Research map / Architectures and efficiency / RNNs and normalization
Dataset Distillation using Neural Feature Regression
Yongchao Zhou, Ehsan Nezhadarya, Jimmy Ba
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 |
|---|---|---|---|
| Use FRePo rather than MTT for one-image-per-class distillation | Dataset distillation · CIFAR100, 1 Img/Cls | test accuracy: 24.3 ± 0.3 -> 28.7 ± 0.1 | Improved |
| Use FRePo rather than MTT for one-image-per-class distillation | Dataset distillation · Tiny ImageNet, 1 Img/Cls | test accuracy: 8.8 ± 0.3 -> 15.4 ± 0.3 | Improved |
| Use FRePo rather than MTT for one-image-per-class distillation | Dataset distillation · CUB-200, 1 Img/Cls | test accuracy: 2.2 ± 0.1 -> 12.4 ± 0.2 | Improved |
| Use FRePo rather than MTT for 50 images per class | Dataset distillation · CIFAR100, 50 Img/Cls | test accuracy: 47.7 ± 0.2 -> 44.3 ± 0.2 | Hurt |
| Use FRePo rather than MTT for 10 images per class | Dataset distillation · CIFAR100, 10 Img/Cls | test accuracy: 40.1 ± 0.4 -> 42.5 ± 0.2 | Improved |
| Use FRePo instead of a random real subset | ImageNet distillation · ImageNet (64x64), 1 Img/Cls | Top1 accuracy: 1.1 ± 0.1 -> 7.5 ± 0.3 | Improved |
| Use a model pool rather than only one online model (ablation) | FRePo · CIFAR100, 1 Img/Cls | test accuracy | Improved |
| Use an intermediate number of online updates rather than too few or too many (ablation) | FRePo · CIFAR100, 1 Img/Cls | test accuracy | Mixed |
| Increase meta-gradient batch size (ablation) | FRePo · CIFAR100, 1 Img/Cls | test accuracy | Improved |
| Use TBPTT instead of the FRePo kernel-regression meta-gradient (ablation) | Dataset distillation · CIFAR100, 1 Img/Cls | test accuracy | Hurt |
| Compute the meta-gradient before rather than after the online model update (ablation) | FRePo · CIFAR100, 1 Img/Cls | test accuracy | No clear effect |
Closest papers by meaning
- Large Scale Dataset Distillation with Domain Shift (ICML 2024 · similarity 0.86)
- Boost Self-Supervised Dataset Distillation via Parameterization, Predefined Augmentation, and Approximation (ICLR 2025 · similarity 0.86)
- Embarrassingly Simple Dataset Distillation (ICLR 2024 · similarity 0.86)
- Provable and Efficient Dataset Distillation for Kernel Ridge Regression (NeurIPS 2024 · similarity 0.86)
- Dataset Distillation for Pre-Trained Self-Supervised Vision Models (NeurIPS 2025 · similarity 0.85)
- Rethinking Dataset Distillation for Classification: Do Distilled Sets Outperform Coresets? (NeurIPS 2026 · 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.
