Given the swift progress of deep learning in fields like computer vision and natural language processing, the fidelity of dataset annotations has ascended to a position of paramount importance. In real-world scenarios, the construction of large-scale datasets often incurs substantial annotation costs, resulting in frequent noisy labels. These noisy labels can significantly increase the risk of overfitting and impair the generalization performance of deep neural networks. Numerous methods have been proposed and widely applied to real-world tasks such as image classification and text annotation to address this challenge, where they have demonstrated strong robustness and practical effectiveness. Label correction, in the context of these methodologies, is geared toward enhancing the caliber of annotations through the identification and subsequent rectification of instances marred by incorrect labeling.This paper presents a comprehensive analysis of this methodology by examining its underlying principles and model frameworks, exploring their development trajectories and advantages. Furthermore, it summarises the key challenges and promising future directions in developing label correction techniques.
Building similarity graph...
Analyzing shared references across papers
Loading...
Zhaowen Zhang
Applied and Computational Engineering
Northeastern University
Building similarity graph...
Analyzing shared references across papers
Loading...
Zhaowen Zhang (Tue,) studied this question.
www.synapsesocial.com/papers/689521de9f4f1c896c427f40 — DOI: https://doi.org/10.54254/2755-2721/2025.bj25820