A Comprehensive Review on Deep Learning With Noisy Labels: Challenges, Solutions, and Frontiers

Deep learning has been widely adopted across various industrial domains due to its powerful learning capacity and effective feature extraction capabilities. However, the development of reliable deep learning models depends heavily on the availability of large-scale, high-quality annotated training data. In practice, the labels provided by human experts or generated through automated annotation tools are often prone to errors, either due to human oversight or inherent limitations in the labeling algorithms. These inaccuracies give rise to the challenge known as the noisy label problem. In recent years, learning with noisy labels has garnered significant attention within the research community. This review presents a comprehensive and up-to-date overview of recent advancements in this area. In particular, two primary categories of strategies are reviewed: methods aimed at improving the quality of training data (e.g., sample selection and label correction) and approaches focused on developing robust training mechanisms to mitigate the impact of label noise. A structured tabular comparison of reviewed LNL approaches is also presented. Furthermore, the practical applications of these techniques are examined in representative domains, including image analysis and natural language processing. Finally, several emerging research directions are identified and discussed to guide future exploration in this evolving field.