Multi-granularity episodic contrastive learning for few-shot learning

Few-shot learning (FSL) aims at fast adaptation to novel classes with few training samples. Among FSL methods, meta-learning and transfer learning-based methods are the most powerful ones. However, most of them rely to some extent on cross-entropy loss, which leads to representations that are overly concerned with the classes already seen, and in turn leads to sub-optimal generalization on novel classes. In this study, we are inspired by meta-learning and transfer learning-based methods and believe good feature representations are vital for FSL. To this end, we propose a new multi-granularity episodic contrastive learning method (MGECL) that introduces contrastive learning into the episode training process. In particular, by enforcing our proposed contrastive loss on both class and instance granularities, the model is able to extract category-independent discriminative patterns and learn richer and more transferable feature representations. Extensive experiments demonstrate that our proposed method achieves state-of-the-art performance on three popular few-shot benchmarks. Our code is available at https://github.com/z1358/MGECL_PR. (C) 2022 Elsevier Ltd. All rights reserved.