Research map / Label-efficient and robust learning / Recognition and long-tail
Dual Focal Loss for Calibration
Linwei Tao, Minjing Dong, Chang Xu
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 |
|---|---|---|---|
| replace Focal Loss with Dual Focal Loss using the largest logit below the ground-truth logit | ResNet-50 · CIFAR-10 | ECE (%): 1.55 -> 0.46 | Improved |
| replace Focal Loss with Dual Focal Loss using the largest logit below the ground-truth logit | ResNet-50 · CIFAR-100 | ECE (%): 4.5 -> 1.08 | Improved |
| replace Focal Loss with Dual Focal Loss using the largest logit below the ground-truth logit | ResNet-50 · Tiny-ImageNet | ECE (%): 1.76 -> 1.50 | Improved |
| replace Focal Loss with Dual Focal Loss | ResNet-50 · CIFAR-10 | classification error (%): 4.98 -> 5.17 | Hurt |
| add a fixed dual logit of +0.1 to Focal Loss (ablation) | ResNet-50 · CIFAR-10 | ECE (%): 1.55 -> 1.15 | Improved |
| add a fixed dual logit of +0.2 to Focal Loss (ablation) | ResNet-50 · CIFAR-10 | ECE (%): 1.55 -> 2.10 | Hurt |
| use the 2nd largest logit below q_gt as the dual logit (ablation) | Dual Focal Loss · CIFAR-10, ResNet-50 | ECE (%): 1.55 -> 1.12 | Improved |
| use the 3rd largest logit below q_gt as the dual logit (ablation) | Dual Focal Loss · CIFAR-10, ResNet-50 | ECE (%): 1.55 -> 1.34 | Improved |
| use the mean of the 1st and 2nd largest logits below q_gt (ablation) | Dual Focal Loss · CIFAR-10, ResNet-50 | ECE (%): 1.55 -> 0.8 | Improved |
| use the mean of the 1st, 2nd, and 3rd largest logits below q_gt (ablation) | Dual Focal Loss · CIFAR-10, ResNet-50 | ECE (%): 1.55 -> 0.61 | Improved |
| use the mean of all logits below q_gt (ablation) | Dual Focal Loss · CIFAR-10, ResNet-50 | ECE (%): 1.55 -> 0.52 | Improved |
| combine AdaFocal's sample-dependent gamma schedule with Dual Focal | ResNet-50 · CIFAR-10 | ECE (%): 0.46 -> 0.43 | Improved |
Closest papers by meaning
- Calibrating Deep Neural Networks using Focal Loss (NeurIPS 2020 · similarity 0.92)
- Uncertainty Weighted Gradients for Model Calibration (CVPR 2025 · similarity 0.89)
- AdaFocal: Calibration-aware Adaptive Focal Loss (NeurIPS 2022 · similarity 0.87)
- Bridging Precision and Confidence: A Train-Time Loss for Calibrating Object Detection (CVPR 2023 · similarity 0.87)
- A Benchmark Study on Calibration (ICLR 2024 · similarity 0.84)
- Enhancing Noise-Robust Losses for Large-Scale Noisy Data Learning (AAAI 2025 · 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.
