Scientists have accelerated the development of AI to work with particle accelerators

Researchers from the HSE Center for AI have developed a way to create stable artificial intelligence (AI) models that work reliably and predictably, and at the same time, model learning has become eight times faster. The study was published in the scientific journal EEE Xplore.

Modern artificial intelligence learns from big data, but not all models behave in the same stable way. Some give accurate predictions every time, others "float" — the result depends on the random distribution of data or initial settings. To solve important tasks, such as determining the energy and direction of particles in accelerators, scientists need reliable models that give predictable results under any conditions.

Russian researchers have proposed a method that allows you to train many models and automatically select the most reliable ones among them. Each model is trained several times on slightly modified data and with different initial parameters, after which the system evaluates how stable the results are. This approach reduces the number of necessary attempts by eight times compared to the usual full iteration of models.

Special intelligence: Putin instructed to create a headquarters for the development of AI

What developments based on new technologies were shown to the president and the public at the AI Journey-2025 conference

The method was developed by a group of scientists led by Fedor Ratnikov, a leading researcher at the Institute of AI and Digital Sciences at the Higher School of Economics (Moscow), to accelerate the creation of neural networks and classical machine learning systems working with data from colliders. Currently, such algorithms are actively used, including at the Large Hadron Collider, where they help to process primary experimental data faster. Checking the algorithm on a set of data from accelerator sensors showed that the optimal AI model can be selected in 41.5 thousand attempts, which is eight times less than with a full search.

The new approach also makes model training more predictable. The researchers found that models with additional "hints" about the data learn faster, require fewer training samples, and produce more stable results. This is especially important for tasks where you need to get equally accurate results with repeated training and when working with different batches of data. In the future, the method will accelerate the development of neural networks for solving problems in particle physics and other fields of science where the stability and reliability of AI are critically important.

Fantastic Beasts: Mutating AI Trojans Take Over the Web

New generation viruses break down the usual methods of protection and can remotely connect to generative AI models.

Alexey Likhachev, head of the Rosatom State Corporation, announced at the first International Symposium on AI and Nuclear Energy at the headquarters of the International Atomic Energy Agency (IAEA) on December 3 that the use of artificial intelligence (AI) technologies would contribute an additional 110 billion euros to the Russian economy by 2030. According to Likhachev, a full technological cycle is currently being created in Russia, which includes the production of energy and computing infrastructure. Along with this, the creation of algorithms and obtaining results is noted.

All important news is on the Izvestia channel in the MAX messenger.