Research map / Label-efficient and robust learning / Recognition and long-tail
The Effects of Regularization and Data Augmentation are Class Dependent
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 |
|---|---|---|---|
| introduce random crop data augmentation (alpha in [100,8]) | ResNet-50 ImageNet · ImageNet barn spider class | class test accuracy: 68% -> 46% | Hurt |
| increase random crop strength (tau decreased from 100 to 8) (ablation) | ResNet-50 ImageNet · ImageNet academic gown class | class test accuracy: 62% -> 40% | Hurt |
| increase random crop strength (ablation) | ResNet-50 ImageNet · ImageNet all classes average | average test accuracy | Helped |
| add weight decay regularization with varying coefficient (ablation) | ResNet-50 ImageNet · ImageNet per-class accuracy | per-class test accuracy | Mixed |
| apply random crop DA during ImageNet pre-training then linear probe | ResNet-50 transfer to INaturalist · INaturalist class 8889 | class test accuracy: 70% -> 30% | Hurt |
| add random crop DA | ResNet-50 ImageNet · ImageNet 1000 classes, t-test | share of classes with significantly lower accuracy: 4.4% of classes at 95% confidence; 2.1% at 99% | Mixed |
Closest papers by meaning
- Understanding the detrimental class-level effects of data augmentation (NeurIPS 2023 · similarity 0.90)
- Time Matters in Regularizing Deep Networks: Weight Decay and Data Augmentation Affect Early Learning Dynamics, Matter Little Near Convergence (NeurIPS 2019 · similarity 0.89)
- Deconstructing the Regularization of BatchNorm (ICLR 2021 · similarity 0.89)
- Why Do We Need Weight Decay in Modern Deep Learning? (NeurIPS 2024 · similarity 0.88)
- Understanding deep learning requires rethinking generalization (ICLR 2017 · similarity 0.87)
- Mirror, Mirror of the Flow: How Does Regularization Shape Implicit Bias? (ICML 2025 · 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.