Researchers from the University of Western Macedonia, Kingston University, University of Macedonia, Harokopio University of Athens, Sidroco Holdings Ltd., and MetaMind Innovations – MINDS have conducted an in-depth comparative study of three key personalization approaches—Active Learning, Knowledge Distillation, and Local Memorization—within federated learning environments.
This significant research investigates the performance of these existing techniques when applied to the challenges of optimizing models for individual nodes in federated domains. The study specifically examines how these approaches can enable the development of smaller models that require fewer computational resources, while effectively incorporating local insights. This directly addresses the often time-consuming nature of training and deploying complex Machine Learning (ML) and Deep Learning (DL) models.
The researchers performed a thorough comparison analysis in both local and federated contexts. The results demonstrate promising outcomes for model optimization and personalization, highlighting the potential of these techniques for advancing Next-Generation IoT (NG-IoT) applications.
Research Team: Ilias Siniosoglou, Vasilis Argyriou, George Fragulis, Panagiotis Fouliras, Georgios Th. Papadopoulos, Anastasios Lytos, Panagiotis Sarigiannidis
Published in IEEE Xplore.
You can read the full paper at: https://arxiv.org/abs/2409.06904
