Research map / Label-efficient and robust learning / Knowledge distillation
Soft Label Pruning and Quantization for Large-Scale Dataset Distillation
Lingao Xiao, Yang He
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 class-wise matching (ablation) | SRe2L baseline · ImageNet-1K IPC50 (ResNet-18), 50x labels | top-1 accuracy (%): 34.8 -> 37.7 | Improved |
| add class-wise BN supervision (ablation) | SRe2L + class-wise matching · ImageNet-1K IPC50 (ResNet-18), 50x labels | top-1 accuracy (%): 37.7 -> 39.7 | Improved |
| add batch-level label pruning (ablation) | LPQLD pipeline · ImageNet-1K IPC50 (ResNet-18), 50x labels | top-1 accuracy (%): 39.7 -> 43.1 | Improved |
| add dynamic knowledge reuse (temperature annealing) (ablation) | LPQLD pipeline · ImageNet-1K IPC50 (ResNet-18), 50x labels | top-1 accuracy (%): 43.1 -> 47.8 | Improved |
| add label quantization (top-k logits) (ablation) | LPQLD pipeline · ImageNet-1K IPC50 (ResNet-18), 50x labels | top-1 accuracy (%): 47.8 -> 49.7 | Improved |
| add calibrated student-teacher alignment (ablation) | LPQLD pipeline · ImageNet-1K IPC50 (ResNet-18), 50x labels | top-1 accuracy (%): 49.7 -> 51.3 | Improved |
| add class-wise matching alone at 100x labels (ablation) | SRe2L baseline · ImageNet-1K IPC50, 100x labels | top-1 accuracy (%): 25.4 -> 22.6 | Hurt |
| use class-wise batching and BN supervision for synthesis | SRe2L / CDA images · ImageNet-1K, cosine similarity (IPC50) | feature cosine similarity (lower better): 0.841 -> 0.796 | Improved |
| add 10x label pruning plus 8x quantization instead of pruning only at ~80x | LPLD vs LPQLD · ImageNet-1K IPC10 | top-1 accuracy (%): 20.1 -> 27.3 | Improved |
| use quantization after 10x pruning rather than 40x pruning only | LPLD · ImageNet-1K IPC50 | top-1 accuracy (%): 46.7 -> 51.3 | Improved |
| reduce top-k from 250 to 10 with 10x pruning (ablation) | LPQLD · ImageNet-1K IPC10 | top-1 accuracy (%): 32.2 -> 20.0 | Hurt |
| replace FKD Marginal Smoothing top-5 quantization with LPQLD pruning+quantization | FKD quantization · ImageNet-1K IPC10 | top-1 accuracy (%): 7.5 -> 29.6 | Improved |
Closest papers by meaning
- Are Large-scale Soft Labels Necessary for Large-scale Dataset Distillation? (NeurIPS 2024 · similarity 0.94)
- GIFT: Unlocking Full Potential of Labels in Distilled Dataset at Near-zero Cost (ICLR 2025 · similarity 0.90)
- A Label is Worth A Thousand Images in Dataset Distillation (NeurIPS 2024 · similarity 0.89)
- Rectifying Soft-Label Entangled Bias in Long-Tailed Dataset Distillation (NeurIPS 2025 · similarity 0.89)
- Label-Efficient Dataset Pruning via Semi-Supervised Pseudo-Labeling (NeurIPS 2026 · similarity 0.88)
- Exploring Learning Complexity for Efficient Downstream Dataset Pruning (ICLR 2025 · 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.
