Graph-Based Prototypical Network for Few-Shot Learning

Few-shot learning (FSL) is a technique for learning new class concepts with a limited number of labeled samples, is a key step towards human-level intelligence. Among existing few-shot learning methods, prototypical network shows to be promising in solving the critical problem of overfitting. However, due to the simplicity of average operation in building the prototype representation for each class, the inter- and intra-class relationships among the samples in the support set are not fully exploited, resulting in deviation of the prototype representation from the true class distribution. In this paper, we propose graph-based prototypical network (GPN) model to overcome this problem. In GPN, a fully learnable message passing graph module is proposed to refine the feature embedding vector of each sample. The refined features are then fed into prototypical network to obtain the robust prototype representations of classes. According to experimental results, the proposed method achieves competitive classification accuracy against state-of-the-art ones.