Wave system realizes error robustness and efficient modulation as similar to a human brain
Designed platform provides a building block for future photonic artificial intelligence
From left, Professor Namkyoo Park, Professor Sunkyu Yu, and Doctor Xianji Piao (Dept. of Electrical and Computer Engineering, SNU)
SNU College of Engineering (Dean Kookheon Char) announced that a research team led by Professor Namkyoo Park and Professor Sunkyu Yu of the Department of Electrical and Computer Engineering succeeded in designing a brain-mimicking hardware platform by utilizing the structural properties of deep-learning artificial neural networks. The proposed system enables very effective signal processing for quantum or classical waves, similar to signal behaviors inside brain networks.
A human brain is a very complex network structure in which hundreds of billions of neurons are connected through thousands of synapses each. The connectivity of these brain neural networks is known to exist in an intermediate regime that is neither completely regular nor completely random. This interesting regime has been characterized as a 'scale-free network.'
Scale-free networks have unequal structural characteristics, which lead to a few hub nodes that have preferential connectivity and are thus particularly sensitive to change. This feature leads to robustness to accidental errors, but at the same time, results in fragility to intended attacks to the network, which verifies the efficient signal transport inside the brain. Therefore, one can assume that the characteristics of a scale-free network will have an advantage when developing brain-mimicking artificial intelligence hardware. However, because of the tremendous number of candidates in configuring disordered materials or complex structures, identifying a scale-free platform in a real space, especially in integrated systems, has been a very challenging task.
Professor Namkyoo Park and Sunkyu Yu's research team have demonstrated that deep-learning artificial neural networks can be applied to identify scale-free wave systems that have brain-like structural characteristics and functional wave behaviors.
Professor Namkyoo Park, the corresponding author of the paper, stated that "After training the deep-learning neural network that describes the interaction between waves and systems, we found that the network provides scale-free characteristics similar to a brain. When the medium was inversely designed using this trained network, it was highly interesting to note that the scale-free nature of the neural network was successfully projected onto the structural properties of the designed hardware platform. This demonstrates that it is possible to design and organize "brain-like hardware (real-space materials or structural systems)" with "brain-like software (deep-learning artificial neural networks)." In other words, it is possible to design neuromorphic hardware using neuromorphic software."
In addition, "One of the ultimate goals of researchers in photonics is to realize a computational system that is operated by light.
Based on the global view on real-space complex networks treated in our recent review article, which was published in <Nature Reviews Materials
> and the research about photonic neurons published last year in <Advanced Science
>, I wish to achieve the development of light-operated artificial intelligence. Ultimately, the long-term goal is to present a new methodology for the design of high-speed, power-efficient artificial intelligence (Photonic Brain), which has recently been intensively developed by MIT, Stanford, and various startups, as well as to develop the Photonic Quantum Brain based on quantum waves.
Professor Sunkyu Yu, the first author of the paper, said, "The scale-free hardware system achieved signal behaviors that are robust against random defects and very sensitive to intended modulation at the same time, similar to the natures of a scale-free network. In other words, it can be applied to efficient switches, logic devices, memory, photonic deep-learning systems, etc., because we can easily control wave behaviors by altering a few hub neurons while minimizing the impact of defects or noises in experiments or signal processing."
This study was conducted with the support from the National Research Foundation of Korea (NRF) through the Global Frontier Program, the Basic Science Research Program, and the Korea Research Fellowship Program, all funded by the Korean government., and was published on September 24 of <Nature Communications
>, a world-renowned journal.
Conceptual diagram of designing scale-free wave systems.
(Left) The human brain with the scale-free characteristics, which is the result of ‘slow biological evolution.’ (Middle) The deep learning artificial neural network that mimics the brain neural network in a software aspect. It also has characteristics of a scale-free network, which results from ‘fast numerical optimization.’ (Right) A scale-free wave system that is inversely engineered using neuromorphic software: deep-learning artificial neural networks. It can be applied to the realization of high-performance neuromorphic hardware mimicking a human brain.
Characteristics of a scale-free wave system.
(left, top) The road network has equal importance for each connection. (right, top) Meanwhile, the airline network has a few hub airports that are much more important in comparison to other airports, making the network stronger against accidental accidents. The hub airports govern the operation of the entire system, resulting in the fragility of the network to the intended attack. (left, down) A typical disordered system. Most particles have "equal" sensitivity. (right, down) A scale-free wave system with “unequal” atomic sensitivity. Certain particles are more sensitive to errors, defects, or modulations. In this case, the system is robust to accidental defects and is fragile to intended modulation. To summarize, the neural network of our brain has unequal structural characteristics, and the medium mimicking the brain is very suitable for signal processing and learning functionalities.
For further information, please contact Prof. Namkyoo Park (email@example.com).