One Adapter, Many Tasks: Task-Conditioned Feature Transformations for Continual Learning
FACET uses one shared adapter that changes its feature transformation by task, instead of storing or merging separate adapters.
The paper targets class-incremental learning, where a model must add new classes without revisiting earlier training data. Its method shapes the adapter feature space into a mixture of task-specific components with reduced overlap, then uses a replay-free consistency loss to limit forgetting. The authors report stronger results on both long task sequences, including 200 tasks, and shorter 20-task settings while using fewer trainable parameters and GFLOPs. Code is promised after acceptance. ArXiv · AI/CL/LG's note
The paper targets class-incremental learning, where a model must add new classes without revisiting earlier training data. Its method shapes the adapter feature space into a mixture of task-specific components with reduced overlap, then uses a replay-free consistency loss to limit forgetting. The authors report stronger results on both long task sequences, including 200 tasks, and shorter 20-task settings while using fewer trainable parameters and GFLOPs. Code is promised after acceptance. ArXiv · AI/CL/LG's note
score 4