Research map / Label-efficient and robust learning / Recognition and long-tail
Two-Way Multi-Label Loss
Takumi Kobayashi
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 |
|---|---|---|---|
| two-way loss (sample-wise plus class-wise) instead of single-direction losses (ablation) | multi-label loss on ResNet-50 · MSCOCO | mAP@class: 74.11 vs 73.06 (class-only) and 67.18 (sample-only) | Improved |
| positive-side temperature T_P=4 in the log-sum-exp margin loss (vs T_P=1) (ablation) | multi-label loss on ResNet-50 · MSCOCO | mAP@class: 74.11 vs 72.30 | Improved |
| setting hard positive min via T_P→0 (no log-sum-exp smoothing) (ablation) | multi-label loss on ResNet-50 · MSCOCO | mAP@class: 71.03 vs 74.11 best | Hurt |
| logit bias (multi-similarity style) instead of temperature margin (ablation) | multi-label loss on ResNet-50 · MSCOCO | mAP@class: 73.44 vs 74.11 with temperature | Hurt |
| replace softmax-style loss with proposed loss in two-way form (softmax two-way comparison) (ablation) | softmax loss in two ways · MSCOCO | mAP@class: 69.19 vs 74.11 | Hurt |
| joint-way loss over all class and sample entries instead of two-way (ablation) | multi-label loss on ResNet-50 · MSCOCO | mAP@class: 71.82 vs 74.11 | Hurt |
| multi-similarity loss instead of proposed loss (ablation) | multi-label loss on ResNet-50 · MSCOCO | mAP@class: 71.27 vs 74.11 | Hurt |
| bucket size M reduced from 128 to 16 (ablation) | class-wise loss in two-way loss · MSCOCO | mAP@class: 70.76 vs 74.11 | Hurt |
| proposed two-way loss vs BCE/Focal/ASL/softmax on multi-label benchmarks (ResNet-50, ResNeXt50, DenseNet169, RegNetY32gf) | multi-label classification · MSCOCO, VISPR, VAW, WIDER, VOC2007, VOC2012 | mAP@class and mAP@sample: e.g. MSCOCO mAP@class 74.11 vs ASL 70.92 (ResNet-50) | Improved |
| ASL instead of proposed loss on ImageNet-transfer | ResNet-50 ImageNet-trained features · Transfer (Aircraft, 10 datasets) | accuracy: Aircraft 27.42 vs 44.01 ours | Hurt |
| two-way loss vs softmax on single-label ImageNet training | ResNet-50 ImageNet-pretrained features · ImageNet top-1 and transfer | top-1 accuracy: 76.29 vs 76.32 top-1; transfer gains on Aircraft (44.01 vs 39.18) | No clear effect |
| logit bias on the multi-similarity-style loss vs no bias (ablation) | multi-label loss on ResNet-50 · MSCOCO | mAP@sample: 86.19-86.25 across eps 0.5-2 vs 86.66 ours | No clear effect |
Closest papers by meaning
- Distributionally Robust Loss for Long-Tailed Multi-Label Image Classification (ECCV 2024 · similarity 0.92)
- Multi-Label Learning with Stronger Consistency Guarantees (NeurIPS 2024 · similarity 0.87)
- Semi-supervised Multi-label Learning with Balanced Binary Angular Margin Loss (NeurIPS 2024 · similarity 0.86)
- Forming Auxiliary High-confident Instance-level Loss to Promote Learning from Label Proportions (CVPR 2025 · similarity 0.86)
- Neural Collapse in Multi-label Learning with Pick-all-label Loss (ICML 2024 · similarity 0.86)
- Double-Layer Hybrid-Label Identification Feature Selection for Multi-View Multi-Label Learning (AAAI 2024 · similarity 0.86)
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.
