Hybrid Attention Deep Adaptive Residual Graph Convolution Network for Few-shot Classification

Few-shot learning is a challenging task in the field of machine learning that aims to acknowledge novel class with a few amount of labeled samples. To address this problem, researchers have proposed several methods, with metric-based methods being one of the most effective approaches. These methods learn a transferable embedding space for classification by computing the similarity between samples. In this context, Graph Neural Networks (GNNs) have been employed to describe the association among support samples and query samples. However, existing GNN-based methods face limitations in their capability to achieve deeper layers, which restricts their ability to effectively transport information from the support images to the query images. To overcome the limitation, we propose a deep adaptive residual graph convolution network with deeper layers that better explores the relationship between support and query sets. Additionally, we design a hybrid attention module to learn the metric distributions, which helps to alleviate the over-fitting problem that can occur with few samples. The proposed method has been shown to be effective through comprehensive experimentation on five benchmark datasets.