PLATE: Plasticity-Tunable Efficient Adapters for Geometry-Aware Continual Learning
Researchers have developed a new continual learning method called PLATE, designed for adapting pretrained models without needing access to old task data. This approach leverages the geometric redundancy found in pretrained networks to create update subspaces that minimize functional drift and improve data retention. PLATE achieves this by parameterizing layer updates with a structured low-rank matrix, where only a portion of the matrix is trained on new tasks, allowing for controlled plasticity-retention trade-offs. AI
IMPACT Enables more efficient adaptation of foundation models by removing the need for old task data, potentially speeding up deployment in data-scarce scenarios.