Research map / Federated learning and privacy / Personalized and heterogeneous FL
Soft Separation and Distillation: Toward Global Uniformity in Federated Unsupervised Learning
Hung-Chieh Fang, Hsuan-Tien Lin, Irwin King, Yifei 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 |
|---|---|---|---|
| add projector distillation alone (ablation) | FedAlignUniform · CIFAR10 cross-silo (K=10) | linear probing accuracy: 80.84 -> 80.74 | No clear effect |
| add dimension-scaled regularization alone (ablation) | FedAlignUniform · CIFAR10 cross-silo (K=10) | linear probing accuracy: 80.84 -> 81.05 | Improved |
| add dimension-scaled regularization alone (ablation) | FedAlignUniform · CIFAR10 cross-silo (K=10) | negative uniformity loss: 3.79 -> 3.81 | Improved |
| add projector distillation on top of DSR (ablation) | FedAlignUniform + DSR · CIFAR10 cross-silo (K=10) | linear probing accuracy: 81.05 -> 81.32 | Improved |
| add projector distillation on top of DSR (ablation) | FedAlignUniform + DSR · CIFAR10 cross-silo (K=10) | negative uniformity loss: 3.81 -> 3.84 | Improved |
| add projector distillation on top of DSR (ablation) | FedAlignUniform + DSR · CIFAR10 cross-silo (K=10) | fine-tuning 1% accuracy: 69.77 -> 70.74 | Improved |
| add DSR when projector is removed (ablation) | FedAlignUniform · CIFAR10 cross-silo (K=10) no projector | linear probing accuracy: 73.16 -> 76.14 (+2.98) | Improved |
| add DSR when projector is kept (ablation) | FedAlignUniform · CIFAR10 cross-silo (K=10) | linear probing accuracy: 80.84 -> 81.05 (+0.21) | Improved |
| keep the projector (vs removing it) with DSR (ablation) | FedAlignUniform + DSR · CIFAR10 cross-silo (K=10) | linear probing accuracy: 76.14 -> 81.05 | Improved |
| use MSE instead of KL for projector distillation (ablation) | SSD · CIFAR10 cross-silo (K=10) | linear probing accuracy: 81.32 -> 81.88 | No clear effect |
| replace FedAlignUniform with SSD | FedAlignUniform · CIFAR100 cross-device (K=50) | linear probing accuracy: 43.03 -> 45.21 | Improved |
| replace FedAlignUniform with SSD | FedAlignUniform · CIFAR100 to CIFAR10 transfer | linear probing accuracy: 77.66 -> 78.48 | Improved |
Closest papers by meaning
- Rethinking the Representation in Federated Unsupervised Learning with Non-IID Data (CVPR 2024 · similarity 0.90)
- Convergence Analysis of Split Federated Learning on Heterogeneous Data (NeurIPS 2024 · similarity 0.87)
- (FL)$^2$: Overcoming Few Labels in Federated Semi-Supervised Learning (NeurIPS 2024 · similarity 0.87)
- FedUV: Uniformity and Variance for Heterogeneous Federated Learning (CVPR 2024 · similarity 0.86)
- Resource-Aware Federated Self-Supervised Learning with Global Class Representations (NeurIPS 2024 · similarity 0.86)
- A Mutual Information Perspective on Federated Contrastive Learning (ICLR 2024 · 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.
