Research map / Label-efficient and robust learning / Noisy labels
Over-Training with Mixup May Hurt Generalization
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 |
|---|---|---|---|
| Train with Mixup for many epochs instead of stopping earlier | ResNet18 on CIFAR10 · CIFAR10 test set | testing accuracy | Hurt |
| Train with Mixup for many epochs instead of stopping earlier | ResNet34 on CIFAR100 · CIFAR100 test set | testing accuracy | Hurt |
| Train with Mixup for many epochs instead of stopping earlier | ResNet18 on 10% CIFAR10 with data augmentation · CIFAR10 test set | testing error | Hurt |
| Train with Mixup for many epochs instead of stopping earlier | ResNet34 on 10% CIFAR100 with data augmentation · CIFAR100 test set | testing error | Hurt |
| Use a smaller training subset (ablation) | Mixup · CIFAR10 and SVHN | epoch of testing-performance turning point | Hurt |
| Add data augmentation (ablation) | Mixup · CIFAR10 and CIFAR100 | epoch of testing-performance turning point | Helped |
| Switch from Mixup to ERM at an appropriate training time (ablation) | Mixup training · CIFAR10 and SVHN | testing performance | Helped |
| Switch from Mixup to ERM too early (ablation) | Mixup training · CIFAR10 and SVHN | testing performance | Hurt |
| Switch from Mixup to ERM too late (ablation) | Mixup training · CIFAR10 and SVHN | testing performance | Hurt |
| Use fixed λ=0.5 instead of λ drawn from Beta(1, 1) (ablation) | Mixup in the teacher-student regression setting · Toy regression test set | epoch when testing loss begins increasing | Hurt |
| Use RegMixup with η=2 instead of standard Mixup (ablation) | ResNet18 · CIFAR10 | testing performance under over-training | No clear effect |
| Use RegMixup with η=0.1 instead of standard Mixup (ablation) | ResNet18 · CIFAR10 | testing performance under over-training | Helped |
Closest papers by meaning
- C-Mixup: Improving Generalization in Regression (NeurIPS 2022 · similarity 0.91)
- Harnessing Hard Mixed Samples with Decoupled Regularizer (NeurIPS 2023 · similarity 0.90)
- Selective Mixup Helps with Distribution Shifts, But Not (Only) because of Mixup (ICML 2024 · similarity 0.89)
- Pushing Boundaries: Mixup's Influence on Neural Collapse (ICLR 2024 · similarity 0.88)
- Tailoring Mixup to Data for Calibration (ICLR 2025 · similarity 0.87)
- Catch-Up Mix: Catch-Up Class for Struggling Filters in CNN (AAAI 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.