RT3: Edge Computing for AI

Theme Leaders: Dr Tomasz Szydlo & Dr Blesson Varghese

Researchers: Dr Jennifer Williams, Prof Julie McCann, Prof Gopal Ramchurn, Prof Qi Wang and Prof Jose M. Alcaraz Calero

Edge AI brings AI processing closer to where data is generated, but edge devices and networks vary widely in computing power, memory, energy requirements, and connectivity. RT3 develops methods to make AI models easier to deploy and operate across these diverse environments, even when network conditions change, or systems face cyber threats. 

Our research focuses on two priorities: evaluating the performance and resource requirements of AI models across different edge devices and developing approaches that enable models to adapt to changing operating conditions. Tiny machine learning (TinyML) is an important part of this work, enabling learning and inference on devices with very limited resources. 

Our goal is to make it easier for developers to build, deploy, and use Edge AI models across diverse edge computing systems. We are developing tools and techniques to support this process, helping to make Edge AI more accessible, reliable, and effective in a wider range of applications. 

The projects below demonstrate our research findings in areas including wireless human sensing, embedded-device machine learning, memory-efficient model training, audio deepfake detection, and interpretable, logic-based AI. 

Edge AI for Wireless Human Sensing 

We are developing scalable edge AI for wireless human sensing, particularly human activity recognition (HAR) and pose estimation using Wi-Fi and other radio-frequency signals. This camera-free, device-free approach can reduce reliance on cameras and wearables. To support practical deployment in homes, workplaces and care settings, our research focuses on recognising activity across multiple people and environments, reducing the need for labelled data through cross-modal and self-supervised learning, and making model decisions more interpretable. Efficient learning and inference methods are designed to bring these capabilities to resource-constrained edge devices. 

Read more about machine learning for resource constrained edge devices.