Offloading AI's Computational Burden Onto Light

Koç University researchers have developed two systems that perform computation using light itself rather than conventional electronic circuitry. Described by the researchers as “photonic neural networks,” these systems go beyond serving merely as auxiliary components for electronic AI hardware: they represent early steps toward new computing architectures in which light performs the core computational operations.

As artificial intelligence systems grow, so do their demand for processing power and energy. Recognizing an image, for example, may require a computer to perform millions of operations. Unlike electronic signals, light can carry and process multiple streams of information simultaneously, enabling rapid and highly parallel computation. By harnessing light instead of relying exclusively on electronic circuits, researchers can therefore perform some of the computationally intensive operations required by AI applications with considerable efficiency.

Two studies by Assistant Professor Uğur Teğin and his students from Koç University’s Computational Sciences and Engineering Program and Department of Electrical and Electronics Engineering, published in two separate Nature Portfolio journals, present complementary approaches to light-based computing.

Processing Color Information with Light—Without Losing It
In the first study, published in the Nature Portfolio journal Communications Engineering, Assistant Professor Uğur Teğin and master’s student Fatma Nur Kılınç developed a programmable photonic neural network that uses light to transform the information in color images into representations that are easier to classify, without discarding their color content.

Many image-processing systems convert color images to grayscale, potentially losing information contained in the red, green, and blue channels. The new system preserves these channels and processes them simultaneously.

The data are first encoded onto a programmable device that controls the propagation of light. As the different color channels travel together through the optical system, the device combines and transforms them in a way tailored to the task at hand. Passing through the optical setup accentuates features—such as color, shape, and texture—that help distinguish one image from another. The resulting optical pattern is then recorded by a camera and used to predict the category to which the image belongs.

A key advantage of the photonic neural network is that it can operate without requiring a perfectly accurate computer model of the physical system. In a real optical setup, minor misalignments, manufacturing imperfections, and other physical effects can cause experimental results to deviate from idealized simulations. By learning directly from experimental measurements, however, the system can adapt to these non-ideal conditions.

In a dermatological disease classification task, the system increased accuracy from 68 percent with raw color pixels to 98 percent after optical processing. The researchers also demonstrated that the method could be applied not only to images but also to the classification of other forms of multidimensional data.

Turning a Laser into a Physical Computing System
In the second study, published in the Nature Portfolio journal Communications Physics, Assistant Professor Uğur Teğin, master’s students Dilem Eşlik and Fatma Nur Kılınç, and undergraduate student Bahadır Utku Kesgin used a multimode fiber laser as a physical processor for machine learning. Operations that would ordinarily be carried out by computer software were instead performed directly within the laser through the dynamics of light.

The researchers first encoded images into patterns that modulate the wavefront of the light. When these information-bearing patterns were processed nonlinearly inside the laser cavity, they were transformed into distinctive optical output patterns that could be more readily separated. This process enhanced similarities among images belonging to the same category while making differences between categories more pronounced, thereby simplifying the subsequent classification task.

The system was tested using images of recyclable waste, weather and satellite imagery, retinal scans, and skin lesions. Following optical processing by the laser, classification accuracies ranging from 85 to 99 percent were achieved across the different datasets using only a simple classification method. The system also required far fewer trainable parameters than conventional deep-learning models because the laser itself performed a substantial share of the computation.

Toward More Energy-Efficient Artificial Intelligence
The first study focuses on processing multidimensional information, such as color, without discarding potentially valuable data. The second uses the physical dynamics inside a laser to reveal complex distinctions between images.

Although both systems still rely on electronic components, advances in high-speed light-control devices and compact, chip-integrated optical systems could enable light to assume a greater share of the workload currently handled by electronic processors while consuming substantially less energy. Such developments could help pave the way for faster and more energy-efficient artificial intelligence technologies.