Robust Training of Deep Neural Networks with Noisy Labels by Graph Label Propagation

Recent developments in technology, such as crowdsourcing and web crawling, have made it easier to train machine learning models that require big data. However, the data collected by non-experts may contain noisy labels, and training a classification model on the data will result in poor generalization performance. In particular, Deep Neural Networks (DNNs) tend to over-fit to the noisy labels more significantly due to the large number of parameters. In this study, we propose a novel method to train DNNs robustly against the noisy labels by updating the network parameters with the labels corrected by graph label propagation on the similarity graph of training samples. The effectiveness of the proposed method is confirmed by comparing it with baseline MLP and CNNs on the noisy MNIST and CIFAR-10 datasets. Experimental results prove that the proposed method successfully corrects the noisy labels and trains DNNs more robustly than the baseline models.