Improving metric-based few-shot learning with dynamically scaled softmax loss

doi:10.1016/j.imavis.2023.104860

CSpace

	Improving metric-based few-shot learning with dynamically scaled softmax loss
	Zhang, Yu 1; Zuo, Xin 1; Zheng, Xuxu 2; Gao, Xiaoyong 1; Wang, Bo 3,4,7; Hu, Weiming 3,5,6
	2023-12-01
发表期刊	IMAGE AND VISION COMPUTING
ISSN	0262-8856
卷号	140 页码:15
摘要	The metric-based learning framework has been widely used in data-scarce few-shot visual classification. However, the current loss function limits the effectiveness of metric learning. One issue is that the nearest neighbor classification technique used greatly narrows the value range of similarity between the query and class prototypes, which limits the guiding ability of the loss function. The other issue is that the episode-based training setting randomizes the class combination in each iteration, which reduces the perception of the traditional softmax losses for effective learning from episodes with various data distributions.To solve these problems, we first review some variants of the softmax loss from a unified perspective, and then propose a novel DynamicallyScaled Softmax Loss (DSSL). By adding a probability regulator (for scaling probabilities) and a loss regulator (for scaling losses), the loss function can adaptively adjust the prediction distribution and the training weights of the samples, which forces the model to focus on more informative samples. Finally, we found the proposed DSSL strategy for few-shot classifiers can achieve competitive results on four generic benchmarks and a fine-grained benchmark, demonstrating the effectiveness of improving the distinguishability (for base classes) and generalizability (for novel classes) of the learned feature space.
关键词	Few-shot learning Metric-based learning framework Softmax loss improvement
DOI	10.1016/j.imavis.2023.104860
收录类别	SCI
语种	英语
资助项目	Beijing Natural Science Foundation[62036011] ; Beijing Natural Science Foundation[62192782] ; Beijing Natural Science Foundation[61721004] ; Beijing Natural Science Foundation[U2033210] ; Major Projects of Guangdong Edu-cation Department for Foundation Research and Applied Research[L223003] ; Guangdong Provincial University Innovation Team Project[2017KZDXM081] ; Guangdong Provincial University Innovation Team Project[2018KZDXM066] ; [2020KCXTD045]
WOS研究方向	Computer Science ; Engineering ; Optics
WOS类目	Computer Science, Artificial Intelligence ; Computer Science, Software Engineering ; Computer Science, Theory & Methods ; Engineering, Electrical & Electronic ; Optics
WOS记录号	WOS:001110250500001
出版者	ELSEVIER
引用统计	被引频次：2[WOS] [WOS记录] [WOS相关记录]
文献类型	期刊论文
条目标识符	http://119.78.100.204/handle/2XEOYT63/38080
专题	中国科学院计算技术研究所
通讯作者	Wang, Bo
作者单位	1.China Univ Petr, Beijing 102249, Peoples R China 2.Chinese Acad Sci, Data Intelligence Syst Res Ctr, Inst Comp Technol, Beijing 100190, Peoples R China 3.Chinese Acad Sci, Inst Automat, Natl Lab Pattern Recognit, Beijing 100190, Peoples R China 4.Peking Univ, Sch Software & Microelect, Beijing 100871, Peoples R China 5.Univ Chinese Acad Sci, Sch Artificial Intelligence, Beijing 100190, Peoples R China 6.ShanghaiTech Univ, Sch Informat Sci & Technol, Shanghai 201210, Peoples R China 7.Chinese Acad Sci, Natl Lab Pattern Recognit, Inst Automat, 95 Zhongguancun East Rd, Beijing 100190, Peoples R China
推荐引用方式 GB/T 7714	Zhang, Yu,Zuo, Xin,Zheng, Xuxu,et al. Improving metric-based few-shot learning with dynamically scaled softmax loss[J]. IMAGE AND VISION COMPUTING,2023,140:15.
APA	Zhang, Yu,Zuo, Xin,Zheng, Xuxu,Gao, Xiaoyong,Wang, Bo,&Hu, Weiming.(2023).Improving metric-based few-shot learning with dynamically scaled softmax loss.IMAGE AND VISION COMPUTING,140,15.
MLA	Zhang, Yu,et al."Improving metric-based few-shot learning with dynamically scaled softmax loss".IMAGE AND VISION COMPUTING 140(2023):15.