Research map / Label-efficient and robust learning / Knowledge distillation
Understanding Dataset Distillation via Spectral Filtering
Deyu Bo, Songhua Liu, 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 |
|---|---|---|---|
| add signal-matching loss to filter-matching loss (ablation) | CFM · CIFAR-100 IPC=50 | accuracy: 64.95 -> 67.32 | Improved |
| add classification loss on top of filter+signal losses (ablation) | CFM · CIFAR-100 IPC=50 | accuracy: 67.32 -> 72.48 | Improved |
| add classification loss on top of filter+signal losses (ablation) | CFM · ImageNet-1k IPC=50 | accuracy: 55.76 -> 57.32 | Improved |
| add signal-matching loss (ablation) | CFM · Tiny-ImageNet IPC=50 | accuracy: 51.20 -> 51.92 | Improved |
| use curriculum high-pass filter (CFM) vs fixed filters (ablation) | CFM · Tiny-ImageNet / ImageNet-1k | accuracy | Improved |
| use low-pass filter instead of high-pass (ablation) | CFM · Tiny-ImageNet / ImageNet-1k | accuracy | Hurt |
| use extremely high-frequency filter beta=1e-4 (ablation) | CFM · Tiny-ImageNet / ImageNet-1k | accuracy | Hurt |
| CFM vs RDED at low budget | CFM · ImageNet-1k IPC=10 (ResNet-18) | accuracy: 42.0 (RDED) vs 40.6 (CFM) | Hurt |
| CFM vs RDED at low budget | CFM · Tiny-ImageNet IPC=50 (ResNet-18) | accuracy: 58.2 (RDED) vs 58.0 (CFM) | Hurt |
| CFM vs FrePo/MTT on small dataset | CFM · CIFAR-10 IPC=10 (ConvNet-128) | accuracy: 65.5 (FrePo) vs 52.1 (CFM) | Hurt |
| CFM vs DWA | CFM · CIFAR-100 IPC=10 (ConvNet-128) | accuracy: 47.6 -> 58.3 | Improved |
| CFM vs G-VBSM across architectures | CFM · ImageNet-1k IPC=50, DenseNet-121 | accuracy: 58.47 -> 60.91 | Improved |
Closest papers by meaning
- Towards Universal Dataset Distillation via Task-Driven Diffusion (CVPR 2025 · similarity 0.87)
- Large Scale Dataset Distillation with Domain Shift (ICML 2024 · similarity 0.87)
- Breaking Class Barriers: Efficient Dataset Distillation via Inter-Class Feature Compensator (ICLR 2025 · similarity 0.87)
- Rethinking Dataset Distillation for Classification: Do Distilled Sets Outperform Coresets? (NeurIPS 2026 · similarity 0.87)
- Emphasizing Discriminative Features for Dataset Distillation in Complex Scenarios (CVPR 2025 · similarity 0.87)
- DD-Ranking: Rethinking the Evaluation of Dataset Distillation (NeurIPS 2026 · similarity 0.87)
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.
