<ul><li>Deep networks are prone to catastrophic forgetting during sequential task learning, leading to the loss of knowledge about old tasks.</li><li>Existing continual learning methods focus on protecting parameters of previous tasks, which can be impractical due to linear increase in parameter size.</li><li>A novel CL framework called learning without isolation (LwI) is introduced to protect pathways in the whole networks, rather than just individual parameters, inspired by neuroscience and physics.</li><li>LwI leverages model fusion through graph matching and adapts pathways for new tasks to address catastrophic forgetting in a parameter-efficient manner, as proven by experiments on benchmark datasets.</li></ul>

Learning without Isolation: Pathway Protection for Continual Learning

Discover more