Research map / Label-efficient and robust learning / Recognition and long-tail
Tailoring Mixup to Data for Calibration
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Similarity Kernel Mixup, which warps Beta interpolation coefficients per pair based on the embedding (classification) or label (regression) distance, improving accuracy and calibration over Mixup variants at lower cost.
Problem
Vanilla Mixup mixes arbitrary pairs with the same interpolation distribution, causing manifold mismatch and noisy labels that hurt calibration; prior calibration-driven Mixup methods are costly.
Approach
Warp the Beta(1,1) interpolation coefficients with an inverse-Beta-CDF warping parameterized by a Gaussian similarity kernel on the rescaled squared distance between paired samples, so similar pairs are mixed more strongly and dissimilar pairs weakly, while keeping in-batch mixing and lookup-table efficiency.
Key results
- On CIFAR10 with ResNet34, SK Mixup (tau_std=0.2) reaches 96.29% accuracy and 1.81 ECE vs 94.69% and 3.65 ECE for ERM.
- On CIFAR100-C with ResNet101, SK RegMixup achieves 57.98% accuracy and 6.89 ECE, versus 55.52% and 16.95 for RegMixup.
- SK Mixup is about 1.5x faster than MIT-A and about 3x faster than RankMixup per the efficiency comparison on CIFAR10/CIFAR100/Tiny-ImageNet with ResNet50.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Similarity Kernel Mixup instead of ERM | ResNet101 · CIFAR100 | ECE: 14.33 $\pm$ 0.17 -> 2.23 $\pm$ 0.39 | Helped |
| Combine SK Mixup with RegMixup | ResNet101 · CIFAR10 | accuracy: 95.04 $\pm$ 0.07 -> 96.02 $\pm$ 0.04 | Helped |
| Use SK RegMixup rather than RegMixup | ResNet101 · CIFAR100-C | ECE: 16.95 $\pm$ 0.37 -> 6.89 $\pm$ 1.38 | Helped |
| Use SK RegMixup rather than RegMixup | ResNet50 · ImageNet-R | accuracy: 25.70 -> 25.92 | Helped |
| Use SK Mixup rather than ERM | regression models · Airfoil | MAPE: 1.720 { $\pm$ 0.219} -> 1.609 { $\pm$ 0.137} | Helped |
| Mix only pairs with lower distances rather than higher distances | Mixup · CIFAR10 | ECE: Lower 0.5: 3.04 $\pm$ 0.4; Higher 0.5: 4.55 $\pm$ 0.63 | Helped |
| Increase τstd from 0.2 to 0.4 (ablation) | SK Mixup · CIFAR10 | accuracy: 96.29 $\pm$ 0.07 -> 96.42 $\pm$ 0.24 | No clear effect |
| Use warping with lookup tables rather than warping without lookup tables (ablation) | SK Mixup · CIFAR10 | total training time (min): 57.36 $\pm$ 0.31 -> 55.59 $\pm$ 0.4 | Helped |
| Remove warping functions (ablation) | SK Mixup · CIFAR10 | accuracy: 96.42 $\pm$ 0.24 -> 96.3 $\pm$ 0.1 | No clear effect |
Surprising findings
- Mixing only pairs with lower-than-median distance improves calibration but mixing only dissimilar pairs can degrade accuracy, and Mixup can worsen calibration after temperature scaling relative to ERM.
Limitations
- Calibration analysis does not account for the effect of temperature scaling theoretically, which is left for future work.
- Introduces two hyperparameters (tau_max, tau_std) that need cross-validation.
- Results on regression use only tabular/time-series datasets and not segmentation or structured tasks.
Experimental setup
Datasets: CIFAR10 CIFAR100 Tiny-ImageNet ImageNet Airfoil Exchange-Rate Electricity
Benchmarks: CIFAR10 CIFAR100 Tiny-ImageNet ImageNet CIFAR10-C CIFAR100-C ImageNet-R ImageNet-A Airfoil Exchange-Rate Electricity
Metrics: accuracy ECE AECE Brier score NLL MAPE UCE ENCE RMSE
Baselines: ERM Mixup Mixup IO Manifold Mixup RegMixup RankMixup MIT-A C-Mixup
Builds on: Mixup Manifold Mixup RegMixup C-Mixup RankMixup
Closest papers by meaning
- On the Pitfall of Mixup for Uncertainty Calibration (CVPR 2023 · similarity 0.90)
- When and How Mixup Improves Calibration (ICML 2022 · similarity 0.90)
- Beyond One-Hot Labels: Semantic Mixing for Model Calibration (ICML 2025 · similarity 0.87)
- Selective Mixup Helps with Distribution Shifts, But Not (Only) because of Mixup (ICML 2024 · similarity 0.86)
- Who’s the (Multi-)Fairest of Them All: Rethinking Interpolation-Based Data Augmentation Through the Lens of Multicalibration (AAAI 2025 · similarity 0.86)
- First-Order Manifold Data Augmentation for Regression Learning (ICML 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.