FeatsFlow: Traceable representation learning based on normalizing flows

doi:10.1016/j.engappai.2023.107151

CORC > 自动化研究所 > 中国科学院自动化研究所 > 多模态人工智能系统全国重点实验室

	FeatsFlow: Traceable representation learning based on normalizing flows
	Zhang, Wenwen 4; Pei, Zhao 4; Wang, Fei-Yue 1,2,3
刊名	ENGINEERING APPLICATIONS OF ARTIFICIAL INTELLIGENCE
	2023-11-01
卷号	126 页码:13
关键词	Representation learning Distribution transformation Traceable features Normalizing flows
ISSN号	0952-1976
DOI	10.1016/j.engappai.2023.107151
通讯作者	Zhang, Wenwen(2021136@snnu.edu.cn)
英文摘要	This paper studies effective traceable feature representation learning in the view of distribution transformation, termed FeatsFlow, by proposing a distribution-aware learning framework combining the discriminating model with a normalizing flow-based model. The process can be regarded as a series of feature distribution transformations, from the input images to the expected results. Focusing on the learned representation of the target model, we take full advantage of the invertible nature of normalizing flows and learn the practical and traceable feature representation for target goals. Considering that it is difficult to model the traceable process for feature extraction, we propose an effective model by combining a general discriminating model with normalizing flows for traceable feature extraction. The normalizing flows module is added to the original model in a plug-in mode, which is convenient to make it available for effective and traceable feature learning. Thus we can obtain an effective and traceable representation distribution. Extensive experiments are conducted on our proposed representation learning model for the image classification task, and the experimental results illustrate that our proposed model is adequate for traceable representation learning. The most important is that we present a distribution-aware representation learning approach, which makes it possible to conduct and understand feature representation learning at the feature level.
资助项目	Natural Science Foundation for Young Scientists in Shaanxi Province of China[2023-JC-QN-0729] ; Fundamental Research Funds for the Central Universities[GK202207008]
WOS研究方向	Automation & Control Systems ; Computer Science ; Engineering
语种	英语
出版者	PERGAMON-ELSEVIER SCIENCE LTD
WOS记录号	WOS:001081735900001
资助机构	Natural Science Foundation for Young Scientists in Shaanxi Province of China ; Fundamental Research Funds for the Central Universities
内容类型	期刊论文
源URL	[http://ir.ia.ac.cn/handle/173211/52968]
专题	多模态人工智能系统全国重点实验室
通讯作者	Zhang, Wenwen
作者单位	1.Univ Chinese Acad Sci, Sch Artificial Intelligence, Beijing 100049, Peoples R China 2.Macau Univ Sci & Technol, Inst Syst Engn, Macau 999078, Peoples R China 3.Chinese Acad Sci, Inst Automat, Beijing 100190, Peoples R China 4.Shaanxi Normal Univ, Sch Comp Sci, Xian 710119, Peoples R China
推荐引用方式 GB/T 7714	Zhang, Wenwen,Pei, Zhao,Wang, Fei-Yue. FeatsFlow: Traceable representation learning based on normalizing flows[J]. ENGINEERING APPLICATIONS OF ARTIFICIAL INTELLIGENCE,2023,126:13.
APA	Zhang, Wenwen,Pei, Zhao,&Wang, Fei-Yue.(2023).FeatsFlow: Traceable representation learning based on normalizing flows.ENGINEERING APPLICATIONS OF ARTIFICIAL INTELLIGENCE,126,13.
MLA	Zhang, Wenwen,et al."FeatsFlow: Traceable representation learning based on normalizing flows".ENGINEERING APPLICATIONS OF ARTIFICIAL INTELLIGENCE 126(2023):13.