Contextual similarity-based multi-level second-order attention network for semi-supervised few-shot learning.

[Display omitted] We think that there are mainly two issues to be considered in few-shot learning, one is how to learn a generalized representation and the other is how to alleviate over-fitting. For the first issue, in contrast to current few-shot learning models which directly take VGG or ResNet as an embedding network, we propose to design a novel embedding network to improve the generalization ability on unseen classes. For the second issue, we propose to deal with few-shot learning problem in a semi-supervised setting. The unlabeled set is used to compensate the lack of few-shot training samples. As illustrated in Graphical abstract figure, the samples in the unlabeled set are from not only known classes but also distracter classes, which is more realistic. Instead of taking pair-wise distance measures to select unlabeled samples, we propose a contextual similarity considering both local pair-wise distance and the global task manifold to pick and label known unlabeled samples. Then, we train an episodic linear classifier based on the combination of the support set and the picked known unlabeled samples. The parameters of the proposed multi-level second-order attention network will be updated by minimizing the loss of the query set. • To improve the generalization ability of the embedding network to unseen classes, we design a novel multi-level second-order attention network to capture the correlations of the intermediate layers. • To pick and label the quasi-known unlabeled instances, we propose a novel and effective contextual similarity measure jointly considering the pair-wise similarity and global similarity. Extensive experiments on four datasets (Omniglot, miniImageNet, tieredImageNet, and CUB-200-2011) show the proposed method is superior in performance. In this paper, we tackle the few-shot learning problem in a semi-supervised setting where a limited number of labeled data-points and a number of low-cost unlabeled samples are assumed to be available. In particular, some of the unlabeled samples share the same label space with the support set, referring to as known samples, while some of them are from distracter classes, referring to as unknown ones. The keys are how to learn a powerful representation and how to pick and label unlabeled known instances to construct discriminative classifiers. We address both issues by learning multi-level second-order attention representation followed by a contextual similarity. We first develop a novel trainable multi-level second-order attention network(MSAN) to adaptively learn more powerful feature representation by using second-order feature statistics. Our proposed MSAN is able to better represent the samples while the parameter is not increased. Then we introduce a contextual measure that considers not only the pair-wise relationship but also the task-specific condition, to calculate the similarity between unlabeled samples and each support class, thus to label and pick the quasi-known samples. The hypothesis is that the known unlabeled sample should not only be strongly similar to one particular class, but also be significantly dissimilar to other classes. With the combination of support set and pseudo-labeled set, we train an episodic linear classifier for each episode and the parameters of multi-level second-order attention network are updated by minimizing the loss of the query set. Extensive experiments on four popular benchmarks (Omniglot, miniImageNet, tieredImageNet, and CUB-200-2011) demonstrate that our simple yet effective approach can achieve competitive accuracy compared to the state-of-the-art methods. [ABSTRACT FROM AUTHOR]