De-Redundancy Distillation And Feature Shift Correction For Cross-Domain Few-Shot Learning

Recently, there has been increasing interest in Cross-Domain Few-Shot Learning (CD-FSL), which aims to solve the traditional Few-Shot Learning (FSL) problem across different domains. The core challenge of the CD-FSL is that the huge domain gap between the source and target domains often leads to poor generalization to the target new tasks. Although many existing methods have achieved certain accomplishments, none of them take into account the negative effects of redundant information during training. To overcome this problem, we propose a de-redundancy distillation (DRD) method that introduces a small amount of unlabeled target data to reduce domain differences. DRD utilizes a de-redundancy based self-supervised learning method to encourage the model to learn more concise and clear target-relevant features. And using a curriculum learning strategy to balance the learning of both source and target information. Moreover, it is common for models to excessively focus on some irrelevant information when generalizing. We propose a simple feature shift correction module (FSC) to allow it to focus on more useful generic target information, which further improves the quality of features. We conduct extensive experiments on the BSCD-FSL benchmark that includes four different target datasets. The experimental results demonstrate the effectiveness of our method.