Uniform Sampling over Episode Difficulty

  • 2021-08-03 17:58:54
  • Sébastien M. R. Arnold, Guneet S. Dhillon, Avinash Ravichandran, Stefano Soatto
  • 4

Abstract

Episodic training is a core ingredient of few-shot learning to train modelson tasks with limited labelled data. Despite its success, episodic trainingremains largely understudied, prompting us to ask the question: what is thebest way to sample episodes? In this paper, we first propose a method toapproximate episode sampling distributions based on their difficulty. Buildingon this method, we perform an extensive analysis and find that samplinguniformly over episode difficulty outperforms other sampling schemes, includingcurriculum and easy-/hard-mining. As the proposed sampling method is algorithmagnostic, we can leverage these insights to improve few-shot learningaccuracies across many episodic training algorithms. We demonstrate theefficacy of our method across popular few-shot learning datasets, algorithms,network architectures, and protocols.

 

Quick Read (beta)

loading the full paper ...