Research map / Learning theory and algorithms / Prediction and decision losses
Beyond probability partitions: Calibrating neural networks with semantic aware grouping
Jia-Qi Yang, De-Chuan Zhan, Le Gan
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 learned semantic-aware groups with group-specific temperature scaling instead of global temperature scaling | GC+TS · CIFAR10-Resnet152 constructed binary setting | ECE: 0.0157 -> 0.0118 | Improved |
| Use learned semantic-aware groups with group-specific calibration instead of global calibration | GC+ETS · CIFAR10, CIFAR100, Imagenet | top-label ECE: -8.92% | Improved |
| Use learned semantic-aware groups with group-specific temperature scaling instead of global temperature scaling | GC+TS · CIFAR10, CIFAR100, Imagenet | top-label ECE: -5.03% | Improved |
| Use learned group partitions and group-specific calibration | GC+TS · CIFAR10-Resnet152 constructed binary setting | PCE: 0.0174 -> 0.0141 | Improved |
| Increase the number of partitions (ablation) | GC+TS · CIFAR10-Resnet152 | calibration performance | Improved |
| Decrease the number of groups within each partition (ablation) | GC+TS · CIFAR10-Resnet152 | NLL and ECE | Hurt |
| Set grouping-function regularization too high (value of 1) (ablation) | GC+TS · CIFAR10-Resnet152 | calibration performance: 1 | Hurt |
| Set grouping-function regularization too low (ablation) | GC+TS · CIFAR10-Resnet152 | calibration performance | Hurt |
| Calibrate using the validation set rather than the hold-out calibration set | existing calibration methods · calibration experiments | calibration performance | Hurt |
| Train the grouping function on the validation set and transfer it to the hold-out set | GC+TS · calibration experiments | calibration performance | Improved |
Closest papers by meaning
- A Benchmark Study on Calibration (ICLR 2024 · similarity 0.89)
- Reassessing How to Compare and Improve the Calibration of Machine Learning Models (ICLR 2025 · similarity 0.88)
- Towards Calibrated Deep Clustering Network (ICLR 2025 · similarity 0.87)
- Divide et Calibra: Multiclass Local Calibration via Vector Quantization (NeurIPS 2026 · similarity 0.86)
- Beyond One-Hot Labels: Semantic Mixing for Model Calibration (ICML 2025 · similarity 0.85)
- Improving GNN Calibration with Discriminative Ability: Insights and Strategies (AAAI 2024 · similarity 0.85)
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.
