Tripartite: Tackling Realistic Noisy Labels with More Precise Partitions

Samples in large-scale datasets may be mislabeled for various reasons, and deep models are inclined to over-fit some noisy samples using conventional training procedures. The key solution is to alleviate the harm of these noisy labels. Many existing methods try to divide training data into clean and noisy subsets in terms of loss values. We observe that a reason hindering the better performance of deep models is the uncertain samples, which have relatively small losses and often appear in real-world datasets. Due to small losses, many uncertain noisy samples are divided into the clean subset and then degrade models' performance. Instead, we propose a Tripartite solution to partition training data into three subsets, uncertain, clean and noisy according to the following criteria: the inconsistency of the predictions of two networks and the given labels. Tripartite considerably improves the quality of the clean subset. Moreover, to maximize the value of clean samples in the uncertain subset and minimize the harm of noisy labels, we apply low-weight learning and a semi-supervised learning, respectively. Extensive experiments demonstrate that Tripartite can filter out noisy samples more precisely and outperforms most state-of-the-art methods on four benchmark datasets and especially real-world datasets.