Research map / Label-efficient and robust learning / Recognition and long-tail
RankMixup: Ranking-Based Mixup Training for Network Calibration
Jongyoun Noh, Hyekang Park, Junghyup Lee, Bumsub Ham
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 |
|---|---|---|---|
| increase MRL margin m beyond 2.0 on CIFAR10 (ablation) | RankMixup (MRL) · CIFAR10 | ECE (%): 1.01 -> 13.13 (m=2.0 -> 5.0) | Hurt |
| increase number of augmented samples Q in M-NDCG (ablation) | RankMixup (M-NDCG) · Tiny-ImageNet | ECE (%): 3.25 -> 1.30 (Q=2 -> Q=6) | Improved |
| increase number of augmented samples Q in M-NDCG (ablation) | RankMixup (M-NDCG) · CIFAR10 | ECE (%): 3.01 -> 2.24 (Q=2 -> Q=6) | Improved |
| use flat Beta(alpha,alpha) with large alpha (2 on CIFAR10, 1 on Tiny-ImageNet) (ablation) | RankMixup (M-NDCG) · CIFAR10, Tiny-ImageNet | ECE/AECE | Improved |
| replace vanilla mixup with Manifold-Mixup in RankMixup (ablation) | RankMixup (M-NDCG) · Tiny-ImageNet | ECE (%): 1.49 -> 1.20 | Improved |
| replace vanilla mixup with Manifold-Mixup in RankMixup (ablation) | RankMixup (M-NDCG) · CIFAR10 | ECE (%): 2.59 -> 2.92 | Hurt |
| replace label-mixture supervision with RankMixup ranking losses vs RegMixup variants | RankMixup with Manifold-Mixup/CutMix · Tiny-ImageNet | ECE (%): 4.76 -> 1.20 (Manifold-Mixup) | Improved |
| apply TS post-hoc scaling on top of RankMixup | RankMixup (M-NDCG) · CIFAR10 (ResNet-50) | ECE (%): 2.59 -> 0.57 | Improved |
| use M-NDCG instead of MRL on long-tailed data (ablation) | RankMixup · CIFAR100-LT rho=100 | ECE (%): 5.28 -> 18.74 | Hurt |
| use RankMixup ranking losses instead of mixup label mixtures | Mixup baseline · CIFAR100-LT rho=10 | ECE (%): 19.69 -> 9.99 | Improved |
| use M-NDCG instead of MRL (ablation) | RankMixup · ImageNet (ResNet-50) | ECE (%): 5.70 -> 3.93 | Improved |
| train with M-NDCG for OOD detection vs mixup | RankMixup (M-NDCG) · OOD CIFAR10 ID vs SVHN | AUROC (%): 79.72 -> 92.62 | Improved |
Closest papers by meaning
- Beyond One-Hot Labels: Semantic Mixing for Model Calibration (ICML 2025 · similarity 0.86)
- Tailoring Mixup to Data for Calibration (ICLR 2025 · similarity 0.86)
- Reassessing How to Compare and Improve the Calibration of Machine Learning Models (ICLR 2025 · similarity 0.85)
- Understanding and Mitigating Under-Confidence in GNNs from the Final Layer (NeurIPS 2026 · similarity 0.85)
- Pushing Boundaries: Mixup's Influence on Neural Collapse (ICLR 2024 · similarity 0.84)
- Improving GNN Calibration with Discriminative Ability: Insights and Strategies (AAAI 2024 · similarity 0.84)
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.
