Training Robust Deep Neural Networks on Noisy Labels Using Adaptive Sample Selection With Disagreement

Learning with noisy labels is one of the most practical but challenging tasks in deep learning. One promising way to treat noisy labels is to use the small-loss trick based on the memorization effect, that is, clean and noisy samples are identified by observing the network’s loss during training. Co-teaching+ is a state-of-the-art method that simultaneously trains two networks with small-loss selection using the “update by disagreement” strategy; however, it suffers from the problem that the selected samples tend to become noisy as the number of iterations increases. This phenomenon means that clean small-loss samples will be biased toward agreement data, which is the set of samples for which the two networks have the same prediction. This paper proposes an adaptive sample selection method to train deep neural networks robustly and prevent noise contamination in the disagreement strategy. Specifically, the proposed method calculates the threshold of the small-loss criterion by considering the loss distribution of the whole batch at each iteration. Then, the network is backpropagated by extracting samples below this threshold from the disagreement data. Combining the disagreement and agreement data of the two networks can suppress the degradation of the true-label rate of training data in a mini batch. Experiments were conducted using five commonly used benchmarks, MNIST, CIFAR-10, CIFAR-100, NEWS, and T-ImageNet to verify the robustness of the proposed method to noisy labels. The results show the proposed method improves generalization performance in an image classification task with simulated noise rates of up to 50%.