KNOT: Resource-aware Knowledge Transfer Methods for Machine Learning Hardware in At-the-Edge Applications

Modern information and communication technologies produce electronic devices that are deployed at the edge, such as for example smart sensors acquiring information about environmental conditions.

Such devices are increasingly expected not only to perform simple data conversion from one form to another but also perform computations such as data analysis, classification and even decision making.

Collectively such devices are called AI-at-the-Edge, which arguably form the fastest growing electronics technology in the UK and the World (CAGR 20.64% in the next five years).

A key challenge of enabling such devices with intelligence is the fact they are limited in resources, such as compute power, energy budget, physical accessibility. So, the main question is how to equip such resource-limited devices with machine learning (ML) capabilities?

To tackle this challenge this project will develop low-cost (e.g. energy-efficient) mechanisms for sharing knowledge between edge devices.

The project's success will be measured in terms of its new methods capable to deliver knowledge transfer between edge devices helping scale up their ML accuracy with at least 3-4 orders of magnitude energy efficiency compared to existing AI-at-the-Edge systems.

The project outcomes in theory and design methodology will be validated by means of extensive simulations, prototyping, and testing, and, ultimately, via an embodiment of the proposed solutions into a concrete IoT-scale application, such as environmental monitoring and electrical battery safety control.