How does transfer learning apply specifically to computer vision tasks, and why is it so effective there?