多示例学习算法综述  

Review of Multi-instance Learning Algorithms

作　　者：赵璐 袁立明 郝琨 ZHAO Lu;YUAN Li-ming;HAO Kun(School of Computer and Information Engineering,Tianjin Chengjian University,Tianjin 300384,China;School of Computer Science and Engineering,Tianjin University of Technology,Tianjin 300384,China)

机构地区：[1]天津城建大学计算机与信息工程学院,天津300384 [2]天津理工大学计算机科学与工程学院,天津300384

出　　处：《计算机科学》2022年第S01期93-99,共7页Computer Science

基　　金：天津市科学技术局技术创新引导专项(21YDTPJC00250);国家自然科学基金(61902273);天津市教委社会科学重大项目(2019JWZD02);天津市新一代人工智能科技重大专项(18ZXZNGX00150);计算机视觉与系统教育部重点实验室开放基金(TJUT-CVS20170001)。

摘　　要：多示例学习是一种典型的弱监督学习框架,其样本示例包是一种集合类型数据,学习过程只需要包的粗粒度类别标记,能较好适应较难获得细粒度标记的应用问题。随着近几年深度学习的快速发展,深度多示例学习逐渐引起了研究者的兴趣。对多示例学习算法的研究进展进行综述,首先依据算法的层次结构将其划分为浅层模型和深度模型;然后对两类模型的相关算法进行回顾和总结,重点分析深度多示例学习模型在池化方式上的差别,并阐述以集合型数据为训练样本的模型所需满足的对称函数基本定理及其在深度多示例学习中的应用;最后通过实验对比分析不同算法的性能,且着重剖析其可解释性,并指明未来有待深入研究的问题。Multi-instance learning(MIL)is a typical weakly supervised learning framework,where every training example,called bag,is a set of instances.Since the learning process of an MIL algorithm depends on only the labels of bags rather than those of any individual instances,MIL can fit well with applications in which instance labels are difficult to get.Recently,deep multi-instance learning methods attract widespread attention,so deep MIL has become a major research focus.This paper reviews some research progress of MIL.Firstly,MIL algorithms are divided into shallow and deep models according to their hierarchical structure.Secondly,various algorithms are reviewed and summarized in these two categories,and then different pooling methods of deep MIL models are analyzed.Moreover,the fundamental theorem of symmetric functions for models with set-type data as training samples and its application in deep MIL are expounded.Finally,the performance of different algorithms is compared and analyzed through experiments,and their interpretability is analyzed thoroughly.After that,problems to be further investigated are discussed.

关 键 词：机器学习 多示例学习 深度学习 多示例池化 对称函数基本定理 可解释性 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]