Transductive Graph-Attention Network for Few-shot Classification

Abstract

Few-shot Classification aims to use a small number of labeled samples to learn a general model that can solve the problem of image classification. At present, Few-shot Classification methods can be divided into three categories: Data Augmentation, Model Optimization and Gradient Descent Algorithm Optimization. However, most of the proposed work still has the problem of insufficient feature extraction due to too little data, and it is difficult for the model to fully operate the extracted features. In order to address these problems, we propose a Transductive Graph-Attention Network for Few-shot Classification (TGAN). In particular, TGAN uses attention to construct the similarity graph structure between samples to describe the relationship between labeled samples and unlabeled samples, and then uses label propagation algorithm to accurately classify. Experiments show that TGAN has excellent performance on MiniImageNet datasets and TieredImageNet datasets.