Why use two FFN layer without activated function inner?

[Qznan]

In DeepBiaffine Parser, is that two FFN layer without activated function equivalent to one FFN layer？ Why choose to use two?

[FranxYao]

I guess 2 is from the original biaffine paper. Activation .. I forget 🤣🤣 but I think adding activation to both biaffine and CRF will not change their performance order 😃. Actually. The paper get the highest score in this task changes the feed forward to a BiLSTM. So if you aim for score, maybe do an LSTM. If you aim for verifying the effectiveness of CRF over Biaffine, the current one without activation should suffice Cheers

…

On Wed, Aug 25, 2021 at 10:45 AM Qznan ***@***.***> wrote: In DeepBiaffine Parser, is that two FFN layer without activated function equivalent to one FFN layer？ Why choose to use two? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#6>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AEHHCHPOGYJU6SU4OHEYHHLT6S3UVANCNFSM5CYVRWFA> . Triage notifications on the go with GitHub Mobile for iOS <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android <https://play.google.com/store/apps/details?id=com.github.android&utm_campaign=notification-email> .