面向标签噪声的联合训练框架
A joint training framework for learning with noisy labels
魏琦 1孙皓亮 1马玉玲 2尹义龙1
作者信息
- 1. 山东大学软件学院,济南 250101
- 2. 山东建筑大学计算机科学与技术学院,济南 250101
- 折叠
摘要
当前面向标签噪声的鲁棒性学习通常依赖样本选择和标签修正两种策略,但是这两类方法均存在缺陷.基于样本选择的方法忽略了被过滤掉的样本中的有效信息,进而降低了模型的性能.基于标签修正的方法常使用自标签技术而引起模型的错误积累问题.对此,本文提出了一个集成样本选择、标签修正的联合训练框架.针对样本选择模块,本文设计了一种新的选择标准,通过在线选择的方法对所挑选的样本集合进行更新.相较于现有选择标准,本文提出的标准可保留更多边界样本,提升了模型对决策边界的学习性能,增强了模型的泛化性能.针对标签修正模块,本文提出了一种联合标签修正策略.相比于传统的自标签修正技术,该模块通过联合特征空间视角,对噪声样本进行多视角的标签修正,解决了传统自标签技术的错误累积问题.此外,本文引入对比学习正则化项,提升了标签修正效果和模型表征学习能力.本文方法在4个测试基准上取得了当前最好分类效果,验证了所提训练框架的有效性.
Abstract
Sample selection and label correction are two effective strategies for learning with noisy labels(LNL).Both of these strategies have limitations for LNL tasks.Sample selection strategies usually ignore discriminative information in discarded samples,thereby degrading the performance of the algorithm.Label correction strategies commonly leverage self-labeling techniques,resulting in notorious error accumulation.In this paper,we propose a new learning framework that combines sample selection with label correction.Specifically,a novel selection criterion is designed to update the selected set online.Compared with the existing criterion,our proposal retains more boundary samples for the decision and can improve the generalization ability of the learning algorithm.For the label correction phase,we designed a joint label correction function.Compared with the conventional self-label strategy,a multiview label correction is proposed to combine multiple feature space views,which can alleviate the effect of error accumulation.In addition,we propose a contrastive learning regularization term to enhance the learning of modified label quality and model representation from the feature perspective.Our framework achieves the new state of the art on four challenging benchmarks,demonstrating its great effectiveness in LNL.
关键词
标签噪声学习/样本选择/标签修正/对比学习Key words
learning with noisy labels/sample selection/label correction/contrastive learning引用本文复制引用
基金项目
国家自然科学基金(62106129)
国家自然科学基金(62176139)
国家自然科学基金(62177031)
山东省自然科学基金(ZR2021QF053)
山东省自然科学基金(ZR2021ZD15)
中国博士后科学基金(2021TQ0195)
中国博士后科学基金(2021M701984)
出版年
2024
NETL
NSTL
万方数据