Research map / Label-efficient and robust learning / Recognition and long-tail
On Mixup Regularization
Luigi Carratino, Moustapha Cissé, Rodolphe Jenatton, Jean-Philippe Vert
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| apply the derived test-time rescaling to Mixup-trained models (ablation) | Mixup (ResNet-50, ImageNet) · ImageNet standard (alpha=0.5) | ECE: 0.0707 -> 0.0280 | Improved |
| apply the derived test-time rescaling to Mixup-trained models (ablation) | Mixup (ResNet-50, ImageNet) · ImageNet standard (alpha=0.5) | test CE-loss: 0.9482 -> 0.8852 | Improved |
| apply the derived test-time rescaling to Mixup-trained models (ablation) | Mixup (ResNet-50, ImageNet) · ImageNet standard (alpha=0.25) | accuracy: 0.7754 -> 0.7764 | Improved |
| apply the derived test-time rescaling to Mixup-trained models (ablation) | Mixup (ResNet-50, ImageNet) · ImageNet standard (alpha=0.25) | ECE: 0.0236 -> 0.0291 | Hurt |
| apply the derived test-time rescaling to Mixup-trained models (ablation) | Mixup (ResNet-50, ImageNet) · ImageNet-A (alpha=0.5) | ECE: 0.3161 -> 0.4318 | Hurt |
| apply the derived test-time rescaling to Mixup-trained models (ablation) | Mixup (ResNet-50, ImageNet) · ImageNet-C (alpha=0.5) | accuracy: 0.4761 -> 0.4719 | Hurt |
| apply the derived test-time rescaling to Mixup-trained models (ablation) | Mixup (ResNet-50, ImageNet) · ImageNet-V2 (alpha=0.5) | accuracy: 0.6575 -> 0.6593 | Improved |
| apply the derived test-time rescaling to Mixup-trained models (ablation) | Mixup (ResNet-50, ImageNet) · ObjectNet (alpha=0.5) | accuracy: 0.2825 -> 0.2837 | Improved |
| apply the derived test-time rescaling to Mixup-trained models (ablation) | Mixup (ResNet-50, ImageNet) · CIFAR-10-C and CIFAR-100-C | accuracy, CE-loss, ECE: gap grows in favor of standard Mixup as noise intensity grows | Hurt |
| apply the Mixup data transformation to an ERM-trained model (ablation) | Mixup (ResNet-50, ImageNet) · CIFAR-10, CIFAR-100, ImageNet | accuracy, loss, ECE: performance always deteriorates | Hurt |
| apply the test-time transformation to Mixup with larger alpha (ablation) | Mixup (ResNet-50, ImageNet) · CIFAR-10, CIFAR-100, ImageNet | accuracy, loss, ECE: bigger improvements for bigger alpha | Improved |
Closest papers by meaning
- Over-Training with Mixup May Hurt Generalization (ICLR 2023 · similarity 0.92)
- Towards Understanding the Data Dependency of Mixup-style Training (ICLR 2022 · similarity 0.91)
- Selective Mixup Helps with Distribution Shifts, But Not (Only) because of Mixup (ICML 2024 · similarity 0.91)
- On the Pitfall of Mixup for Uncertainty Calibration (CVPR 2023 · similarity 0.90)
- Tailoring Mixup to Data for Calibration (ICLR 2025 · similarity 0.88)
- Pushing Boundaries: Mixup's Influence on Neural Collapse (ICLR 2024 · similarity 0.88)
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.
