Tweeted By @svlevine
Adaptive risk minimization (ARM) can adapt to distributional shift without labels. A model is meta-trained to perform well on a new distribution after observing a few unlabeled examples from that distribution. Check out the blog post by Marvin Zhang:https://t.co/3GReH3Z6vA
— Sergey Levine (@svlevine) November 9, 2020
-> pic.twitter.com/gggcLBfIPi