Large margin meta-learning for few-shot classification

Abstract

This paper proposes a large margin principle to overcome the problem of data scarcity in training meta-learning models for few-shot classification. To realize it, we develop an efficient framework that can be easily incorporated in existing metric-based meta-learning models. We demonstrate that the large margin principle can improve the generalization capacity of state-of-the-art meta-learning methods and lead to consistent and considerable improvements on few-shot classification.