SSL-DC: Improving Transductive Few-Shot Learning via Self-Supervised Learning and Distribution Calibration

Few-shot learning, aiming to distinguish unseen classes by training with few labeled samples, is still challenged by the overfitting problem. The transductive few-shot learning paradigm enables us to reduce overfitting by training a highly discriminative feature representation via self-supervised learning since the entire unlabeled samples are allowed to be accessed. In this paper, we propose a simple but efficient approach based on self-supervised pre-training and nearest class prototype search, which can obtain a significant improvement in the performance of transductive few-shot learning tasks without external samples. However, since the class prototype is obtained through limited support samples, it is easily affected by biased samples. Therefore, we propose to train a conditional generative adversarial network to estimate the distribution of features instead of assuming it follows Gaussian distribution as previous arts. Thus, we can generate features that are closed to real features from the estimated distribution to calibrate the distribution of the class prototype. Finally, more detailed experiments show that our method can exceed plenty of recent transductive few-shot learning methods significantly and achieve 9.83% and 4.38% improvements over the existing best method under the transductive 5-way 1-shot and 5-shot settings with ResNet-12 on the miniImageNet.