Dynamic Knowledge Path Learning for Few-Shot Learning

作     者：Jingzhu Li Zhe Yin Xu Yang Jianbin Jiao Ye Ding 

作者机构：School of Integrated CircuitsUniversity of Chinese Academy of SciencesBeijing 100049China School of Computer Science and TechnologyHarbin Institute of Technology(Shenzhen)Shenzhen 518055China School of ElectronicElectrical and Communication EngineeringUniversity of Chinese Academy of SciencesBeijing 100049China School of Computer Science and TechnologyDongguan University of TechnologyDongguan 523808China 

出 版 物：《Big Data Mining and Analytics》 (大数据挖掘与分析(英文))

年 卷 期：2025年第8卷第2期

页      面：479-495页

核心收录：

学科分类：12[管理学] 1201[管理学-管理科学与工程(可授管理学、工学学位)] 081104[工学-模式识别与智能系统] 08[工学] 0835[工学-软件工程] 0811[工学-控制科学与工程] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：supported in part by the National Natural Science Foundation of China(No.62227808) the Shenzhen Science and Technology Program(No.ZDSYS20210623091809029) 

主　　题：data mining few-shot learning image classification 

摘      要：Few-shot learning is a challenging task that aims to train a classifier with very limited training *** existing few-shot learning methods mainly focus on mining knowledge from limited training samples as much as possible and ignore the learning *** by human learning,people select useful knowledge and follow a learning path to enhance their learning *** this *** propose a novel few-shot learning model called dynamic knowledge path learning(DKPL)to guide the few-shot learning task to learn useful selected knowledge with suitable learning ***,we simultaneously consider the importance,direction,and diversity of knowledge and propose a dynamic path learning strategy in the dynamic path construction ***,we design a new learner to absorb knowledge,step by step,according to each class’s learning path in the knowledge path propagation *** far as we know,this is the first few-shot learning work to consider dynamic path learning to improve classification *** and visual case studies demonstrate the effectiveness and superiority of the DKPL model on four real-world image datasets.