Multi-Branch Network for Few-shot Learning

Few-shot learning aims provide precise predictions for unseen data through learning from only one or few labelled samples of each class. However, it often suffers from the overfitting problem because of insufficient training data. In this paper, we propose a novel metric-based few-shot learning method, multi-branch network (MBN), with a new data augmentation module to improve the generalization ability of the model. Specifically, we generate different types of noise contaminated data through multiple branches in the network to simulate the real-world scenarios when noisy images are obtained. Following this novel data augmentation module, the feature embedding and similarities between the support and query samples are learned simultaneously through the embedding and metric modules, respectively. Moreover, to consider more details in the feature maps, we propose to utilize the average-pooling layer in the metric module rather than the commonly adopted max-pooling layer. The network is trained from end to end by the Kullback- Leibler (KL) divergence, to minimize the difference between the distributions of the ground truths and predictions. Extensive experiments on Standford-Dogs, Standford-Cars, CUB-200-2011 and mini-ImageNet in the 1-shot and 5-shot tasks demonstrate the superior classification performance of MBN.