Collaborative Learning with Pseudo Labels for Robust Classification in the Presence of Noisy Labels

Abstract

Supervised learning depends on labels of dataset to train models with desired properties. Therefore, data containing mislabeled samples (a.k.a. noisy labels) can deteriorate supervised learning performance significantly as it makes models to be trained with wrong targets. There are technics to train models in the presence of noise in data labels, but they usually suffer from the data inefficiency or overhead of additional steps. In this work, we propose a new way to train supervised learning models in the presence of noisy labels. The proposed approach effectively handles noisy labels while maintaining data efficiency by replacing labels of large-loss instances that are likely to be noise with newly generated pseudo labels in the training process. We conducted experiments to demonstrate the effectiveness of the proposed method with public benchmark datasets: CIFAR-10, CIFAR-100 and Tiny-ImageNet. They showed that our method successfully identified correct labels and performed better than other state-of-the-art algorithms for noisy labels.