近年來，深度學習在圖像分類任務中取得了出色的表現，尤其是在使用大型資料集的情況下。但是這些大型數據同時也必須包含高質量的標註，使得這些資料集很昂貴。因此另外還存在一種替代的廉價資料集，這些資料集的來源是從索引擎、社交媒體網站的圖片並且根據周圍的文字來當作標註、或是每份資料僅被單一個人標註。這些數據集通常包含不准確的標籤，這會導致當我們訓練模型時，模型過度擬合那些不正確的標籤使得模型在預測測試資料時準確度降低。
為了克服這個問題，我們提出了基於集成學習的標籤校正模型訓練演算法。與現在直接使用一個函數增加每個子模型之間預測差異性的方法不同，我們提出了一種基於資料流型局部結構的資料分割法，將訓練數據集分為多個不相交的子集來訓練每個子模型。每個子集都包含部分幾個局部流形，這些流形可能是乾淨或是不正確標註的資料。與直接使用完整的資料集訓練相比，每個子集僅包含一部分不正確標註的資料，因此對於每個子模型來說，只會擬合一部份不正確標註的資料，不會受到其他不正確標註的資料的影響。因此我們可以使用集成學習法來過濾掉那些有不正確標註資料的流形，因為每個不正確標註資料的流形只會被單個子模型學習，當我們利用多數決來修正有標籤時，可以將其過濾掉。在實驗方面，我們使用了現實世界的錯誤標籤資料集進行實驗。結果顯示與以前的研究相比，我們的方法具有一定的優勢。

Recently, deep learning has achieved great performance in image classifications tasks, especially with large-scale dataset. However, it also depends on high-quality annotations, which is time-consuming and expensive to manual label it. Thus, there exist alternative inexpensive dataset like search engine, social media with surrounding text or tags as label, or single annotator to each data. These dataset usually contain inaccurate and noisy labels, which will cause low-performance because model overfit to those noisy labels.
To overcome this problem, we propose ensemble learning based training algorithm with label correction. Different from prior works which directly add a loss function to increase the variance of predictions between each sub-network, we propose a local structure based data manifold splitting method to separate training dataset into several disjoint subsets to train each sub-network. And each subset contains several local manifolds, which may be clean or noisy. Compared to full training dataset, each subset only contain part of noisy manifold, so each sub-network only can overfit on those incomplete noisy manifold. And we can use ensemble leaning to filter out those noisy manifold because each noisy manifold only learned by single sub- network, which can be ignored when we correct noisy label with label propagation and majority decision. We conduct extensive experiments on real-world noisy label dataset. The results show that the advantage of our method compared to previous works.

摘要 i
Abstract ii
Content iii
Chapter 1 Introduction 1
Chapter 2 Related Work 5
2.1 Noise Robust Loss 5
2.2 Label Correction 6
2.3 Ensemble learning 7
2.4 Label Propagation 8
Chapter 3 Proposed Method 9
3.1 Overview 9
3.2 Local Structure based Data Manifold Splitting 11
3.3 Multi-Graph and Label Propagation 13
3.4 Label Correction with Ensemble Learning Method 16
3.5 Loss Function 16
3.6 Iterative Training 18
Chapter 4 Experiments 19
4.1 Datasets 19
4.2 Experimental Setup 19
4.3 Clohting1M 20
4.4 Food 101-N 21
4.5 Ablation Study 23
Chapter 5 Conclusion 29
Reference 30

