Research map / Learning theory and algorithms / Prediction and decision losses
An Alternative Probabilistic Interpretation of the Huber Loss
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 |
|---|---|---|---|
| Replace publication hyperparameters with the implementation hyperparameters | Faster R-CNN · MS-COCO 2014 val | mAP @ 0.5 IoU: 42.8 -> 44.7 | Helped |
| Replace publication hyperparameters with the implementation hyperparameters | Faster R-CNN · MS-COCO 2014 val | mAP @ 0.75 IoU: 18.7 -> 23.1 | Helped |
| Use Experiment A hyperparameters instead of implementation hyperparameters | Faster R-CNN · MS-COCO 2014 val | mAP @ 0.5-0.95 IoU: 23.8 -> 24.2 | Helped |
| Use Experiment B hyperparameters instead of implementation hyperparameters | Faster R-CNN · MS-COCO 2014 val | mAP @ 0.75 IoU: 23.1 -> 25.0 | Helped |
| Use Experiment B hyperparameters instead of implementation hyperparameters | Faster R-CNN · MS-COCO 2014 val | mAP @ 0.5 IoU: 44.7 -> 44.2 | Hurt |
| Use Experiment C hyperparameters instead of implementation hyperparameters | Faster R-CNN · MS-COCO 2014 val | mAP @ 0.5-0.95 IoU: 23.8 -> 24.7 | Helped |
| Use Experiment A hyperparameters instead of implementation hyperparameters | RetinaNet · MS-COCO 2017 val | mAP @ 0.5-0.95 IoU: 40.4 -> 40.8 | Helped |
| Use Experiment B hyperparameters instead of implementation hyperparameters | RetinaNet · MS-COCO 2017 val | mAP @ 0.5 IoU: 60.1 -> 58.9 | Hurt |
| Replace log width/height targets with percentage-change targets (ablation) | Faster R-CNN · MS-COCO 2014 val | mAP @ 0.5-0.95 IoU: 23.8 -> 23.7 | Hurt |
| Replace Huber loss with D_{α/2,1/α} (ablation) | Faster R-CNN · MS-COCO 2014 val | mAP @ 0.75 IoU: 23.1 -> 23.3 | Helped |
Closest papers by meaning
- Beyond Spectra: Eigenvector Overlaps in Loss Geometry (ICLR 2026 · similarity 0.84)
- Learning with Fitzpatrick Losses (NeurIPS 2024 · similarity 0.83)
- Enhancing Noise-Robust Losses for Large-Scale Noisy Data Learning (AAAI 2025 · similarity 0.83)
- Decoupled Kullback-Leibler Divergence Loss (NeurIPS 2024 · similarity 0.83)
- Binary Losses for Density Ratio Estimation (ICLR 2025 · similarity 0.83)
- Variance-Reducing Couplings for Random Features (ICLR 2025 · 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.