A General Curriculum for Meta-Learning

Abstract

Meta-learning is an effective method for few-shot learning, building transferable machine learning models that generalize to new contexts with relatively little new data. Given this limited data, a natural path to improve performance would be to intelligently sequence the training examples, a field of research called curriculum learning. Despite this intuition, little research exists to provide curricula for meta-learning. This thesis presents an original, general method for curriculum learning that can be used for any meta-learner in any context. At each training iteration, our curriculum uses a a novel task-based curriculum learning method to both accelerate training at the beginning and increase robustness of the final model. We demonstrate that this curriculum robustly improves the performance of state-of-the-art meta-learning models across a suite of benchmark datasets in different domains, suggesting its applicability more generally.