Abstract:In machine learning, the label quality of the training samples seriously affects the final effect of the classification algorithms. Although the effect of a clean label is relatively good, it takes time and effort to collect and use. Therefore, in order to save costs and make the model adapt to the general situation, researchers gradually began to learn from ordinary data, that is, data with label noise. In recent years, although some works have been devoted to label noise, they lack comprehensive analysis. Based on this, this paper first introduces the label noise briefly and comprehensively, then analyzes the learning algorithms of tag noise in recent years explicitly and implicitly, and summarizes them. Finally, we look forward to the future research on label noise.