Research map / Label-efficient and robust learning / Recognition and long-tail
Leveraging Hierarchical Feature Sharing for Efficient Dataset Condensation
Haizhong Zheng, Jiachen Sun, Shutong Wu, Bhavya Kailkhura, Zhuoqing Morley Mao, Chaowei Xiao, Atul Prakash
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 image container with HMN under the same gradient matching loss | data parameterization for dataset condensation · CIFAR10 10 IPC | test accuracy (%): 67.8 (LinBa) -> 73.7 | Improved |
| add over-budget condensation with double-end pruning of instance-level memories (ablation) | HMN · CIFAR10 50 IPC | test accuracy (%): 75.4 -> 76.9 | Improved |
| prune easy examples only instead of double-end pruning (ablation) | HMN · CIFAR10 1 IPC | test accuracy (%): 65.7 -> 65.3 | Hurt |
| add over-budget pruning versus in-budget condensation at 1 IPC (ablation) | HMN · CIFAR10 1 IPC | test accuracy (%): 65.1 -> 65.7 | Improved |
| increase instance-level memory size (GIPC falls from 85 to 13) (ablation) | HMN · CIFAR10 1 IPC | test accuracy (%): 65.1% -> 48.2% | Hurt |
| prune 30% of instances in 1 IPC HMN (ablation) | HMN · CIFAR10 1 IPC | test accuracy (%): 66.2% -> 62.2% | Hurt |
| prune 30% of instances in 50 IPC HMN (ablation) | HMN · CIFAR10 50 IPC | test accuracy (%): 76.7% -> 76.5% | No clear effect |
| prune 10% highest-AUM images of HaBa-generated data (ablation) | HaBa · CIFAR10 10 IPC | test accuracy (%): 69.5 -> 69.5 | No clear effect |
| replace cosine annealing with multi-step LR scheduler (ablation) | HMN · CIFAR10 | test accuracy (%): 73.7 -> 73.4 (10 IPC) | No clear effect |
| replace trajectory loss with gradient matching loss in HaBa and LinBa (ablation) | HaBa/LinBa · CIFAR10 10 IPC | test accuracy (%): 69.9 -> 61.8 (HaBa) | Hurt |
| use HMN condensed with ConvNet to train other architectures | HMN vs IDC · CIFAR10 1 IPC cross-architecture (VGG16) | test accuracy (%): 28.7 -> 58.5 | Improved |
| use HMN with batch-based loss instead of HaBa with trajectory loss | HMN · CIFAR10 50 IPC GPU memory | GPU memory: 48276M -> 10426M | Improved |
Closest papers by meaning
- Hierarchical Features Matter: A Deep Exploration of Progressive Parameterization Method for Dataset Distillation (CVPR 2025 · similarity 0.88)
- Elucidating the Design Space of Dataset Condensation (NeurIPS 2024 · similarity 0.88)
- Boost Self-Supervised Dataset Distillation via Parameterization, Predefined Augmentation, and Approximation (ICLR 2025 · similarity 0.86)
- Are Large-scale Soft Labels Necessary for Large-scale Dataset Distillation? (NeurIPS 2024 · similarity 0.86)
- Multisize Dataset Condensation (ICLR 2024 · similarity 0.86)
- M3D: Dataset Condensation by Minimizing Maximum Mean Discrepancy (AAAI 2024 · 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.
