Work

Communication Design for Dynamic Distributed Learning Systems

Year: 2026
Type: dissertation
Abstract: Distributed learning leverages distributed data across multiple entities to collaboratively train a shared model on their collective data, without requiring local data to leave its respective entity. It has gained attention compared to non-distributed traditional learning frameworks, as it enhances privacy by keeping data local and enables efficient parallelized computation required for model training.Despite its advantages, distributed learning relies on frequent signal exchange and communication between the server and the data-holding entities throughout the learning process, making \textit{communication} a critical factor. Therefore, the design of methods and parameters that influence this communication is crucial, as it directly affects learning performance, the utilization of available resources, and the privacy of data stored locally on each entity. Furthermore, in real-world environments, system characteristics are often dynamic and uncertain. As a result, communication designs must be adaptive and capable of operating effectively under unknown future conditions. In this dissertation, we aim to design communication strategies for dynamic distributed learning systems to enhance efficiency in terms of learning performance, wireless resource utilization, and the privacy of data stored locally on participating entities. Our initial focus is on a Federated Learning (FL) system with over-the-air aggregation for uplink transmission from data-holding devices to a multi-antenna server. We first propose joint designs for receive beamforming and device selection to enhance convergence of the training loss, subject to limited device transmit power. Next, we improve device transmit power efficiency while ensuring training convergence by jointly designing the receive and transmit beamforming under two scenarios: when the server has perfect channel state information (CSI) and when it has imperfect CSI. We then consider a digital uplink transmission scheme with dynamic and diverse communication conditions across devices, where we design adaptive sparsification levels to minimize communication delay while maintaining a desired training convergence level. Finally, for an FL system with over-the-air aggregation and time-varying channels, we adaptively design receive scaling factors to enhance the privacy of local data while guaranteeing training convergence under transmit power constraints. We conduct both theoretical and numerical evaluations of the proposed solutions. Our numerical results demonstrate their superiority over state-of-the-art benchmarks.
Language: English
Open Access status: green
OpenAlex ID: W7203960192
Cites: 98
Cited by:
Related to: 10