Research map / Learning theory and algorithms / Kernels and regression theory
GeoDM: Geometry-aware Distribution Matching for Dataset Distillation
Xuhui Li, Zhengquan Luo, Zihui Cui, Kai Zhao, zhiqiang xu
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 |
|---|---|---|---|
| add product space (Euclidean+hyperbolic+spherical) instead of Euclidean-only matching (ablation) | NCFM-based distribution matching · CIFAR-10 IPC=10 | accuracy: 71.8 -> 73.5 | Improved |
| add learnable curvature and weights to product space (ablation) | GeoDM · CIFAR-10 IPC=10 | accuracy | Improved |
| add geometry-aware OT loss to the full product model (ablation) | GeoDM · CIFAR-10 IPC=10 | accuracy: 73.5 -> 74.4 | Improved |
| replace NCFM objective with simple DM objective inside product space (ablation) | GeoDM · CIFAR-10 IPC=10 | accuracy: up to 2.79% over single-space baselines | Improved |
| replace NCFM objective with DSDM inside product space (ablation) | GeoDM · CIFAR-10 IPC=10 | accuracy: around 2% | Improved |
| use product space with curvature and weights instead of Euclidean NCFM, at matched accuracy (ablation) | NCFM · CIFAR-10 IPC=10 total cost | GFLOPs: 270.0 -> 181.7 | Improved |
| add OT loss to product space with curvature and weights (ablation) | GeoDM · CIFAR-10 IPC=10 total cost | GFLOPs: 181.7 -> 242.4 | Hurt |
| transfer distilled data to ResNet-18 and AlexNet | NCFM · CIFAR-10 IPC=10 cross-architecture | accuracy: ResNet-18 67.3±0.4 -> 69.1±0.5 | Improved |
| apply GeoDM to higher-resolution data | NCFM · Imagenette IPC=10 | accuracy: 77.3 -> 78.8 | Improved |
| apply GeoDM at high IPC | NCFM · CIFAR-100 IPC=50 | accuracy: 54.7 -> 55.0 | No clear effect |
| apply GeoDM at low IPC | NCFM · CIFAR-100 IPC=1 | accuracy: 34.4 -> 38.0 | Improved |
| replace 50% of CIFAR-10 with Gaussian noise | GeoDM learned curvature · CIFAR-10 noise test | learned curvature: spherical 1.41 -> 0.84; hyperbolic -0.79 -> -0.48 | Hurt |
Closest papers by meaning
- Optimizing Distributional Geometry Alignment with Optimal Transport for Generative Dataset Distillation (NeurIPS 2025 · similarity 0.89)
- Hyperbolic Dataset Distillation (NeurIPS 2025 · similarity 0.89)
- Dataset Distillation of 3D Point Clouds via Distribution Matching (NeurIPS 2025 · similarity 0.88)
- Learning Distances from Data with Normalizing Flows and Score Matching (ICML 2025 · similarity 0.85)
- OPTICAL: Leveraging Optimal Transport for Contribution Allocation in Dataset Distillation (CVPR 2025 · similarity 0.85)
- Categorical Flow Matching on Statistical Manifolds (NeurIPS 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.
