计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (7): 162-166.DOI: 10.3778/j.issn.1002-8331.2108-0185

• 模式识别与人工智能 • 上一篇    下一篇

融合动态残差的多源域自适应算法研究

王斌,李昕   

  1. 中国石油大学(华东) 计算机科学与技术学院,山东 青岛 266580
  • 出版日期:2022-04-01 发布日期:2022-04-01

Research on Multi-Source Domain Adaptive Algorithm Integrating Dynamic Residuals

WANG Bin, LI Xin   

  1. College of Computer Science and Technology, China University of Petroleum(East China), Qingdao, Shangdong 266580, China
  • Online:2022-04-01 Published:2022-04-01

摘要: 多源域自适应问题通常是指拥有多个源域与单个目标域的场景。常见做法是依据域标签两两对齐源域与目标域分布,通过减小域间距离,将分布映射到共同隐空间内,去预测未知目标域的数据分类。源数据集通常需要域标签,且模型在经过训练阶段后,参数固定,这就很难达到拟合未知目标域分布的目的。基于动态残差块的多源域自适应算法不是从域的角度而是从数据自身特征映射生成神经网络参数,不需要域标签,将多源域自适应问题转化为单源域问题。而且动态残差块能够跨阶段的根据输入数据特征改变网络参数,更好地让网络参数拟合未经训练的目标域数据分布,简化了多源域自适应的模型设计复杂程度,减少了数据准备工作量。实验结果表明,在模型中引入动态残差块,与静态模型相比准确率提高了8.1%,同时也节约了模型运行的时间和空间。

关键词: 域自适应, 动态残差块, 多源域自适应, 迁移学习, 深度学习

Abstract: Multi-source domain adaptation problem usually refers to a scene with multiple source domains and a single target domain. The common approach is to align the distribution of source domain and target domain according to the domain label, and map the distribution to the common hidden space by reducing the distance between domains to predict the data classification of unknown target domain. The source data set usually needs domain label, and the parameters of the model are fixed after the training stage, which is difficult to fit the distribution of unknown target domain. The multi-source domain adaptive algorithm based on dynamic residual block generates neural network parameters not from the perspective of domain, but from the feature mapping of data itself. Without domain label, the multi-source domain adaptive problem is transformed into a single source domain problem. Moreover, the dynamic residual block can change the network parameters across stages according to the characteristics of the input data, better let the network parameters fit the untrained target domain data distribution, simplify the complexity of multi-source domain adaptive model design, and reduce the workload of data preparation. The experimental results show that the accuracy is improved by 8.1% compared with the static model, and the running time and space of the model are saved.

Key words: domain adaption, dynamic residual block, multi-source domain adaption, transfer learning, deep learning