HyperAIHyperAI超神经
首页算力平台文档资讯论文教程数据集百科SOTALLM 模型天梯GPU 天梯顶会
全站搜索
关于
HyperAIHyperAI超神经
3 个月前

基于图模型的实例相关噪声标签学习

查看论文详情
Arpit GargCuong NguyenRafael FelixThanh-Toan DoGustavo Carneiro
基于图模型的实例相关噪声标签学习

摘要

在深度学习生态系统中，噪声标签不可避免且极具挑战性，因为模型极易对噪声标签产生过拟合。标签噪声存在多种类型，包括对称噪声、非对称噪声以及实例相关噪声（Instance-Dependent Noise, IDN），其中仅IDN依赖于图像本身的特征信息。由于标签错误在很大程度上源于图像中视觉类别信息的不足或模糊，IDN对图像信息的依赖性使其成为亟需深入研究的关键噪声类型。为有效应对IDN问题，本文提出一种新型图模型方法——InstanceGM，该方法融合了判别式模型与生成式模型的优势。InstanceGM的主要贡献包括：i）采用连续伯努利分布（continuous Bernoulli distribution）来训练生成式模型，显著提升了训练效率与稳定性；ii）引入当前最先进的噪声标签判别分类器，用于从实例相关噪声标签样本中生成干净标签。实验结果表明，InstanceGM在当前主流的噪声标签学习方法中具有较强竞争力，尤其在基于合成数据与真实世界数据的IDN基准测试中，多数实验场景下其分类准确率均优于现有对比方法。

代码仓库

arpit2412/InstanceGM
pytorch
GitHub 中提及

基准测试

基准方法指标
image-classification-on-clothing1mInstanceGM
Accuracy: 74.40%
image-classification-on-red-miniimagenet-20InstanceGM
Accuracy: 58.38
image-classification-on-red-miniimagenet-20InstanceGM-SS
Accuracy: 60.89
image-classification-on-red-miniimagenet-40InstanceGM-SS
Accuracy: 56.37
image-classification-on-red-miniimagenet-40InstanceGM
Accuracy: 52.24
image-classification-on-red-miniimagenet-60InstanceGM
Accuracy: 47.96
image-classification-on-red-miniimagenet-60InstanceGM-SS
Accuracy: 53.21
image-classification-on-red-miniimagenet-80InstanceGM
Accuracy: 39.62
image-classification-on-red-miniimagenet-80InstanceGM-SS
Accuracy: 44.03
learning-with-noisy-labels-on-animalInstanceGM
Accuracy: 84.6
ImageNet Pretrained: NO
Network: Vgg19-BN
learning-with-noisy-labels-on-animalInstanceGM with ConvNeXt
Accuracy: 84.7
ImageNet Pretrained: NO
Network: ConvNeXt
learning-with-noisy-labels-on-animalInstanceGM with ResNet
Accuracy: 82.3
ImageNet Pretrained: NO
Network: ResNet
learning-with-noisy-labels-on-cifar-10InstanceGM
Test Accuracy: 95.9
learning-with-noisy-labels-on-cifar-100InstanceGM
Test Accuracy: 77.19
learning-with-noisy-labels-on-redInstanceGM
Test Accuracy: 58.38
learning-with-noisy-labels-on-redInstanceGM-SS
Test Accuracy: 60.89
learning-with-noisy-labels-on-red-1InstanceGM-SS
Test Accuracy: 56.37
learning-with-noisy-labels-on-red-1InstanceGM
Test Accuracy: 52.24
learning-with-noisy-labels-on-red-2InstanceGM
Test Accuracy: 47.96
learning-with-noisy-labels-on-red-2InstanceGM-SS
Test Accuracy: 53.21
learning-with-noisy-labels-on-red-3InstanceGM-SS
Test Accuracy: 44.03
learning-with-noisy-labels-on-red-3InstanceGM
Test Accuracy: 39.62

用 AI 构建 AI

从想法到上线——通过免费 AI 协同编程、开箱即用的环境和市场最优价格的 GPU 加速您的 AI 开发

AI 协同编程
即用型 GPU
最优价格
立即开始

Hyper Newsletters

订阅我们的最新资讯
我们会在北京时间 每周一的上午九点 向您的邮箱投递本周内的最新更新
邮件发送服务由 MailChimp 提供
基于图模型的实例相关噪声标签学习 | 论文 | HyperAI超神经