Conceptual
Browse Instruct Curate

Continual Test-Time Adaptation with Sample Partitioning and Anti-Forgetting Regularization

A continual test-time adaptation framework (SPARNet) that partitions incoming unlabeled samples into reliable and unreliable groups, handling each differently with a mean-teacher consistency loss, and adds a regularization term restricting changes to important parameters, so a deployed model adapts to a long sequence of shifting domains without error accumulation or catastrophic forgetting.