Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Update Estimation and Scheduling for Over-the-Air Federated Learning with Energy Harvesting Devices

Furkan Bagci, Busra Tegin, Mohammad Kazemi, Tolga M. Duman

TL;DR

This paper tackles OTA-FL with energy harvesting devices over a fading MAC under highly non-iid data. It introduces two scheduling strategies to select diverse users: an entropy-based method using known data distributions and a least-squares–based update-representation estimation method for unknown distributions, leveraging OTA aggregation and multiple PS antennas without CSIT. A formal convergence bound is derived, highlighting the roles of energy harvesting, channel noise, and partial participation, and guiding scheduling to minimize the error term between full and partial participation. Numerical results on MNIST, FMNIST, and CIFAR-10 demonstrate that the proposed scheduling and update estimation substantially improve learning performance, especially under strong data heterogeneity, with potential extensions to clustered FL on inferred user groups.

Abstract

We study over-the-air (OTA) federated learning (FL) for energy harvesting devices with heterogeneous data distribution over wireless fading multiple access channel (MAC). To address the impact of low energy arrivals and data heterogeneity on global learning, we propose user scheduling strategies. Specifically, we develop two approaches: 1) entropy-based scheduling for known data distributions and 2) least-squares-based user representation estimation for scheduling with unknown data distributions at the parameter server. Both methods aim to select diverse users, mitigating bias and enhancing convergence. Numerical and analytical results demonstrate improved learning performance by reducing redundancy and conserving energy.

Update Estimation and Scheduling for Over-the-Air Federated Learning with Energy Harvesting Devices

TL;DR

This paper tackles OTA-FL with energy harvesting devices over a fading MAC under highly non-iid data. It introduces two scheduling strategies to select diverse users: an entropy-based method using known data distributions and a least-squares–based update-representation estimation method for unknown distributions, leveraging OTA aggregation and multiple PS antennas without CSIT. A formal convergence bound is derived, highlighting the roles of energy harvesting, channel noise, and partial participation, and guiding scheduling to minimize the error term between full and partial participation. Numerical results on MNIST, FMNIST, and CIFAR-10 demonstrate that the proposed scheduling and update estimation substantially improve learning performance, especially under strong data heterogeneity, with potential extensions to clustered FL on inferred user groups.

Abstract

We study over-the-air (OTA) federated learning (FL) for energy harvesting devices with heterogeneous data distribution over wireless fading multiple access channel (MAC). To address the impact of low energy arrivals and data heterogeneity on global learning, we propose user scheduling strategies. Specifically, we develop two approaches: 1) entropy-based scheduling for known data distributions and 2) least-squares-based user representation estimation for scheduling with unknown data distributions at the parameter server. Both methods aim to select diverse users, mitigating bias and enhancing convergence. Numerical and analytical results demonstrate improved learning performance by reducing redundancy and conserving energy.

Paper Structure

This paper contains 12 sections, 5 theorems, 27 equations, 3 figures.

Table of Contents

  1. Introduction
  2. System Model
  3. EH Devices
  4. OTA FL with EH Devices with Unit Battery
  5. Convergence Analysis
  6. Preliminaries
  7. Convergence Rate
  8. Update Estimation and User Scheduling
  9. Entropy-based User Scheduling with Known Data Distributions
  10. User Clustering and Scheduling with Unknown Data Distribution
  11. Numerical Results
  12. Conclusions

Key Result

Theorem 1

For $0 < \eta(t) \leq \min \left\{ 1, \frac{1}{\mu \tau} \right\}, \forall t$. We have with for some constant $c \geq 0$.

Figures (3)

  • Figure 1: The mean test accuracy of entropy-based scheduling for CIFAR-10 with $M = 100$, $\left| \mathcal{B}_{m} \right| = 500$ and $p_e^{m}(t) = 0.1$, $\forall m, t$.
  • Figure 2: The mean test accuracy for MNIST with $M = 40$, $\left| \mathcal{B}_{m} \right| = 1250$ and $p_e^{m}(t) = 0.25$, $\forall m, t$.
  • Figure 3: The mean test accuracy for MNIST and FMNIST.

Theorems & Definitions (10)

  • Theorem 1
  • proof
  • Lemma 1
  • proof
  • Lemma 2
  • proof
  • Lemma 3
  • proof
  • Lemma 4
  • proof