Deep learning AI, a software simulation of neural networks
Development of hardware-simulated AI neuromorphic chips is essential
Unit neuron computational function utilizing metamaterials
Successfully simulated and reproduced using optical signal processing

Domestic researchers have succeeded in designing a neuromorphic optical neuron device for AI implementation.

The Ministry of Science and ICT announced on the 4th that a research team led by Professor Park Nam-kyu, Dr. Sun Sun-kyu, and Dr. Park Hyun-hee of Seoul National University succeeded in mimicking the operation of neurons, the basic units of the brain, using the flow of light to implement high-speed computing AI.

Neurons in the human brain can be described as the unit processors of the nervous system, acting like transistors in electronic circuits. The brain's learning and memory abilities are realized through the interconnected signal processing functions of individual neurons via a complex neural network.

Deep learning-based AI technology is a software simulation of the brain's neural network through computer programs. To implement AI stably and efficiently, in addition to a software approach, it is essential to develop AI-dedicated neuromorphic chips that hardware-simulate the operation of neurons and the network itself.

Neuromorphic technology refers to a technology that simulates signal processing in biological nervous systems through hardware systems in fields such as semiconductors and optics.

As nano-processing in semiconductor electronic circuits has recently been miniaturized to the nanometer level, limitations in heat generation and speed are becoming fundamental constraints on the performance improvement of semiconductor devices, and this is expected to become an ultimate limiting factor in the implementation of neuromorphic semiconductor electronic circuits.

On the other hand, neuromorphic devices that perform computations using light rather than electrons are being researched at research institutions around the world, including MIT and Stanford University, because they enable low-power and ultra-high-speed operation without heat generation.


The research team utilized a metamaterial that adds nonlinearity controlling time symmetry to an amplification and loss material satisfying a peculiar physical symmetry called parity-time symmetry, We succeeded in simulating and reproducing the various computational functions of a unit neuron using optical signal processing.

Parity-time symmetry is a symmetry that describes a system that remains the same as the original system when simultaneously subjected to spatial inversion and temporal reversal. It is known that if this symmetry is satisfied, the overall energy state can remain stable even if there are energy gains or losses within the system.

Nonlinearity is a characteristic in which the input and output values are not in a proportional relationship. Light has linear characteristics according to Maxwell's equations, which determine the characteristics of electromagnetic waves, but it can have nonlinearity depending on the characteristics of the light propagation medium and the intensity of the light.

Metamaterials are materials designed to possess properties not found in nature through the arrangement of artificial media structures, and are applied in applications such as invisibility cloaks.

The research team developed a metamaterial with nonlinearity in which input and output values change depending on the intensity of light, and by correlating it with sodium and potassium channels similar to neurons in the brain, they succeeded in implementing neural signal processing in neuromorphic optical devices at the speed of light.

It was also theoretically confirmed that various functions required for neuromorphic and brain-mimicking memory devices, such as electrical signals maintaining stable strength without fluctuation despite external noise, can be implemented using the flow of light; this establishes the design concept for an ultra-high-speed neuron-mimicking optical device capable of operating at the speed of light for the first time in the world.

Professor Park Nam-kyu, who led the research, stated, “In this study, we applied a multidisciplinary approach to interpret the operating principles of biological structures through physical symmetry and design new optical devices using this.”

Furthermore, “By enabling the functions of neurons, the unit elements of neuromorphic circuits, to be driven using light as a signal carrier, it serves as a turning point in the development of ultra-high-speed neuromorphic devices and AI.” "Of course, each implemented function can also be applied to lasers with high stability," they stated.