Research map / Optimization / Distributed and compressed
Distributed Event-Based Learning via ADMM
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes an event-triggered over-relaxed ADMM for distributed learning that communicates only when local variables change beyond a threshold, with convergence analysis (accelerated in convex case), communication-failure robustness via periodic resets, and MNIST/CIFAR-10 experiments against FedAvg, FedProx, SCAFFOLD and FedADMM.
Problem
Distributed learning over a network with non-i.i.d. data suffers from high communication overhead and sensitivity to communication failures, and existing random agent selection can miss important updates.
Approach
Combine over-relaxed ADMM with sent-on-delta style event-triggered communication (vanilla and randomized triggering) and periodic resets, analyzed as a perturbed dynamical system with a Lyapunov function that accounts for communication-induced errors, making convergence agnostic to data heterogeneity.
Key results
- On CIFAR-10 the event-based method reaches 78% accuracy with 18,376 communication events versus 27,000 for FedADMM.
- On MNIST the randomized event-based variant reaches 80% with 629 communication events versus 800 for FedAvg and FedADMM.
- Communication savings of more than 35% are reported with less than 1% loss in solution accuracy.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use event-based communication instead of FedADMM | Event-Based ADMM (Randomized) · CIFAR-10 78% target accuracy | communication events: 27000 -> 18376 | Helped |
| use randomized instead of vanilla event-based communication (ablation) | Event-Based ADMM · MNIST 80% target accuracy | communication events: 816 -> 629 | Helped |
| use randomized instead of vanilla event-based communication (ablation) | Event-Based ADMM · MNIST 85% target accuracy | communication events: 1285 -> 693 | Helped |
| use randomized instead of vanilla event-based communication (ablation) | Event-Based ADMM · MNIST 90% target accuracy | communication events: 1710 -> 1723 | Hurt |
| use randomized instead of vanilla event-based communication (ablation) | Event-Based ADMM · CIFAR-10 70% target accuracy | communication events: 12214 -> 12531 | Hurt |
| use randomized instead of vanilla event-based communication (ablation) | Event-Based ADMM · CIFAR-10 75% target accuracy | communication events: 14780 -> 13422 | Helped |
| use randomized instead of vanilla event-based communication (ablation) | Event-Based ADMM · CIFAR-10 77% target accuracy | communication events: 14780 -> 15008 | Hurt |
| use randomized instead of vanilla event-based communication (ablation) | Event-Based ADMM · CIFAR-10 78% target accuracy | communication events: 20690 -> 18376 | Helped |
Surprising findings
- Without periodic resets, inter-agent errors accumulate rapidly under packet drops and prevent convergence.
- Randomized event triggering often improves on vanilla event triggering in the communication-accuracy trade-off.
Limitations
- Does not account for adversarial attacks such as gradient poisoning or unreliable nodes.
- Not analyzed for differential privacy.
- Strong convexity needed for linear and accelerated rates; nonconvex result only gives sublinear rates.
Experimental setup
Datasets: MNIST CIFAR-10
Benchmarks: MNIST CIFAR-10
Metrics: validation accuracy number of communication events
Baselines: FedAvg FedProx SCAFFOLD FedADMM
Builds on: ADMM sent-on-delta FedADMM SCAFFOLD
Closest papers by meaning
- A Delay-tolerant Proximal-Gradient Algorithm for Distributed Learning (ICML 2018 · similarity 0.88)
- Tight analyses of first-order methods with error feedback (NeurIPS 2025 · similarity 0.86)
- B-ary Tree Push-Pull Method is Provably Efficient for Distributed Learning on Heterogeneous Data (NeurIPS 2024 · similarity 0.85)
- Approximate Gradient Coding for Distributed Learning with Heterogeneous Stragglers (NeurIPS 2025 · similarity 0.85)
- Does Worst-Performing Agent Lead the Pack? Analyzing Agent Dynamics in Unified Distributed SGD (NeurIPS 2024 · similarity 0.84)
- Federated Learning under Periodic Client Participation and Heterogeneous Data: A New Communication-Efficient Algorithm and Analysis (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.