MIPT, FIC IU RAS, AIRI: Machine learning algorithm for improving the memory of robots

Main article: Machine Learning

2025: Introducing a biologically similar memory algorithm for AI

Russian scientists from the Moscow Institute of Physics and Technology, the Federal Research Center "Informatics and Management" of the Russian Academy of Sciences and the AIRI Institute have developed a biologically similar memory algorithm for artificial intelligence, which significantly increases the efficiency of training robots in a noise environment. MIPT announced this on January 23, 2025. The method is based on the principles of the work of dendrites - the processes of brain neurons responsible for signaling. This algorithm allows AI to process information faster and find links between data, while reducing the cost of computing resources.

This machine learning algorithm will allow robotic systems to generalize data and find relationships between them. This will significantly reduce the time and computational resources spent on information processing and help more effective AI training.

The basis of the development was the idea of ​ ​ using mathematical models that resemble dendrites - the processes of neurons in the brain, which play a key role in the transmission of information. In artificial models, they help to recognize and classify objects. In the process of learning, the model adapts, "growing" and expanding its knowledge by creating new "dendrites."

However, existing models face the problem of overgrowth. This often happens in "noisy" environments, where each new deviation requires the creation of additional segments, which increases the complexity and resource capacity of the system.

The solution that scientists propose is to change the machine learning algorithm in such a way that the computational model recognizes not the entire object, but certain parts of it. To do this, a "soft adapter" was introduced into the algorithm. This function allows existing segments to recognize new objects by partial similarity. The results of the experiments showed that this approach significantly slows down the growth of "dendrites" without a significant loss of recognition quality.

Experiments show that such a method significantly reduces the growth of "dendrites." At the same time, the segments already grown use much wider. Moreover, it turned out that this does not lead to a noticeable drop in the quality of recognition, but allows you to work in noise conditions and reduce the time and power required to process information. Thus, AI learns to generalize data by certain features and find relationships between them. These actions, in principle, can already be attributed to primitive thought operations, - said Petr Kuderov, assistant at the Center for Cognitive Modeling at MIPT and junior researcher at the AIRI Institute.

The development of such an algorithm not only minimizes overgrowth, but also increases the ability of AI to generalize data, finding relationships even in conditions of increased noise. This adaptability helps systems adapt to different levels of complexity and noise.

Regardless of the nature of the data - whether text, image, or physical objects - the computational model effectively handles recognition, opening up perspectives for creating realistic pictures of the world. With this technology, AI can become a step closer to implementing primitive forms of thinking.

However, the proposed algorithm has adaptability, which is expressed in the ability to adjust the level of recognition accuracy. This gives the machine brain the ability, like a radio receiver, to fine-tune to a given noise range.