Edge computing devices, devices located in proximity to the source of data instead of in large data centers, could perform computations locally. This could reduce latency, particularly in real-time applications, as it would minimize the need to transfer data from the cloud.

Implementing deep learning algorithms on edge devices has so far proved challenging, in part due to their power constraints and limited computational resources. Fuzzy logic systems, computational frameworks that rely on approximate reasoning as opposed to binary logic processes, could help to overcome these challenges.

Researchers at the University of Southern California, Northwestern University, University of Hong Kong, Chinese Academy of Science, and other institutes recently developed a new multi-gate van der Waals interfacial junction transistor that could be used to create reconfigurable fuzzy logic hardware. This transistor, presented in a paper in Nature Electronics,

"Artificial neural networks are powerful tools driving the current AI revolution," Han Wang at University of Hong Kong, senior author of the paper, told Tech Xplore. "However, their implementation demands highly complex hardware with significant power consumption, which limits their applicability in edge devices that process information locally and in real-time. In contrast, fuzzy logic systems operate on simple rules, require fewer hardware resources, and can effectively handle many tasks."

Van der Waals materials, layered materials that are held together by weak van der Waals forces, have proved to be promising for the fabrication of more energy-efficient membership function generators. These are the most power-intensive components of fuzzy logic hardware, which are responsible for creating so-called membership functions (i.e., functions that define the extent to which an input belongs into distinct fuzzy sets).

Building on previous research efforts, Wang and his colleagues thus set out to develop a new transistor based on van der Waals materials that could be used to develop efficient membership function generators. The transistor they created is based on molybdenum disulfide (MoS₂), a transition metal dichalcogenide widely used in the development of electronics.

"The van der Waals interfacial junction transistor (vdW-IJT) is built on a MoS₂ homojunction with varied carrier concentration in different regions, exhibiting either current amplification or division behaviors controlled by multiple graphene gate terminals," explained Hefei Liu, first author of the paper.

"Its primary advantage is the ability to intrinsically generate Gaussian or π-shaped membership functions within a single device, whereas traditional CMOS technology requires tens of transistors to achieve this. As a result, vdW-IJTs enable more compact and energy-efficient membership function generators."

As part of their study, Wang and his colleagues integrated their transistors with peripheral circuits to create reconfigurable fuzzy logic hardware that can control nonlinear systems. This hardware was then used to run a simple convolutional neural network (CNNs) trained to complete image segmentation tasks.

"We discovered the significant potential of emerging vdW materials in enabling novel device concepts and computational architectures, such as fuzzy neural networks, within intelligent systems that achieve complex functionality with low power consumption," said Jiangbin Wu, a key researcher involved in this work.

"This advancement could shift information processing from data centers to local devices, providing real-time responses and extending battery life for applications like robotic motion control and autonomous vehicles."

The researchers found that the fuzzy logic system they developed by combining their transistors with a CNN achieved remarkable accuracy on image segmentation tasks. In the future, their proposed design could inspire the development of similar electronic components aimed at enhancing the ability of edge devices to run deep learning algorithms.

"Our future studies will focus on large-scale implementation of vdW-IJT-based fuzzy logic systems, addressing scalable fabrication, variation control, and integration with neural network hardware," added Mark Hersam from Northwestern University, another lead researcher in this work. "These efforts aim to deliver more capable and energy-efficient intelligent edge devices for real-world applications."

© 2024 Science X Network