Tweeted By @svlevine
Few-shot learning problems can be ambiguous. Now MAML can handle the ambiguity by sampling multiple classifiers via a Bayesian formulation of meta-learning. We call it PLATIPUS: https://t.co/oOHQl1tuj4
— Sergey Levine (@svlevine) June 8, 2018
w/ @chelseabfinn and @imkelvinxu
some ambiguous celeba task: pic.twitter.com/oGakQW1Tr5