Research map / Architectures and efficiency / RNNs and normalization
Dataset Distillation by Matching Training Trajectories
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 long-range trajectory matching instead of short-range matching | Trajectory Matching · CIFAR-100, 1 image/class; longer real-training ranges | normalized parameter-space MSE | Helped |
| Use long-range trajectory matching rather than short-range matching | CIFAR-10 and CIFAR-100 dataset distillation · CIFAR-10 and CIFAR-100 | test accuracy | Helped |
| Add ZCA whitening (ablation) | Trajectory Matching · CIFAR-100, 1 image/class | accuracy: 19.1 $\pm$ 0.6 -> 24.3 $\pm$ 0.3 | Helped |
| Add differentiable augmentation (ablation) | Trajectory Matching · CIFAR-100, 1 image/class | accuracy: 20.7 $\pm$ 0.7 -> 24.3 $\pm$ 0.3 | Helped |
| Initialize synthetic images from real samples instead of Gaussian noise (ablation) | Trajectory Matching · CIFAR-100, 1 image/class | accuracy: 22.9 $\pm$ 0.4 -> 24.3 $\pm$ 0.3 | Helped |
| Use ZCA normalization (ablation) | Trajectory Matching · CIFAR distillation | distillation performance | No clear effect |
| Increase synthetic images per class from 1 to 10 | Trajectory Matching · CIFAR-10 | test accuracy: 46.3% to 65.3% | Helped |
| Increase synthetic images per class from 10 to 50 | Trajectory Matching · CIFAR-10 | test accuracy: 65.3% to 71.5% | Helped |
| Use more expert trajectories (ablation) | Trajectory Matching · CIFAR-100, 1 image/class | validation accuracy | Helped |
| Bound the maximum starting timestep to exclude very early or very late trajectory points (ablation) | Trajectory Matching · CIFAR-100, 1 image/class | validation accuracy | Helped |
| Use a learnable synthetic learning rate (ablation) | Trajectory Matching · CIFAR distillation | performance across learning-rate initializations | Helped |
| Match expert and student output logits instead of using parameter distance (ablation) | Trajectory Matching · Dataset distillation | distillation performance | No clear effect |
Closest papers by meaning
- Towards Lossless Dataset Distillation via Difficulty-Aligned Trajectory Matching (ICLR 2024 · similarity 0.92)
- What is Dataset Distillation Learning? (ICML 2024 · similarity 0.91)
- Data Distillation Can Be Like Vodka: Distilling More Times For Better Quality (ICLR 2024 · similarity 0.90)
- Dataset Distillation via Knowledge Distillation: Towards Efficient Self-Supervised Pre-training of Deep Networks (ICLR 2025 · similarity 0.90)
- Large Scale Dataset Distillation with Domain Shift (ICML 2024 · similarity 0.90)
- Distilling Long-tailed Datasets (CVPR 2025 · similarity 0.90)
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.