of a digital neural network.

The researchers encoded data onto a programmable spatial light device. Light then passed through an optical arrangement that mixed and transformed the channels. A camera captured the resulting pattern, which supplied features to a simple digital classifier.

This distinction matters. The optical equipment performs feature extraction, meaning it converts raw input into a representation that separates categories more clearly. The final decision still happens electronically.

According to the multichannel study, the architecture also learns from physical measurements instead of depending entirely on an idealized simulation. That helps it adjust to misalignment, optical aberrations, manufacturing variation, dust, and thermal drift.

Those imperfections often separate a convincing photonics simulation from a usable device. An optical model can perform well on paper while the physical version behaves differently. Direct hardware optimization makes that mismatch part of the learning process.

On the HAM10000 skin-lesion data set, the researchers reported 98% classification accuracy after optimized RGB processing. The preprint’s detailed results show a 67% raw-pixel baseline, producing a 31 percentage-point improvement.

The system also reached 90% on the STL-10 natural-image benchmark after channel mixing optimization. RGB processing consistently exceeded grayscale processing by five to six percentage points under the same optical architecture.

The second study moved a different operation into physics. Teğin, Dilem Eşlik, Kılınç, and Bahadır Utku Kesgin used a continuous-wave multimode fiber laser as a nonlinear optical processor.

Nonlinearity lets a processor create relationships more complex than a direct weighted sum. Modern neural networks depend on nonlinear transformations to separate patterns that simple linear methods cannot distinguish.

The researchers encoded images as phase patterns before sending them into the laser cavity. Multimode interference and gain saturation transformed the input into stable intensity patterns. A camera recorded those patterns for a linear classifier.

That setup reached 85% to 99% accuracy across benchmarks involving recyclable waste, weather, satellite imagery, retinal scans, and skin lesions. The fiber-laser paper says it required orders of magnitude fewer trainable parameters than deep neural networks.

The two systems therefore tackle complementary problems. One preserves and optimizes multiple information channels. The other supplies nonlinear feature transformation through the behavior of a laser cavity.

Together, they show that optical hardware can do more than accelerate a predetermined matrix operation. It can shape information into a form that makes a downstream classification task easier.

Why Sustainable AI Hardware Needs More Than Faster GPUs

The pressure behind photonic AI comes from electricity, heat, and data movement, not from a sudden failure of electronic chips.

AI infrastructure remains dominated by graphics processing units and other electronic accelerators. These systems are programmable, supported by mature software, and manufactured through established supply chains.

Their weakness appears when computation grows across larger models and clusters. Moving data between memory, processors, and network interfaces consumes energy. Electrical resistance also creates heat, which adds cooling requirements.

The International Energy Agency estimated that data centers consumed 415 terawatt-hours of electricity in 2024. That represented about 1.5% of worldwide electricity consumption.

Its energy outlook projects roughly 945 terawatt-hours by 2030. AI is the largest driver of that growth, although other digital services also contribute.

The forecast does not mean AI alone will consume the entire projected total. It does show why hardware efficiency now affects grid planning, data-center construction, operating costs, and deployment schedules.

The IEA also estimates that a typical AI-focused data center uses as much electricity as 100,000 households. The largest facilities under construction could use 20 times that amount.

Local constraints can arrive before global electricity consumption becomes dominant. Data centers cluster around available fiber, land, customers, and power infrastructure. New transmission lines often require four to eight years in advanced economies.

Photonic processing targets part of this pressure by exploiting the physical behavior of light. Photons have no electrical charge, so moving them through an optical structure avoids some resistive losses found in electronic circuits.

Light also supports parallelism across wavelength, space, phase, and polarization. A single optical path can carry several information channels without processing each channel through an identical sequence of electronic operations.

Those properties make image data attractive for early photonic systems. The original signal is already optical before a camera sensor converts it into electronic pixels. Processing some information before, or near, that conversion can reduce later digital work.

Possible settings include machine vision, medical imaging, industrial inspection, remote sensing, and autonomous systems. Each needs fast decisions from visual inputs, sometimes under strict power or latency limits.

However, greater optical parallelism does not automatically produce a lower system-level energy bill. Lasers require electricity. Spatial light modulators need control signals. Cameras must capture outputs, while electronic hardware prepares inputs and interprets results.

Koç University’s two prototypes should therefore be viewed as workload offloading systems. They transfer selected transformations from digital software into optical physics.

That approach resembles the wider history of accelerators. GPUs did not eliminate central processors. They took over operations that matched their parallel design. Specialized AI chips later moved additional matrix calculations into purpose-built hardware.

Photonic accelerators seek another division of labor. They aim to handle operations where parallel light propagation offers a meaningful advantage. Electronic processors would keep control flow, memory management, general software, and tasks requiring high numerical flexibility.

The strongest sustainability case will depend on total workload energy, not the efficiency of the optical step alone. Researchers must count data preparation, conversions, light generation, detection, cooling, and supporting electronics.

A laboratory apparatus can still matter before those measurements exist. It can identify which operations map cleanly onto light and which hardware problems block deployment.

