Tweeted By @seb_ruder

This post by the Snorkel team gives a great overview of ingredients that make up a state-of-the-art approach on GLUE:
1) Traditional supervision
2) Transfer learning
3) Multi-task learning
4) Dataset slicing (motivated by error analysis)
5) Ensemblinghttps://t.co/AcVTfjdiPF pic.twitter.com/3i6gu6X88w

— Sebastian Ruder (@seb_ruder) April 11, 2019