Focuses on parameter isolation methods for continual learning, where each task uses separate parameter masks or subnetworks to prevent forgetting. Implements Hard Attention to the Task (HAT), Supermask Superposition (SupSup), and Piggyback, with visualization tools and metrics for task overlap and capacity usage.
pytorch neural-networks representation-learning hat incremental-learning lifelong-learning continual-learning piggyback parameter-isolation supsup mask-learning
-
Updated
Oct 15, 2025 - Python