Conceptual
Browse Instruct Curate

Domain-Shift-Aided Contrastive Prototypes for Class-Incremental Learning

A plug-and-play approach to class-incremental learning built on the finding that inter-task domain shift reduces catastrophic forgetting by separating per-task feature distributions. It imposes task-level and class-level contrastive regularization on a lightweight prototype pool and adds a cross-task contrastive distillation loss, keeping representations distinguishable so existing CIL methods forget less.