Research map / Federated learning and privacy / Personalized and heterogeneous FL
Divergence-aware Federated Self-Supervised Learning
Weiming Zhuang, Yonggang Wen, Shuai Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use divergence-aware EMA to blend local client parameters with the global model | FedBYOL · CIFAR-10 and CIFAR-100, non-IID linear evaluation | top-1 accuracy: 79.44 ± 0.99 -> 83.34 ± 0.39; 57.51 ± 0.09 -> 61.78 ± 0.14 | Improved |
| Use divergence-aware EMA to blend local client parameters with the global model | FedBYOL · CIFAR-10 and CIFAR-100, non-IID semi-supervised evaluation | top-1 accuracy: 70.48 ± 0.30 -> 73.44 ± 0.22; 76.95 ± 0.46 -> 79.49 ± 0.34; 30.21 ± 0.40 -> 33.04 ± 0.23; 47.07 ± 0.14 -> 50.48 ± 0.11 | Improved |
| Update both online and target encoders with the global model, removing retained local target-encoder knowledge (ablation) | FedBYOL · CIFAR-10, non-IID 2 or 4 classes per client | top-1 accuracy: 79.44 -> 74.50; 82.82 -> 78.77 | Hurt |
| Remove EMA updates of the target encoder (ablation) | FedBYOL · CIFAR-10 non-IID linear evaluation | top-1 accuracy: 79.44 -> 50.20 | Hurt |
| Remove both EMA and stop-gradient (ablation) | FedBYOL · CIFAR-10 non-IID linear evaluation | top-1 accuracy: 79.44 -> 11.97 | Hurt |
| Remove EMA and stop-gradient, but update both encoders from the global model (ablation) | FedBYOL · CIFAR-10 non-IID linear evaluation | top-1 accuracy: 11.97 -> 68.75 | Improved |
| Apply FedEMA only to the predictor instead of to encoder and predictor (ablation) | FedEMA · CIFAR-10, linear evaluation | top-1 accuracy: 83.34 -> 81.13 | Hurt |
| Apply FedEMA only to the encoder instead of to encoder and predictor (ablation) | FedEMA · CIFAR-10, linear evaluation | top-1 accuracy: 83.34 -> 82.39 | Hurt |
| Use a constant EMA decay instead of divergence-aware dynamic decay (ablation) | FedEMA · CIFAR-10 non-IID | top-1 accuracy | Hurt |
| Use λ > 1, causing the decay μ to reach 1 at the start (ablation) | FedEMA · CIFAR datasets | top-1 accuracy | Hurt |
| Remove stop-gradient from FedBYOL (ablation) | FedBYOL · Non-IID CIFAR-10 | kNN testing accuracy | No clear effect |
| Remove the predictor (ablation) | FedBYOL and FedSimSiam · Non-IID CIFAR-10 | kNN testing accuracy | Hurt |
Closest papers by meaning
- Fed-QSSL: A Framework for Personalized Federated Learning under Bitwidth and Data Heterogeneity (AAAI 2024 · similarity 0.89)
- FedSC: Provable Federated Self-supervised Learning with Spectral Contrastive Objective over Non-i.i.d. Data (ICML 2024 · similarity 0.88)
- (FL)$^2$: Overcoming Few Labels in Federated Semi-Supervised Learning (NeurIPS 2024 · similarity 0.88)
- FedQS: Optimizing Gradient and Model Aggregation for Semi-Asynchronous Federated Learning (NeurIPS 2025 · similarity 0.87)
- Convergence Analysis of Split Federated Learning on Heterogeneous Data (NeurIPS 2024 · similarity 0.87)
- Combating Data Imbalances in Federated Semi-supervised Learning with Dual Regulators (AAAI 2024 · similarity 0.87)
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.
