[论文阅读笔记58]Learning from Noisy Labels with Deep Neural Networks：A Survey

1.题目

Learning from Noisy Labels with Deep Neural Networks: A Survey
作者团队：韩国科学技术院（KAIST）
Song H , Kim M , Park D , et al. Learning from Noisy Labels with Deep Neural Networks: A Survey. 2020.

2. 摘要

1. 重述问题：从监督学习的角度来描述使用标签噪声学习的问题；

2. 方法回顾：对57种最先进的鲁棒训练方法进行了全面的回顾，并根据其方法差异分为5组；然后系统地比较了六种属性来评估它们的优越性；

3. 评估：对噪声率估计进行了深入的分析，并总结了典型使用的评估方法，包括公共噪声数据集和评估指标；

4. 总结：我们提出了几个很有前途的研究方向。

3. 研究Noise Learn的意义

对比了三种情况，第一种是clean数据的训练情况；第二种是noise数据没有加入Reg(正则化)的情况；第三种是noise数据加入了Reg的情况；一般情况下我们想方设法去处理正则的内容（data augmentation，weight decay，dropout， batch normalization），可是noise的影响也是很大的，像图中的Gap.

4. 目前的相关综述

Frenay and Verleysen [12] —Classifification in the presence of label noise: A survey-2013–经典的监督学习（说明了noise的定义，来源等等相关内容。Bayes,SVM）—讲述了noise学习的统计学习；

Zhang et al. [27]—Learning from crowdsourced labeled data: A survey–2016-- 讨论众包数据方法（expectation maximization (EM) algorithms）—其实这是不错的一篇综述，工程上挺有用的，众包情景就是一个好的情境。

这篇论文可以结合弱监督的论文来看，特征在后面的推理Ground truth那里。

工具：The project CEKA is available at: http://ceka.sourceforge.net/

Nigam et al.[28] -Impact of noisy labels in learning techniques: A survey- 2020-- 局限在 the loss function and sample selection 两方面

Hanet al. [29] --A survey of label-noise representation learning: Past,present and future-- 2020-- 总结了带有噪声标签的robust学习的基本组成部分，但它们的分类与我们的哲学分类完全不同；从机器学习的定义出去去讲述Noise学习的问题，对于理解机器学习有很帮助的一篇文章，之前也看到篇相关的，忘记哪篇了，也是从机器学习定义出发去讲述一件事情；

1. 给出LNRL的定义；
2. 从学习理论的视角对Noise训练有更深一层的理解；
3. 从数据、目标，优化算子的角度进行了分类；并分析了各类的优缺点；
4. 提出了新的研究方法；
5. https://github.com/bhanML/label-noise-papers

这篇文章写得也比较清淅的：

第1节，引言；写作动机与贡献，文章组织情况；

第2节写了Lable-Noise学习的相关文献，完整版本见附录1(早期（1988开始）–Emerging Stage（2015）-- Flourished Stage（2019）)；

第3节综述的概述，包括LNRL的正式定义、核心问题，以及根据数据、目标和优化对现有工作的分类；

第4节针对利用噪声转换矩阵来求解LNRL的方法；

第5节是关于修改目标函数以使LNRL可行的方法；

第6节是关于利用深度网络的特性来解决LNRL问题的方法；

在第7节中，我们提出了LNRL的未来发展方向。除了LNRL之外，该调查还揭示了几个很有前途的未来方向；

在第8节，总结；

对于数据，主要是一个Noise transition matrix T，T提示了clean标注与noise标注的关系；使用三种方法去使用T来处理Noise标注；

5. 预备知识

这篇综述主要是在系统的方法论上，[29]关注的是一般的视角上（ input data, objective functions, optimization policies）；

这个综述对存在的robust训练方法作了一个对比；

5.0 lable-noise的监督学习；

5.1 标签噪声分类

​ a.独立于实例的标签噪声;

​ b.依赖实例的标签噪声;

5.2 非深度学习方法—分成四类

​ a. 数据清洗；

​ b. Surrogate Loss（代理损失函数）

​ c. 概率方法

​ d. 基于模型方法

5.3 理论基础

​ a.Label Transition: 从数据的角度来看，noise是来源于label的转移矩阵；这个转移矩阵可以发现其中的内在关系；

​ b. Risk Minimization

​ c. Memorization Effect

5.4 Regression with Noisy Labels

6. 深度学习方法

深度学习的robust训练（分为5类）：

它的关注点是深度学习在监督学习过程中更robust.

