标签噪声鲁棒学习算法研究综述  

A Survey of Label Noise Robust Learning Algorithms

作　　者：宫辰 张闯 王启舟 Gong Chen;Zhang Chuang;Wang Qizhou(Key Lab of Intelligent Perception and Systems for High-Dimensional Information of Ministry of Education,School of Computer Science and Engineering,Nanjing University of Science and Technology,Nanjing 210094,China;Jiangsu Key Lab of Image and Video Understanding for Social Security,School of Computer Science and Engineering,Nanjing University of Science and Technology,Nanjing 210094,China)

机构地区：[1]南京理工大学计算机科学与工程学院高维信息智能感知与系统教育部重点实验室,南京210094 [2]南京理工大学计算机科学与工程学院江苏省社会安全图像与视频理解重点实验室,南京210094

出　　处：《航空兵器》2020年第3期20-26,共7页Aero Weaponry

基　　金：国家自然科学基金项目(61973162);江苏省自然科学基金项目(BK20171430)。

摘　　要：在机器学习领域,监督学习算法在理论层面和工程应用中均取得了丰硕的成果,但此类算法的效果严重依赖训练样本的标签质量,在实际问题中获取具有高质量标签的训练样本通常费时费力。为节省人力物力,网络爬虫、众包方法等替代方法被用于对训练数据的采集。不幸的是,这些替代方法获取的数据往往存在大量的错误标注,即标签噪声,由此带来了很多潜在的问题。因此,对标签噪声鲁棒学习算法的研究,在推广机器学习工程应用、降低机器学习算法部署成本方面具有重要的意义。本文对标签噪声鲁棒学习算法的最新研究成果进展进行了全面综述,分别从标签噪声的产生、影响、分类等方面进行了详细的总结,对每类标签噪声的处理方法进行了介绍,并对每类处理方法的优缺点进行分析。In the field of machine learning,supervised learning algorithm has achineved fruitful results both in theory and engineering application.However,such fully supervised learning algorithms are severely dependent on the label quality of the training sample,and reliably labeled data are often expensive and time consuming to obtain in real-world applications.Some surrogate approaches such as web crawler and crowd-sourcing methods,are widely used to collect training data.Unfortunately,there are usually lots of misannotations(i.e.label noise)in the data obtained by these surrogate methods,which result in many potential negative consequences.Therefore,the research on label noise robust learning algorithm is of great significance in promoting the application of machine learning engineering and reducing the deployment cost of machine learning algorithm.In this paper,the latest research progress of label noise robust learning algorithm is comprehensively reviewed.The generation,influence and classification of label noise are summarized in detail.The processing methods of each kind of label noise are introduced,and the advantages and disadvantages of each kind of processing methods are analyzed.

关 键 词：人工智能 机器学习 弱监督学习 标签噪声 深度学习 鲁棒学习算法 

分 类 号：TJ760[兵器科学与技术—武器系统与运用工程] TP18[自动化与计算机技术—控制理论与控制工程]