当前位置:网站首页>CVPR 2022 | 应对噪声标签,西安大略大学、字节跳动等提出对比正则化方法
CVPR 2022 | 应对噪声标签,西安大略大学、字节跳动等提出对比正则化方法
2022-06-12 21:28:00 【智源社区】
噪声标签(Noisy labels)随着深度学习研究的深入得到广泛的关注,因为在众多实际落地的场景模型的训练都离不开真实可靠的标签信息。由于人工标注误差(专业性不足等问题)、数据原始噪声,带噪声的数据不可避免,清洗数据的工作也是更加困难。
在有监督的图像分类问题中,经典的 cross-entropy (CE) 损失函数是最为广泛应用的函数之一。当数据集不存在任何的噪声标签的时候,它往往能带来非常不错的效果。然而,当数据集中存在噪声标签的时候,它会导致模型对噪声标签过拟合,使模型的泛化性变差。本文从对比学习的角度研究了如何通过约束图像的特征来防止模型对噪声标签的过拟合。
现有的解决噪声标签的问题有基于 robust regularization, label correction, loss reweighting, 和 robust loss functions 等。本文的研究动机源于 robust loss functions。为了防止模型对噪声标签过拟合,现有的对噪声鲁棒的损失函数(mean absolute error (MAE)[1,2],reverse cross-entropy loss (RCE) [3] 等)在一定程度上解决了噪声标签过拟合的同时,也存在对数据欠拟合的问题[4,5]。在实际应用中,这些对噪声鲁棒的损失函数是结合 CE 一起使用的,而 CE 容易造成对噪声数据的过拟合问题。因此我们思考,能否仅仅通过约束图像的特征,使整个模型仍然可以用 CE 训练且不受噪声标签的影响。
近期,来自西安大略大学,纽约大学以及字节跳动的学者们研究了如何从带有噪声标签的数据集中学到可靠的模型,这一基础且重要的问题。本工作主要由西安大略大学统计及精算系的易立完成,通讯作者为西安大略大学计算机系的助理教授王博予。

论文地址:https://arxiv.org/abs/2203.01785
本文主要回答了两个问题:(1)基于对比学习得到的图像特征能给在标签噪声中学习带来什么好处; (2)如何从噪声数据中学到基于对比学习的图像特征。在之后的实验部分,我们也展示了此方法可以和现有的带噪学习的方法相结合,能进一步提升模型的表现。这项研究已被 CVPR2022 接收。
边栏推荐
- Experiment 7-2-6 print Yanghui triangle (20 points)
- Vs2017 environmental issues
- 初步了解認識正則錶達式(Regex)
- USB mechanical keyboard changed to Bluetooth Keyboard
- 字符串基础知识
- #141 Linked List Cycle
- Lua pattern matching
- Digital intelligence data depth | Bi goes down the altar? It's not that the market has declined, it's that the story has changed
- 同花顺能开户吗,在APP上可以直接开通券商安全吗
- ASCII code comparison table
猜你喜欢

Solution of multi machine room dynamic loop status network touch screen monitoring

GPU giant NVIDIA suffered a "devastating" network attack, and the number one malware shut down its botnet infrastructure | global network security hotspot on February 28

Lombok package is successfully installed, but the runtime prompts that get, set method and constructor solution cannot be found

Sorting out the knowledge points of primary and secondary indicators

Integrated monitoring solution for power environment of small and medium-sized computer rooms

Lake shore PT-100 platinum resistance temperature sensor

Data visualization - broken line area chart

#141 Linked List Cycle

GNS安装与配置

leetcode:210. 課程錶 II
随机推荐
How to design a message box through draftjs
MySql主从复制
Risk control modeling X: Discussion on problems existing in traditional modeling methods and Exploration on improvement methods
Composer version degradation
Digital intelligence data depth | Bi goes down the altar? It's not that the market has declined, it's that the story has changed
金融信创爆发年!袋鼠云数栈DTinsight全线产品通过信通院信创专项测试
Research Report on market supply and demand and strategy of hydraulic operating table industry in China
Product Manager: "click here to jump to any page I want to jump" -- decoupling efficiency improving artifact "unified hop routing"
Draw according to weight
CUDA out of memory
Lua pattern matching
zgc的垃圾收集的主要階段
阅读笔记 Deep Hough Voting for 3D Object Detection in Point Clouds
Graphics2d class basic use
linux备份mysql
drf 接收嵌套数据并创建对象, 解决:drf NOT NULL constraint failed
remote: Support for password authentication was removed on August 13, 2021
Yanghui triangle code implementation
Common error in script execution: build sh: caller: not found
选择排序