Research map / Label-efficient and robust learning / Knowledge distillation
Rectified Decoupled Dataset Distillation: A Closer Look for Fair and Comprehensive Evaluation
Xinhao Zhong, Shuoyang Sun, Xulin Gu, Chenyang Zhu, Bin Chen, Yaowei Wang
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 |
|---|---|---|---|
| re-evaluate all methods under a unified post-evaluation protocol (ablation) | decoupled dataset distillation methods · ImageNet-1K IPC=10 | top-1 accuracy: SRe2L 21.3 -> 40.2; EDC 48.6 -> 46.9 | Improved |
| add multi-teacher hybrid soft labels (ablation) | SRe2L, RDED, D4M, EDC · ImageNet-1K IPC=1 | top-1 accuracy: SRe2L 4.1 -> 12.2; RDED 7.6 -> 15.6 | Improved |
| add multi-teacher hybrid soft labels (ablation) | EDC · ImageNet-1K IPC=50 | top-1 accuracy: 60.1 -> 57.1 | Hurt |
| use Gaussian noise instead of real/RDED initialization (ablation) | EDC · ImageNet-1K IPC=10 | top-1 accuracy: 46.9 -> 36.5 | Hurt |
| initialize with random real images instead of noise (ablation) | G-VBSM · ImageNet-1K IPC=10 | top-1 accuracy: 41.5 -> 46.4 | Improved |
| replace KL loss with MSE+GT loss (ablation) | decoupled DD methods · ImageNet-1K IPC=1/10/50/100 | top-1 accuracy: EDC IPC=100 63.2 -> 64.1; SRe2L IPC=1 4.1 -> 3.6 | Mixed |
| random sampling with soft labels as distilled set | random baseline vs distillation methods · ImageNet-1K IPC=10 | top-1 accuracy: random 45.8 vs SRe2L 40.2 | Improved |
| use soft labels instead of hard labels for random images (ablation) | random sampling · ImageNet-1K IPC=10 | top-1 accuracy: 6.1 -> 45.8 | Improved |
| smaller learning-rate smoothing factor with larger eval models (ablation) | evaluation protocol · ImageNet-1K IPC=10 | top-1 accuracy | Mixed |
| use diffusion-based generation (D4M) instead of optimization (SRe2L) | evaluation on ViT-B-16 · ImageNet-1K IPC=50 | top-1 accuracy: 53.4 -> 63.7 | Improved |
Closest papers by meaning
- Embarrassingly Simple Dataset Distillation (ICLR 2024 · similarity 0.93)
- DD-Ranking: Rethinking the Evaluation of Dataset Distillation (NeurIPS 2026 · similarity 0.90)
- Towards Lossless Dataset Distillation via Difficulty-Aligned Trajectory Matching (ICLR 2024 · similarity 0.89)
- Rethinking Dataset Distillation for Classification: Do Distilled Sets Outperform Coresets? (NeurIPS 2026 · similarity 0.89)
- Large Scale Dataset Distillation with Domain Shift (ICML 2024 · similarity 0.89)
- A Label is Worth A Thousand Images in Dataset Distillation (NeurIPS 2024 · similarity 0.89)
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.
