Hybrid Graph Neural Networks for Few-Shot Learning

Graph neural networks (GNNs) have been used to tackle the few-shot learning (FSL) problem and shown great potentials under transductive setting. However inductive setting, existing GNN based methods are less competitive. This is because they use an instance as a label propagation/classification module, which jointly meta-learned with feature embedding network. design problematic classifier needs adapt quickly new tasks while does not. To overcome this problem, in paper we propose novel hybrid (HGNN) model consisting of two GNNs, prototype GNN. Instead propagation, act adaptation modules for quick tasks. Importantly designed deal fundamental yet often neglected challenge FSL, that is, only handful shots per class, any would be sensitive badly sampled either outliers or can cause inter-class distribution overlapping. Extensive experiments show our HGNN obtains state-of-the-art on three FSL benchmarks. The code models available at https://github.com/TianyuanYu/HGNN.