Research map / Label-efficient and robust learning / Semi-supervised learning
Semi-supervised Multi-label Learning with Balanced Binary Angular Margin Loss
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes S2ML2-BBAM, a self-training semi-supervised multi-label learning method that uses a balanced binary angular margin loss (Gaussian-transformed feature angles) and prototype-based negative sampling to reduce positive/negative variance bias.
Problem
Semi-supervised multi-label learning suffers from variance bias between feature distributions of positive and negative samples per label when training on labeled plus pseudo-labeled data with binary losses and negative sampling.
Approach
Extends the binary angular margin loss with per-label Gaussian linear transformations of feature angles so positive and negative angle distributions share a balanced variance, iteratively re-estimated during training; pairs this with class-aware pseudo-labeling and prototype-nearest negative sampling.
Key results
- On VOC at 5% labeled, S2ML2-BBAM reaches Micro-F1 0.7897 and mAP 0.7866 versus CAP's 0.6162 Micro-F1 at the same setting.
- On COCO at 5% labeled, Micro-F1 0.6830 vs CAP 0.5629; ablation without BBAM drops COCO mAP from 0.6354 to 0.5894 at 5%.
- On AAPD at 5% labeled, Micro-F1 0.7057 vs best baseline DRML 0.4160.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use S2ML2-BBAM instead of the strongest listed baseline | SSMLL models · VOC, COCO, AWA, Ohsumed, and AAPD, π = 5% | Micro-F1: VOC: 0.6542 -> 0.7897; COCO: 0.6071 -> 0.6830; AWA: 0.6868 -> 0.7213; Ohsumed: 0.5562 -> 0.6671; AAPD: 0.5722 -> 0.7057 | Helped |
| Add BBAM loss instead of BAM loss (ablation) | S2ML2-BBAM · VOC, π = 5%, 10%, 15%, and 20% | Micro-F1: 0.7845 -> 0.7897; 0.8206 -> 0.8401; 0.8301 -> 0.8443; 0.8318 -> 0.8458 | Helped |
| Add BBAM loss instead of BAM loss (ablation) | S2ML2-BBAM · VOC, π = 5%, 10%, 15%, and 20% | Macro-F1: 0.7247 -> 0.7306; 0.7789 -> 0.8015; 0.7988 -> 0.8124; 0.7967 -> 0.8141 | Helped |
| Add BBAM loss instead of BAM loss (ablation) | S2ML2-BBAM · VOC, π = 5%, 10%, 15%, and 20% | mAP: 0.7881 -> 0.7866; 0.8204 -> 0.8345; 0.8274 -> 0.8454; 0.8282 -> 0.8458 | Mixed |
| Add BBAM loss instead of BAM loss (ablation) | S2ML2-BBAM · COCO, π = 5%, 10%, 15%, and 20% | Micro-F1: 0.6691 -> 0.6830; 0.6952 -> 0.7074; 0.7052 -> 0.7150; 0.7143 -> 0.7246 | Helped |
| Add BBAM loss instead of BAM loss (ablation) | S2ML2-BBAM · COCO, π = 5%, 10%, 15%, and 20% | Macro-F1: 0.5885 -> 0.6144; 0.6264 -> 0.6480; 0.6424 -> 0.6594; 0.6530 -> 0.6726 | Helped |
| Add BBAM loss instead of BAM loss (ablation) | S2ML2-BBAM · COCO, π = 5%, 10%, 15%, and 20% | mAP: 0.5894 -> 0.6354; 0.6316 -> 0.6741; 0.6520 -> 0.6886; 0.6628 -> 0.7023 | Helped |
| Vary the rescaled norm s from 1 to 10, with m fixed at 0.4 (ablation) | S2ML2-BBAM · VOC2012, π = 5% | performance | Helped |
| Vary the rescaled norm s from 10 to 50, with m fixed at 0.4 (ablation) | S2ML2-BBAM · VOC2012, π = 5% | performance | No clear effect |
| Vary cosine margin m; best accuracy is reported at m = 0.4 with s fixed at 20 (ablation) | S2ML2-BBAM · VOC2012, π = 5% | accuracy | Helped |
Surprising findings
- On VOC at 5% labeled, the BAM-only ablation achieves slightly higher mAP (0.7881) than the full method (0.7866).
- MIME (single-positive MLL) performs far below SSMLL methods, attributed to losing single positive observations on unlabeled data.
Limitations
- mAP slightly declines on VOC and COCO as the labeled proportion increases, per the authors.
- Reported results have no error bars; the authors state variance is too small to visualize.
- Theoretical analysis is on a linear SSBC model under a mixture-Gaussian assumption, which is a strong simplification.
Experimental setup
Datasets: VOC2012 MS-COCO 2014 AWA Ohsumed AAPD
Benchmarks: VOC2012 MS-COCO 2014 AWA Ohsumed AAPD
Metrics: Micro-F1 Macro-F1 mAP Hamming Loss One Loss
Baselines: SoftMatch FlatMatch MIME DRML CAP
Builds on: binary angular margin loss class-aware pseudo-labeling FixMatch-style self-training
Closest papers by meaning
- Distribution Aligning Refinery of Pseudo-label for Imbalanced Semi-supervised Learning (NeurIPS 2020 · similarity 0.89)
- LaSSL: Label-Guided Self-Training for Semi-supervised Learning (AAAI 2022 · similarity 0.88)
- Correlation-Induced Label Prior for Semi-Supervised Multi-Label Learning (ICML 2024 · similarity 0.88)
- CDMAD: Class-Distribution-Mismatch-Aware Debiasing for Class-Imbalanced Semi-Supervised Learning (CVPR 2024 · similarity 0.87)
- Bi-Level Optimization for Semi-Supervised Learning with Pseudo-Labeling (AAAI 2025 · similarity 0.87)
- Forming Auxiliary High-confident Instance-level Loss to Promote Learning from Label Proportions (CVPR 2025 · similarity 0.87)
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.