Research map / Label-efficient and robust learning / Recognition and long-tail
AdaFocal: Calibration-aware Adaptive Focal Loss
Arindam Ghosh, Thomas Schaaf, Matthew R. Gormley
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 |
|---|---|---|---|
| use AdaFocal adaptive per-bin gamma instead of FLSD-53 heuristic gamma | focal loss calibration · CIFAR-10 ResNet50 | ECE_EM (%): 1.63 -> 0.66 | Improved |
| use AdaFocal instead of FLSD-53 on a dataset where FLSD-53 is under-confident | focal loss calibration · ImageNet ResNet50 | ECE_EM (%): 16.77 -> 1.87 | Improved |
| use AdaFocal instead of cross entropy | classifier training · CIFAR-100 ResNet50 | ECE_EM (%): 17.17 -> 1.36 | Improved |
| use AdaFocal instead of cross entropy | classifier training · CIFAR-10 ResNet50 | test error (%): 4.95 -> 5.30 | Hurt |
| use AdaFocal instead of cross entropy | classifier training · CIFAR-100 ResNet50 | test error (%): 23.30 -> 22.60 | Improved |
| use AdaFocal instead of cross entropy | classifier training · 20Newsgroup CNN | test error (%): 26.68 -> 28.53 | Hurt |
| add temperature scaling after AdaFocal training | AdaFocal · CIFAR-10 ResNet50 | ECE_EM (%): 0.66 -> 0.44 | Improved |
| use 15 rather than 5 equal-mass bins for gamma statistics (ablation) | AdaFocal · CIFAR-10 ResNet50 | ECE_EM (%): 0.63 -> 0.56 | Mixed |
| use 50 rather than 15 bins for gamma statistics (ablation) | AdaFocal · CIFAR-10 ResNet50 | ECE_EM (%): 0.56 -> 1.10 | Hurt |
| update gamma more frequently (every few mini-batches instead of per epoch) (ablation) | AdaFocal · 20Newsgroup CNN/BERT | calibration over training | Improved |
| switch to inverse-focal loss when gamma falls below threshold (ablation) | AdaFocal · ImageNet ResNet50 | ECE_EM (%) | Improved |
Later papers that compared with this method
1 later accepted paper reports AdaFocal as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Model Calibration in Dense Classification with Adaptive Label Perturbation (ICCV 2023 · better on 4 of 4 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Calibrating Deep Neural Networks using Focal Loss (NeurIPS 2020 · similarity 0.90)
- Dual Focal Loss for Calibration (ICML 2023 · similarity 0.87)
- On Focal Loss for Class-Posterior Probability Estimation: A Theoretical Perspective (CVPR 2021 · similarity 0.85)
- Uncertainty Weighted Gradients for Model Calibration (CVPR 2025 · similarity 0.84)
- Understanding and Mitigating Miscalibration in Prompt Tuning for Vision-Language Models (ICML 2025 · similarity 0.82)
- AdaBM: On-the-Fly Adaptive Bit Mapping for Image Super-Resolution (CVPR 2024 · similarity 0.82)
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.
