On federated learning over wireless channels with over-the-air aggregation

Date
2022-07
Editor(s)
Advisor
Duman, Tolga Mete
Supervisor
Co-Advisor
Co-Supervisor
Instructor
Source Title
Print ISSN
Electronic ISSN
Publisher
Bilkent University
Volume
Issue
Pages
Language
English
Journal Title
Journal ISSN
Volume Title
Series
Abstract

A decentralized machine learning (ML) approach called federated learning (FL) has recently been at the center of attention since it secures edge users’ data and decreases communication costs. In FL, a parameter server (PS), which keeps track of the global model orchestrates local training and global model aggregation across a set of mobile users (MUs). While there exist studies on FL over wireless channels, its performance on practical wireless communication scenarios has not been investigated very well. With this motivation, this thesis considers wireless FL schemes that use realistic channel models, and analyze the impact of different wireless channel effects. In the first part of the thesis, we study hierarchical federated learning (HFL) where intermediate servers (ISs) are utilized to make the server-side closer to the MUs. Clustering approach is used where MUs are assigned to ISs to perform multiple cluster aggregations before the global aggregation. We first analyze the performance of a partially wireless approach where the MUs send their gradients through a channel with path-loss and fading using over-the-air (OTA) aggregation. We assume that there is no inter-cluster interference and the gradients from the ISs to the PS are sent error-free. We show through numerical and experimental analysis that our proposed algorithm offers a faster convergence and lower power consumption compared to the standard FL with OTA aggregation. As an extension, we also examine a fully-wireless HFL setup where both the MUs and ISs send their gradients through OTA aggregation, taking into account the effect of inter-cluster interference. Our numerical and experimental results reveal that utilizing ISs results in a faster convergence and a better performance than the OTA FL without any IS while using less transmit power. It is also shown that the best choice of cluster aggregations depends on the data distribution among the MUs and the clusters. In the second part of the thesis, we study FL with energy harvesting MUs with stochastic energy arrivals. In every global iteration, the MUs with enough energy in their batteries perform local SGD iterations, and transmit their gradients using OTA aggregation. Before sending the gradients to the PS, the gradients are scaled with respect to the idle time and data cardinality of each MU, through a cooldown multiplier, to amplify the importance of the MUs that send less frequent local updates. We provide a convergence analysis of the proposed setup, and validate our results with numerical and neural network simulations under different energy arrival profiles. The results show that the OTA FL with energy harvesting devices performs slightly worse than the OTA FL without any energy restrictions, and that utilizing the excess energy for more local SGD iterations gives a better convergence rate than simply increasing the transmit power.

Course
Other identifiers
Book Title
Citation
Published Version (Please cite this version)