Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SegOTA: Accelerating Over-the-Air Federated Learning with Segmented Transmission

Chong Zhang, Min Dong, Ben Liang, Ali Afana, Yahia Ahmed

TL;DR

SegOTA tackles high uplink latency in over-the-air federated learning by partitioning model updates into segments and assigning each segment to a device group for simultaneous transmission. By formulating a per-round online optimization that jointly groups devices and designs transmit-receive beamformers, and by deriving a tractable upper bound on the learning gap, SegOTA achieves fast convergence with significantly reduced communication rounds. The method leverages spherical k-means for spatially aware grouping and a block-coordinate optimization yielding closed-form solutions, with theoretical convergence guarantees. Simulations on MNIST show SegOTA outperforming full-model OTA and competitive baselines, approaching ideal performance for moderate segment counts and confirming practical latency reductions in multi-antenna wireless FL systems.

Abstract

Federated learning (FL) with over-the-air computation efficiently utilizes the communication resources, but it can still experience significant latency when each device transmits a large number of model parameters to the server. This paper proposes the Segmented Over-The-Air (SegOTA) method for FL, which reduces latency by partitioning devices into groups and letting each group transmit only one segment of the model parameters in each communication round. Considering a multi-antenna server, we model the SegOTA transmission and reception process to establish an upper bound on the expected model learning optimality gap. We minimize this upper bound, by formulating the per-round online optimization of device grouping and joint transmit-receive beamforming, for which we derive efficient closed-form solutions. Simulation results show that our proposed SegOTA substantially outperforms the conventional full-model OTA approach and other common alternatives.

SegOTA: Accelerating Over-the-Air Federated Learning with Segmented Transmission

TL;DR

SegOTA tackles high uplink latency in over-the-air federated learning by partitioning model updates into segments and assigning each segment to a device group for simultaneous transmission. By formulating a per-round online optimization that jointly groups devices and designs transmit-receive beamformers, and by deriving a tractable upper bound on the learning gap, SegOTA achieves fast convergence with significantly reduced communication rounds. The method leverages spherical k-means for spatially aware grouping and a block-coordinate optimization yielding closed-form solutions, with theoretical convergence guarantees. Simulations on MNIST show SegOTA outperforming full-model OTA and competitive baselines, approaching ideal performance for moderate segment counts and confirming practical latency reductions in multi-antenna wireless FL systems.

Abstract

Federated learning (FL) with over-the-air computation efficiently utilizes the communication resources, but it can still experience significant latency when each device transmits a large number of model parameters to the server. This paper proposes the Segmented Over-The-Air (SegOTA) method for FL, which reduces latency by partitioning devices into groups and letting each group transmit only one segment of the model parameters in each communication round. Considering a multi-antenna server, we model the SegOTA transmission and reception process to establish an upper bound on the expected model learning optimality gap. We minimize this upper bound, by formulating the per-round online optimization of device grouping and joint transmit-receive beamforming, for which we derive efficient closed-form solutions. Simulation results show that our proposed SegOTA substantially outperforms the conventional full-model OTA approach and other common alternatives.

Paper Structure

This paper contains 14 sections, 3 theorems, 33 equations, 4 figures.

Table of Contents

  1. Introduction
  2. Wireless FL System Model
  3. Uplink Segmented Over-the-Air Aggregation
  4. SegOTA Design Optimization
  5. Joint Optimization Formulation
  6. SegOTA Training Convergence Analysis
  7. Beamforming Optimization for SegOTA
  8. Device grouping via spherical k-means
  9. Joint transmit-receive beamforming
  10. Simulation Results
  11. Simulation Setup
  12. Performance Comparison
  13. Conclusion
  14. Proof of Lemma \ref{['lemma2']}

Key Result

Lemma 1

Consider SegOTA described in Section sec:SegOTA and the ideal centralized training described in Section sec:SegOTA_convrg_analysis. For $\eta_t<\frac{1}{L}$, $\forall t\in{\cal T}$, under Assumption assump_smooth, ${\mathbb{E}}[\|\tilde{{\bf v}}^{J}_{t} - \tilde{\boldsymbol \theta}^\star\|^2]$ is up

Figures (4)

  • Figure 1: Uplink analog OTA aggregation for wireless FL. Left: traditional full-model OTA approach; Right: proposed SegOTA (Each colored box represents a model segment that consists of $I_t$ parameters; different colors indicate different segments).
  • Figure 2: Test accuracy vs. number of model segments $S_t$ ($(N,K) = (32,50)$).
  • Figure 3: Test accuracy vs. number of model segments $S_t$ ($(N,K) = (64,50)$).
  • Figure 4: Test accuracy vs. number of devices $K$ for $N=32$ and $S_t=10$.

Theorems & Definitions (3)

  • Lemma 1: Bounding ${{\mathbb{E}}[\|\tilde{{\bf v}}^{J}_{t} - \tilde{\boldsymbol \theta}^\star\|^2] }$
  • Lemma 2: Bounding ${{\mathbb{E}}[\|\tilde{\boldsymbol \alpha}_{t}\|^2]}$, ${{\mathbb{E}}[\|\tilde{\boldsymbol \beta }_{t}\|^2]}$, and ${{\mathbb{E}}[\|\tilde{\boldsymbol \delta }_{t}\|^2]}$
  • Proposition 1