文献类型：期刊文献

英文题名：Unsupervised few-shot image classification via one-vs-all contrastive learning

作者：Zheng, Zijun[1,2];Feng, Xiang[1,2];Yu, Huiqun[1,2];Li, Xiuquan[3];Gao, Mengqi[1,2]

机构：[1]East China Univ Sci & Technol, Dept Comp Sci & Engn, Shanghai 200237, Peoples R China;[2]Shanghai Engn Res Ctr Smart Energy, Shanghai 200237, Peoples R China;[3]Chinese Acad Sci & Technol Dev, Beijing 100038, Peoples R China

年份：2023

卷号：53

期号：7

起止页码：7833

外文期刊名：APPLIED INTELLIGENCE

收录：;EI(收录号：20223012405498);WOS:【SCI-EXPANDED(收录号:WOS:000828440700001)】；

基金：This work was supported in part by the Key Program of the National Natural Science Foundation of China under Grant No. 62136003, the National Natural Science Foundation of China under Grant Nos. 61772200 and 61772201, Shanghai Pujiang Talent Program under Grant No. 17PJ1401900, Shanghai Economic and Information Commission "Special Fund for Information Development" under Grant No. XX-XXFZ-02-20-2463.

语种：英文

外文关键词：Few-shot image classification; One-vs-all contrastive loss; Instance invariance; Neural projection

摘要：Human beings innately possess the ability to perceive novel concepts from only a few samples. As a setting to imitate the learned ability of human beings, few-shot image classification (FSIC) has recently aroused a research boom. FSIC aims to distinguish the novel class when given scarce samples. However, most of the existing few-shot methods build on the assumption that an adequate labeled dataset is provided in the source domain, ie, the base class dataset. Due to the expensive burden of labeled samples in the base class, this assumption may not be practical in a real-world application. To solve this labeled burden, in the paper, we propose a novel unsupervised few-shot image classification via One-vs-All contrastive learning. In this approach, to generate positive pairs in each instance, we first adopt a data augmentation technique to build the instance invariance. With the positive pairs, a neural projection with only a fully connected layer is then applied to maintain the structure consistent with the corresponding features of positive pairs. Finally, the One-Vs-All (OVA) contrastive learning is devised to pull one positive pair close while pushing all negative pairs away in a minibatch. By doing this OVA contrastive learning, the model can effectively acquire the discriminative feature and improve the generalization ability to recognize the novel class. We also further develop a theory of the generalized upper bound of the model for the OVA contrastive loss. Our experimental analyses suggest that the proposed approach achieves better performance compared to most existing few-shot methods, and various modules in the approach demonstrate their utility by conducting ablation studies.