[1] T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollar, and C. L. Zitnick. Microsoft coco:Common objects in context. In European conference on computer vision, pages 740{755. Springer, 2014.
[2] DENG, Jia, et al. Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. Ieee, 2009. p. 248-255.
[3] Benoˆıt Frenay and Michel Verleysen. Classification in the ´ presence of label noise: A survey. IEEE Transactions on Neural Networks and Learning Systems, 25(5):845–869, 2014
[4] Sainbayar Sukhbaatar, Joan Bruna, Manohar Paluri, Lubomir Bourdev, and Rob Fergus. Training convolutional networks with noisy labels. In ICLR Workshop, 2015.
[5] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning requires rethinking generalization. In ICLR, 2017.
[6] LECUN, Yann, et al. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 1998, 86.11: 2278-2324.
[7] KRIZHEVSKY, Alex, et al. Learning multiple layers of features from tiny images. 2009.
[8] PATRINI, Giorgio, et al. Making deep neural networks robust to label noise: A loss correction approach. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2017. p. 1944-1952.
[9] Dan Hendrycks, Mantas Mazeika, Duncan Wilson, and Kevin Gimpel. Using trusted data to train deep networks on labels corrupted by severe noise. In Advances in Neural Information Processing Systems, pages 10477–10486, 2018.
[10] XIA, Xiaobo, et al. Are Anchor Points Really Indispensable in Label-Noise Learning?. In: Advances in Neural Information Processing Systems. 2019. p. 6838-6849.
[11] SHARMA, Karishma, et al. NoiseRank: Unsupervised Label Noise Reduction with Dependence Models. arXiv preprint arXiv:2003.06729, 2020.
[12] HAN, Jiangfan; LUO, Ping; WANG, Xiaogang. Deep self-learning from noisy labels. In: Proceedings of the IEEE International Conference on Computer Vision. 2019. p. 5138-5147.
[13] Daiki Tanaka, Daiki Ikami, Toshihiko Yamasaki, and Kiyoharu Aizawa. Joint optimization framework for learning with noisy labels. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5552– 5560, 2018.
[14] Zhilu Zhang and Mert Sabuncu. Generalized cross entropy loss for training deep neural networks with noisy labels. In Advances in Neural Information Processing Systems, pages 8792–8802, 2018.
[15] WEI, Hongxin, et al. Combating noisy labels by agreement: A joint training method with co-regularization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020. p. 13726-13735.
[16] LAINE, Samuli; AILA, Timo. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016.
[17] TARVAINEN, Antti; VALPOLA, Harri. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In: Advances in neural information processing systems. 2017. p. 1195-1204.
[18] LI, Junnan, et al. Learning to learn from noisy labeled data. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2019. p. 5051-5059.
[19] LI, Junnan; SOCHER, Richard; HOI, Steven CH. Dividemix: Learning with noisy labels as semi-supervised learning. arXiv preprint arXiv:2002.07394, 2020.
[20] SONG, Hwanjun, et al. Learning from Noisy Labels with Deep Neural Networks: A Survey. arXiv preprint arXiv:2007.08199, 2020.
[21] R. Wang, T. Liu, and D. Tao, “Multiclass learning with partially corrupted labels,” IEEE Transactions on Neural Networks and Learning Systems, vol. 29, no. 6, pp. 2568–2580, 2017
[22] Florian Schroff, Dmitry Kalenichenko, and James Philbin, “Facenet: A unified embedding for face recognition and clustering,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2015, pp. 815–823.
[23] E. Arazo, D. Ortego, P. Albert, N. E. O’Connor, and K. McGuinness, “Unsupervised label noise modeling and loss correction,” in Proc. ICML, 2019.
[24] NGUYEN, Duc Tam, et al. Self: Learning to filter noisy labels with self-ensembling. arXiv preprint arXiv:1910.01842, 2019.
[25] A. Ghosh, H. Kumar, and P. Sastry, “Robust loss functions under label noise for deep neural networks,” in Proc. AAAI, 2017.
[26] Z. Zhang and M. Sabuncu, “Generalized cross entropy loss for training deep neural networks with noisy labels,” in Proc. NeurIPS, 2018, pp. 8778–8788.
[27] Y. Wang, X. Ma, Z. Chen, Y. Luo, J. Yi, and J. Bailey, “Symmetric cross entropy for robust learning with noisy labels,” in Proc. ICCV, 2019, pp. 322–330.
[28] Y. Yan, Z. Xu, I. W. Tsang, G. Long, and Y. Yang, “Robust semisupervised learning through label aggregation,” in Proc. AAAI, 2016.
[29] ZHOU, Dengyong, et al. Learning with local and global consistency. In: Advances in neural information processing systems. 2004. p. 321-328.
[30] ISCEN, Ahmet, et al. Label propagation for deep semi-supervised learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. 2019. p. 5070-5079.
[31] Matthijs Douze, Arthur Szlam, Bharath Hariharan, and Herve J ´ egou. Low-shot learning with large-scale diffusion. ´ In CVPR, 2018.
[32] Aykut Erdem and Marcello Pelillo. Graph transduction as a noncooperative game. Neural Computation, 24, 2012.
[33] BONTONOU, Myriam, et al. Introducing graph smoothness loss for training deep learning architectures. arXiv preprint arXiv:1905.00301, 2019.
[34] Tong Xiao, Tian Xia, Yi Yang, Chang Huang, and Xiaogang Wang. Learning from massive noisy labeled data for image classification. In CVPR, 2015
[35] Kuang-Huei Lee, Xiaodong He, Lei Zhang, and Linjun Yang. Cleannet: Transfer learning for scalable image classfier training with label noise. In CVPR, 2017
[36] Lukas Bossard, Matthieu Guillaumin, and Luc J. Van Gool. Food-101 - mining discriminative components with random forests. In ECCV, 2014.
[37] HE, Kaiming, et al. Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. 2016. p. 770-778.
[38] Florian Schroff, Dmitry Kalenichenko, and James Philbin, “Facenet: A unified embedding for face recognition and clustering,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2015, pp. 815–823.
[39] LIU, Sheng, et al. Early-Learning Regularization Prevents Memorization of Noisy Labels. arXiv preprint arXiv:2007.00151, 2020.
[40] ZHANG, Zizhao, et al. Distilling Effective Supervision from Severe Label Noise. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020. p. 9294-9303.
[41] Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. Making deep neural networks robust to label noise: A loss correction approach. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1944–1952, 2017.
[42] ZHANG, Weihe; WANG, Yali; QIAO, Yu. Metacleaner: Learning to hallucinate clean representations for noisy-labeled visual recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2019. p. 7373-7382.