iNLC: Iterative Noisy Label Correction

초록

Convolutional neural networks have achieved remarkable success in image classification, but the presence of noisy labels in the training dataset can significantly hinder their achievement. Creating clean labeled datasets is time-consuming; therefore, learning with noisy datasets is a practical approach to solving real-world problems. In this paper, we propose a novel method called iterative Noisy Label Correction (iNLC) that employs gradual data refining to mitigate the impact of incorrect labels in practice. iNLC consists of three main components: robust prior learning, ensemble strategy, and gradual data refining. The robust prior learning trains the prior model via semi-supervised learning to make it more robust against noise and facilitate subsequent gradual refinement. The ensemble strategy improves performance by combining different augmentation strategies, and the gradual data refining process progressively incorporates additional data into the training using fine-grained learning schedules based on data volume in order to prevent overfitting and underfitting. Our method achieved the classification accuracy of 93.98%, 92.96%, and 75.32% for noise ratios of 0.2, 0.4, and 0.8, respectively, on PreAct-ResNet-18 on CIFAR-10.

키워드