Semantic-Guided Dynamic Sparsification for Pre-Trained Model-based Class-Incremental Learning
Semantic-Guided Dynamic Sparsification (SGDS) is a method for class-incremental learning that prevents interference by guiding activation patterns into orthogonal, sparse subspaces instead of rigidly constraining model parameters. It dynamically decides whether to reuse existing features or create new ones based on semantic similarity, leading to superior stability and performance across long learning sequences.