Towards prior gap and representation gap for long-tailed recognition

doi:10.1016/j.patcog.2022.109012

CASIA OpenIR > 多模态人工智能系统全国重点实验室 > 模式分析与学习

	Towards prior gap and representation gap for long-tailed recognition
	Zhang, Ming-Liang 1,2; Zhang, Xu-Yao1,2 ; Wang, Chuang1,2 ; Liu, Cheng-Lin1,2
发表期刊	PATTERN RECOGNITION
ISSN	0031-3203
	2023
卷号	133 页码:12
通讯作者	Zhang, Ming-Liang(zhangmingliang2018@ia.ac.cn)
摘要	Most deep learning models are elaborately designed for balanced datasets, and thus they inevitably suf-fer performance degradation in practical long-tailed recognition tasks, especially to the minority classes. There are two crucial issues in learning from imbalanced datasets: skew decision boundary and unrep-resentative feature space. In this work, we establish a theoretical framework to analyze the sources of these two issues from Bayesian perspective, and find that they are closely related to the prior gap and the representation gap, respectively. Under this framework, we show that existing long-tailed recognition methods manage to remove either the prior gap or the presentation gap. Different from these methods, we propose to simultaneously remove the two gaps to achieve more accurate long-tailed recognition. Specifically, we propose the prior calibration strategy to remove the prior gap and introduce three strate-gies (representative feature extraction, optimization strategy adjustment and effective sample modeling) to mitigate the representation gap. Extensive experiments on five benchmark datasets validate the supe-riority of our method against the state-of-the-art competitors.(c) 2022 Elsevier Ltd. All rights reserved.
关键词	Long-tailed learning Prior gap Representation gap Image recognition
DOI	10.1016/j.patcog.2022.109012
关键词[WOS]	NEURAL-NETWORK CLASSIFICATION
收录类别	SCI
语种	英语
资助项目	National Key Research and Development Program[2018AAA010 040 0] ; National Natural Science Foundation of China (NSFC)[U20A20223] ; National Natural Science Foundation of China (NSFC)[62076236] ; National Natural Science Foundation of China (NSFC)[61721004]
项目资助者	National Key Research and Development Program ; National Natural Science Foundation of China (NSFC)
WOS研究方向	Computer Science ; Engineering
WOS类目	Computer Science, Artificial Intelligence ; Engineering, Electrical & Electronic
WOS记录号	WOS:000863094500012
出版者	ELSEVIER SCI LTD
引用统计	被引频次：7[WOS] [WOS记录] [WOS相关记录]
文献类型	期刊论文
条目标识符	http://ir.ia.ac.cn/handle/173211/50309
专题	多模态人工智能系统全国重点实验室_模式分析与学习
通讯作者	Zhang, Ming-Liang
作者单位	1.Chinese Acad Sci, Natl Lab Pattern Recognit NLPR, Inst Automat, Beijing 100190, Peoples R China 2.Univ Chinese Acad Sci, Sch Artificial Intelligence, Beijing 100049, Peoples R China
第一作者单位	模式识别国家重点实验室
通讯作者单位	模式识别国家重点实验室
推荐引用方式 GB/T 7714	Zhang, Ming-Liang,Zhang, Xu-Yao,Wang, Chuang,et al. Towards prior gap and representation gap for long-tailed recognition[J]. PATTERN RECOGNITION,2023,133:12.
APA	Zhang, Ming-Liang,Zhang, Xu-Yao,Wang, Chuang,&Liu, Cheng-Lin.(2023).Towards prior gap and representation gap for long-tailed recognition.PATTERN RECOGNITION,133,12.
MLA	Zhang, Ming-Liang,et al."Towards prior gap and representation gap for long-tailed recognition".PATTERN RECOGNITION 133(2023):12.