An Artificial Brain Learns to Regulate Itself: Russian Research Moves Neuromorphic Computing Closer to Reality

The invention is based on the principle of homeostasis – the ability of biological systems to maintain internal balance. The approach is designed for spiking neural networks (SNNs), which reproduce selected principles of how the human brain works.

Unlike conventional neural networks, spiking systems transmit data through brief electrical pulses known as spikes. Neurons activate only when they receive the appropriate signals, potentially reducing energy consumption significantly.

The proposed mechanism monitors the activity of each neuron and adjusts the strength of incoming signals. If a neuron fires too frequently, the system reduces the input; if its activity falls too low, the system increases it. The approach also incorporates short-term synaptic depression, a temporary weakening of signal transmission when activity becomes excessive. This prevents prolonged bursts of excitation that can leave a network unable to process new information.

Computer simulations demonstrated that the mechanism can maintain stable network activity and enable synchronized responses to external signals, even in the presence of interference.

New Possibilities for Robotics

For now, the work is limited to a scientific model of a spiking neural network, not a finished neuromorphic processor. The experiment does not demonstrate that the mechanism could stop today's large language models from hallucinating. Nor have the researchers reported tests on commercially manufactured hardware.

Nevertheless, the research has significant scientific and technological implications. In the longer term, it could give AI developers additional ways to build neural networks that regulate the intensity of their own computational activity. Microelectronics manufacturers could draw on the findings when designing specialized neuromorphic processors. For robotics, the research points toward autonomous systems capable of processing signals directly on the device.

The innovation will not have an immediate impact on everyday gadget users: the research is not aimed at producing a new mass-market software application. Over time, however, technologies of this kind could help make intelligent devices more energy-efficient.

Five Years of Progress in Russian Neuromorphic Technology

The work by MIPT and Lobachevsky State University is part of a broader effort by Russian researchers to advance neuromorphic computing.

In October 2022, Kaspersky Lab and Motive NT introduced the Kaspersky Neuromorphic Platform, combining software tools for developing spiking neural networks with the Russian-made Altay (Altai) neuromorphic processor.

In May 2024, researchers from MIPT, Lobachevsky State University and Southern Federal University incorporated a mathematical model of a memristor into a biophysical model of a neural network. The model reproduces the transmission of inhibitory signals between nerve cells.

In October 2025, ITMO University announced the development of a perovskite nanomemristor. The device can withstand more than 1,500 rewrite cycles, consumes 70–80 nanowatts of power, and has an active element measuring approximately 130–160 nanometers.

Research has been advancing internationally as well. In October 2023, IBM unveiled its NorthPole neuromorphic processor. In April 2024, Intel launched Hala Point, a research system comprising 1,152 Loihi 2 processors and capable of modeling the activity of up to 1.15 billion artificial neurons.

Limitations and Future Prospects

Moving from a scientific model to practical technology will require several challenges to be addressed. Researchers must adapt the algorithm to run on actual neuromorphic hardware, test network stability as the number of neurons increases, measure real-world power consumption, and verify its advantages over conventional processors.

Training spiking neural networks presents a particular challenge because no universal, efficient methods for doing so are currently available. Even so, further advances in neuromorphic processors can be expected over the medium term, from 2027 to 2029. Researchers at MIPT and Lobachevsky State University have already identified the next step in their work: developing more complex networks that can independently form new connections to replace those that have been lost.

Progress in artificial intelligence depends not only on building larger neural networks and more powerful processors, but also on finding fundamentally new ways to process information. The central question is not whether researchers have managed to reproduce the human brain in full, but whether the proposed mechanisms can deliver measurable improvements in the reliability and energy efficiency of computing systems. Their practical applicability will determine the development's long-term significance for Russia's microelectronics, robotics and AI industries.

At some point, neuroscientists discovered that neurons in the human brain do not work with continuous signals, as they do in classical neural network models. Instead, they transmit short pulses, or spikes, only when needed. That discovery prompted mathematicians and electronics engineers alike to turn to a different branch of development, which they called ‘neuromorphic’ – meaning ‘modeled on the way the brain works.’ The problem with the classical approach is that the computing unit is physically separate from memory, and memory is tens of times slower than the processor itself. Neuromorphic devices, by contrast, move computation directly into memory

Oleg Telminov

Head of the Neuromorphic Systems Research Laboratory at the Research Institute of Molecular Electronics (NIIME), Scientific Secretary of the Russian Academy of Sciences' Scientific Council on Quantum Technologies