Meta-Teaching: Curriculum Generation for Lifelong Learning

Abstract

Meta-learning will be crucial to creating lifelong, generalizable AI. In practice, however, it is hard to define the meta-training task distribution that is used to train meta-learners. If made too small, tasks are too similar for a model to meaningfully generalize. If made too large, generalization becomes incredibly difficult. We argue that both problems can be alleviated by introducing a teacher model that controls the sequence of tasks that a meta-learner is trained on. This teacher model is incentivized to start the student meta-learner on simple tasks then adaptively increase task difficulty in response to student progress. While this approach has been previously studied in curriculum generation, our main contribution is in extending it to meta-learning.

Tasks

Reproductions