Bias oriented unbiased data augmentation for cross-bias representation learning

doi:10.1007/s00530-022-01013-6

	Bias oriented unbiased data augmentation for cross-bias representation learning
	Li, Lei 1,2; Tang, Fan 1,2; Cao, Juan 1,2; Li, Xirong 3; Wang, Danding 1,2
	2022-10-25
发表期刊	MULTIMEDIA SYSTEMS
ISSN	0942-4962
页码	14
摘要	The biased cues in the training data may build strong connections between specific targets and unexpected concepts, leading the learned representations could not be applied to real-world data that does not contain the same biased cues. To learn cross-bias representations which can generalize on unbiased datasets by only using biased data, researchers focus on reducing the influence of biased cues through unbiased sampling or augmentation on the basis of artificial experience. However, the distributions of biased cues in the dataset are neglected, which limits the performance of these methods. In this paper, we propose a bias oriented data augmentation to enhance the cross-bias generalization by enlarging "safety" and "unbiasedness" constraints in the training data without manual prior intervention. The safety constraint is proposed to maintain the class-specific information for augmentation while the unbiasedness constraint reduces the statistical correlation of bias information and class labels. Experiments under different biased proportions on four synthetic/real-world datasets show that the proposed approach could improve the performance of other SOTA debiasing approaches (colored MNIST: 0.35-26.14%, corrupted CIFAR10: 3.14-8.44%, BFFHQ: 1.50% and BAR: 1.72%).
关键词	Cross-bias generalization Data augmentation Unbiased representation
DOI	10.1007/s00530-022-01013-6
收录类别	SCI
语种	英语
WOS研究方向	Computer Science
WOS类目	Computer Science, Information Systems ; Computer Science, Theory & Methods
WOS记录号	WOS:000871828500001
出版者	SPRINGER
引用统计
文献类型	期刊论文
条目标识符	http://119.78.100.204/handle/2XEOYT63/19771
专题	中国科学院计算技术研究所期刊论文
通讯作者	Tang, Fan
作者单位	1.Chinese Acad Sci, Inst Comp Technol, Beijing 100080, Peoples R China 2.Univ Chinese Acad Sci, Beijing 100049, Peoples R China 3.Renmin Univ China, Beijing 100872, Peoples R China
推荐引用方式 GB/T 7714	Li, Lei,Tang, Fan,Cao, Juan,et al. Bias oriented unbiased data augmentation for cross-bias representation learning[J]. MULTIMEDIA SYSTEMS,2022:14.
APA	Li, Lei,Tang, Fan,Cao, Juan,Li, Xirong,&Wang, Danding.(2022).Bias oriented unbiased data augmentation for cross-bias representation learning.MULTIMEDIA SYSTEMS,14.
MLA	Li, Lei,et al."Bias oriented unbiased data augmentation for cross-bias representation learning".MULTIMEDIA SYSTEMS (2022):14.