Parameter Regularization

CWR (Copy Weight with Reinit)

Strategy that duplicates model weights after learning each task and selectively reinitializes certain weights for learning the new task. CWR maintains a copy of important weights while allowing adaptation for new knowledge.

← Indietro