Physicists from St Petersburg University develop architecture for next-generation high-speed neural networks
Scientists from St Petersburg University have developed polariton neurons for high-speed neuromorphic systems, capable of recognising handwritten digits and voice commands more efficiently than existing counterparts.
Artificial neural networks have rapidly become one of the key drivers of technological advancement worldwide. They are attracting increasing attention from specialists and are being widely applied due to their ability to perform pattern recognition, decision-making, and large-scale data analysis. These advancements are not only stimulating but also significantly accelerating progress in healthcare, autonomous robotics, and other industries.
The research findings are published in the scientific journal Light: Science & Applications.
Traditional neural networks rely on the von Neumann architecture, i.e. a computer architecture in which data and programs are stored in a shared memory, and information is processed sequentially by the central processor. This system has several limitations, the most significant of which is the physical separation of the computational and memory units — the central processor and RAM. This separation slows down data exchange and increases power consumption. As a result, scientists believe that the full potential of neural networks has yet to be realised.
Alexey Kavokin, Head of the Spin Optics Laboratory at St Petersburg University, and Evgeny Sedov, Senior Research Associate in the same laboratory, have developed a neuromorphic architecture based on two-dimensional lattices of "light droplets", capable of rapidly and reliably recognising handwritten digits and voice commands.
The fundamental building blocks of this system are bosonic condensates of exciton-polaritons (or simply polaritons — particles formed by the strong coupling of photons and excitons, which are quasiparticles in semiconductors). These polaritons exhibit both light-like properties and interactions characteristic of matter, earning them the name "liquid light quanta".
Alexey Kavokin, Head of the Spin Optics Laboratory at St Petersburg University, Senior Research Associate at the University
"When polaritons come together, they can form a special collective state known as a Bose-Einstein condensate. When two such condensates are placed side by side, they begin to exchange particles, creating a pattern of alternating light and dark bands. A laser beam can modify this pattern, transforming light bands into dark bands and vice versa. This modulation serves as an output signal, enabling these structures to function as artificial neurons within a neural network," explained Alexey Kavokin.
The Uraltsev Spin Optics Laboratory was established in 2011 with the support of a mega-grant from the Government of the Russian Federation. Currently, the laboratory comprises more than 40 researchers, the majority of whom are early-career researchers, including eight PhD candidates from St Petersburg University. The laboratory is led by Alexey Kavokin, who ranks among the top 2% of the most cited scientists worldwide and is a recipient of multiple international awards. The scientists are engaged in one of the University’s most significant research areas, a field that was highlighted in the documentary film The First University.
The architecture proposed by the physicists from St Petersburg University belongs to the class of Binary Neural Networks (BNNs), which operate with binary input and output signals from neurons. Unlike traditional solutions that use continuous variables, BNNs process information faster and require less memory than conventional neural networks. As a result, such networks consume less power and scale better. These features make them suitable for devices with limited power resources, such as the Internet of Things (IoT) or edge computing, where speed and efficiency are important but high accuracy is less critical.
This architecture was tested by solving two problems. The first was to recognise handwritten digits from the largest MNIST database. The recognition accuracy of the system developed by the scientists from St Petersburg University exceeded 97.5%. To assess the network’s adaptability and versatility in working with different types of data, it was tested on a set of Speech Commands for recognising voice commands. Here, the development of the scientists from St Petersburg University also outperformed all existing solutions and methods.
St Petersburg University, the oldest university in Russia, was founded on 28 January (8 February) 1724. This is the day when Peter the Great issued a decree establishing the University and the Russian Academy of Sciences. Today, St Petersburg University is an internationally recognised centre for education, research and culture. In 2024, St Petersburg University celebrated its 300th anniversary.
The plan of events during the celebration of the anniversary of the University was approved at the meeting of the Organising Committee for the celebration of St Petersburg University’s 300th anniversary. The meeting was chaired by Dmitry Chernyshenko, Deputy Prime Minister of the Russian Federation. Among the events are: the naming of a minor planet in honour of St Petersburg University; the issuance of bank cards with a special design; and the branding of the aircraft of the Rossiya Airlines to name just a few. To mark the 300th anniversary of St Petersburg University, a postage stamp depicting the Twelve Collegia building and the monument to Count Sergey Uvarov was issued. Also, a Soyuz rocket bearing the symbols of the University was launched from the Baikonur Cosmodrome.
By the decision of the Governor of St Petersburg Alexander Beglov, 2024 was a year of the 300th anniversary of St Petersburg University in St Petersburg. On the day of the University’s 300th anniversary torches were lit on the Rostral Columns on the Spit of Vasilyevsky Island. St Petersburg University flags were raised on the Palace Bridge. The city public transport was decorated with the University’s symbols. During St Petersburg’s City Day celebrations in May 2024, St Petersburg University acted as a participating venue. Additionally, the University launched a website dedicated to the anniversary. The website contains information about outstanding University staff, students, and alumni; scientific achievements; and details of events held as part of the celebration of the 300th anniversary of the University.