undefined | Better HN

0 pointsbuboard6y ago0 comments

google seemed to make a genuine effort to make a model that is useful rather than record-breaking with bert. But i think it's wrong to consider it the "final" model upon which everything else will be built.

0 comments

3 comments · 1 top-level

bitL6y ago· 2 in thread

BERT is already outdated, but still useful as you need only 1 Titan RTX to retrain its BERT_large model via transfer learning.

turnersr6y ago

What methods make BERT outdated? Do you have pointers to other options?

bitL6y ago

e.g. XLNet:

https://arxiv.org/abs/1906.08237

1 more reply

j / k navigate · click thread line to collapse