Task-adaptive class division for transductive few-shot learning

Abstract

The key challenge of few-shot learning is to recognize novel classes with a few examples. Most existing few-shot learning models represent a class as a single prototype to train a model which can adapt to novel classes. However, these methods make a model to ignore the detailed characteristics of the class. In this paper, we propose MPLNet that task-adaptively divides the class into account the detailed features of the class. Our key idea is to employ each prototype of a divided sub-class as a class prototype to represent a class as multiple prototypes and utilize unlabeled data, not only labeled data. In order to extract information from unlabeled data, we introduce pseudo-labeling for multimodal distribution to assign a pseudo-label to unlabeled data. Experimental results on miniImagenet and tieredImagenet show that our method is comparable to or even outperforms state-of-the-art methods.