RSCC: Robust Semi-supervised Learning with Contrastive Learning and Augmentation Consistency Regularization

Semi-supervised learning (SSL) can effectively take advantage of unlabeled data. Aiming at the poor performance of existing SSL methods in the case of only a very small number of labels and the problem of pseudo-label confirmation bias, we propose a novel SSL method, RSCC, which combines the powerful representation learning capability of contrast learning with augmentation consistency regularization methods and introduces a symmetric cross-entropy learning to mitigate the impact of noisy pseudo-labels on model performance. RSCC consists of two key steps. We first perform self-supervised pre-training on unlabeled data using contrast learning to extract meaningful representations from the data, and then perform tuning training based on SSL methods of augmentation consistency regularization and symmetric cross-entropy learning. We conduct rich experiments, which show that RSCC achieves state-of-the-art accuracy on multiple datasets, such as CIFAR-10 and CIFAR-100, especially when labeled data is extremely scarce. This underscores its cutting-edge and effective performance.