S-learning: a joint supervised and self-supervised approach for robust learning with noisy labels
Futian Wang, Yitao Chen, Hao Hao · 2025
Noisy labels, resulting from manual labeling errors or unreliable data sources, can lead to overfitting and reduced generalization performance in neural networks when used for direct training. Self-supervised learning, which operates independently of labels, can mitigate the negative effects of noisy labels. Inspired by joint training that combines supervised and self-supervised learning, we propose an efficient method named S-learning for dealing with noisy labels. This method dynamically generates a clean sample set for supervised learning via three selection strategies: small-loss selection within class, high-confidence selection, and high-similarity selection. Meanwhile, other samples are utilized for self-supervised learning. Comprehensive experiments conducted on both synthetic and real-world noisy datasets show that S-learning surpasses numerous state-of-the-art methods.