Research map / Continual learning / Class-incremental
Specifying What You Know or Not for 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 HCP, a multi-label class-incremental learning framework that purifies class-specific features, recalls old classes with per-class distribution-based pseudo-labels, and synthesizes an unknown "prospective" class by interpolating absent-class features.
Problem
Multi-label class-incremental learning suffers from contradictory learning objectives because images contain unlabeled past and future classes, causing feature aliasing and catastrophic forgetting.
Approach
Clarify known classes via class-embedding attention (dynamic feature purification) and per-class confidence-based pseudo-labels (recall enhancement), and probe unknown classes by interpolating absent-class features into a synthetic prospective class that compacts known-class representations.
Key results
- On MS-COCO B0-C10 with no replay buffer, HCP reaches 77.9% average mAP and 71.2% last mAP, versus KRT's 74.6% and 65.9% with buffer size 0 setting in the table.
- On VOC B10-C2, HCP achieves 92.1% Avg. Acc with 2/class buffer; without buffer it reaches 90.1% Avg. Acc.
- Ablation on VOC B10-C2 shows feature purification, recall enhancement, and probing unknown each add to Avg. Acc, reaching 90.05% combined.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HCP instead of KRT without a replay buffer | MLCIL · MS-COCO B0-C10, buffer 0 | Avg. Acc (mAP, %): 74.6 -> 77.9 | Helped |
| Use HCP instead of KRT without a replay buffer | MLCIL · MS-COCO B0-C10, buffer 0 | Last Acc (mAP, %): 65.9 -> 71.2 | Helped |
| Use HCP instead of KRT without a replay buffer | MLCIL · MS-COCO B40-C10, buffer 0 | Avg. Acc (mAP, %): 77.8 -> 78.9 | Helped |
| Use HCP instead of KRT without a replay buffer | MLCIL · MS-COCO B40-C10, buffer 0 | Last Acc (mAP, %): 74.0 -> 75.3 | Helped |
| Use HCP instead of KRT with a 2/class buffer | MLCIL · VOC B0-C4 | Avg. Acc: 90.7 -> 93.5 | Helped |
| Use HCP instead of KRT with a 2/class buffer | MLCIL · VOC B0-C4 | Last Acc: 83.4 -> 89.2 | Helped |
| Use HCP instead of KRT with a 2/class buffer | MLCIL · VOC B10-C2 | Avg. Acc: 87.7 -> 92.1 | Helped |
| Use HCP instead of KRT with a 2/class buffer | MLCIL · VOC B10-C2 | Last Acc: 80.5 -> 86.3 | Helped |
| Add Feature Purification to the fine-tuning baseline (ablation) | HCP · VOC B10-C2 | Avg. Acc: 72.87 -> 82.09 | Helped |
| Add Recall Enhancement to Feature Purification (ablation) | HCP · VOC B10-C2 | Avg. Acc: 82.09 -> 86.00 | Helped |
| Add Probing Unknown to Feature Purification (ablation) | HCP · VOC B10-C2 | Avg. Acc: 82.09 -> 85.47 | Helped |
| Add Probing Unknown to Feature Purification and Recall Enhancement (ablation) | HCP · VOC B10-C2 | Avg. Acc: 86.00 -> 90.05 | Helped |
Limitations
- Attention of absent classes may fall on present classes due to inter-class similarity, so synthetic unknown features can be noisy.
- Pseudo-label distribution queues must be updated each session, adding bookkeeping.
Experimental setup
Datasets: MS-COCO 2014 PASCAL VOC 2007 ImageNet-21k
Benchmarks: MS-COCO B0-C10 MS-COCO B40-C10 PASCAL VOC B0-C4 PASCAL VOC B10-C2
Metrics: Avg. Acc (mAP) Last Acc (mAP) CF1 OF1
Baselines: FT PODNet AGCN KRT TPCIL DER++ iCaRL BiC PRS OCDM
Builds on: KRT PODNet ASL
Closest papers by meaning
- Rebalancing Multi-Label Class-Incremental Learning (AAAI 2025 · similarity 0.91)
- L3A: Label-Augmented Analytic Adaptation for Multi-Label Class Incremental Learning (ICML 2025 · similarity 0.88)
- Adaptive Discovering and Merging for Incremental Novel Class Discovery (AAAI 2024 · similarity 0.86)
- Generative Multi-modal Models are Good Class Incremental Learners (CVPR 2024 · similarity 0.86)
- Distilling Reliable Knowledge for Instance-Dependent Partial Label Learning (AAAI 2024 · similarity 0.85)
- Towards Macro-AUC Oriented Imbalanced Multi-Label Continual Learning (AAAI 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.