摘要
大数据时代为机器智能化提供了更丰富的数据信息和更全面的学习依据。但由于获取精准的高质量数据标签信息需要花费大量的人力和物力,学习过程中能够高效利用的带标签数据依然很有限。大量的未标注样本或低质量的标注样本所包含的信息存在较大的不确定性。弱监督是指监督信息具有不确定性,弱监督学习则是监督信号的不确定性形式化后训练和推理的范式。该文从不确定性建模的视角分析了弱监督学习的问题描述和相应的解决方法,综述了部分弱监督学习范式与不确定性建模之间的关系。
In the era of big data,richer data information and more comprehensive learning basis has been provided for machine intelligence.However,the number of labeled instances that can be utilized effectively is still limited due to the case that acquiring high-quality labeled data needs tons of manpower and materials.As a result,there is a large number of unlabeled or low-quality labeled instances with quite a lot of uncertain information.Weak supervision means that the supervision information is uncertain,whereas weakly supervised learning is defined as a training and inference paradigm based on the formalization of the uncertainty that exists in supervision signals.This work analyzes the problem and the corresponding solutions of weakly supervised learning and summarizes the relationship between the paradigm of several kinds of weakly supervised learning methods and uncertainty quantification from the perspective of uncertainty modeling.
作者
周欣蕾
王熙照
ZHOU Xinlei;WANG Xizhao(Department of Computer Science and Software Engineering,Shenzhen University,Shenzhen 518060,China)
出处
《西北大学学报（自然科学版）》
CAS
CSCD
北大核心
2022年第5期813-823,共11页
Journal of Northwest University（Natural Science Edition）
基金
国家自然科学基金(61976141,61732011)。
关键词
监督信息
弱监督学习
不确定性建模
supervision
weakly supervised learning
uncertainty modeling
