Research map / Label-efficient and robust learning / Recognition and long-tail
How Much Data Are Augmentations Worth? An Investigation into Scaling Laws, Invariance, and Implicit Regularization
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 |
|---|---|---|---|
| diverse but inconsistent augmentation (TrivialAug) vs consistent flips at small sizes (ablation) | TrivialAug vs horizontal flips · CINIC-10 in-domain, 48000 base samples | exchange rate | Mixed |
| adding random crops to flips (Flips&Crops) vs flips alone (ablation) | ResNet-18 · CINIC-10 in-domain | exchange rate: 2.94 vs 1.79 at 4x | Helped |
| TrivialAug augmentation evaluated on CIFAR-10-C (out-of-domain) vs in-domain | TrivialAug&Flips&Crops · CIFAR-10-C | exchange rate | Helped |
| horizontal flips as augmentation at larger repetition counts on CIFAR-10-C (ablation) | Horiz. Flips · CIFAR-10-C | exchange rate: 0.84 at 1x down to 0.52 at 32x | Hurt |
| invariant architectures (flip-invariant, E2CNN) vs random flip augmentation at small/medium sizes | ResNet-18 invariant architectures · CINIC-10 / CIFAR-10 rotated invariance experiments | test accuracy | Hurt |
| random augmentation applied per batch (vs same-batch or fixed views) for flips and crops (ablation) | Flips & Crops ResNet-18 · CIFAR-10 end-of-training gradient std and flatness | gradient std / flatness: Grad std 24.7397 random vs 10.263 fixed views+same batch | Helped |
| random augmentation sampling vs no augmentation (ablation) | ResNet-18 with TrivialAug&Flips&Crops · CIFAR-10 flatness | flatness: Flatness 20.1336 vs 11.4192 no augmentation | Helped |
Closest papers by meaning
- A Data-Augmentation Is Worth A Thousand Samples: Analytical Moments And Sampling-Free Training (NeurIPS 2022 · similarity 0.89)
- A Group-Theoretic Framework for Data Augmentation (NeurIPS 2020 · similarity 0.89)
- Tradeoffs in Data Augmentation: An Empirical Study (ICLR 2021 · similarity 0.89)
- Characterizing the Generalization Error of Random Feature Regression with Arbitrary Data-Augmentation (NeurIPS 2026 · similarity 0.88)
- Understanding when Dynamics-Invariant Data Augmentations Benefit Model-free Reinforcement Learning Updates (ICLR 2024 · similarity 0.85)
- Revisiting Data Augmentation in Deep Reinforcement Learning (ICLR 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.