Hierarchy SeparateEMD for Few-Shot Learning

Abstract

AbstractFew-shot learning methods are studied for the problem of insufficient samples in neural network tasks. Taking advantage of the excellent feature extraction capabilities of neural networks, meta-learning was proposed and became the mainstream for few-shot learning. Among the few shot learning methods, the metric-based method has the characteristics of simple mode and no need for iterative training in the inference process, which is more in line with the original intention of few-shot learning task and has been widely studied. The metric-based method with Euclidean distance has established a relatively complete training and inference system, but the accuracy has gradually stagnated. The Earth Mover’s Distance has gradually emerged in the field of few-shot learning, and greatly surpassed Euclidean distance methods in terms of accuracy. Aiming at the problem that the current EMD method can only be trained with 1-shot learning and cannot be trained with multi-shot, we proposed a SeparateEMD that can train in multi-shot mode. In order to make up for the lack of global information of the SeparateEMD, Hierarchy Attention Module for proto is constructed to increase the intra-category correlation and intra-category global information of the support samples, and also improve the sample discrimination. We build the Half Pyramid Merge Module for query to increase the correlation between local information and global information of a single query sample. The Global Sampling is constructed on the basis of the Random Sampling of the EMD, which can not only increase the randomness of the samples, but also ensure the stability of the information. Experiments show that our proposed method can achieve 5-shot training of EMD-like method and outperform the current state-of-the-art few-shot learning methods on popular benchmark datasets.KeywordsFew shot learningSeparateEMDProto