(P1) Flexibility，(P2) No Pre-training，(P3) Full Exploration，(P4) No Supervision，(P5) Heavy Noise，(P6) Complex Noise

圆圈：完全支持，叉：不支持，三角：支持但不完全支持

6.1 Robust框架

在DNN上增加一个Noise适应层去学习label transition，或开了一个专用架构来处理；

6.1.A Noise Adaptation Layer

​

这个方法的原理：

论文：Training deep neural-networks using a noise adaptation layer,” in Proc. ICLR, 2017.

这论文采用了EM算法来处理，理论性学是比较强的。

A.1. Noise Adaptation Layer

Year	Venue	Title	Implementation
2015	ICCV	Webly supervised learning of convolutional networks	Official (Caffe)
2015	ICLRW	Training convolutional networks with noisy labels	Unofficial (Keras)
2016	ICDM	Learning deep networks from noisy labels with dropout regularization	Official (MATLAB)
2016	ICASSP	Training deep neural-networks based on unreliable labels	Unofficial (Chainer)
2017	ICLR	Training deep neural-networks using a noise adaptation layer	Official (Keras)

A.2. Dedicated Architecture（专门架构）

Year	Venue	Title	Implementation
2015	CVPR	Learning from massive noisy labeled data for image classification	Official (Caffe)	管理了两个独立的网络
2018	NeurIPS	Masking: A new perspective of noisy supervision	Official (TensorFlow)	人工辅助的方法
2018	TIP	Deep learning from noisy image labels with quality embedding	N/A
2019	ICML	Robust inference via generative classifiers for handling noisy labels	Official (PyTorch)

6.2 Robust正则化

B.1. Explicit Regularization

Year	Venue	Title	Implementation
2018	ECCV	Deep bilevel learning	Official (TensorFlow)
2019	CVPR	Learning from noisy labels by regularized estimation of annotator confusion	Official (TensorFlow)
2019	ICML	Using pre-training can improve model robustness and uncertainty	Official (PyTorch)
2020	ICLR	Can gradient clipping mitigate label noise?	Unofficial (PyTorch)
2020	ICLR	Wasserstein adversarial regularization (WAR) on label noise	N/A
2021	ICLR	Robust early-learning: Hindering the memorization of noisy labels	Official (PyTorch)

B.2. Implicit Regularization

Year	Venue	Title	Implementation
2015	ICLR	Explaining and harnessing adversarial examples	Unofficial (PyTorch)
2017	ICLRW	Regularizing neural networks by penalizing confident output distributions	Unofficial (PyTorch)
2018	ICLR	Mixup: Beyond empirical risk minimization	Official (PyTorch)

C. Robust Loss Function

Year	Venue	Title	Implementation
2017	AAAI	Robust loss functions under label noise for deep neural networks	N/A
2017	ICCV	Symmetric cross entropy for robust learning with noisy labels	Official (Keras)
2018	NeurIPS	Generalized cross entropy loss for training deep neural networks with noisy labels	Unofficial (PyTorch)
2020	ICLR	Curriculum loss: Robust learning and generalization against label corruption	N/A
2020	ICML	Normalized loss functions for deep learning with noisy labels	Official (PyTorch)
2020	ICML	Peer loss functions: Learning from noisy labels without knowing noise rates	Official (PyTorch)

6.D 损失函数调整

改进损失函数；

D.1. Loss Correction

Year	Venue	Title	Implementation
2017	CVPR	Making deep neural networks robust to label noise: A loss correction approach	Official (Keras)
2018	NeurIPS	Using trusted data to train deep networks on labels corrupted by severe noise	Official (PyTorch)
2019	NeurIPS	Are anchor points really indispensable in label-noise learning?	Official (PyTorch)
2020	NeurIPS	Dual T: Reducing estimation error for transition matrix in label-noise learning	N/A

D.2. Loss Reweighting

Year	Venue	Title	Implementation
2017	TNNLS	Multiclass learning with partially corrupted labels	Unofficial (PyTorch)
2017	NeurIPS	Active Bias: Training more accurate neural networks by emphasizing high variance samples	Unofficial (TensorFlow)

