Boosting Few-Shot Classification with Lie Group Contrastive Learning

Few-shot learning can alleviate the issue of sample scarcity, however, there remains a certain degree of overfitting. There have been solutions for this problem by combining contrastive learning with few-shot learning. In previous works, sample pairs are usually constructed with traditional data augmentation. The fitting of traditional data augmentation methods to real sample distributions poses difficulties. In this paper, our method employs Lie group transformations for data augmentation, resulting in the model learning more discriminative feature representations. Otherwise, we consider the congruence between contrastive learning and few-shot learning with respect to classification objectives. We also incorporate an attention mechanism into the model. Utilizing the attention module obtained through contrastive learning, the performance of few-shot learning can be improved. Inspired by the loss function of contrastive learning, we incorporate a penalty term into the loss function for few-shot classification. This penalty term serves to regulate the similarity between classes and non-classes. We conduct experiments with two different feature extraction networks on the standard few-shot image classification benchmark datasets, namely miniImageNet and tieredImageNet. The experimental results show that the proposed method effectively improves the performance of the few-shot classification.