L3A: Label-Augmented Analytic Adaptation for Multi-Label Class Incremental Learning

Class-incremental learning (CIL) enables models to learn new classes continually without forgetting previously acquired knowledge. Multi-label CIL (MLCIL) extends CIL to a real-world scenario where each sample may belong to multiple classes, introducing several challenges: label absence, which leads to incomplete historical information due to missing labels, and class imbalance, which results in the model bias toward majority classes. To address these challenges, we propose Label-Augmented Analytic Adaptation (L3A), an exemplar-free approach without storing past samples. L3A integrates two key modules. The pseudo-label (PL) module implements label augmentation by generating pseudo-labels for current phase samples, addressing the label absence problem. The weighted analytic classifier (WAC) derives a closed-form solution for neural networks. It introduces sample-specific weights to adaptively balance the class contribution and mitigate class imbalance. Experiments on MS-COCO and PASCAL VOC datasets demonstrate that L3A outperforms existing methods in MLCIL tasks. Our code is available at https://github.com/scut-zx/L3A.

Humans learn throughout their lives, building on past experiences without forgetting earlier knowledge. The class-incremental learning (CIL) research aims to teach AI models to do the same, to keep learning new classes in samples over time without needing to retrain from scratch.In many real-world scenarios, many samples contain multiple classes. This leads to a more complex challenge called multi-label class-incremental learning (MLCIL), where the AI models must learn to recognize many classes in a sample, even when not all of them are labeled during training. Over time, the model should be able to recognize all previously seen classes.We propose a new method called L3A, which learns new classes without forgetting the old ones and without needing to store past data. We use a technique to generate missing labels and assign different importance to each training sample, so the system doesn’t favor only the most common classes. This results in more accurate and fairer AI models that can continuously learn in complex environments.