Data fusing and joint training for learning with noisy labels

It is well known that deep learning depends on a large amount of clean data. Because of high annotation cost, various methods have been devoted to annotating the data automatically. However, a larger number of the noisy labels are generated in the datasets, which is a challenging problem. In this paper, we propose a new method for selecting training data accurately. Specifically, our approach fits a mixture model to the per-sample loss of the raw label and the predicted label, and the mixture model is utilized to dynamically divide the training set into a correctly labeled set, a correctly predicted set, and a wrong set. Then, a network is trained with these sets in the supervised learning manner. Due to the confirmation bias problem, we train the two networks alternately, and each network establishes the data division to teach the other network. When optimizing network parameters, the labels of the samples fuse respectively by the probabilities from the mixture model. Experiments on CIFAR-10, CIFAR-100 and Clothing1M demonstrate that this method is the same or superior to the state-of-the-art methods.