Towards guarantees for parameter isolation in continual learning