Research map / Label-efficient and robust learning / Knowledge distillation
Dataset Distillation by Automatic Training Trajectories
Dai Liu, Jindong Gu, Hu Cao, Carsten Trinitis, Martin Schulz
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 |
|---|---|---|---|
| replace fixed trajectory length with automatic (minimum-distance) trajectory length | MTT/FTD · CIFAR-10 IPC=10 cross-architecture (VGG11) | test accuracy: 47.0 -> 54.2 | Improved |
| replace fixed trajectory length with automatic trajectory length | FTD · CIFAR-10 IPC=50 cross-architecture (AlexNet) | test accuracy: 53.8 -> 60.0 | Improved |
| replace fixed trajectory length with automatic trajectory length | FTD · CIFAR-10 IPC=10 (ConvNet) | test accuracy: 66.6 -> 67.7 | Improved |
| replace fixed trajectory length with automatic trajectory length | FTD · ImageMeow IPC=10 | test accuracy: 43.3 -> 45.5 | Improved |
| replace fixed trajectory length with automatic trajectory length | FTD · Tiny ImageNet IPC=10 | test accuracy: 24.5 -> 25.8 | Improved |
| use ATT instead of KIP on CIFAR-10 IPC=1 | KIP · CIFAR-10 IPC=1 | test accuracy: 49.9 -> 48.3 | Hurt |
| use ATT instead of fixed trajectory length under learning-rate perturbations (ablation) | vanilla LDD · parameter variation (lr img, lr step size) | successful runs out of 20 | Improved |
| use ATT instead of fixed trajectory length | vanilla LDD · CIFAR-10 IPC=1 | time per iteration: 0.8s vs 0.8s | No clear effect |
| increase trajectory bound N_S for ATT (30, 50, 70, up to 130) (ablation) | ATT · CIFAR-10 IPC=1 | test accuracy | Improved |
| add error tolerance gamma to fixed-length matching (ablation) | FTL · CIFAR-10 IPC=1 (N_S=50) | step selection errors | No clear effect |
Closest papers by meaning
- Towards Lossless Dataset Distillation via Difficulty-Aligned Trajectory Matching (ICLR 2024 · similarity 0.92)
- Towards Stable and Storage-efficient Dataset Distillation: Matching Convexified Trajectory (CVPR 2025 · similarity 0.91)
- Distilling Long-tailed Datasets (CVPR 2025 · similarity 0.88)
- Embarrassingly Simple Dataset Distillation (ICLR 2024 · similarity 0.87)
- Path-Guided Flow Matching for Dataset Distillation (NeurIPS 2026 · similarity 0.87)
- SelMatch: Effectively Scaling Up Dataset Distillation via Selection-Based Initialization and Partial Updates by Trajectory Matching (ICML 2024 · similarity 0.86)
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.
