Edge AI breakthrough: Faster learning with less energy

Researchers unlock faster, greener AI training for Edge devices

A team of researchers—Zhiyuan Zhai, Wei Ni, and Xin Wang—is examining ways to boost learning performance in Edge AI systems. Their work focuses on optimising energy use, time efficiency, and communication in decentralised machine learning. A key area of study is Federated Learning, where models train on data spread across multiple devices without centralising it. Federated Learning helps train AI models on devices like phones or sensors while keeping raw data local. This approach improves privacy and reduces the need for high-bandwidth transfers. However, challenges remain, particularly with non-IID (non-independent and identically distributed) data, where information varies across devices.

The team has developed a detailed model of energy and time demands in Edge AI. Their system accounts for data collection, computation, and communication costs. By analysing these factors, they aim to allocate resources more efficiently and cut energy use. One innovation under investigation is over-the-air computation, which aggregates model updates during transmission. This method could lower communication delays, speeding up training. Other techniques, such as model compression and quantization, are also being explored to streamline data exchange. To encourage participation, the researchers are designing incentive mechanisms. These ensure fair compensation for clients contributing data and computing power. The team has framed this as a system-wide optimisation problem: maximising learning performance while respecting time and energy limits. Recent projects, like Fraunhofer IIS's SEC-Learn, support these goals by developing neuromorphic chips for Spiking Neural Networks (SNNs). These chips reduce power consumption in edge devices. Similarly, ECSEL initiatives ANDANTE and TEMPO focus on optimising chips, algorithms, and tools to minimise energy use in smart devices. Industry trends highlight the use of edge servers with AI accelerators and Time-Sensitive Networking (TSN) for reliable, low-bandwidth communication. Meanwhile, Ferdinand Heinrich's award-winning work improves Federated Learning scalability by generating synthetic clients through time series augmentation. This helps optimise communication in large-scale deployments.

The research provides a clearer picture of how data volume and training rounds affect learning in Edge AI. By refining resource allocation and communication methods, the team aims to make decentralised AI more efficient and sustainable. These advancements could lead to faster, lower-energy AI training across a wide range of devices.