摘要
现实世界中训练数据和测试数据往往存在分布差异,导致基于独立同分布假设的模型丧失鲁棒性.无监督域自适应是一种重要解决方法,极具应用价值.鉴于此,国内外研究者进行大量理论基础和方法技术的研究,促进了很多应用领域的发展,包括自动驾驶、智慧医疗等.但是,目前主流的方法仍存在一些问题:源域和目标域的概率分布距离是否能真正代表它们之间的差异,以及如何更准确地度量2个分布之间的差异,仍然是一个值得探讨的问题.同时,如何更有效地利用伪标签,也是一个值得继续探索的问题.提出了反向伪标签最优化传输(backward pseudo-label and optimal transport,BPLOT),不仅利用瓦瑟斯坦距离和格罗莫夫-瓦瑟斯坦距离,从最优化特征-拓扑传输的角度更准确地计算了2个分布之间的差异;而且提出了反向验证伪标签的模块来更有效地利用伪标签,在训练过程中验证伪标签的质量.将所提出的方法在多个无监督域自适应的数据集上进行了实验验证.实验结果表明,BPLOT模型的效果超过了所有对比的基准方法.
The distribution shift in the real world will degrade the model performance,so unsupervised domain adaptation is a valuable solution that is of great importance with high application value.Training data XS and the model’s test data XT are assumed to be independent and identically distributed(IID)in traditional machine learning.But distribution shift can be found in the training and test data in most cases,leading to the loss of robustness of the models based on the IID assumption.In view of this,researchers at home and abroad have conducted a lot of research on the theoretical basis and methodological techniques in the direction of unsupervised domain adaptation,which has led to the development of many application areas,including autonomous driving,intelligent medical care,etc.It is worth noting that there are some problems such as whether the distance obtained by the existing methods can truly represent the difference between the two distributions,and the ways of more accurately measuring the difference between the two distributions in the mainstream unsupervised domain adaptation approaches.Moreover,how to verify the model’s ability to transfer knowledge during training with the help of pseudo labeling is also an issue worthy of continuously exploring.Hence,we propose a backward pseudo-label and optimal transport for unsupervised domain adaptation(BPLOT)in order to explicitly reduce the distance between the two distributions in the feature space by calculating the difference between the two distributions more accurately from the perspective of optimal transport.More than that,a backward pseudo-label verification method is also proposed to verify the mode’s ability of knowledge transfer and guide model training by means of verifying the quality of pseudo-label during training.At last,the proposed network is experimentally verified on a plurality of datasets for unsupervised domain adaptation.The BPLOT model is superior to all compared baseline methods in terms of the effect.
作者
孙昊
韩忠义
王帆
尹义龙
Sun Hao;Han Zhongyi;Wang Fan;Yin Yilong(School of Software,Shandong University,Jinan 250000)
出处
《计算机研究与发展》
EI
CSCD
北大核心
2023年第8期1696-1710,共15页
Journal of Computer Research and Development