arxivcs.CV2026-07-02
DRDN: Decoupled Representation Dynamic Network for From-Scratch ViT Class-Incremental Learning
Bingchen Huang, Yifu Chen, Zhiling Wang, Yuanchao Du
Dynamic expansion methods for class-incremental learning (CIL) protect task-specific knowledge by growing dedicated tokens or subnetworks, yet our analyses suggest that classification supervision alone does not sufficiently preserve task-agnostic shared backbone representations o…