Research map / Label-efficient and robust learning / Recognition and long-tail
Rebalancing Multi-Label Class-Incremental Learning
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 RebLL, which rebalances positive-negative label imbalance in multi-label class-incremental learning at the loss level (asymmetric knowledge distillation) and the label level (online relabeling of memory samples), reaching SOTA with a vanilla CNN on PASCAL VOC and MS-COCO.
Problem
Multi-label class-incremental learning under task-level partial labels suffers from a positive-negative imbalance: most missing labels are effectively negative, causing false positives, forgetting, and loss gradients dominated by positives.
Approach
Asymmetric knowledge distillation down-weights positive loss terms in the classification loss and overconfident old-task predictions in the distillation loss (focal-style adaptive decay tied to class count), while online relabeling fills missing labels in replay memory using the past and current models with a threshold.
Key results
- On VOC B4-C2 with 2/class buffer, RebLL (AKD+OR) reaches 81.0 last mAP and 88.5 average mAP, versus CSC's 73.8 and 81.9.
- On MS-COCO B20-C4 with 5/class buffer, RebLL reaches 62.8 last mAP and 71.2 average mAP using a vanilla CNN, exceeding CSC (60.6 / 69.8).
- AKD alone (no buffer) reduces false positive rate from 31.9% (KD) to 2.7% on VOC B4-C2; OR lowers FPR from 6.7% to 2.4%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add AKD and OR to the CNN-based method | RebLL · PASCAL VOC B4-C2 | last mAP: 73.8 -> 81.0 | Helped |
| Add AKD and OR to the CNN-based method | RebLL · PASCAL VOC B5-C3 | last mAP: 81.3 -> 84.0 | Helped |
| Add AKD and OR to the CNN-based method | RebLL · MS-COCO B20-C4 | last mAP: 60.6 -> 62.8 | Helped |
| Add AKD and OR to the CNN-based method | RebLL · MS-COCO B0-C5 | last mAP: 63.4 -> 65.5 | Helped |
| Emphasize negative labels in classification loss (ablation) | Fine-Tuning · PASCAL VOC B4-C2 | last mAP: 22.6 -> 36.3 | Helped |
| Use asymmetric classification and distillation losses instead of BCE and KD (ablation) | KD · PASCAL VOC B4-C2 | last mAP: 58.2 -> 73.1 | Helped |
| Add replay with BCE loss to AKD (ablation) | AKD · PASCAL VOC B4-C2 | last mAP: 73.1 -> 74.2 | Helped |
| Add online relabeling to AKD (ablation) | AKD · PASCAL VOC B4-C2 | last mAP: 73.1 -> 77.3 | Helped |
| Add online relabeling and the OR replay loss to AKD (ablation) | AKD · PASCAL VOC B4-C2 | last mAP: 73.1 -> 81.0 | Helped |
| Replace the fixed decay factor with an adaptive decay factor (ablation) | AKD · PASCAL VOC B4-C2 | FPR: 7.2 -> 2.7 | Helped |
| Use AKD instead of KD | Knowledge distillation · MLCIL prediction statistics | FPR: 31.9% to 2.7% | Helped |
| Add online relabeling to replay memory | Classical replay with AKD · MLCIL memory samples | FPR: 6.7% to 2.4% | Helped |
Surprising findings
- Naively adding classical replay on top of AKD increased false positive rate from 2.7% to 6.7%, because memory samples carry imbalanced labels.
- Replay contributes only minimally to gains compared with AKD and OR components, according to the authors.
Limitations
- Gains may be less pronounced in simple CIL settings with minimal forgetting, fewer tasks, or weaker imbalance.
- Relabeling relies on a threshold n and on the reliability of past/current models for labeling memory samples.
Experimental setup
Datasets: PASCAL VOC 2007 MS-COCO 2014
Benchmarks: PASCAL VOC 2007 MS-COCO 2014
Metrics: mAP CF1 OF1 Avg.Acc Last Acc FPR
Baselines: Fine-Tuning Joint LwF AGCN KRT CSC ER PODNet DER++ L2P
Builds on: LwF focal loss reservoir sampling CSC KRT
Closest papers by meaning
- Specifying What You Know or Not for Multi-Label Class-Incremental Learning (AAAI 2025 · similarity 0.91)
- Towards Macro-AUC Oriented Imbalanced Multi-Label Continual Learning (AAAI 2025 · similarity 0.90)
- Gradient Reweighting: Towards Imbalanced Class-Incremental Learning (CVPR 2024 · similarity 0.90)
- L3A: Label-Augmented Analytic Adaptation for Multi-Label Class Incremental Learning (ICML 2025 · similarity 0.89)
- Toward Robustness in Multi-Label Classification: A Data Augmentation Strategy against Imbalance and Noise (AAAI 2024 · similarity 0.88)
- Distilling Reliable Knowledge for Instance-Dependent Partial Label Learning (AAAI 2024 · 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.