Multi-distance metric network for few-shot learning

Few-shot learning aims to make classification when few samples are available. In general, metric-based methods map images into a space by learning the embedding function. However, conventional metric-based methods rely on a single distance value, which does not pay attention to the shallow features. In this paper, we propose a multi-distance metric network (MDM-Net) by employing a multi-output embedding network to map samples into different feature spaces. In addition, we maximize the inter-class distance which is popular in metric learning field to improve the performance of few-shot classifier. Furthermore, we design a task-adaptive margin to adjust the distance between different sample pairs, and we found that the distance loss combined with cross-entropy loss is beneficial to achieve better results in meta-task training. The proposed method is verified by tests on miniImageNet and FC100 these two benchmarks for 5-way 1-shot classification task and 5-way 5-shot classification task with competitive results.