• 分享主题：Transfer Learning, Unsupervised Domain Adaptation, Explicit Feature Distribution Alignment, CV
  • 论文标题：Adversarial-Learned Loss for Domain Adaptation
  • 论文链接：https://arxiv.org/pdf/2001.01046.pdf

    1.Summary

    This is a paper about classification tasks. Now there is a source domain and a target domain. The source domain data is labeled and the target domain data is not labeled. So this is a unsupervised domain adaptation problem. Because there is no label in the target domain, the previous methods directly align the features from the target domain without using the label information. Therefore, this paper proposes Adversarial-Learned Loss for Domain Adaptation (ALDA) model. ALDA treats the result of the label classifier as a pseudo label. The function of the discriminator is no longer to discriminate domain labels, but to generate noise vector and construct a confusing matrix. Then, the confusing matrix is multiplied by the result of the label classifier. For the source domain, the loss is calculated directly from the result and the real label. For the target domain, it is optimized in the opposite direction to the pseudo label. In this way, the discriminator can distinguish the characteristics of the source domain and the target domain. Experiments show that this is effective. In order to deepen my understanding of this paper, I can read some papers about domain adaptation.

    2.你对于论文的思考

    这篇文章是关于UDA的，因为目标域上没有标签，所以以往的工作没有关注这一点，于是本文提出了ALDA模型，利用伪标签来帮助迁移效果的提升。此外，ALDA中判别器的作用也不再是直接判断域标签，而是生成一个noise vector，再用这个noise vector构造confusing matrix，与标签分类器矩阵相乘后，如果是源域数据，那就直接和真实标签计算损失函数，如果是目标域数据，那就向着伪标签的反方向进行优化，也就是源域数据和目标域数据的优化方向时相反的，这样就可以起到对齐源域和目标域的目的。

    3. 其他

    3.1 需要解决的问题

    利用目标域上的伪标签，提高域对抗过程中的迁移效果。

    3.2 ALDA

    20220714155942.png

    3.2.1 判别器损失

    判别器的损失如下：
    20220714160237.png
    第一部分相当于域对抗损失：
    20220714160313.png
    20220714160329.png
    20220714160339.png
    第二部分相当于正则化：
    20220714160415.png

    3.2.2 标签分类器损失

    标签分类器损失如下：
    20220714160500.png
    20220714160539.png

    3.2.3 特征提取器（生成器）损失

    特征提取器损失如下：
    20220714160634.png

    3.3 实验

    （1）数据集：Office-31
    20220714160859.png
    （2）数据集：Office-Home
    20220714160920.png
    （3）数据集：VisDA-2017
    20220714161005.png
    （4）特征可视化（利用了t-SNE降到了二维）
    图中蓝色和红色代表不同的域，ALDA效果很明显。
    20220714161050.png