A transfer-based few-shot classification approach via masked manifold mixup and fuzzy memory contrastive learning

Few-shot learning studies the problem of classifying unseen images by learning only a small number of samples in these categories with the assistance of a large amount of data in other classes. In recent studies, the idea of transfer learning is an effective method to solve the problem of few-shot classification. However, the insufficient generalization ability of the model still restricts the performance of these transfer-based methods. This paper proposes a masked manifold mixup and fuzzy memory contrastive learning (M3FM) method for transfer-based few-shot learning to improve the generalization ability. We design a regularization technique that enhances the model's learning of local features by masking and mixing the data manifold in the hidden states of neural networks. Then, a momentum updated fuzzy memory is adopted in contrastive learning with the masked mixup manifold to help the model learn the specific distinctions of different categories. Experimental results show that the proposed method outperforms previous baseline methods on miniImageNet, CUB-200, and CIFAR-FS benchmarks. Further adaptation research demonstrates that our method can be generalized to complex few-shot classification tasks and cross-domain scenarios. Ablation studies verify the effectiveness of masked manifold mixup and fuzzy memory contrastive learning.