As AI moves from the cloud to billions of connected devices, a major challenge emerges: how can these devices learn together without constantly sending data to a central server?

FCP researchers from the Singapore University of Technology and Design (SUTD), together with collaborators, have developed a new approach called 𝗠𝘂𝗹𝘁𝗶-𝗖𝗹𝘂𝘀𝘁𝗲𝗿 𝗗𝗲𝗰𝗲𝗻𝘁𝗿𝗮𝗹𝗶𝘇𝗲𝗱 𝗙𝗲𝗱𝗲𝗿𝗮𝘁𝗲𝗱 𝗘𝗱𝗴𝗲 𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴 (𝗠𝗗-𝗙𝗘𝗘𝗟) that helps large networks of devices collaboratively train AI models while preserving data privacy.

Instead of relying on a single central server, the system allows devices to collaborate directly with one another. The researchers introduced an adaptive clustering mechanism that automatically groups devices with similar data patterns and strong connectivity. By organising devices into smaller cooperative communities, the framework reduces communication overhead, accelerates learning, and improves resilience against network failures.

The team further developed an asynchronous learning algorithm that integrates the contributions of both faster and slower devices. This addresses a common real-world challenge where devices have different computing capabilities, from smartphones and drones to autonomous vehicles and industrial sensors.

Experiments on benchmark datasets showed that the proposed approach achieved faster convergence and better generalisation performance than existing synchronous benchmark algorithms. These findings represent an important step towards scalable, privacy-preserving AI systems for future smart cities, industrial IoT, autonomous systems, and next-generation communication networks.

𝘑. 𝘛𝘢𝘯𝘨, 𝘟. 𝘓𝘪𝘢𝘯𝘨, 𝘔. 𝘚𝘪𝘦𝘸, 𝘢𝘯𝘥 𝘛. 𝘘. 𝘚. 𝘘𝘶𝘦𝘬, “𝘈𝘥𝘢𝘱𝘵𝘪𝘷𝘦 𝘊𝘭𝘶𝘴𝘵𝘦𝘳𝘪𝘯𝘨-𝘌𝘯𝘢𝘣𝘭𝘦𝘥 𝘓𝘢𝘳𝘨𝘦-𝘚𝘤𝘢𝘭𝘦 𝘋𝘦𝘤𝘦𝘯𝘵𝘳𝘢𝘭𝘪𝘻𝘦𝘥 𝘍𝘦𝘥𝘦𝘳𝘢𝘵𝘦𝘥 𝘓𝘦𝘢𝘳𝘯𝘪𝘯𝘨,” 𝘐𝘌𝘌𝘌 𝘛𝘳𝘢𝘯𝘴. 𝘔𝘰𝘣𝘪𝘭𝘦 𝘊𝘰𝘮𝘱𝘶𝘵𝘪𝘯𝘨, 𝘷𝘰𝘭. 25, 𝘯𝘰. 4, 𝘱𝘱. 5350–5363, 𝘈𝘱𝘳. 2026.

#SUTD #FCP #IMDA #NRF #EdgeAI #MachineLearning #IoT #IEEE