HSE Researchers Introduce Novel Symmetry-Aware Neural Network Architecture
Researchers at the HSE Laboratory for Geometric Algebra and Applications have developed a new neural network architecture that can accelerate and streamline data analysis in physics, biology, and engineering. The scientists presented their solution on July 16 in Vancouver at ICML 2025, one of the world's leading conferences on machine learning. Both the paper and the source code are publicly available.
Many physical systems—from robots to molecules to charged particles—exhibit the same behaviour when rotated or mirrored. Modern equivariant neural networks are specifically designed to recognise and preserve such symmetries in data, making them especially valuable for scientific and technological applications ranging from modelling chemical compounds and analysing physical processes to image recognition.
However, these models have a drawback: their high accuracy comes at the cost of increased complexity. They often require a large number of trainable parameters, making them resource-intensive and prone to overfitting—especially when training data is limited.
Ekaterina Filimoshina and Dmitry Shirokov, researchers at the Department of Mathematics and the Laboratory for Geometric Algebra and Applications of the HSE Faculty of Economic Sciences, have developed GLGENN (Generalized Lipschitz Group Equivalent Neural Networks), a neural network architecture designed to overcome these limitations. It enables models to preserve data symmetries using significantly fewer parameters. The authors based their design on the well-established mathematical framework of Clifford (geometric) algebras and introduced a novel weight-sharing technique that respects the underlying algebraic structures.
'We set out to design a model that is both intelligent and lightweight,' says Ekaterina Filimoshina, Research Assistant at the HSE FES Laboratory for Geometric Algebra and Applications. 'GLGENN demonstrates that equivariant neural networks don’t need to be large and complex—they can learn effectively, even from limited data, without compromising quality.'
The model has been tested on a variety of tasks, ranging from simulating physical processes to handling geometric data, and has demonstrated results comparable to or better than existing methods. Additionally, GLGENN operates faster and more efficiently due to its reduced number of trainable parameters, making it more accessible for practical applications.
Dmitry Shirokov
'These results could mark a significant step toward the development of new neural-network tools for science and technology' adds Dmitry Shirokov, Head of the HSE FES Laboratory for Geometric Algebra and Applications. 'We are confident that the geometric-algebra approach will find applications across a wide range of fields, including bioinformatics, robotics, and geoinformatics.'
Participation in ICML served as recognition of the high level of machine learning research conducted at HSE University. The scientists plan to further develop the GLGENN architecture by expanding its capabilities to handle new types of data and exploring its potential applications in physics, robotics, and computer vision.
The study was supported by the HSE Mirror Laboratories project titled 'Quaternions, geometric algebras, and applications.'
See also:
HSE Scientists Develop Method to Compress Large Language Models Without Losing Quality
Researchers from the AI and Digital Science Institute at the HSE Faculty of Computer Science have developed a new compression method for large language models such as GPT and LLaMA that reduces their size by 25–36% without additional training or significant loss of accuracy. This is the first approach to use mathematical transformations—specifically, rotations of model weights—to make models more amenable to compression with structured matrices. The study results have been published in ACL Findings 2025. The code is available on GitHub.
Machine Learning Models Can Help Reduce Volatility and Boost Stock Market Returns
The use of machine learning models makes it possible to achieve greater accuracy in predicting risks in the Russian stock market compared to classical econometric approaches. The predictive power of these models increases by 23%, while the average investor’s return can reach up to 13% per annum. These conclusions were drawn by Nikita Lysenok from the Department of Financial Market Infrastructure at the HSE Faculty of Economic Sciences. The paper has been published in Fundamental and Applied Mathematics.
HSE Study Reveals Imbalance in the Generative AI Market
Researchers at HSE University analysed how effectively the global generative artificial intelligence market converts investment into real revenue, concluding that AI is currently developing faster than it is paying off. The results have been published in the journal Foresight and STI Governance.
Living with Volcanoes and Tsunamis
It is not easy to get university graduates and students interested in living and working in Kamchatka, a remote region with harsh natural conditions and an extreme climate. Mirror Laboratories, a joint project of the HSE Institute of Education and Kamchatka State University (KamGU), encourages the development of ‘explorer’ competencies and the modernisation of the region’s education system. The HSE News Service discussed the project with Taras Pashchenko, Chief Expert of the HSE Laboratory for Curriculum Design.
Electronics of the Future: Why Superconductors and Spintronics Work Together
It was once believed that superconductivity and magnetism avoided each other like the devil avoids holy water. However, modern nanostructures prove the opposite. A Russian theoretical physicist and Indian experimentalists have joined forces to create the electronics of the future—free from energy losses. Nataliya Pugach, Professor at the School of Electronic Engineering at HSE MIEM and Leading Research Fellow at the Quantum Nanoelectronics Laboratory, explains how a long-standing acquaintance in Cambridge grew into a mirror laboratory project with the Indian Institute of Technology Bombay (IIT Bombay), how superconducting spintronics works, and what surprises a researcher in India beyond the university campus.
HSE University Presents Research Results at AI Conference in Oman
In April 2026, the International Conference on Intelligent Systems and Artificial Intelligence Applications (ISAA 2026) was held at the University of Nizwa in the Sultanate of Oman. The event was co-organised by HSE University, the University of Nizwa, and the University of Technology and Applied Sciences–Ibri. Researchers from HSE University were among the key speakers at the conference.
HSE University Installs Geoscan Station at IIT Bombay
A Russian ground station for receiving SONIKS satellite data has been installed on the campus of the Indian Institute of Technology Bombay (IIT Bombay). Developed by Geoscan, the system will become part of a mirror laboratory project run jointly by HSE University and one of India’s leading universities.
How the Brain Processes a Word: HSE Researchers Compare Reading Routes in Adults and Children
Researchers from the HSE Center for Language and Brain used magnetoencephalography to study how the brains of adults and children respond to words during reading. They showed that in children the brain takes longer to process words that are frequently used in everyday speech, while rare words and pseudowords are processed in the same way—slowly and in parts. With age, the system is reorganised: high-frequency words shift to a fast route, whereas new letter combinations are still analysed slowly. The study was published in the journal Psychophysiology.
How Neural Networks Detect and Interpret Wordplay: New Insights from HSE Researchers
An international team including researchers from the HSE Faculty of Computer Science has presented KoWit-24, an annotated dataset of 2,700 Russian-language Kommersant news headlines containing wordplay. The dataset enables an assessment of how artificial intelligence detects and interprets wordplay. Experiments with five large language models show that even advanced systems still make mistakes, and that interpreting wordplay is more challenging for them than detecting it. The results were presented at the RANLP conference; the paper is available on Arxiv.org, and the dataset and the code for reproducing the experiments are available on GitHub.
HSE Physicists Propose Unified Theory for Describing Electric Double Layer
To develop more efficient batteries and catalysts, it is essential to understand the processes occurring at the metal–solution interface in the electric double layer (EDL). Physicists at HSE MIEM have proposed a unified theoretical model of the EDL that simultaneously accounts for selective adsorption of ions on the surface and partial charge transfer between ions and the metal—phenomena that had previously been described separately. The model’s predictions are consistent with experimental data. In the future, it may be used in the development of batteries, supercapacitors, and catalysts. The study has been published in Electrochimica Acta.