1 paper · 1 filter
Pulkit Gopalani, Ekdeep Singh Lubana, Wei Hu
Recent analysis on the training dynamics of Transformers has unveiled an interesting characteristic: the training loss plateaus for a significant number of training steps, and then…