Research map / Label-efficient and robust learning / Self-supervised and contrastive
Understanding the Robustness of Distributed Self-Supervised Learning Frameworks Against Non-IID Data
Xuanyu Chen, NAN YANG, Shuai Wang, Dong Yuan
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 MAE (MIM) instead of SimSiam (CL) pre-training | distributed ViT pre-training · CIFAR-100 label non-IID | IID-to-non-IID accuracy gap: SimSiam 5.11 drop vs MAE 1.09 drop | Improved |
| use MAE instead of SimSiam | distributed CNN pre-training · CIFAR-10 label non-IID | accuracy gap: 1.70 drop -> 0.31 drop | Improved |
| use MAE instead of SimSiam | distributed ViT pre-training · ImageNet feature non-IID | accuracy gap: 2.84 drop -> 0.18 drop | Improved |
| add MAR loss to MAE | MAE + ViT · local-global weight distance | l2-norm difference: 36.10 -> 35.75 (local 1 vs local 100) | Improved |
| replace SimSiam by MAE | ViT · local-global weight distance | l2-norm difference: 45.37 -> 36.10 (local 1 vs local 100) | Improved |
| increase average connectivity from about 5 to about 10 | decentralized MIM · CIFAR-100 | accuracy: 46.62 -> 52.28 (data size weights) | Improved |
| use federated instead of decentralized learning (connectivity about 5) | distributed MIM · CIFAR-100 | accuracy: 45.22-46.62 -> 52.72 | Improved |
| change consensus matrix scheme (data size, degree normalized, doubly stochastic, push sum) (ablation) | decentralized MIM · CIFAR-100 | accuracy: 52.28, 52.08, 52.18, 52.22 at connectivity about 10 | No clear effect |
| add A-MMD (mean sigma) instead of cosine similarity (ablation) | MAR loss · CIFAR-100 | accuracy: 52.47 -> 54.39 (gamma 1e-2) | Improved |
| use mean instead of median pairwise distance for A-MMD bandwidth (ablation) | MAR loss · CIFAR-100 | accuracy: 54.13 -> 54.39 (gamma 1e-2) | Improved |
| add alignment term (gamma 1e-2, A-MMD mean) to vanilla MIM (ablation) | MAE · CIFAR-100 | accuracy: 51.45 -> 54.39 | Improved |
| set regularization weight gamma to 1 instead of 1e-2 (ablation) | MAR loss · CIFAR-100 | accuracy: 54.39 -> 51.50 | Hurt |
Closest papers by meaning
- SemiDFL: A Semi-Supervised Paradigm for Decentralized Federated Learning (AAAI 2025 · similarity 0.87)
- (FL)$^2$: Overcoming Few Labels in Federated Semi-Supervised Learning (NeurIPS 2024 · similarity 0.87)
- Combating Data Imbalances in Federated Semi-supervised Learning with Dual Regulators (AAAI 2024 · similarity 0.87)
- Certifiably Robust Model Evaluation in Federated Learning under Meta-Distributional Shifts (ICML 2025 · similarity 0.86)
- On the Importance of Embedding Norms in Self-Supervised Learning (ICML 2025 · similarity 0.86)
- Decentralized Sporadic Federated Learning: A Unified Algorithmic Framework with Convergence Guarantees (ICLR 2025 · similarity 0.86)
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.
