Distributed Event-Based Learning via ADMM

Event-triggered communication protocol for distributed learning

Overview

We introduce an event-triggered communication protocol for distributed learning, built on the ADMM framework. Instead of communicating at every iteration, each agent transmits updates only when significant changes occur. This approach cuts down communication without sacrificing convergence or accuracy, and is resilient to heterogeneous data across agents. The algorithm supports convex and nonconvex objectives and accommodates a wide range of learning problems.

This work was conducted by Güner Dilsad Er and Michael Muehlebach at the Max-Planck Institute for Intelligent Systems, and Sebastian Trimpe at RWTH Aachen University.

Paper accepted to ICML 2025 — Read on arXiv

The authors also thank Michael Cummins for his valuable contribution to the codebase.

Code & Implementation

View on GitHub

Open-source Python implementation featuring:

  • Complete ADMM framework with event-triggered communication
  • Benchmark experiments on MNIST and CIFAR-10
  • Comparison with FedAvg, FedProx, SCAFFOLD, and FedADMM
  • Comprehensive documentation and reproducible experiments
  • Ready-to-use code for distributed learning research

📋 Read the Documentation for setup instructions, usage examples, and API reference.

Presentation

Key Features

  • Event-triggered ADMM: Communication occurs only when necessary, reducing overhead by over 35% on benchmark tasks
  • Global convergence: Proven linear rates in strongly convex settings and sublinear rates in general nonconvex problems
  • Robustness to failures: The method remains stable under packet drops, aided by a simple periodic reset mechanism
  • Strong baseline performance: Outperforms FedAvg, FedProx, SCAFFOLD, and FedADMM in both accuracy and communication efficiency

Results

Our experiments on MNIST and CIFAR-10 confirm that our method achieves target accuracy with fewer communication rounds. The method remains effective even in extreme data heterogeneity, such as one-class-per-agent scenarios.

Total communication events to reach accuracy targets:

Validation accuracy vs. communication rounds:

Code & Implementation

The full implementation is available on GitHub.

Citation

@article{er2024admm,
  title={Distributed Event-Based Learning via ADMM},
  author={Er, G. D. and Trimpe, S. and Mühlebach, M.},
  year={2025},
  journal={42nd International Conference on Machine Learning (ICML)},
  url={https://proceedings.mlr.press/v267/er25a.html}
}