PLATE: Plasticity-Tunable Efficient Adapters for Geometry-Aware Continual Learning

The article presents PLATE (Plasticity-Tunable Efficient Adapters), a continual learning method that enables adaptation of pretrained models without requiring access to old-task data. PLATE leverages geometric redundancy within pretrained networks to construct protected update subspaces and restricts updates to a subset of redundant neurons, which enhances retention of old-task performance while allowing for new-task learning. The method employs a structured low-rank update mechanism, where only the adaptation matrix \(A\) is trained, ensuring efficient updates with controlled plasticity and retention trade-offs.