Research map / Detection and segmentation / CNN features
HEM: a margin-based loss for visual categorisation tasks
Michael Spratling, Heiko Schütt
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 |
|---|---|---|---|
| take mean of above-zero errors only (maz) (ablation) | MM loss · CIFAR100 ResNet18 | clean accuracy: 70.13 -> 74.94 | Improved |
| take mean of above-zero errors only (maz) (ablation) | MM loss · CIFAR10 ResNet18 | clean accuracy: 93.79 -> 93.81 | No clear effect |
| zero errors below the per-sample mean (thres) (ablation) | MM loss · CIFAR100 ResNet18 | clean accuracy: 70.13 -> 73.13 | Improved |
| zero errors below the mean (thres) (ablation) | MM loss · CIFAR10 ResNet18 | clean accuracy: 93.79 -> 93.78 | No clear effect |
| add thres on top of maz (ablation) | MM + maz · CIFAR100 ResNet18 | clean accuracy: 74.94 -> 74.95 | No clear effect |
| replace CE loss with HEM | ResNets etc. standard data · unknown class rejection | AUROC: +4.52% | Improved |
| replace CE loss with HEM | standard data · AutoAttack rejection | DAR: +17.17% | Improved |
| replace CE loss with HEM | standard balanced data · clean accuracy | accuracy: -1.19% | Hurt |
| replace CE loss with HEM | long-tailed data · unknown class rejection | AUROC: +9.33% | Improved |
| replace CE loss with HEM | continual learning · accuracy | accuracy: +1.27% | Improved |
| replace CE loss with HEM | semantic segmentation · segmentation performance | metric: +5.31% | Improved |
| use shared margin (HEM-) instead of class-specific margins (HEM) (ablation) | HEM · semantic segmentation | performance: HEM- better than HEM | Hurt |
Closest papers by meaning
- Learning Towards The Largest Margins (ICLR 2022 · similarity 0.87)
- Large Margin Deep Networks for Classification (NeurIPS 2018 · similarity 0.86)
- Are All Losses Created Equal: A Neural Collapse Perspective (NeurIPS 2022 · similarity 0.86)
- Large-Margin Softmax Loss for Convolutional Neural Networks (ICML 2016 · similarity 0.86)
- Penalizing Top Performers: Conservative Loss for Semantic Segmentation Adaptation (ECCV 2018 · similarity 0.86)
- Complex Wavelet Mutual Information Loss: A Multi-Scale Loss Function for Semantic Segmentation (ICML 2025 · 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.
