Curriculum Meta Learning: Learning to Learn from Easy to Hard

Meta-learning is a machine learning paradigm that extracts crosstask knowledge by learning a large number of subtasks, to fast adapt to new tasks. Many meta-learning methods are widely applied in few-shot classification. These methods adopt an episodic training strategy, and the learning subtasks are sampled uniformly from the task distribution. In this paper, we explore the effect of the order of training subtasks on the performance of different meta-learning algorithms and propose a curriculum learning framework to improve the generalization performance. We define the hardness of subtasks at the class level and guide the model to learn training subtasks from easy to hard. We evaluate our curriculum learning framework on two few-shot classification benchmarks (mini-ImageNet and FC100), and it achieves improvements across different meta-learning algorithms and datasets. In the cross-domain scenario, we compare the performance of different meta learning algorithms under three curriculum settings. The results show that our CL approach improves significantly the generalization performance of different meta-learning methods.