Over-the-Air Federated Learning with Energy Harvesting Devices

TL;DR

This paper explores federated learning with energy harvesting devices over wireless channels, proposing a weighted aggregation scheme that accounts for heterogeneous energy availability, with analysis and experiments showing robustness and some performance loss.

Contribution

It introduces an energy-aware weighted averaging method for over-the-air federated learning with energy harvesting devices, along with convergence analysis and experimental validation.

Findings

Robustness against heterogeneous energy arrivals in error-free scenarios.

5-10% performance loss due to energy harvesting constraints.

Effective convergence behavior demonstrated through numerical experiments.

Abstract

We consider federated edge learning (FEEL) among mobile devices that harvest the required energy from their surroundings, and share their updates with the parameter server (PS) through a shared wireless channel. In particular, we consider energy harvesting FL with over-the-air (OTA) aggregation, where the participating devices perform local computations and wireless transmission only when they have the required energy available, and transmit the local updates simultaneously over the same channel bandwidth. In order to prevent bias among heterogeneous devices, we utilize a weighted averaging with respect to their latest energy arrivals and data cardinalities. We provide a convergence analysis and carry out numerical experiments with different energy arrival profiles, which show that even though the proposed scheme is robust against devices with heterogeneous energy arrivals in error-free scenarios, we observe a 5-10% performance loss in energy harvesting OTA FL.