This computer-vision paper addresses multi-label class-incremental learning (MLCIL), a harder and less-studied variant of continual learning in which a model must keep recognizing previously learned object categories while new categories are introduced over successive training sessions, and where each training image is labeled only for a subset of the objects it actually contains. The authors observe that directly applying a vision-language model such as CLIP to MLCIL causes catastrophic forgetting. Their method combines two ingredients to curb this forgetting: an improved, confidence-based data-replay mechanism that selects representative past samples to rehearse, and a prompt loss (prompt regularization) that stabilizes learned prompt information so the model adapts to multi-label classification without overwriting prior knowledge. Across multiple benchmark datasets the approach substantially improves MLCIL performance over prior methods, validating the combination of enhanced prompt tuning, confidence-based replay, and prompt-loss regularization.
This computer-vision paper addresses multi-label class-incremental learning (MLCIL), a harder and less-studied variant of continual learning in which a model must keep recognizing previously learned …