Research map / Learning theory and algorithms / Optimal transport
Dataset Distillation via the Wasserstein Metric
Haoyang Liu, Yijiang Li, Tiancheng Xing, Peiran Wang, Vibhu Dalal, Luwei Li, Jingrui He, Haohan 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 |
|---|---|---|---|
| add Per-Class BatchNorm (PCBN) instead of standard BN (ablation) | WMDD (Wasserstein loss) · ImageNette 10 IPC | top-1 accuracy: 60.7 -> 64.7 | Improved |
| add PCBN instead of standard BN (ablation) | WMDD (Wasserstein loss) · ImageNet-1K 10 IPC | top-1 accuracy: 26.8 -> 38.1 | Improved |
| add PCBN instead of standard BN (ablation) | CE feature loss · ImageNette 10 IPC | top-1 accuracy: 54.2 -> 63.5 | Improved |
| add PCBN instead of standard BN (ablation) | CE feature loss · Tiny ImageNet 10 IPC | top-1 accuracy: 38.0 -> 41.0 | Improved |
| use Wasserstein barycenter loss instead of cross-entropy (ablation) | WMDD with PCBN · ImageNet-1K 10 IPC | top-1 accuracy: 36.4 -> 38.1 | Improved |
| use Wasserstein barycenter loss instead of cross-entropy (ablation) | WMDD with PCBN · ImageNette 10 IPC | top-1 accuracy: 63.5 -> 64.7 | Improved |
| use Wasserstein loss instead of cross-entropy with standard BN (ablation) | WMDD with BN · Tiny ImageNet 10 IPC | top-1 accuracy: 38.0 -> 36.6 | Hurt |
| replace Wasserstein metric with sliced Wasserstein (ablation) | WMDD · Tiny ImageNet 50 IPC | top-1 accuracy: 59.4 -> 58.3 | No clear effect |
| use features from layer 18 instead of layer 5 (ablation) | WMDD · Tiny ImageNet | top-1 accuracy: 2.4 -> 41.8 | Improved |
| replace Wasserstein metric with MMD (ablation) | WMDD · Tiny ImageNet and ImageNet-1K | top-1 accuracy | Hurt |
| use WMDD instead of CDA for ViT-Tiny evaluation | WMDD · ImageNet-1K 50 IPC cross-architecture | top-1 accuracy: 31.22 -> 34.25 | Improved |
Closest papers by meaning
- Flowing Datasets with Wasserstein over Wasserstein Gradient Flows (ICML 2025 · similarity 0.88)
- Distilling Long-tailed Datasets (CVPR 2025 · similarity 0.87)
- Breaking Class Barriers: Efficient Dataset Distillation via Inter-Class Feature Compensator (ICLR 2025 · similarity 0.87)
- Large Scale Dataset Distillation with Domain Shift (ICML 2024 · similarity 0.86)
- Hyperbolic Dataset Distillation (NeurIPS 2025 · similarity 0.86)
- Going Beyond Feature Similarity: Effective Dataset distillation based on Class-aware Conditional Mutual Information (ICLR 2025 · similarity 0.85)
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.
