Task Level Data Augmentation for Meta-Learning

Jialin Liu, Fei Chao, Chih-Min Lin

Keywords: Meta-learning, few-shot learning, data augmentation

TL;DR: We propose a data augmentation approach for meta-learning and prove that it is valid.

Abstract: Data augmentation is one of the most effective approaches for improving the accuracy of modern machine learning models, and it is also indispensable to train a deep model for meta-learning. However, most current data augmentation implementations applied in meta-learning are the same as those used in the conventional image classification. In this paper, we introduce a new data augmentation method for meta-learning, which is named as ``Task Level Data Augmentation'' (referred to Task Aug). The basic idea of Task Aug is to increase the number of image classes rather than the number of images in each class. In contrast, with a larger amount of classes, we can sample more diverse task instances during training. This allows us to train a deep network by meta-learning methods with little over-fitting. Experimental results show that our approach achieves state-of-the-art performance on miniImageNet, CIFAR-FS, and FC100 few-shot learning benchmarks. Once paper is accepted, we will provide the link to code.