Research map / Label-efficient and robust learning / Recognition and long-tail
On Focal Loss for Class-Posterior Probability Estimation: A Theoretical Perspective
Nontawat Charoenphakdee, Jayakorn Vongkulbhisal, Nuttapong Chairatanakul, Masashi Sugiyama
Check 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 focal loss with any γ ≥ 0 instead of an uncalibrated surrogate | Multiclass classification risk minimization · Multiclass classification | Bayes classification risk | Improved |
| Use focal loss with γ > 0 without transforming its output | Focal-loss classifier · Class-posterior probability estimation | Probability estimation correctness | Hurt |
| Apply the proposed Ψ^γ transformation to focal-loss scores | Focal-loss classifier · Synthetic 3-class classification | KLD | Improved |
| Apply the proposed Ψ^γ transformation to focal-loss scores | Focal-loss classifier · Synthetic 3-class classification | Classification error | No clear effect |
| Apply Ψ^γ after the softmax layer | ResNet models trained on CIFAR10 and SVHN · CIFAR10 and SVHN | ECE | Mixed |
| Apply Ψ^γ after the softmax layer | ResNet110 trained with label smoothing or focal-loss temperature scaling · CIFAR10 and SVHN | ECE | Improved |
| Use focal loss without Ψ^γ rather than cross-entropy | Neural-network classifiers · 30 datasets | ECE | Hurt |
| Use focal loss without Ψ^γ rather than cross-entropy | Neural-network classifiers · Some benchmark settings | ECE | Mixed |
| Increase γ for focal-loss training in binary classification | Focal-loss classifier · Binary classification | Maximum-score underconfidence | Hurt |
| Use temperature scaling rather than Ψ^γ to post-process focal-loss scores | Focal-loss classifier · Synthetic 3-class classification | Class-posterior estimation | Hurt |
| Increase focal-loss γ under label smoothing or focal-loss temperature scaling | ResNet110 · CIFAR10 and SVHN | ECE | Hurt |
| Use Ψ^γ when the focal risk minimizer is estimated less accurately, as with less training data or a more complex model | Focal-loss classifier · CIFAR10, CIFAR10-s, and SVHN | ECE | Hurt |
Closest papers by meaning
- Generalized Focal Loss: Learning Qualified and Distributed Bounding Boxes for Dense Object Detection (NeurIPS 2020 · similarity 0.86)
- Generalized Focal Loss: Towards Efficient Representation Learning for Dense Object Detection (TPAMI 2022 · similarity 0.85)
- AdaFocal: Calibration-aware Adaptive Focal Loss (NeurIPS 2022 · similarity 0.85)
- Focal Loss for Dense Object Detection (ICCV 2017 · similarity 0.85)
- Calibrating Deep Neural Networks using Focal Loss (NeurIPS 2020 · similarity 0.84)
- Bridging Precision and Confidence: A Train-Time Loss for Calibrating Object Detection (CVPR 2023 · 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.