D.3. Label Refurbishment

Year	Venue	Title	Implementation
2015	ICLR	Training deep neural networks on noisy labels with bootstrapping	Unofficial (Keras)
2018	ICML	Dimensionality-driven learning with noisy labels	Official (Keras)
2019	ICML	Unsupervised label noise modeling and loss correction	Official (PyTorch)
2020	NeurIPS	Self-adaptive training: beyond empirical risk minimization	Official (PyTorch)
2020	ICML	Error-bounded correction of noisy labels	Official (PyTorch)
2021	AAAI	Beyond class-conditional assumption: A primary attempt to combat instancedependent label noise	Official (PyTorch)

D.4. Meta Learning

Year	Venue	Title	Implementation
2017	NeurIPSW	Learning to learn from weak supervision by full supervision	Unofficial (TensorFlow)
2017	ICCV	Learning from noisy labels with distillation	N/A
2018	ICML	Learning to reweight examples for robust deep learning	Official (TensorFlow)
2019	NeurIPS	Meta-Weight-Net: Learning an explicit mapping for sample weighting	Official (PyTorch)
2020	CVPR	Distilling effective supervision from severe label noise	Official (TensorFlow)
2021	AAAI	Meta label correction for noisy label learning	Official (PyTorch)

6.4 样本选择

通过多网络或多轮学习，从有噪声的训练数据中识别true-labeled的样本。

E.1. Multi-network Learning – 多网络学习

Year	Venue	Title	Implementation
2017	NeurIPS	Decoupling when to update from how to update	Official (TensorFlow)
2018	ICML	MentorNet: Learning data-driven curriculum for very deep neural networks on corrupted labels	Official (TensorFlow)
2018	NeurIPS	Co-teaching: Robust training of deep neural networks with extremely noisy labels	Official (PyTorch)
2019	ICML	How does disagreement help generalization against label corruption?	Official (PyTorch)

E.2. Multi-round Learning–多轮学习方法

Year	Venue	Title	Implementation
2018	CVPR	Iterative learning with open-set noisy labels	Official (Keras)
2019	ICML	Learning with bad training data via iterative trimmed loss minimization	Official (GluonCV)
2019	ICML	Understanding and utilizing deep neural networks trained with noisy labels	Official (Keras)
2019	ICCV	O2U-Net: A simple noisy label detection approach for deep neural networks	Unofficial (PyTorch)
2020	ICMLW	How does early stopping can help generalization against label noise?	Official (Tensorflow)
2020	NeurIPS	A topological filter for learning with label noise	Official (PyTorch)

E.3. Hybrid Learning

Year	Venue	Title	Implementation
2019	ICML	SELFIE: Refurbishing unclean samples for robust deep learning	Official (TensorFlow)
2020	ICLR	SELF: Learning to filter noisy labels with self-ensembling	N/A
2020	ICLR	DivideMix: Learning with noisy labels as semi-supervised learning	Official (PyTorch)
2021	ICLR	Robust curriculum learning: from clean label detection to noisy label self-correction	N/A

7. 数据集

8. 总结

其实弱监督学习，noise学习，主动学习，出发点都是想去解决语料的问题。弱监督是想在没有标准的数据上进行自动标注，然后对这些标注进行软合并；noise学习，解决标注出来数据的noise问题；主动学习，就是用机器到已标注的数据进行学习，对未标注的样本进行估计，目前是想用直可能标注的样本数据来代替整个样本集的内容。

可是发现，很多领域都是在处理图像的，自然语言是否可以考虑？

9. 参考

论文：https://arxiv.org/pdf/2007.08199.pdf
相关资料： https://github.com/songhwanjun/Awesome-Noisy-Labels

附：

LNRL： Label-Noise Representation Learning

LNSL： label-noise statistical learning

surrogate loss function：代理损失函数或者称为替代损失函数，一般是指当目标函数非凸、不连续时，数学性质不好，优化起来比较复杂，这时候需要使用其他的性能较好的函数进行替换。

ICCV 的全称是 IEEE International Conference on Computer Vision，即国际计算机视觉大会

ICDM（国际数据挖掘会议）
IEEE国际声学、语言和信号处理会议(ICASSP)
国际学习表征会议(International Conference On Learning Representations)