[1]姜高霞,张　尧,王文剑.面向实例依赖标签噪声学习的动态混合噪声识别方法[J].郑州大学学报(工学版),2026,47(3):67-75.[doi:10.13705/j.issn.1671-6833.2025.06.009]
　JIANG Gaoxia,ZHANG Yao,WANG Wenjian.A Dynamic Mixture Noise Identification Method for Learning with Instance-dependent Label Noise[J].Journal of Zhengzhou University (Engineering Science),2026,47(3):67-75.[doi:10.13705/j.issn.1671-6833.2025.06.009]
点击复制

面向实例依赖标签噪声学习的动态混合噪声识别方法()
分享到：

《郑州大学学报(工学版)》[ISSN:1671-6833/CN:41-1339/T]

卷:
47
期数:
2026年3期
页码:
67-75
栏目:
出版日期:
2026-05-27

文章信息/Info

Title:
A Dynamic Mixture Noise Identification Method for Learning with Instance-dependent Label Noise
文章编号:
1671-6833(2026)03-0067-09
作者:
姜高霞1, 张　尧1, 王文剑1,2
1.山西大学 计算机与信息技术学院,山西 太原 030006;2.数据智能与认知计算山西省重点实验室,山西 太原 030006
Author(s):
JIANG Gaoxia1, ZHANG Yao1, WANG Wenjian1,2
1.School of Computer and Information Technology, Shanxi University, Taiyuan 030006, China; 2.Key Laboratory of Data Intelligence and Cognitive Computing of Shanxi Province, Taiyuan 030006, China
关键词:
实例依赖噪声;  标签噪声学习;  类重心;  动态混合模型;  半监督学习
Keywords:
instance-dependent noise;  learning with noisy label;  class centroid;  dynamic mixture model;  semi-supervised learning
分类号:
TP181:TP183
DOI:
10.13705/j.issn.1671-6833.2025.06.009
文献标志码:
A
摘要:
在实例依赖标签噪声IDN学习中,半监督方法能缓解噪声干扰并利用特征信息,但其效果依赖于准确的噪声识别,易受识别方法的影响。为解决噪声识别不准确的问题,设计了鲁棒特征重心以弱化不可靠数据的干扰,并提出了一种基于特征相似度的分布自适应动态混合模型DMM,通过提取特征相似度、结合高斯混合模型GMM与Beta混合模型BMM拟合分布并动态融合,实现更准确的噪声识别,最终结合半监督策略完成训练。在人工加噪的CIFAR-10/100数据集上,所提方法均达到了最优性能。在真实世界噪声数据集Animal-10N和Clothing1M上的最高分类准确率分别为84.21%和75.80%,优于现有代表性方法,验证了所提方法在实例依赖标签噪声学习任务中的有效性与适用性。
Abstract:
In learning with instance-dependent label noise (IDN), semi-supervised methods could mitigate noise interference and leverage feature information, but their effectiveness depended on accurate noise identification and was susceptible to the choice of recognition technique. To address this limitation, a robust feature-centroid mechanism was designed to weaken the influence of unreliable samples and a distribution-adaptive dynamic mixture model (DMM) was proposed based on feature similarity. Pairwise feature similarities was extracted, both Gaussian Mixture Models (GMM) and Beta Mixture Models (BMM) were used to fit these similarity distributions, and dynamically to fuse their outputs to achieve more accurate noise identification. A semi-supervised learning strategy was then integrated to complete the training process. On artificially corrupted CIFAR-10 and CIFAR-100 datasets, our method achieved state-of-the-art performance. On real-world noisy benchmarks Animal-10N and Clothing1M, it attained classification accuracies of 84.21% and 75.80%, respectively, outperforming representative existing approaches and demonstrating the effectiveness and applicability of our approach for IDN learning tasks.

参考文献/References:

[1]SONG H, KIM M, PARK D, et al. Learning from noisy labels with deep neural networks: a survey[J]. IEEE Transactions on Neural Networks and Learning Systems, 2023, 34(11): 8135-8153.

