Exploring the Potential of Programmable Neural Networks Based on Spoof Plasmonic Devices

Friday - May 19 2023, 02:06 UTC - 1 year ago

A research team, led by Prof. Tie Jun Cui, in China has develpoed a new programmable neural network based on a so-called spoof surface plasmon polariton. This SPNN architecture can detect and process microwaves, which could be useful for wireless communication and other technological applications. SPNNs have adjustable weights and activation functions, and are able to perform an image classification task in a tenth of the time taken by conventional digital methods.

AI tools based on artificial neural networks (ANNs) are being introduced in a growing number of settings, helping humans to tackle many problems faster and more efficiently. While most of these algorithms run on conventional digital devices and computers, electronic engineers have been exploring the potential of running them on alternative platforms, such as diffractive optical devices.

A research team led by Prof. Tie Jun Cui at Southeast University in China has recently developed a new programmable neural network based on a so-called spoof surface plasmon polariton (SSPP), which is a surface electromagnetic wave that propagates along planar interfaces. This newly proposed surface plasmonic neural network (SPNN) architecture, introduced in a paper in Nature Electronics, can detect and process microwaves, which could be useful for wireless communication and other technological applications.

"In digital hardware research for the implementation of artificial neural networks, optical neural networks and diffractive deep neural networks recently emerged as promising solutions," Qian Ma, one of the researchers who carried out the study, told Tech Xplore. "Previous research focusing on optical neural networks showed that simultaneous high-level programmability and nonlinear computing can be difficult to achieve. Therefore, these ONN devices usually have been limited to specific tasks without programmability, or only applied for simple recognition tasks (i.e., linear problems)." .

The primary objective of these researchers' recent work was to further improve the performance of neural networks on complex nonlinear problems, while also making them suitable for a broad variety of applications. SPNN, their proposed architecture, can be programmed for different weight configurations, which means that it should theoretically generalize well across different tasks.

The research team led by Prof. Cui has been developing programmable spoof surface plasmonic devices and exploring their use for electromagnetic regulations for several years. Inspired by their previous findings, they thus set out to develop a neural network with programmable weights and activation functions based on one of these plasmonic devices. In principle, the architecture they proposed could achieve remarkable processing speeds, approaching the speed of light.

"The SPNN was created in a layer-by-layer fashion, where each layer consists of multiple programmable SSPP supercells," Ma explained. "Each supercell with a four-in and four-out fully connected network is composed of eight programmable SSPP cells. We design a three-dimensional composite structure, which cleverly realizes the characteristics of the full connection." .

Each of the programmable supercells that the researchers used to create their platform is made up of a SSPP power divider and a coupler. This unique design allows it to robustly manipulate electromagnetic waves and then use them to realize plasmonic neural networks.

"The weight parameters of neural networks are adjusted by changing the voltages of varactors loading on the couplers," Ma said. "More importantly, the activation function can be customized by detecting the input intensity using detectors and feeding back the threshold to an amplifier. The SPNN can perform an image classification task in one-tenth of the time taken by conventional digital methods and is surprisingly accurate." .

Overall, this study could pave the way for future applications and enable wider accessibility for alternative platforms based on spoof plasmonic devices, such as SPNN neural networks.