Trusted-Data-Guided Label Enhancement on Noisy Labels
Ning Xu, Jiayu Li, Yunpeng Liu, Xin Geng
Abstract
Label distribution covers a certain number of labels, representing the degree to which each label describes the instance. Label enhancement (LE) is a procedure of recovering the label distribution from the logical labels in the training data, the purpose of which is to better depict the label ambiguity through label distribution. However, data annotation inevitably introduces label noise, and it is extremely challenging to implement LE on corrupted labels. To deal with this problem, one way to recover the label distribution from the corrupted labels is to be guided by a small batch of trusted data. In this article, a novel LE method named TALEN is proposed via recovering and progressively refining label distribution guided by trusted data. Specifically, an LE process is applied to the untrusted data to select samples with a clean label. In addition, a combined loss function is designed to train the predictive model for classification. Experiments on datasets with synthetic label noise validate the feasibility of identifying clean labels via the recovered label distribution. Furthermore, experimental results on both synthetic label noise and real-world label noise on image datasets and additional experiments on text datasets show a clear advantage of TALEN over several existing noise-robust learning methods.