Domain-adaptive graph neural network for few-shot learning

Abstract

The task of few-shot classification (FSC) is to build a model to discriminate novel categories that do not present in training categories with limited labeled samples. However, in many applications, the training categories and novel categories are assumed to come from the same domain. Existing few-shot classification algorithms achieve promising performance in a single domain, but often fail to generalize to unseen domains because of the different feature distribution across domains. The main goal of this work is to propose an effective recognition model that can work on various image domains with domain shift. Specifically, we propose to construct a new domain-attention mechanism with some adapters based on the squeeze-and-excitation network architectures. The proposed network processes multiple domains simultaneously and all parameters are shared across domains. In addition, because of the large discrepancy in feature distributions among different domains, similarity transformation layers are used to reduce the differences in the feature distributions in the training stage. Extensive experiments were conducted to validate the domain generalization capability (GC) of this model on four FSC datasets: mini-ImageNet, Cars, CUB-200-2011, and Simpsons Characters Data. The results show that the presented method has excellent performances on various datasets across diverse domains.