[2]林楠, 唐凯鹏, 牛勇鹏, 等. 基于双阶段特征提取网络的ECG降噪分类算法[J]. 郑州大学学报(工学版), 2024,45(5): 61-68.
LIN N, TANG K P, NIU Y P, et al. An ECG denoising and classification algorithm based on two-stage feature extraction network[J]. Journal of Zhengzhou University (Engineering Science), 2024, 45(5): 61-68.
[3]CHEN P F, YE J J, CHEN G Y, et al. Beyond classconditional assumption: a primary attempt to combat instance-dependent label noise[C]∥Proceedings of the AAAI Conference on Artificial Intelligence. Virtual Conference: AAAI Press, 2021: 11442-11450.
[4]LIU Y. Understanding instance-level label noise: disparate impacts and treatments[C]∥International Conference on Machine Learning. Vienna: PMLR, 2021: 6725-6735.
[5]BERTHON A, HAN B, NIU G, et al. Confidence scores make instance-dependent label-noise learning possible[EB/OL].(2021-02-22)[2025-03-01].https:∥arxiv.org/abs/2001.03772v2.
[6]LI J Z, SUN H L, LI J Y. Beyond confusion matrix: learning from multiple annotators with awareness of instance features[J]. Machine Learning, 2023, 112(3): 1053-1075.
[7]CHO Y, KIM W J, HONG S, et al. Part-based pseudo label refinement for unsupervised person re-identification[C]∥2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2022: 7298-7308.
[8]XIA X, LIU T, HAN B, et al. Part-dependent label noise: towards instance-dependent label noise[C]∥Advances in Neural Information Processing Systems. Virtual Conference: Curran Associates, Inc., 2020: 7597-7610.
[9]SHEN Y Q, XU L W, YANG Y Z, et al. Self-distillation from the last mini-batch for consistency regularization[C]∥2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway:IEEE, 2022: 11933-11942.
[10] SHU J, XIE Q, YI L, et al. Meta-weight-net: Learning an explicit mapping for sample weighting[EB/OL].(2019-12-08)[2025-03-01]. https:∥arxiv.org/pdf/1902.07379.
[11] ZHANG Y H, WANG C R, LING X, et al. Learn from all: erasing attention consistency for noisy label facial expression recognition[C]∥Computer Vision-ECCV 2022. Cham: Springer, 2022: 418-434.
[12] JIANG L, ZHOU Z Y, LEUNG T, et al. MentorNet: learning data-driven curriculum for very deep neural networks on corrupted labels[EB/OL]. (2018-08-13)[2025-03-01]. https:∥arxiv.org/abs/1712.05055v2.
[13] REN M Y, ZENG W Y, YANG B, et al. Learning to reweight examples for robust deep learning[EB/OL]. (2019-05-05)[2025-03-02]. https:∥arxiv. org/abs/1803.09050v3.
[14] HAN B, YAO Q M, YU X R, et al. Co-teaching: robust training of deep neural networks with extremely noisy labels[EB/OL].(2018-10-30)[2025-03-01]. https:∥arxiv.org/abs/1804.06872v3.
[15] YAO Q M, YANG H S, HAN B, et al. Searching to exploit memorization effect in learning fromcorrupted labels[EB/OL].(2020-09-18)[2025-03-01]. https:∥arxiv.org/abs/1911.02377v5.
[16] LI J, SOCHER R, HOI S C H. DivideMix: Learning with Noisy Labels as Semi-supervised Learning[EB/OL].(2020-02-18)[2025-03-01]. https:∥arxiv.org/abs/2002.07394.
[17] SONG H, KIM M, LEE J G. Selfie: Refurbishing unclean samples for robust deep learning[C]∥International Conference on Machine Learning. Long Beach: PMLR, 2019: 5907-5915.
[18] LYU Y M, TSANG I W. Curriculum loss: robust learning and generalization against label corruption[EB/OL].(2020-02-21)[2025-03-01]. https:∥arxiv. org/abs/1905.10045v3.
[19] LU Y, BO Y, HE W. Noise attention learning: Enhancing noise robustness by gradient scaling[C]∥ 36th Conference on Neural Information Processing Systems, NeurIPS 2022. New Orleans:CNIPS,2022, 35: 2316423177.
[20] XU Y L, CAO P, KONG Y Q, et al. L_DMI: a novel information-theoretic loss function for training deep nets robust to label noise[EB/OL].(2019-12-08)[202503-01]. https:∥specialsci. cn/detail/cf61b99e-91cf427a-84a1-064b84330db2? resourceType=0.
[21] ZHANG Z L, SABUNCU M R. Generalized cross entropy loss for training deep neural networks with noisy labels[C]∥Advances in Neural Information Processing Systems. Montréal: Curran Associates, Inc., 2018: 8792-8802.
[22] YU X, HAN B, YAO J, et al. How does disagreement help generalization against label corruption? [J] Statistics, 2019,2: 7164-7173.
[23]WEI H X, FENG L, CHEN X Y, et al. Combating noisy labels by agreement: a joint training method with co-regularization[C]∥2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2020: 13726-13735.
[24] LIU Y, GUO H Y. Peer loss functions: learning from noisy labels without knowing noise rates[EB/OL].(202008-14) [2025-03-10]. https:∥arxiv. org/abs/1910.03231v7.
[25] CHENG H, ZHU Z, LI X, et al. Learning with InstanceDependent Label Noise: A Sample Sieve Approach[EB/OL].(2021-03-22)[2025-03-01]. https:∥arxiv.org/abs/2010.02347.
[26] ZHU Z W, LIU T L, LIU Y. A second-order approach to learning with instance-dependent label noise[C]∥2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2021: 1011310123.
[27] ZHAO G L, LI G B, QIN Y P, et al. Centrality and consistency: two-stage clean samples identification for learning with instance-dependent noisy labels[C]∥17th European Conference on Computer Vision-ECCV 2022. Cham: Springer, 2022: 21-37.
[28] LIANG X, JI Y L, ZHENG W S, et al. SV-learner: support-vector contrastive learning for robust learning with noisy labels[J]. IEEE Transactions on Knowledge and Data Engineering, 2024, 36(10): 5409-5422.
[29] THULASIDASAN S, BHATTACHARYA T, BILMES J, et al. Combating label noise in deep learning using abstention[EB/OL]. (2019-08-01)[2025-02-10].https:∥arxiv.org/abs/1905.10964v2.
[30] XIAO T, XIA T, YANG Y, et al. Learning from massive noisy labeled data for image classification[C]∥2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway:IEEE, 2015: 2691-2699.
[31] ZHANG Y K, ZHENG S Z, WU P X, et al. Learning with feature-dependent label noise: a progressive approach[EB/OL] .(2021-03-27)[2025-03-01]. https:∥arxiv.org/abs/2103.07756v3.
[32] CHEN Y Y, SHEN X, HU S X, et al. Boosting coteaching with compression regularization for label noise[C]∥2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). Piscataway:IEEE, 2021: 2688-2692.

更新日期/Last Update: 2026-05-27