A Mathematical Theory of Reusable Neural Bases for Network Compression