Driven by the evolving field of artificial intelligence (AI), a host of tools utilizing artificial neural networks (ANNs) have been proliferating, offering enhanced solutions to diverse problems. While traditionally executed on conventional digital platforms, recent explorations have centered on deploying these tools via unconventional platforms, like diffractive optical devices. A pioneering contribution to this exploration comes from a team of researchers led by Prof. Tie Jun Cui at Southeast University in China. They’ve introduced a novel programmable neural network with an intriguing new architecture that could revolutionize wireless communications and more.
Advanced Neural Network for Microwave Processing
Prof. Cui’s team, driven by a quest for improved performance and broader applicability of neural networks, has created a unique neural network architecture based on a phenomenon known as spoof surface plasmon polariton (SSPP). This SSPP-based surface plasmonic neural network (SPNN) displays the capacity to detect and process microwaves—an advancement with significant implications for wireless communication and other technological applications.
Building on the SSPP Paradigm: From Linear to Nonlinear
While optical neural networks and diffractive deep neural networks have made their mark in the realm of digital hardware research for neural network implementation, they have shown certain limitations. Issues around achieving simultaneous high-level programmability and nonlinear computing have been obstacles, often confining these devices to specific, usually linear tasks. The SSPP-based SPNN, in contrast, promises more versatility, capable of tackling complex nonlinear problems with customizable weight configurations.
Engineered for Speed and Efficiency: SPNN Architecture
Drawing on years of expertise in developing programmable SSPP devices, the team established a neural network with programmable weights and activation functions. In a bid to maximize processing speeds, the architecture was designed to approach the speed of light. This is accomplished through a layer-by-layer construction of programmable SSPP supercells, which form the backbone of the neural network. Each supercell’s complex design enables it to manipulate electromagnetic waves effectively, facilitating the robust performance of the plasmonic neural networks.
Broader Applicability and Enhanced Performance
A key attribute of the SPNN architecture lies in its programmable weights and activation functions. This feature increases its task versatility, overcoming the limitations of previous neural networks that used phase change materials. The newly developed SPNN not only modulates and processes electromagnetic waves flexibly but also includes nonlinear activation functions, enabling it to handle more complex problems. Consequently, the SPNN holds considerable potential for detecting and processing microwaves on a large scale—an advancement promising for future 5G and 6G wireless communications.
Moving forward, the team intends to increase the SPNN’s complexity, enabling it to solve more advanced problems and to evaluate its performance on a broader range of tasks.