Year: 2024
Type: dissertation
Abstract: Federated Learning (FL) is a distributed machine learning approach that efficiently trains Deep Neural Networks (DNN) using decentralized data. Despite the advantages, it faces challenges due to substantial data communication, leading to high energy consumption and network congestion. This thesis focuses on Over-the-Air FL, a novel approach that reduces communication delays through analog wireless transmission, and investigates the impact of estimation bias, arising from biased model weight updates and heterogeneous data, on its convergence. We develop a system model for Over-the-Air FL, incorporating a power control scheme with over static channels. Our results provide a closed-form upper bound on the convergence rate and post- convergence error, indicating that the Over-the-Air FL algorithm achieves a linear convergence rate and non-diminishing post-convergence error in the worst case, and the performance improves with decreasing bias in model weight updates and more balanced data. Finally, experimental results on the real data support the theoretical findings.
Author: Wang, Yineng
Institution: KTH Royal Institute of Technology
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7000397787
