Aran Komatsuzaki   @arankomatsuzaki   6/9/2021       

Parameter-efficient Multi-task Fine-tuning for Transformers via Shared Hypernetworks Shows that we can learn adapter parameters for all layers and tasks by generating them using hypernetworks. Improved perf in multi-task learning while adding only 0.29% parameters per task.

 Reply  0     Retweet   15      Like   115





Posted by Aran Komatsuzaki