Research map / Label-efficient and robust learning / Recognition and long-tail
On the Pitfall of Mixup for Uncertainty Calibration
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 (α=1.0) instead of ERM | ResNet models · SVHN, CIFAR-10, CIFAR-100, and Tiny-ImageNet | calibrated ECE: 0.75 -> 1.14; 0.83 -> 0.83; 2.64 -> 3.02; 1.36 -> 2.26 | Hurt |
| Use mixup inference in training with Δλ>1/2 across network blocks | MIT-A · SVHN | calibrated ECE: 0.75 -> 0.70 | Helped |
| Use mixup inference in training with Δλ>1/2 across network blocks | MIT-A · CIFAR-10 | calibrated ECE: 0.83 -> 0.78 | Helped |
| Use mixup inference in training with Δλ>1/2 across network blocks | MIT-A · CIFAR-100 | calibrated ECE: 2.64 -> 2.00 | Helped |
| Use mixup inference in training with Δλ>1/2 across network blocks | MIT-A · Tiny-ImageNet | calibrated ECE: 1.36 -> 1.39 | No clear effect |
| Use mixup inference in training with Δλ>1/2 across network blocks | MIT-A · Tiny-ImageNet | accuracy: 48.5 -> 50.8 | Helped |
| Remove random perturbation and retain only data transformation (Mixup-DT) (ablation) | Mixup · SVHN, CIFAR-10, CIFAR-100, and Tiny-ImageNet | calibrated ECE | Hurt |
| Mix target labels only, retaining transformed labels (Mixup-TO) (ablation) | Mixup · SVHN, CIFAR-10, CIFAR-100, and Tiny-ImageNet | calibrated ECE | Hurt |
| Mix inputs only and train on original one-hot labels (Mixup-IO) (ablation) | Mixup · SVHN, CIFAR-10, CIFAR-100, and Tiny-ImageNet | accuracy | Hurt |
| Mix within the same class and train on the original one-hot label (Mixup-SC) (ablation) | Mixup · SVHN, CIFAR-10, CIFAR-100, and Tiny-ImageNet | accuracy | Hurt |
| Constrain the coefficient gap Δλ>1/2 (ablation) | MIT-A · SVHN, CIFAR-10, CIFAR-100, and Tiny-ImageNet | accuracy | Helped |
| Apply the decoupling process only in the last layer rather than across all blocks (ablation) | Mixup inference in training · SVHN, CIFAR-10, CIFAR-100, and Tiny-ImageNet | accuracy | Mixed |
Closest papers by meaning
- Tailoring Mixup to Data for Calibration (ICLR 2025 · similarity 0.90)
- Beyond One-Hot Labels: Semantic Mixing for Model Calibration (ICML 2025 · similarity 0.89)
- Provable Uncertainty Decomposition via Higher-Order Calibration (ICLR 2025 · similarity 0.87)
- Reassessing How to Compare and Improve the Calibration of Machine Learning Models (ICLR 2025 · similarity 0.86)
- Selective Mixup Helps with Distribution Shifts, But Not (Only) because of Mixup (ICML 2024 · similarity 0.85)
- Pushing Boundaries: Mixup's Influence on Neural Collapse (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.