I guess the practical implication is probably just train on middle layers or increase the corresponding learning rate
I guess the practical implication is probably just train on middle layers or increase the corresponding learning rate