xg15 5 hours ago (2021), still very interesting. Especially the "post-overfitting" training strategy is unexpected. luckystarr an hour ago I remember vaguely that this was observed when training GPT-3 (probably?) as well. Just trained on and on, and the error went up and then down again. Like a phase transition in the model.
luckystarr an hour ago I remember vaguely that this was observed when training GPT-3 (probably?) as well. Just trained on and on, and the error went up and then down again. Like a phase transition in the model.
(2021), still very interesting. Especially the "post-overfitting" training strategy is unexpected.
I remember vaguely that this was observed when training GPT-3 (probably?) as well. Just trained on and on, and the error went up and then down again. Like a phase transition in the model.
The low sample efficiency of RL is well explained.