Koç University’s contribution lies there. Its systems broaden the set of useful optical transformations while confronting real hardware imperfections. They do not yet settle the data-center energy debate.

How Koç University Photonic AI Actually Processes Images

The mechanism works by turning light propagation into feature extraction, then handing the easier classification problem back to electronics.

The multichannel system begins with an image or another multidimensional input. Software maps that data onto a programmable optical device, which controls how different portions of a light field propagate.

The device preserves the RGB channels rather than collapsing them into one grayscale representation. That gives the optical system access to color relationships alongside shapes and textures.

Light passes through multiple controlled transformations. Interference combines waves according to their amplitudes and phases, producing an output pattern that reflects many input relationships at once.

A camera then measures the intensity pattern. Those measurements become numerical features for a ridge classifier, a relatively simple linear model with regularization.

The classifier is not responsible for discovering all useful image structure by itself. The optical transformation has already rearranged the data so categories become easier to separate.

The researchers used Bayesian optimization to adjust six channel-mixing parameters. Bayesian optimization selects promising configurations based on the results of earlier experiments, reducing the need to test every combination.

They also applied a hardware-in-the-loop method inspired by self-organized criticality. This process perturbed a much larger optical phase space and evaluated changes using measurements from the real apparatus.

That second optimization handled 22,500 parameters, according to the study preprint. It converged in roughly 75 iterations on one reported task.

The two methods address different layers. Channel mixing changes how information enters the system. Hardware optimization changes the physical phase configuration through which the light travels.

Each delivered another six to seven percentage points on its respective experiments. That similarity is notable because one method adjusts only six variables, while the other explores thousands.

It suggests that careful input structure can matter as much as a much larger hardware search. It also warns against treating raw optical parallelism as sufficient by itself.

The fiber-laser system uses a different physical effect. A multimode fiber supports many spatial patterns of light at the same time. Inside a laser cavity, those modes interact through interference and gain dynamics.

An input image is encoded as a phase pattern. As that pattern circulates, the cavity maps it into a higher-dimensional distribution of light intensity.

Gain saturation supplies the important nonlinear response. As optical amplification approaches its limit, the output no longer changes in direct proportion to the input.

This nonlinear behavior helps separate complex categories. The resulting pattern can make an image class accessible to a linear decision boundary, even when raw pixels are not easily separable.

The research team used a continuous-wave laser rather than ultrashort, high-power pulses. That choice supports its argument for a more compact and potentially lower-power physical processor.

Still, the published accuracy values need careful interpretation. Accuracy depends on the data set, training split, preprocessing, class balance, and comparison model.

A 98% result on one medical-image experiment does not mean the apparatus diagnoses skin disease with 98% reliability in clinical practice. The work evaluates a machine-learning benchmark, not a deployed medical product.

The first study also compares optical features with raw pixels processed by the same simple classifier. That controls for classifier capacity, but it is not the same comparison as a complete modern vision model.

The authors acknowledge that direct comparison with deep learning is complicated by differences in model capacity, training data, and computational cost. Their main claim concerns optical feature extraction, not universal superiority over digital models.

This is the appropriate frame for the reported results. The optical systems produced valuable representations with limited digital training. They did not defeat every electronic alternative on accuracy, energy, cost, and scale.

The mechanism also reveals where integration work must happen. Data begins in an electronic system unless it comes directly from an optical sensor. It must be encoded into light, processed, detected, and returned to an electronic workflow.

Every boundary adds latency, energy consumption, calibration demands, and potential information loss. A successful product must gain more inside the optical stage than it spends crossing those boundaries.

Conversion Losses Keep Photonic AI Behind General-Purpose GPUs

The most important unresolved question is whether the complete optical system beats optimized electronics after every conversion and control component is counted.

Photonic computing often looks compelling when one optical operation is isolated. Light can propagate quickly, combine many channels, and perform some mathematical operations with little incremental energy.

A deployed computer cannot isolate that operation. It needs memory, software, input encoding, output detection, error management, and interfaces with existing processors.

Maxim Shcherbakov, an electrical engineering professor at the University of California, Irvine, identified electrical-to-optical conversion as a central obstacle. Lasers must be modulated with electronic signals, while detectors convert the result back into electronic data.

His photonics assessment argues that general-purpose photonic processors still trail standard GPUs. He sees stronger near-term prospects in specialized edge applications.

That judgment fits the Koç University experiments. Both systems target structured classification tasks. Neither runs a general large language model or replaces a complete GPU software stack.

Edge systems may offer a cleaner entry point because some inputs already arrive as light. A camera, microscope, medical imager, or remote-sensing instrument can potentially process optical information before creating a full electronic representation.

That reduces the number of conversions. It can also cut data volume before information reaches a general processor, memory system, or network connection.

A visual sensor might identify useful features near the lens, then transmit only a compact representation. The value would come from lower latency and reduced downstream work, not from recreating a GPU in glass.

Cloud data centers present a harder standard. Their hardware must support many models, tensor sizes, numerical formats, and rapidly changing software frameworks.

