Dynamic Reinforcement and Alignment Graph Convolution Networks for Few-Shot Learning

Abstract

The dependence of deep learning models on large-scale labeled training data limits their application in real-world scenarios. To address this problem, researchers have proposed few-shot learning. However, most existing few-shot learning methods tend to ignore the contribution of local detailed information with class characteristics to classification. In this paper, we propose the dynamic reinforcement and alignment graph convolution networks (DRAGCN). Our proposed model can learn to generate the reinforcement basis that contains valuable information of local details with class characters based on experiential knowledge and obtain the reinforced feature maps by solving the neural ordinary differential equations (Neural ODE). These reinforced feature maps of the input images are constructed as graph-structured data, and the node features and edge features of the graph are optimized with the semantic alignment graph convolution networks, which introduces the semantic alignment operation to prevent the over-smoothing phenomenon. Experimental results on two popular datasets show that the proposed DRAGCN outperforms existing methods on few-shot learning tasks.