.png){kind=icon}
Return
Knowledge-aware replay for multi-label class-incremental learning
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2026.133167.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Multi-Label Class-Incremental Learning (MLCIL) aims to adapt multi-label classifiers to evolving classes while mitigating catastrophic forgetting. The replay paradigm, which stores and revisits exemplars in a buffer, is a widely adopted strategy in state-of-the-art MLCIL. However, its effectiveness is severely limited by the label-absence issue: instance annotations are restricted to their original task classes, ignoring potential co-occurring concepts. This causes two critical drawbacks: (i) incompletely annotated exemplars penalize correct predictions as false negatives; and (ii) exemplar selection strategies inherited from single-label settings build biased buffers that overlook multi-label co-occurrence patterns. To address these challenges, we present Knowledge-Aware Replay (KAR), which reconciles the replay paradigm with MLCIL under the constraint of label absence. KAR features two core components. First, Bidirectional Knowledge Transfer (BKT) performs pseudo-labeling on both new data and replayed exemplars to recover missing labels, thereby correcting supervisory signals and reducing false negatives. Second, Knowledge-Informed Selection (KIS) constructs memory-efficient buffers by jointly optimizing representativeness and informativeness. This strategy captures evolving class distributions and prioritizes semantic-bridge instances to increase label density per exemplar. Extensive evaluations demonstrate that KAR consistently outperforms existing MLCIL baselines and exhibits broad versatility. Code is available at: https://github.com/ChengtaiCao/KAR.
Keywords:
Class-incremental learning
Multi-label learning
Continual learning
Lifelong learning
Image recognition
Journal
IF:
7.5
Papers:
3.0W
Citations:
10.2W
Organization
Cited Papers
No cited papers available