ISHIK2023
Title：Few-Shot Learning: A Survey
発表者：张涛, 王圣杰, 梁秋金, 金渝皓, 王铎 （Tao ZHANG, Shengjie WANG, Qiujin LIANG, Yuhao JIN, Duo WANG）
所属：清华大学自动化系,北京100084, 中国 （Department of Automation, Tsinghua University, Beijing 100084, China）
キーワード：
Keywords：Deep learning, Few-shot learning, Metric learning, Image Classification, Object Detection
要旨：
Abstract：
Deep learning methods have achieved remarkable results in various fields such as image classification, object detection, feature recognition, and fault diagnosis. However, in real-world applications, researchers often face constraints in acquiring a large amount of sample data or encounter high costs associated with obtaining samples. As a result, exploring algorithms for learning in few-shot situations has become a prominent research focus. The objective of this survey is to summarize mainstream methods for few-shot learning and their effectiveness in typical application scenarios. Firstly, addressing the issue of overfitting, the advantages and disadvantages of different approaches are discussed, including metric-based methods, memory-based methods, parameter-updating methods, and data augmentation methods. Secondly, we explain the application of existing few-shot learning (FSL) methods in four typical domains: image classification, object detection, semantic segmentation, and fault diagnosis. Finally, the challenges faced by few-shot learning techniques and the future research trends in this field are discussed.