| Title |
Noise Label Correction via Gaussian Mixture Model with Prototype Mixing and Contrastive Learning |
| Authors |
김병일(Byeong-il Kim) ; 고병철(Byoung Chul Ko) |
| DOI |
https://doi.org/10.5573/ieie.2024.61.11.105 |
| Keywords |
Learning with noisy labels; Feature mixing; Gaussian mixture model; Label correction; Contrastive learning |
| Abstract |
Label noise is one of the major factors that degrade the performance of deep learning model training. This paper proposes a novel method to detect and correct label noise by leveraging the mixing of reliable prototypes and data features for each class. First, we train a classification model using contrastive learning and detect noisy labels through the mixing of image and prototype features. Data with low loss are considered informative and are defined as prototypes. Next, we detect noise by mixing the features of prototypes with reassigned labels. Finally, we construct a Gaussian Mixture Model (GMM) using clean-labeled data and prototypes to correct noisy labels through pseudo labels. The proposed method demonstrates superior performance compared to existing methods on the CIFAR-10/100 datasets. Specifically, for the CIFAR-100 dataset with a 90% noise ratio, the proposed method achieved a 9% improvement in accuracy over existing methods. |