4 ms·

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