Russian scientists create biological neural network with exquisite hearing

A biological neural network capable of effectively distinguishing external sound signals is the creative achievement of Saratov National Research University.

In their opinion, this innovation will be able to significantly reduce energy consumption compared to conventional artificial neural networks. The research results have been published in the scientific journal Chaos.

In modern technology, signal recognition methods using second-generation artificial neural networks have been widely included. However, the neurons used to model real neurons are much more complex than those of artificial neural networks . Therefore, the third-generation biological (spiking) neural networks formed from them are quite different from the second-generation networks. In recent years, the scientific community has been increasingly interested in studying spiking neural networks, but there are still many unanswered questions.

Illustration of artificial neural network.

As is well known, neural networks in the human brain consist of groups of neurons that are chemically or functionally connected. Experts describe their behavior using the FitzHugh-Nagumo conceptual mathematical model , proposed in the late 20th century.

Scientists from the National Research University of Saratov named after NG Chernyshevsky set themselves the task of determining the ability of a spiking neural network consisting of neurons to recognize sound signals using the FitzHugh-Nagumo conceptual mathematical modeling method. They hypothesized that networks based on such neurons may possess broader capabilities due to their integrated complexity.

"We focused on studying how such a network of neurons behaves in relation to external signals. The studied network is not very large, but its number of elements is sufficient to achieve the desired effect. We found that the FitzHugh-Nagumo neural connection can exhibit selectivity for signals with different frequency ratios and distinguish external signals by selecting only certain connections between neurons. This led to the conclusion that it is possible to build a special network of neurons to ensure the recognition of sound signal fields," explains Andrei Bukh, Associate Professor, PhD, Department of Radiation Physics and Nonlinear Dynamics of Saratov National Research University.

The researchers say their new discovery will help create efficient neural networks, where efficiency is understood as the ratio between the energy consumed and the complexity of the task being solved.

'It is known that to solve the same problem the human brain consumes less energy than a conventional computer. That means that a spiking neural network can consume significantly less energy than a conventional artificial neural network ,' scientist Saratov emphasized.

He added that due to the non-linearity of the components of a spiking neural network , it would become very complex and the responses of the neurons in it could vary greatly. So measuring the effectiveness would be quite difficult. In his opinion, that can only be done when spiking neural networks start to be applied in practice.

'We studied the question of how well the network can be selectively connected to external signals for the simplest specific neuron and found promising results. But for each specific task, the gains in efficiency will be different. There are other difficulties, the most serious of which is that 'there are only a very small number of methods that can be applied to train spiking neural networks' , notes Associate Professor Bukh.

As he says, the results obtained are mainly ensured by selecting only a few connections between neurons. The rest are disconnected. If all connections between neurons are turned on, the network will not exhibit selectivity. On the other hand, an insufficient number of connections will lead to an almost complete absence of responses in it.

In the future, researchers from Saratov State University plan to investigate whether a single neuron model is capable of 'accumulating' signals and exhibiting different behaviors depending on the 'context' .

"Preliminary results show that the neural model accumulates input signals . That is, the neuron's history affects its current state, it responds to the 'context'. But whether a network based on such neurons becomes a classifier remains an open question," concludes  expert Andrei Bukh.