基于分类差异与信息熵对抗的无监督域适应算法

PDF下载HTML

Title:

Unsupervised domain adaptation algorithm based on classification discrepancy and information entropy

作者:

李庆勇¹, 何军^1,2, 张春晓¹

1. 南京信息工程大学 电子与信息工程学院，江苏 南京 210044;
2. 南京信息工程大学 人工智能学院，江苏 南京 210044

Author(s):

LI Qingyong¹, HE Jun^1,2, ZHANG Chunxiao¹

1. School of Electronics and Information Engineering, Nanjing University of Information Science and Technology, Nanjing 210044, China;
2. School of Artificial Intelligence, Nanjing University of Information Science and Technology, Nanjing 210044, China

关键词:

域适应; 对抗训练; 神经网络; 无监督学习; 迁移学习; 分类差异; 信息熵; 决策边界

Keywords:

domain adaptation; confrontation training; neural network; unsupervised learning; transfer learning; classification discrepancy; information entropy; decision boundary

分类号:

TP391

DOI:

10.11992/tis.202010020

摘要:

采用对抗训练的方式成为域适应算法的主流，通过域分类器将源域和目标域的特征分布对齐，减小不同域之间的特征分布差异。但是，现有的域适应方法仅将不同域数据之间的距离缩小，而没有考虑目标域数据分布与决策边界之间的关系，这会降低目标域内不同类别的特征的域内可区分性。针对现有方法的缺点，提出一种基于分类差异与信息熵对抗的无监督域适应算法(adversarial training on classification discrepancy and information entropy for unsupervised domain adaptation, ACDIE)。该算法利用两个分类器之间的不一致性对齐域间差异，同时利用最小化信息熵的方式降低不确定性，使目标域特征远离决策边界，提高了不同类别的可区分性。在数字标识数据集和Office-31数据集上的实验结果表明，ACDIE算法可以学习到更优的特征表示，域适应分类准确率有明显提高。

Abstract:

The adversarial training method has become the mainstream of the domain adaptation algorithm. The feature distributions of the source and target domains are aligned by a domain classifier to reduce the feature distribution discrepancy among different domains. However, existing domain adaptation methods only reduce the distance between different domain data without considering the relationship between the data distribution of the target domain and decision boundaries, thus decreasing the intradomain distinguishability of different categories in the target domain. Considering the shortcomings of the existing methods, an unsupervised domain adaptation algorithm based on classification discrepancy and information entropy confrontation (ACDIE) is proposed in this study. The algorithm uses the discrepancy and the domain aligning discrepancy between two classifiers and minimizes the information entropy to reduce uncertainty. Consequently, the proposed method makes the target domain feature far away from the decision boundaries and improves the distinguishability of different categories. The experimental results of the digital identification and Office-31 datasets show that the ACDIE algorithm can learn better feature representation. Moreover, the accuracy of the domain adaptation classification is considerably improved.