Dynamic Topology-Preserving Multi-Label Class-Incremental Learning

To address the insufficient stability of the feature space in class-incremental learning, this paper proposes a Dynamic Topology-Preserving Multi-Label Class-Incremental Learning (DTP-MLCIL) method. Specifically, a Neural Gas (NG) network is introduced to construct and maintain an adaptive topological skeleton of the feature space through the Competitive Hebbian Learning (CHL) algorithm, thereby approximating and tracking the underlying feature distribution at the manifold geometry level. Based on this topological representation, an anti-forgetting node loss is designed to constrain the consistency between current features and the historical topology from both node-position and topological-connectivity perspectives, thereby reinforcing regions that encode previously acquired knowledge. Furthermore, a max-min loss is introduced to proactively organize the feature space of newly emerging classes, encouraging new class representations to remain well separated from historical regions while forming compact intra-class clusters. By explicitly preserving the geometric stability of the feature space, the proposed method effectively alleviates structural drift and topological distortion during the incremental learning process. Comprehensive experimental evaluations demonstrate the effectiveness of the proposed method.