@huanghaian 2020-06-17T06:31:31.000000Z 字数 2085 阅读 1103

# 噪声label学习

分类

https://github.com/subeeshvasu/Awesome-Learning-with-Label-Noise

https://zhuanlan.zhihu.com/p/110959020

## aum

arxiv:2001.10528

(1) 提出了aum分值 Area Under the Margin来区分正确和错误标注样本
(2) 由于aum分值是和数据集相关的，阈值切分比较关键，故作者提出指示样本indicator samples的概念来自动找到最合适的aum阈值

Not all data in a typical training set help with generalization

AUM的定义非常简单：

momentum, a learning rate of 0.1, and a batch size of 256。The learning rate is dropped by a factor of 10 at epochs 150 and 225.

When computing the AUM to identify mislabeled data, we train these models up until the first learning rate drop (150 epochs)，并且把batch size变成64，主要是增加sgd训练时候的方差，对统计aum有好处。在移除错误样本后，再采用上述策略进行训练。

## Learning Adaptive Loss for Robust Learning with Noisy Labels

arxiv：2002.06482

(1) ce loss

(2) Generalized Cross Entropy (GCE)

(3) Symmetric Cross Entropy (SL)

(4) Bi-Tempered logistic Loss (Bi-Tempered)

(5) Polynomial Soft Weighting loss (PolySoft)

## 论文

DivideMix: Learning with Noisy Labels as Semi-supervised Learning
Simple and Effective Regularization Methods for Training on Noisily Labeled Data with Generalization Guarantee
SELF: Learning to Filter Noisy Labels with Self-Ensembling

• 私有
• 公开
• 删除