Research map / Architectures and efficiency / RNNs and normalization
Dataset Condensation via Efficient Synthetic-Data Parameterization
Jang-Hyun Kim, Jinuk Kim, Seong Joon Oh, Sangdoo Yun, Hwanjun Song, Joonhyun Jeong, Jung-Woo Ha, Hyun Oh Song
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 multi-formation instead of identity formation (ablation) | IDC on CIFAR-10 · CIFAR-10, 50 images/class, ConvNet-3 | top-1 test accuracy: 69.5 -> 74.5 | Improved |
| Train the matching network on real data rather than synthetic data (ablation) | gradient-matching condensation · CIFAR-10, 50 images/class, ConvNet-3 | top-1 test accuracy: 25.8 -> 67.0 | Improved |
| Use MSE rather than cosine matching with real-trained networks (ablation) | IDC-I · CIFAR-10, 50 images/class, ConvNet-3 | top-1 test accuracy: 63.4 -> 67.0 | Improved |
| Add stronger network regularization (ablation) | Real+MSE condensation · CIFAR-10, 50 images/class, ConvNet-3 | top-1 test accuracy: 67.0 -> 69.5 | Improved |
| Downsample synthetic data after optimization instead of optimizing with the formation function (ablation) | IDC-I-post · CIFAR-10, same stored pixel budget, ConvNet-3 | top-1 test accuracy: 68.8 vs 74.5 | Hurt |
| Raise the multi-formation factor from 2 to 3 or 4 (ablation) | IDC · CIFAR-10, 50 images/class, ConvNet-3 | top-1 test accuracy: 74.5 -> 68.9; 74.5 -> 62.0 | Hurt |
| Raise the multi-formation factor from 2 to 3 or 4 (ablation) | IDC · ImageNet-10, 20 images/class, ResNetAP-10 | top-1 test accuracy: 73.3 -> 76.6 -> 77.5 | Improved |
| Replace uniform formation with multi-scale formation (ablation) | IDC, CIFAR-10 · 10 images/class, ConvNet-3 | top-1 test accuracy: 67.5 -> 69.2 | Improved |
| Replace uniform formation with multi-scale formation (ablation) | IDC, CIFAR-10 · 50 images/class, ConvNet-3 | top-1 test accuracy: 74.5 -> 73.1 | Hurt |
| Replace upsampling with learnable FSRCNN formation at factor 2 (ablation) | IDC, CIFAR-10 · 10 images/class, ConvNet-3 | top-1 test accuracy: 67.5 -> 66.2 | Hurt |
| Replace upsampling with learnable FSRCNN formation at factor 3 (ablation) | IDC, CIFAR-10 · 10 images/class, ConvNet-3 | top-1 test accuracy: 66.7 -> 67.9 | Improved |
| Add ZCA whitening (ablation) | IDC, CIFAR-10 · 50 images/class, ConvNet-3 | top-1 test accuracy: 74.5 -> 72.5 | Hurt |
Closest papers by meaning
- Elucidating the Design Space of Dataset Condensation (NeurIPS 2024 · similarity 0.90)
- M3D: Dataset Condensation by Minimizing Maximum Mean Discrepancy (AAAI 2024 · similarity 0.88)
- Ameliorate Spurious Correlations in Dataset Condensation (ICML 2024 · similarity 0.88)
- CondTSF: One-line Plugin of Dataset Condensation for Time Series Forecasting (NeurIPS 2024 · similarity 0.88)
- Diversity-Driven Synthesis: Enhancing Dataset Distillation through Directed Weight Adjustment (NeurIPS 2024 · similarity 0.87)
- Multisize Dataset Condensation (ICLR 2024 · similarity 0.87)
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.
