Abstract Existing domain adaptation approaches focus on taking advantage of easy samples, i.e, target samples which are easier for adaptation. In previous work, tough, or hard, target samples are generally regarded as outliers or just being left to chance. As a result, the adaptation of tough target samples remains as a challenging problem in the community. In this paper, we report three novel ideas for domain adaptation: 1) splitting target samples into easy and tough ones; 2) deploying different strategies for samples with different adaptation difficulties; 3) leveraging easy samples to facilitate tough ones. Furthermore, we present a novel approach, named challenging tough sample networks (CTSN), to practice the three ideas and tame tough samples. Specifically, in our approach, a CNN with domain adaptation layers is first used to rapidly handle the easy samples and identify the tough ones. Then, a GAN with two classifiers is tailored to adapt the tough samples. The GAN leverages classification discrepancy and easy samples to tame the tough ones. Extensive experiments on both classic and large-scale benchmarks verify that both easy and tough samples do exist in real-world datasets and our approach is able to handle them.

中文翻译：

---

在无监督领域适应中挑战艰难的样本

摘要 现有的领域适应方法侧重于利用容易的样本，即更容易适应的目标样本。在以前的工作中，困难或困难的目标样本通常被视为异常值或只是偶然。因此，艰难目标样本的适应仍然是社区中的一个具有挑战性的问题。在本文中，我们报告了域适应的三个新想法：1）将目标样本分为简单样本和困难样本；2）针对不同适应难度的样本部署不同的策略；3）利用简单的样本促进困难的样本。此外，我们提出了一种名为挑战艰难样本网络（CTSN）的新方法，以实践这三个想法并驯服艰难的样本。具体来说，在我们的方法中，首先使用具有域适应层的 CNN 来快速处理简单的样本并识别困难的样本。然后，定制具有两个分类器的 GAN 以适应困难的样本。GAN 利用分类差异和简单的样本来驯服困难的样本。在经典和大规模基准测试上的大量实验证实，在现实世界的数据集中确实存在简单和困难的样本，并且我们的方法能够处理它们。