CJC-net: A cyclical training method with joint loss and co-teaching strategy net for deep learning under noisy labels

Abstract The recent success of deep convolutional neural networks (CNNs) is mostly due to the availability of large-scale datasets with accurate annotations. However, the collection of such large datasets with clean annotations is time-consuming and not always feasible. In this paper, we propose a novel framework for learning with noisy labels, called the Cyclical training method with Joint loss and Co-teaching strategy net (CJC-net), where the net means our method is insensitive to the structures of the CNN. CJC-net pretrains two networks simultaneously and then performs the cyclical training strategy under an improved co-teaching method based on the two pretrained networks. During the training process, we adjust the learning rates of the two networks to make the network states periodically transfer from overfitting to underfitting. The cumulative loss of each sample under two networks is recorded; and the higher the cumulative loss of a sample is, the higher the probability of identifying it as a noisy label or a hard label. Then, we remove those samples with a high loss and fine-tune the two networks using the remaining data. The experimental results on several datasets demonstrate that CJC-net is superior to many state-of-the-art methods.