Exploration on Inaccurate Supervision Problems Generated by Noisy Labels

Abstract

In the field of machine learning, the problem of noisy labels is still prevalent even though the application of artificial intelligence techniques is becoming more and more mature. Noisy labels, in other words, inaccurate or unreliable data labels, cannot be avoided due to reasons such as the complex and time-consuming process of image labelling. Meanwhile, the existence of noisy labels can have a significant impact on the training of deep learning methods. Therefore, how to effectively recognise noise labels and deal with the problems they bring remains challenging in the field. In this paper, we will first explore the reasons for the formation of noise labels and their impact and then analyse some cutting-edge methods to solve the noise labelling problem, analyse their theoretical foundations, advantages, and limitations, and discuss the future research trends in this field, with the aim of providing a systematic perspective for machine learning researchers and practitioners to deepen their theoretical understanding of the noise problem and explore effective algorithmic frameworks to improve the training efficiency and prediction confidence of machine learning models in the future.