Hyperspectral imaging reveals what is hidden from the human eye, but the price for this "superpower" is terabytes of data that conventional processors take minutes to digest. Scientists at Samara University propose a radically different path: instead of chasing higher clock speeds, they make light itself do the computational work.
Their analog photonic system, developed with the support of the Ministry of Science and Higher Education and Rosatom under the National Center for Physics and Mathematics program, does not add up bits. Instead, it redistributes a stream of photons through a diffractive neural network. Light traverses the system in nanoseconds, though obtaining the final processed result requires waiting a few milliseconds for the photosensitive matrix to register the output.
In 2024, the team led by Professor Roman Skidanov not only created an experimental prototype but also made a seemingly paradoxical leap: they simplified the design by removing redundant optical elements, yet managed to boost recognition accuracy to nearly 100% and increase energy efficiency by 1.5 times. Today, as the world struggles with the massive energy consumption of AI data centers and foreign GPUs become increasingly inaccessible, this development looks not just like an alternative, but a potential answer to the technological challenges of our time.
In an exclusive interview with Kommersant-Nauka, Roman Skidanov, Doctor of Physical and Mathematical Sciences and Professor at the Department of Technical Cybernetics at Samara University, discusses this breakthrough and more.
— The concept of optical computing is nearly 70 years old, and it was largely abandoned in the 1980s. What exactly has changed today, beyond miniaturization, that has brought this technology back to the forefront at the peak of the digital era?
— It was not entirely abandoned in the 80s: optical correlators were widely used right up until the early 21st century. As classical computer performance increased, these systems were gradually replaced. However, given the classified nature of certain industries, we cannot rule out that similar correlators are still in use today. The modern surge of interest is tied to a new interpretation: viewing the passage of light through a system of transparencies as the physical operation of a neural network.
— How does your analog photonic system fundamentally differ from the digital photonic chips currently being developed in the West? Why did you choose the analog path over the digital one?
— Both in the West and the East, systems similar to ours—analog and volumetric—are being developed. As for chip-based configurations, creating a classical-type photonic computer is currently impossible due to the lack of photonic transistors and full-fledged optical RAM. Modern photonic chips are generally just an add-on to classical electronic processors. For example, optics is used to solve the problem of high-speed data transfer within the processor. Our system, however, is exclusively tailored for the implementation of diffractive neural networks. Since a neural network is inherently an analog structure, converting it to a digital format is not strictly necessary.
— What physically happens inside the computer when processing a hyperspectral image? If a digital processor adds up bits, what do photons do to arrive at the result "this is the target object"?
— Inside the computer, photons do the only thing they can: they propagate through an optical system equipped with a special transparency—a diffractive neural network. Its task is to distribute the photon stream in the correct manner at the computer's output, depending on the input signal. At the output, there is a set of registration zones, which can be viewed as the output neural layer. Subsequently, the classification process can be carried out simply by identifying the maximum signal in this output neural layer, or by using this layer as an input for a small, supplementary electronic neural network.
— You transitioned from amplitude modulation to phase modulation, simplifying the design and boosting recognition accuracy by nearly 1%—reaching expert levels. In the digital world, such a gain usually requires massive volumes of data, whereas here you link it to a change in the device's underlying physics. Could you explain in more detail how abandoning additional optical elements during phase input allowed you to bring the real computer closer to the theoretical model, and what exactly drove this leap in quality?
— The goal was not to increase accuracy per se, but to minimize the accuracy drop in the real physical computer compared to the theoretical model. The input spatial light modulator is fundamentally a phase device. To convert it into an amplitude device, two additional optical elements must be added. With phase input, these elements are absent, and consequently, so are the optical distortions associated with them. Thus, the actual physical result becomes substantially closer to the calculated mathematical model.
— You mention that the speed is limited only by input/output devices. Does this mean the system's latency is effectively equal to the photon's flight time (nanoseconds), and have we reached the physical limit of processing speed?
— The time it takes for light to pass through the system is about 1 nanosecond. However, the main delay in the system equals the time the input/output devices spend on a single cycle. For instance, liquid crystal light modulators take 17 milliseconds per cycle, and cameras have a similar registration cycle time. If we replace liquid crystal modulators with micro-mirror modulators, the cycle duration will shrink to approximately 0.1 milliseconds. In effect, computational performance would increase by 170 times. However, micro-mirror modulators are currently produced by only one company in the world: Texas Instruments.
— You are working on a mobile platform. What is the main physical limitation you are overcoming during miniaturization, given that optics is bound by the laws of diffraction and path length requirements?
— In the mobile platform, we plan to use a static diffractive optical element to implement the diffractive neural network, instead of a bulky spatial light modulator. This significantly reduces the physical size of the diffractive neuron. While the modulator used neurons measuring 3.5x3.5 micrometers, our technology allows us to create neurons as small as 1x1 micrometer, and even down to 0.6x0.6 micrometers. The current setup utilizes a 2048x2048 neuron array. The mobile setup is planned to use up to a 14,000x14,000 array, which equals 196 million neurons. Such a system should enable the solving of much more complex tasks.
— Your optical network is essentially "hardwired" for specific recognition tasks. Doesn't such a computer turn into a highly specialized tool that loses to a universal GPU in flexibility, even if it wins in speed?
— The system is not as highly specialized as it might seem. In essence, any multi-parameter task whose parameters can be converted into an image can be analyzed by a diffractive neural network. There is, of course, some narrowing of functionality. However, it is unlikely that our country will be able to organize mass production of advanced GPUs in the near future, and analog devices allow us to solve at least a critical portion of these computational tasks. Moreover, the system possesses another crucial advantage: its energy consumption is substantially lower compared to a GPU.
Source: Kommersant
