Research map / Label-efficient and robust learning / Knowledge distillation
Distill Gold from Massive Ores: Bi-level Data Pruning towards Efficient Dataset Distillation
Yue Xu, Yong-Lu Li, Kaitong Cui, Ziyu Wang, Cewu Lu, Yu-Wing Tai, Chi-Keung Tang
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 |
|---|---|---|---|
| prune real data by empirical loss before distillation (ablation) | DC, CIFAR10 IPC=1 · CIFAR10 IPC=1 (DC) | accuracy (%): 28.3 -> 29.7 at 99% pruning | Improved |
| random pruning of real data before distillation (ablation) | DC, CIFAR10 IPC=1 · CIFAR10 IPC=1 (DC) | maximum pruning ratio: 90% | No clear effect |
| replace random with empirical-loss pruning on DC CIFAR10 IPC=1 (ablation) | DC · CIFAR10 IPC=1 | maximum pruning ratio: 90% -> 99.5% | Improved |
| use empirical-loss pruning on MTT/IDC/IDM at larger IPC (ablation) | MTT, IDC, IDM · CIFAR10 IPC=10 | maximum pruning ratio: IDM 10% -> 10% | Mixed |
| add adaptive ITE-based pruning on top of preemptive pruning (ablation) | BiLP with IDC · CIFAR10 IPC=10 | top-1 accuracy (%): 68.5 -> 69.2 | Improved |
| prune both small and large ITE instead of one direction (ablation) | BiLP · CIFAR100 IPC=1 | top-1 accuracy (%): 29.8 -> 30.1 | Improved |
| add Taylor approximation of ITE (ablation) | ITE computation · CIFAR10 IPC=10 | total iteration time (s/iter): 23.38 -> 0.65 | Improved |
| add lazy selection (update every 10 iterations) (ablation) | BiLP · CIFAR10 IPC=10 | total iteration time (s/iter): 0.65 -> 0.47 | Improved |
| apply BiLP to DC and IDC | DC, IDC · CIFAR10 IPC=1 | top-1 accuracy (%): 50.6 -> 51.5 | Improved |
| use larger multi-formation factor (x3) with BiLP (ablation) | BiLP+IDC · CIFAR100 IPC=1 | top-1 accuracy (%): 30.1 -> 34.0 | Improved |
| apply BiLP preemptive pruning to DM on ImageNet-1K | DM · ImageNet-1K IPC=10 | accuracy (%): 3.86 -> 5.21 | Improved |
| use empirical loss pruning instead of coreset methods CRAIG, GradMatch, GLISTER (ablation) | DC, CIFAR10 IPC=50 · CIFAR10 IPC=50 pruning 70% | accuracy (%): 49.0 -> 54.1 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DREAM — DREAM: Efficient Dataset Distillation by Representative Matching (ICCV 2023) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Rethinking Dataset Distillation for Classification: Do Distilled Sets Outperform Coresets? (NeurIPS 2026 · similarity 0.90)
- Boost Self-Supervised Dataset Distillation via Parameterization, Predefined Augmentation, and Approximation (ICLR 2025 · similarity 0.90)
- Embarrassingly Simple Dataset Distillation (ICLR 2024 · similarity 0.90)
- Large Scale Dataset Distillation with Domain Shift (ICML 2024 · similarity 0.89)
- What is Dataset Distillation Learning? (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.
