基于自适应噪声校正的鲁棒域适应学习

域适应（Domain Adaptation，DA）学习旨在利用标签丰富的源域来帮助标签稀缺的目标域学习。DA方法通常假设源域数据已正确标记，然而现实中通常很难收集到大量带有干净标签的源实例，带有噪声源标签的噪声DA学习可能会降低目标学习性能。为此，本文提出基于自适应标签噪声校正的鲁棒DA学习方法（Robust DA Method through Adaptive Noise Correction，RoDAC）。RoDAC包含两个学习阶段，即自适应噪声标签检测（Adaptive Noise Label Detection，ANLD）和自适应噪声标签校正（Adaptive Noise Label Correction，ANLC）。在ANLD中，使用自适应噪声检测器识别带有噪声标签的源实例，并进一步在ANLC中自适应地校正噪声标签，将其重新投入域适应学习中。与基准数据集进行比较，结果表明RoDAC方法在源域标签存在噪声的域适应场景中取得了显著的性能提升。该学习策略可集成至许多现有的DA方法中，以提升其在噪声标签场景下的学习性能。

Robust Domain Adaptive Learning with Adaptive Noise Correction

Domain Adaptation (DA) learning aims to use lable-rich source domains to help the learning of label-scarce target domain.The DA method usually assumes that the source domain data has been correctly labeled.However, in reality, it is usually difficult to collect a large number of source instances with clean labels.Noise DA learning with noise source labels may reduce the target learning performance.Therefore, this article proposes a Robust DA Method through Adaptive Noise Correction (RoDAC).RoDAC consists of two learning stages, Adaptive Noise Label Detection (ANLD) and Adaptive Noise Label Correction (ANLC).In ANLD, an adaptive noise detector is used to identify the source instance with noise labels, and the noise labels are further adaptively corrected in ANLC and reinvested in domain adaptation learning.Compared with the benchmark data set, the results show that the RoDAC method achieves significant performance improvement in the domain adaptation scenario where the source domain label has noise.This learning strategy can be integrated into many existing DA methods to improve its learning performance in noisy label scenarios.