| 引用本文: | 许亚雲,严华.无监督域适应的表示学习算法[J].哈尔滨工业大学学报,2021,53(2):40.DOI:10.11918/202006133 |
| XU Yayun,YAN Hua.Representation learning for unsupervised domain adaptation[J].Journal of Harbin Institute of Technology,2021,53(2):40.DOI:10.11918/202006133 |
|
| 摘要: |
| 在传统模式识别的训练方法中,拥有大量标签的有监督学习方法在识别准确率上取得了很好的效果.然而在实际生活中样本常常缺失标签,或现存有标签的样本与目标样本具有较大分布差异而不能直接使用.为了解决这些问题,无监督域自适应算法应运而生,借助源域有标签但不同分布的样本去识别无标签的目的域样本.针对目标识别样本与训练样本分布不一致的情况,本文提出了一种探寻两个样本域之间的最优表示学习的无监督域适应算法.通过在共同的子空间上引入两个表示矩阵去更好地减少两个域的分布差异,同时对两个表示矩阵进行各自的最优化约束设计,使得源域和目的域最优地相互表示,缩小两个域之间的分布差异,从而实现无监督跨域学习（即迁移学习）.最后在3个迁移学习常见的无监督域适应数据集上开展实验,实验结果表明:本文算法的识别准确率超过了目前很多优秀的传统迁移学习方法和一些深度方法.实验结果验证了本文提出的无监督域适应的表示学习算法的有效性和鲁棒性. |
| 关键词: 模式识别 迁移学习 无监督域适应 表示学习 分布差异 |
| DOI：10.11918/202006133 |
| 分类号:TP391.4 |
| 文献标识码:A |
| 基金项目:国家自然科学基金(11872069) |
|
| Representation learning for unsupervised domain adaptation |
|
XU Yayun,YAN Hua
|
|
(College of Electronics and Information Engineering, Sichuan University, Chengdu 610065, China)
|
| Abstract: |
| Among the conventional training methods of pattern recognition, the supervised learning methods with abundant labels have achieved significant performance in recognition accuracy. However, in real life, there are problems that samples often lack labels, or existing labeled samples cannot be used directly due to the distributional divergences of the target samples. To resolve these issues, unsupervised domain adaptation is often applied to recognize samples of unlabeled target domain by taking advantage of the data of the source domain with sufficient labels but different distributions. Considering the situation that the distributions of the target recognition samples and the source training samples are different, an optimal representation learning method for unsupervised domain adaption was proposed. Two representation matrices were introduced to the common subspace of the domain samples to better reduce the distributional divergence of the domains. Then, optimization constraints were implemented on the two representation matrices, so as to make the source domain and the target domain optimally represent each other, thereby reducing the distributional divergence between the domains. In this way, the unlabeled target domain samples could be recognized by the fully labeled source domain samples (i.e., transfer learning). Experiments on three common unsupervised domain adaption datasets show that the proposed method outperformed the conventional transfer learning methods and deep learning methods in recognition accuracy, which verifies the validity and robustness of the proposed algorithm. |
| Key words: pattern recognition transfer learning unsupervised domain adaptation representation learning distributional divergence |
