A Representation Learning Perspective on the Importance of
            Train-Validation Splitting in Meta-Learning

                     Nikunj Saunshi 1 Arushi Gupta 1 Wei Hu 1

             Abstract                from many “train” tasks to learn a useful prior that can
    An effective approach in meta-learning is to uti-     help solve new “test” tasks. The goal is to do well on test
    lize multiple “train tasks” to learn a good initial-    tasks with way fewer samples per task than would  ...