A Discriminative Approach To Unsupervised Domain Adaptation in Coarse-To-Fine Classifiers
A Discriminative Approach To Unsupervised Domain Adaptation in Coarse-To-Fine Classifiers
复制标题
DOI：
10.1109/mlsp55844.2023.10285908
复制
发表时间：
2023-09
期刊：
影响因子：
--
通讯作者：
Ismail R. Alkhouri;Akram S. Awad;Connor Hatfield;George Atia
中科院分区：
文献类型：
--
作者：
Ismail R. Alkhouri;Akram S. Awad;Connor Hatfield;George Atia
当在来自与训练数据不同的概率分布的数据上进行测试时，传统的机器学习模型通常表现出较差的性能。这种现象被称为“域转移”。为了克服这一问题，领域自适应(DA)方法被提出以减小两个领域之间的差异并提高在目标领域上的性能。然而，现有的大多数DA分类技术都集中在一级分类器(OLC)上，这不适合于遵循分层分类结构的数据。在本文中，我们提出了一种区分DA方法用于从粗到精(C2F)的分类器，该方法同时提供了粗细标签。我们的方法在粗略和精细两个层次上学习跨域数据的不变特征表示。在已知数字数据集上的实验结果表明，该算法在目标领域数据的分类精度方面优于基本的C2F模型。
Conventional machine learning models often exhibit poor performance when tested on data that comes from a different probability distribution than that of the training data. This phenomenon is known as ‘domain shift’. To overcome this problem, domain adaptation (DA) methods have been proposed to reduce the discrepancy between the two domains and improve the performance on the target domain. However, most of the existing DA techniques for classification have focused on One Level Classifiers (OLCs), which are not suitable for data that follows a hierarchical classification structure. In this paper, we propose a discriminative DA approach for coarse-to-fine (C2F) classifiers, which provide both coarse and fine labels. Our approach learns an invariant feature representation of the data across domains for both the coarse and finer levels. Our experimental results on well-known digit datasets demonstrate that the proposed algorithm outperforms the base C2F model in terms of the classification accuracy of the target domain data.