CLC: Noisy Label Correction via Curriculum Learning
Jaeyoon Lee, Hyuntak Lim, Ki‐Seok Chung · 2021 IEEE Symposium Series on Computational Intelligence (SSCI) · 2021
Deep neural networks reveal their usefulness through learning from large amounts of data. However, unless the data is correctly labeled, it may be very difficult to properly train a neural network. Labeling the large set of data is a time-consuming and labor-intensive task. To overcome the risk of mislabeling, several methods that are robust against the label noise have been proposed. In this paper, we propose an effective label correction method called Curriculum Label Correction (CLC). With reference to the loss distribution from self-supervised learning, CLC identifies and corrects noisy labels utilizing curriculum learning. Our experimental results verify that CLC shows outstanding performance especially in a harshly noisy condition, 91.06% test accuracy on CIFAR-10 at a noise rate of 0.8. Code is available at https://github.com/LJY-HY/CLC.