Skip to content

Latest commit

 

History

History
9 lines (9 loc) · 2.55 KB

噪声标签.md

File metadata and controls

9 lines (9 loc) · 2.55 KB
名称 来源 说明 状态 备注
《ehensive Introduction to Label Noise》 ESANN 2014 梳理了三种噪声标签的处理手段(主要是面对分类问题,整体比较水):
3.1 Label Noise-Robust Models;
3.2 Data Cleansing Methods;
3.3 Label Noise-Tolerant Learning Algorithms(没太看懂);
NULL https://www.cs.bham.ac.uk/~axk/ESANNtut.pdf
《Classification in the Presence of Label Noise: a Survey》 IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS 2014 《ehensive Introduction to Label Noise》综述的同作者详细版本:
I. INTRODUCTION;
II. DEFINITION, SOURCES, AND TAXONOMY OF LABEL NOISE;
III. CONSEQUENCES OF LABEL NOISE ON LEARNING;
IV. METHODS TO DEAL WITH LABEL NOISE;
V. LABEL NOISE-ROBUST MODELS;
VI. DATA CLEANSING METHODS FOR LABEL NOISE-POLLUTED DATASETS;
VII. LABEL NOISE-TOLERANT LEARNING ALGORITHMS;
VIII. EXPERIMENTS IN THE PRESENCE OF LABEL NOISE
IX. CONCLUSION;
NULL https://romisatriawahono.net/lecture/rm/survey/machine%20learning/Frenay%20-%20Classification%20in%20the%20Presence%20of%20Label%20Noise%20-%202014.pdf
《Image Classification with Deep Learning in the Presence of Noisy Labels: A Survey》 arXiv 2020 侧重深度学习的噪声标签综述(和图像的关系好像并不大):
I. INTRODUCTION;
II. PRELIMINARIES;
III. NOISE MODEL BASED METHODS;
IV. NOISE MODEL FREE METHODS;
V. CONCLUSION;
NULL NULL
《Confident Learning: Estimating Uncertainty in Dataset Labels》 arXiv2019 cleanlab:
1 Count:估计噪声标签和真实标签的联合分布;
2 Clean:找出并过滤掉错误样本;
3 Re-Training:过滤错误样本后,重新训练(采用Co-Teaching,是因为剩余的训练数据中仍然存在噪声?);
NULL https://zhuanlan.zhihu.com/p/146557232
《Co-teaching: Robust Training of Deep Neural Networks with Extremely Noisy Labels》 NIPS 2018 基于两个网络交互的噪声数据训练方法(cleanlab使用了):
1 创建两个网络;
2 每个网络输入所有数据并选择一些可能干净的标签数据;
3 两个网络相互通信这个批量中应该用于训练的数据;
4 每个网络反向传播另一个网络挑选的数据并更新自己;
NULL https://zhuanlan.zhihu.com/p/65632321
《Learning with Noisy Label-深度学习廉价落地》 知乎 噪声标签综述&资料汇总;
涉及多篇文章,没仔细细看;
NULL NULL
NULL NULL NULL NULL NULL