Research map / Label-efficient and robust learning / Recognition and long-tail
On the Generalization Effects of Linear Transformations in Data Augmentation
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 |
|---|---|---|---|
| add label-invariant transformations (rotation, random cropping, cutout) to training data (ablation) | MLP on MNIST · MNIST | intrinsic error score: baseline 1.52% to 0.77-1.43% | Helped |
| uncertainty-based sampling of transformations instead of RandAugment | Wide-ResNet-28-10 · CIFAR-10 and CIFAR-100 | test accuracy: +0.59% CIFAR-10, +1.24% CIFAR-100 | Helped |
| uncertainty-based sampling instead of RandAugment | ResNet-50 · ImageNet | top-1 accuracy: +1.54% (79.14 vs 77.60) | Helped |
| uncertainty-based sampling instead of RandAugment | Shake-Shake and PyramidNet+ShakeDrop · CIFAR-10 | test accuracy: +0.27% Shake-Shake, +0.16% PyramidNet | Helped |
| remove mixup from the transformation set (ablation) | uncertainty-based sampling on Wide-ResNet-28-10 · CIFAR-10 | test accuracy: still +0.45% over RandAugment without mixup | Mixed |
| increase augmented samples (C=8, S=4) versus C=4, S=1 (ablation) | uncertainty-based sampling on Wide-ResNet-28-10 · CIFAR-100 | test accuracy: 85.02% vs 84.54% | Helped |
| increase number of augmented samples (C=8, S=4) (ablation) | uncertainty-based sampling on Wide-ResNet-28-10 · CIFAR-10 | test accuracy: 98.16% vs 97.89% | Helped |
| mixing same-class digits (increasing same-class mixup fraction) (ablation) | MLP on MNIST · MNIST | instability score: instability decreases as same-class fraction increases | Helped |
| compose rotation with random cropping (ablation) | MLP on MNIST · MNIST | accuracy and intrinsic error: 98.72% acc, 0.86% error vs rotation 98.65% and random crop 98.61% | Helped |
| compose X and Y translations (ablation) | MLP on MNIST · MNIST | accuracy: 98.0% vs 98.38% (TranslateX) and 98.19% (TranslateY) | Hurt |
| uncertainty-based sampling vs RandAugment on IMDb sentiment with BERT-Large | BERT-Large · IMDb | test accuracy: 95.96% vs 95.22% baseline | Helped |
Closest papers by meaning
- A Data-Augmentation Is Worth A Thousand Samples: Analytical Moments And Sampling-Free Training (NeurIPS 2022 · similarity 0.89)
- Tradeoffs in Data Augmentation: An Empirical Study (ICLR 2021 · similarity 0.88)
- Automatic Data Augmentation via Invariance-Constrained Learning (ICML 2023 · similarity 0.88)
- Characterizing the Generalization Error of Random Feature Regression with Arbitrary Data-Augmentation (NeurIPS 2026 · similarity 0.88)
- A Group-Theoretic Framework for Data Augmentation (NeurIPS 2020 · similarity 0.88)
- Effective Data Augmentation With Diffusion Models (ICLR 2024 · similarity 0.86)
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.