In ULMFiT we called this discriminative fine tuning. It's been the default fine tuning method in fastai since v1. 😀

→ View original post on X — @jeremyphoward