Which components are employed in learning from previous models for very similar tasks?