Research map / Label-efficient and robust learning / Recognition and long-tail
Anchor Loss: Modulating Loss Scale Based on Prediction Difficulty
Serim Ryou, Seong-Gyun Jeong, Pietro Perona
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 cross entropy with anchor loss | ResNet-110 · CIFAR-100 | top-1 accuracy: 72.98 -> 74.25 | Improved |
| replace binary cross entropy with anchor loss | ResNet-110 · CIFAR-100 | top-1 accuracy: 73.88 -> 74.25 | Improved |
| replace cross entropy with anchor loss | ResNet-50 · ImageNet val | top-1 accuracy: 76.39 -> 76.82 | Improved |
| use OHEM hard-example sampling instead of CE | ResNet-110 · CIFAR-100 | top-1 accuracy: 72.98 -> 73.03 | No clear effect |
| add CE warmup for first 5 epochs to anchor loss | ResNet-110 anchor loss · CIFAR-100 | top-1 accuracy: 74.25 -> 74.38 | Improved |
| use network-derived dynamic anchor instead of static q*=0.5 (ablation) | Anchor loss · CIFAR-100 | top-1 accuracy: 73.77 -> 74.25 | Improved |
| raise gamma from 0.5 to 2.0 (ablation) | Anchor loss · CIFAR-100 | top-1 accuracy: 74.25 -> 71.86 | Hurt |
| modulate background loss (ii) instead of target loss (i) (ablation) | Anchor loss · CIFAR-100 | top-1 accuracy: 74.06 -> 74.25 | Improved |
| modulate both target and background (iii) instead of background only (ablation) | Anchor loss · CIFAR-100 | top-1 accuracy: 74.25 -> 73.90 | Hurt |
| replace MSE with anchor loss | Hourglass · MPII test | PCK total: 90.9 -> 91.9 | Improved |
| replace MSE with anchor loss, single-scale | Hourglass · MPII val | PCK mean: 89.32 -> 89.93 | Improved |
| set gamma=2 instead of gamma=5 in anchor loss (ablation) | 2-stack Hourglass · MPII val | PCK mean: 88.84 -> 89.11 | Improved |
Later papers that compared with this method
1 later accepted paper reports Anchor Loss as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Fixing Localization Errors to Improve Image Classification (ECCV 2020 · better on 1 of 1 shared result)
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
- Why Do Better Loss Functions Lead to Less Transferable Features? (NeurIPS 2021 · similarity 0.86)
- Unimodal-Concentrated Loss: Fully Adaptive Label Distribution Learning for Ordinal Regression (CVPR 2022 · similarity 0.86)
- Learning Loss for Active Learning (CVPR 2019 · similarity 0.86)
- Are All Losses Created Equal: A Neural Collapse Perspective (NeurIPS 2022 · similarity 0.85)
- Thumb on the Scale: Optimal Loss Weighting in Last Layer Retraining (NeurIPS 2025 · similarity 0.85)
- MaxSup: Overcoming Representation Collapse in Label Smoothing (NeurIPS 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.
