AS-Net: Class-Aware Assistance and Suppression Network for Few-Shot Learning

Abstract

Few-shot learning targets to recognize objects while only limited data is provided for each class. Existing methods tend to solve this problem by mapping the raw inputs into a shared embedding space and averaging them at the class level to form the corresponding prototypes. However, the prototypes are vulnerable to outliers, as they can contribute significantly to the class description when there is little data. In this paper, a class-aware assistance and suppression framework (AS-Net) is proposed to identify the informative patches and outliers as well as to further facilitate the rectification of prototypes. Specifically, we firstly introduce local features commonly shared across the whole set while distinguishable to other classes. These additional features can enhance the image embeddings and narrow the distance between the outliers and the majority of the class. We then attenuate the effect of outliers on the prototype by assigning lower weights to them. During these two stages, each labeled sample assists the rest samples of the same class to recognize their commonality, which provides positive attention for the intra-class consistency, i.e., which area or sample to focus on. At the same time, it suppresses the samples of other classes with similar features, which provides negative attention for the inter-class distinction, i.e., which area or sample not to focus on. Extensive experiments on several benchmark datasets demonstrate the superiority of our proposed method over the state-of-the-art approaches in the few-shot learning task.