Physicists Train the Oscillatory Neural Network to Recognize Images

BEL-based Emotional AI

Physicists Train the Oscillatory Neural Network to Recognize Images

A new paper by physicists from Petrozavodsk State University proposes a novel method for oscillatory/spiking neural network (SNN) to recognize and categorize simple images. The advantage of oscillatory networks with adjustable synchronous state of individual neurons is that their dynamics are more similar to the neuronal networks in the actual brain.

An oscillatory or Spiking Neural Network (SNN) is a complex network of interacting oscillators that receive and send oscillations of a certain frequency. Following receiving inputs of different frequencies from preceding neurons, the oscillator may synchronize its oscillations with the input frequency. Consequently, some of these neurons may become synchronized with each other, and some may not. In this manner, a space-time picture of the synchronization distribution is drawn.

In this paper, physicists from the Department of Electronics and Power Engineering of Petrozavodsk State University have developed a synchronization registration method, based on coupled oscillator networks implemented on vanadium dioxide structures, with high sensitivity and selectivity. Their goal is to create an oscillating neural network that is capable of recognizing images in the same way that neuronal networks in biological systems do.

In the study, researchers created different input images in the form of 3×3 tables. These images were fed to the SNN by changing the supply currents, and consequently, currents changed the oscillation frequencies of oscillators. Hence, the network reacted to each input image with different dynamics. The system is then trained to synchronize only for a specific incoming image, which means to recognize/classify it.

The synchronization state of the output neuron-oscillator relative to the rhythm of the main neuron-oscillator was chosen as the output recorded signal. The result shows that synchronization can be observed not only at the fundamental frequencies, but also at their multiple parts (subharmonics). An increase in the number of synchronous states due to subharmonics is called a high order synchronization effect. Having simultaneously several states of synchronization, the neuron becomes a multilevel neuron.

Using this property, the physicists configured the network in such a way that different input images caused different synchronization patterns of the oscillatory network. The results show that the SNN is capable of classifying 512 visual patterns (as a cell array 3 * 3, distributed by symmetry into 102 classes) into a set of classes with a maximum number of elements up to fourteen. 

According to the lead author Andrei Velichko: “In the future, compact neural network chips with nanoscale oscillators can be created on the basis of these networks. The distinctive feature of the neural network technology that we are developing is a fundamentally new information processing system. The effect of high-order synchronization of pulsed signals allows utilization of multilevel neurons with a high degree of functionality. The advantage of such oscillatory neural networks is the prospect of creating neural networks using a wide variety of physical oscillators, including magnetic and electrical oscillators. At the same time, the trained network no longer needs computer calculations, and operates independently as a separate neural organism.”

In conclusion, an oscillatory network of a small number of neurons can perform complex operations such as speech, image recognition, prediction, optimization and control problems.