Research map / Image generation / Autoregressive and latent generation
Unifying Dataset Pruning and Distillation for Efficient Large-scale Compression
Lingao Xiao, Songhua Liu, Yang He, Xinchao 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 |
|---|---|---|---|
| train on pruned (EL2N) subset instead of random subset, hard labels (ablation) | ImageNet-1K IPC10 · ImageNet-1K ResNet-18 | top-1 accuracy: 4.6 -> 12.2 | Improved |
| add cropping-free combination of pruned images (ablation) | PCA · ImageNet-1K IPC10 hard label | top-1 accuracy: 12.2 -> 14.4 | Improved |
| add constrained augmentation (ablation) | PCA · ImageNet-1K IPC10 hard label | top-1 accuracy: 14.4 -> 22.8 | Improved |
| use soft labels with distilled images vs random subset | SRe2L/CDA · ImageNet-1K IPC10 soft label | top-1 accuracy: 35.8 -> 33.5 | Hurt |
| use pruning subsets instead of random with soft labels | EL2N/AUM · ImageNet-1K IPC10 soft label | top-1 accuracy: 35.8 -> 41.5 | Improved |
| replace noise init with real image init in SRe2L (ablation) | SRe2L · ImageNet-1K IPC10 hard label | top-1 accuracy: 1.5 -> 2.5 | Improved |
| select hard instead of easy samples for pruning (ablation) | EL2N · ImageNet-1K IPC10 hard label | top-1 accuracy: 12.2 -> 0.2 | Hurt |
| enforce class balance in pruning (ablation) | EL2N easy · ImageNet-1K IPC50 hard label | top-1 accuracy: 0.3 -> 31.1 | Improved |
| crop patches before training (dataset cropping) (ablation) | PCA · ImageNet-1K IPC10 hard label | top-1 accuracy: 22.8 -> 17.1 | Hurt |
| restrict training crop ratio to r=1.0 (no crop) (ablation) | PCA · ImageNet-1K IPC10 | top-1 accuracy: 22.8 -> 18.7 | Hurt |
| add CutMix with label mixing at probability 1.0 (ablation) | PCA SGD · ImageNet-1K IPC10 | top-1 accuracy: 25.6 -> 17.4 | Hurt |
| add Cutout (ablation) | PCA SGD · ImageNet-1K IPC10 | top-1 accuracy: 25.6 -> 26.2 | Improved |
Closest papers by meaning
- Are Large-scale Soft Labels Necessary for Large-scale Dataset Distillation? (NeurIPS 2024 · similarity 0.91)
- Exploring Learning Complexity for Efficient Downstream Dataset Pruning (ICLR 2025 · similarity 0.90)
- DD-Ranking: Rethinking the Evaluation of Dataset Distillation (NeurIPS 2026 · similarity 0.90)
- A Label is Worth A Thousand Images in Dataset Distillation (NeurIPS 2024 · similarity 0.89)
- Lightweight Dataset Pruning without Full Training via Example Difficulty and Prediction Uncertainty (ICML 2025 · similarity 0.89)
- GIFT: Unlocking Full Potential of Labels in Distilled Dataset at Near-zero Cost (ICLR 2025 · similarity 0.88)
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.
