Pseudo-margin-based universal domain adaptation

摘要

As a more practical setting for unsupervised domain adaptation, Universal Domain Adaptation (UDA) is recently introduced, where the structure of the target label set is assumed to be unavailable. One of the significant challenges in UDA is identifying the common label set shared by source and target domains without labels in the target domain. Most previous works identify the common label set according to the intra-domain information. This article proposes a novel inter-domain approach to identify the common label set efficiently, where the pseudo-margins of target samples are defined to estimate each source class’s probability belonging to the common label set. Specifically, a pseudo-margin vector (PMV) is proposed to describe each source class’s reliability belonging to the common label set, equipped with a target pseudo-margin register (TPMR), which stores and updates the PMV during training. We propose an improved universal adaptation network (I-UAN) to perform domain alignment on the commonly-labeled samples by incorporating the PMV. In I-UAN, a class-wise weighting mechanism based on TPMR for source classes, combined with the sample-wise weighting for target samples, efficiently improves the UDA’s performance. Our approach is robust to outliers as it applies a class-wise weighting mechanism in the source domain. Moreover, we provide experimental analysis on the adaptability, universality, and robustness of I-UAN. Extensive experimental results demonstrate that I-UAN works well in various UDA settings and outperforms the state-of-the-art methods by large margins.