Research map / Label-efficient and robust learning / Recognition and long-tail
The Devil Is in the Margin: Margin-Based Label Smoothing for Network Calibration
Check 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 |
|---|---|---|---|
| Add a positive logit-distance margin rather than use zero margin (ablation) | MbLS · Tiny-ImageNet, ResNet-50 | ECE: 2.50 -> 1.64 | Helped |
| Add a positive logit-distance margin rather than use zero margin (ablation) | MbLS · Tiny-ImageNet, ResNet-101 | ECE: 1.89 -> 1.62 | Helped |
| Add a positive logit-distance margin rather than use zero margin (ablation) | MbLS · CIFAR-10, ResNet-50 | ECE: 3.72 -> 1.16 | Helped |
| Add a positive logit-distance margin rather than use zero margin (ablation) | MbLS · CIFAR-10, ResNet-101 | ECE: 3.07 -> 1.38 | Helped |
| Add a positive logit-distance margin rather than use zero margin (ablation) | MbLS · CUB-200-2011, ResNet-101 | ECE: 5.11 -> 2.78 | Helped |
| Use a positive margin instead of zero margin (ablation) | MbLS · CUB-200-2011, ResNet-101 | accuracy: 73.92 -> 74.56 | Helped |
| Use zero margin instead of the tuned positive margin (ablation) | MbLS · CIFAR-10, ResNet-50 | accuracy: 94.76 -> 95.25 | Hurt |
| Increase the margin to m=20 rather than use the selected margin (ablation) | MbLS · CIFAR-10, ResNet-50 | ECE: 3.05 | Hurt |
| Increase the margin to m=20 rather than use the selected margin (ablation) | MbLS · Tiny-ImageNet, ResNet-50 | ECE: 2.99 | Hurt |
| Use a positive margin rather than zero margin | MbLS · 20 Newsgroups test | ECE: 22.75 -> 5.40 | Helped |
| Use MbLS rather than the strongest listed calibration baseline | ResNet-101 · CUB-200-2011 | ECE: 5.16 -> 2.78 | Helped |
| Use MbLS rather than the strongest listed calibration baseline | ViT · CIFAR-10 | ECE: 1.13 -> 0.39 | Helped |
Closest papers by meaning
- Dirichlet-Based Prediction Calibration for Learning with Noisy Labels (AAAI 2024 · similarity 0.89)
- Towards Calibrated Multi-label Deep Neural Networks (CVPR 2024 · similarity 0.88)
- MaxSup: Overcoming Representation Collapse in Label Smoothing (NeurIPS 2025 · similarity 0.88)
- Training Deep Neural Networks with Virtual Smoothing Classes (AAAI 2025 · similarity 0.88)
- A Benchmark Study on Calibration (ICLR 2024 · similarity 0.87)
- Improving GNN Calibration with Discriminative Ability: Insights and Strategies (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.