Research map / Label-efficient and robust learning / Knowledge distillation
Dataset Distillation via Factorization
Songhua Liu, Kai Wang, Xingyi Yang, Jingwen Ye, 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 |
|---|---|---|---|
| Factorize distilled data into shared hallucinators and bases | MTT dataset distillation · SVHN, CIFAR10, CIFAR100 | test accuracy | Improved |
| Add hallucinator-basis factorization | DC, DM, and MTT · CIFAR10, ConvNet; 2, 11, and 51 IPC | test accuracy: +2.75, +4.59, +4.67; +2.75, +6.48, +2.41; +6.17, +5.58, +3.44 | Improved |
| Add hallucinator-basis factorization | DD distilled datasets · Cross-architecture CIFAR10 | test accuracy: +1.40 to +17.57 | Improved |
| Remove the cosine-similarity loss L_cos. (ablation) | HaBa · CIFAR10 | test accuracy | Hurt |
| Remove the contrastive loss L_con. (ablation) | HaBa · CIFAR10 | test accuracy | Hurt |
| Remove the task-specific loss L_task (ablation) | HaBa · CIFAR10 | test accuracy | Hurt |
| Use L_con. during downstream training (ablation) | HaBa · CIFAR10 | test accuracy | Improved |
| Use class-independent hallucinators instead of shared hallucinators (ablation) | HaBa · CIFAR10, low BPC | test accuracy | Improved |
| Use class-independent hallucinators instead of shared hallucinators (ablation) | HaBa · CIFAR10, higher BPC | test accuracy | Hurt |
| Use single-channel bases and increase BPC to use the same storage (ablation) | HaBa · CIFAR10, small BPC | test accuracy | Improved |
| Increase hallucinator count above 10 (ablation) | HaBa · CIFAR10, 10 or 50 BPC | test accuracy | No clear effect |
| Remove DSA data augmentation (ablation) | HaBa · CIFAR10, 10 BPC | test accuracy | Hurt |
Closest papers by meaning
- Boost Self-Supervised Dataset Distillation via Parameterization, Predefined Augmentation, and Approximation (ICLR 2025 · similarity 0.89)
- Large Scale Dataset Distillation with Domain Shift (ICML 2024 · similarity 0.89)
- Rethinking Dataset Distillation for Classification: Do Distilled Sets Outperform Coresets? (NeurIPS 2026 · similarity 0.89)
- DD-Ranking: Rethinking the Evaluation of Dataset Distillation (NeurIPS 2026 · similarity 0.88)
- Unlocking Dataset Distillation with Diffusion Models (NeurIPS 2025 · similarity 0.88)
- Hierarchical Features Matter: A Deep Exploration of Progressive Parameterization Method for Dataset Distillation (CVPR 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.
