Research map / Optimization / Implicit bias and generalization
Data-Efficient Augmentation for Training Neural Networks
Read the paperCheck 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 |
|---|---|---|---|
| Select and augment a gradient-space coreset instead of a random subset | CIFAR10, ResNet20 · CIFAR10, 0.1% subset | accuracy: 33.5 -> 37.8 | Helped |
| Select and augment a gradient-space coreset instead of a random subset | Caltech256, ResNet18 · Caltech256, 10% subset | accuracy: 61.8 -> 65.4 | Helped |
| Select and augment a gradient-space coreset instead of a random subset | ImageNet, ResNet50 · ImageNet, 30% subset | accuracy: 70.39 -> 71.99 | Helped |
| Augment a selected 50% coreset instead of the full dataset | CIFAR10 with 50% label noise, ResNet20 · 50% label noise | accuracy: 75.87 -> 76.20 | Helped |
| Use the selected coreset rather than a random subset for augmentation | Caltech256, ResNet18 · 30% subset with CutOut | accuracy: 43.32 -> 55.53 | Helped |
| Use the selected coreset rather than a random subset for augmentation | Caltech256, ResNet18 · 30% subset with AugMix | accuracy: 40.77 -> 52.72 | Helped |
| Use the selected coreset rather than a random subset for augmentation | Caltech256, ResNet18 · 30% subset with perturbation augmentation | accuracy: 48.51 -> 58.29 | Helped |
| Use the selected coreset rather than a random subset for augmentation | CIFAR10, ResNet20 · 30% subset augmentation over full data | accuracy improvement over weak augmentation: 65.6% -> 75.0% | Helped |
| Use the selected coreset rather than a random subset for augmentation | CIFAR10 with 50% label noise, ResNet20 · 50% subset | accuracy: 75.36 ± 0.05 -> 76.20 ± 0.75** | Helped |
| Use max-loss selection instead of gradient-space coreset selection | CIFAR10, ResNet20 · 0.1% subset | accuracy: 37.8 -> <15% | Hurt |
| Use max-loss selection instead of gradient-space coreset selection | CIFAR10, ResNet20 · training on augmented subsets | convergence | Hurt |
| Vary the selected subset size from 20% to 50% (ablation) | TinyImageNet, ResNet50 · TinyImageNet | accuracy: 51.99 -> 55.16 | Helped |
Closest papers by meaning
- Tradeoffs in Data Augmentation: An Empirical Study (ICLR 2021 · similarity 0.90)
- Data Augmentation as Feature Manipulation (ICML 2022 · similarity 0.89)
- Efficient Augmentation via Data Subsampling (ICLR 2019 · similarity 0.89)
- When Dynamic Data Selection Meets Data Augmentation: Achieving Enhanced Training Acceleration (ICML 2025 · similarity 0.88)
- First-Order Manifold Data Augmentation for Regression Learning (ICML 2024 · similarity 0.87)
- Boost Self-Supervised Dataset Distillation via Parameterization, Predefined Augmentation, and Approximation (ICLR 2025 · 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.