Specialists from Yaroslav the Wise Novgorod State University have developed a unique mathematical model that allows dynamic reconfiguration of hardware neural networks in real time.
The technology has no analogues in the world and paves the way for the creation of physical neural networks capable of adapting to changing conditions without manual intervention.
The key problem of modern hardware neural networks is the difficulty of changing the "weights" (connection parameters between neurons) in devices with submicron elements. This limits their flexibility: such systems cannot quickly switch between tasks. The new model solves this problem by externally controlling synapses - the areas of interaction between neurons.
The development is based on the concept of coupled oscillation generators that mimic the operation of artificial neurons. Their synchronization depends on the frequency of external signals, which allows changing the "behavior" of the network. To control the connections between the generators, the scientists used layered magnetoelectric structures. These elements act as a platform where the interaction of neurons is regulated by an external electric field.
"We mathematically linked the field configuration with the interaction areas of the generators," explained Alexander Nikitin. "This made it possible to transfer physical processes to the Python software environment, which simplifies the integration of the model with machine learning algorithms."
Dynamic reconfiguration of neural networks "in hardware" eliminates the need to convert requests into machine code, speeding up calculations by hundreds of times. Such systems will be able to adapt instantly - for example, in autonomous robots operating in extreme conditions, or in real-time systems for analyzing big data. In addition, the technology reduces energy consumption, as it eliminates the stage of software interpretation.
The research is at the stage of mathematical modeling. The next step will be to create a prototype device.
Related materials:
Skoltech AI has learned to predict forest fires with an accuracy of up to 87%
Avito Rabota launches a new language system A-Vibe: will help to create a resume
