Learning class predication distributions with noisy labels

Abstract

In this paper, we propose a simple but very effective approach in the presence of noisy labels. The <i>memorization effects</i> of Deep Neural Networks (DNNs) manifests that they first memorize training data with clean labels and memorize data with noisy labels gradually. Based on this phenomenon, we build Class Prediction Distributions(CPD) for each sample in the initial stage of network training. On the basis of CPD, we use our clean data selection strategy to divide training data into confidently clean data and noisy data. In this selection strategy, we rank the maximum value of CPD. Top-ranked samples are more likely to be clean samples. Finally, noisy labels classification is successfully achieved by using semi-supervised learning. Experiments on benchmark datasets including MNIST, Cifar-10, Cifar-100 and Clothing1M demonstrate that our approach can achieve a competitive performance.