HN Explorer

2 comments

[ 9.3 ms ] story [ 43.0 ms ] thread
NitpickLawyer
Are any of these methods doable on pre-trained models? Like freeze the model and only train these add-ons? Having to redo the training runs with these optimisations doesn't sound too practical, in the great scheme of things.
impossiblefork
It's obviously practical for the next model you train from scratch. The point of research is obviously not to improve existing commercial products.