Photonic hardware often gains efficiency by specializing. Fixed optical behavior can be extremely fast for a matching operation but difficult to repurpose when an algorithm changes.

Programmable modulators improve flexibility, yet they add control complexity. Calibration can drift with temperature, vibration, component aging, or small mechanical changes.

Koç University’s hardware-based optimization directly addresses some of those issues. Learning from measurements lets the system absorb imperfections that an ideal simulation misses.

That is an engineering advantage, but continuous recalibration also consumes time and energy. A commercial evaluation must measure how often adaptation occurs and what happens between calibration cycles.

Precision is another pressure point. Electronic accelerators support well-defined numerical formats and mature error analysis. Analog optical measurements contain noise and device variation.

Some machine-learning workloads tolerate reduced precision. Others amplify small errors across repeated operations, especially when many optical and electronic stages are connected.

The published experiments use an optical transformation followed by a digital classifier. That limited structure reduces the risk of errors accumulating through many optical layers.

Scaling the approach could introduce new tradeoffs. More channels might increase throughput, but they can also create crosstalk. Larger systems need more detectors, control elements, packaging, and alignment.

Manufacturing will matter as much as laboratory physics. A useful accelerator must be repeatable across units, stable over time, compatible with production processes, and serviceable inside a computing system.

Energy evidence remains the largest gap in the current announcement. The fiber-laser paper describes its processor as compact and low-power, while the multichannel paper describes potential efficiency advantages.

Neither public summary provides a complete, independently validated energy comparison against a matched GPU workload. There is no full accounting for the source, modulator, camera, controller, and classifier.

That absence does not invalidate the research. It limits the conclusion readers should draw from it.

The systems demonstrate computation with light and report strong classification results. They do not yet demonstrate lower end-to-end energy per correctly classified sample under commercial conditions.

The primary opponent is therefore not another photonics laboratory. It is the established electronic stack, including its software maturity, manufacturing scale, flexibility, and continuing efficiency improvements.

Photonic AI wins only when it beats that complete stack on a valuable workload. A faster optical kernel is an opening, not the final result.

Three Signals Will Show Whether Optical Computing Can Scale

The next stage requires system-level energy measurements, integrated hardware, and credible tests on workloads beyond small laboratory classifiers.

The first signal is an end-to-end energy benchmark. Researchers should report joules per inference or per processed sample for the entire apparatus.

That measurement must include the laser, optical modulation, camera, digital control, preprocessing, calibration, and downstream classifier. The electronic comparison should perform the same task at a similar accuracy and latency.

If Koç University or an independent team publishes such a result, the sustainable AI claim will become much stronger. If optical savings disappear after conversions, the case will narrow to latency or specialized sensing.

Measurement boundaries must also be explicit. Reporting only the energy used during light propagation would omit components that make the processor operational.

The second signal is physical integration. Both reported systems use laboratory optical arrangements with cameras and control equipment. The researchers identify compact, chip-integrated devices as an important route forward.

Integration can shorten optical paths, reduce alignment sensitivity, and lower packaging requirements. It can also make the technology easier to combine with sensors and electronic processors.

A successful integrated prototype would not need to replace a GPU. It could demonstrate a stable optical front end that plugs into an existing machine-learning workflow.

Watch for operation outside a controlled laboratory, particularly under changing temperatures or vibration. Long-term accuracy and calibration frequency will reveal whether self-optimization solves practical drift.

The third signal is broader workload validation. The current studies cover several image and structured-data benchmarks, which is useful evidence of versatility.

The next tests should include streaming sensor data, larger inputs, latency measurements, and comparisons with modern digital feature extractors. Medical or industrial trials would also need representative data and independent evaluation.

A deployed edge experiment would strengthen the argument more than another isolated accuracy record. It would show how the optical processor interacts with memory, communications, power limits, and changing input conditions.

The likely path is hybrid rather than fully optical. Light will handle selected transformations that match its physics, while electronics retain memory, orchestration, and general-purpose decisions.

That division can still matter. Removing one expensive stage from millions of repeated operations can reduce heat and energy even when the rest of the system remains electronic.

It also changes how developers should assess emerging AI hardware. Peak optical throughput says little without integration costs, numerical accuracy, utilization, and software support.

Enterprise buyers should ask whether a device accelerates their actual workload, not whether light is theoretically faster than electrons. They should also demand measurements at the wall plug.

Developers and researchers face a documentation problem as this field expands. Results often span preprints, journal papers, hardware notes, benchmark details, and later corrections.

A searchable knowledge base can help engineering teams preserve those details and compare claims under consistent evaluation boundaries.

Koç University photonic AI research deserves attention because it moves useful image transformations into physical hardware while preserving color and introducing nonlinear processing.

Its strongest result is not a settled claim about sustainable computing. It is a clearer experimental path toward finding where optical hardware genuinely earns a place beside electronics.

The decisive evidence will come from full-system measurements, integrated prototypes, and real workloads. Until then, the right question is specific: which processing stage saves enough energy when moved into light to justify everything surrounding it?