PARN: Position-Aware Relation Networks for Few-Shot Learning

Few-shot learning presents a challenge that a classifier must quickly adapt to new classes that do not appear in the training set, given only a few labeled examples of each new class. This paper proposes a position-aware relation network (PARN) to learn a more flexible and robust metric ability for few-shot learning. Relation networks (RNs), a kind of architectures for relational reasoning, can acquire a deep metric ability for images by just being designed as a simple convolutional neural network (CNN)[23]. However, due to the inherent local connectivity of CNN, the CNN-based relation network (RN) can be sensitive to the spatial position relationship of semantic objects in two compared images. To address this problem, we introduce a deformable feature extractor (DFE) to extract more efficient features, and design a dual correlation attention mechanism (DCA) to deal with its inherent local connectivity. Successfully, our proposed approach extents the potential of RN to be position-aware of semantic objects by introducing only a small number of parameters. We evaluate our approach on two major benchmark datasets, i.e., Omniglot and Mini-Imagenet, and on both of the datasets our approach achieves state-of-the-art performance. It's worth noting that our 5-way 1-shot result on Omniglot even outperforms the previous 5-way 5-shot results.

1. Introduction

Humans can effectively utilize prior knowledge to easily learn new concepts given just a few examples. Fewshot learning [11], [20], [15] aims to acquire some transferable knowledge like humans, where a classifier is able to generalize to new classes when given only one or a few labeled examples of each class, i. e., one-or few-shot. In this paper, we focus on the ability of learning how to compare, namely metric-based methods. Metric-based method-s [2], [11], [22], [23], [25] often consist of a feature extractor and a metric module. Given an unlabeled query image and a few labeled sample images, the feature extractor first generates embeddings for all input images, and then the metric module measures distances between the query embedding and sample embeddings to give a recognition result. Figure 1:

Two situations where the comparison ability of RN will be limited. The top row shows the two compared images, and the bottom row shows their extracted features, where blue areas represent the response of corresponding semantic objects. (a) The convolutional kernel fails to involve the two objects. (b) The convolutional kernel fails to involve the same fine-grained features.