论文阅读笔记 Improving Robustness of DNNs against Common Corruptions via Gaussian Adversarial Training

最新推荐文章于 2022-03-09 10:42:44 发布
最新推荐文章于 2022-03-09 10:42:44 发布
阅读量227 收藏
点赞数
分类专栏： 深度学习
版权声明：本文为博主原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接和本声明。
本文链接：https://blog.csdn.net/qq_28234239/article/details/111636399
版权

深度学习 模型健壮性 高斯对抗训练 鞍点方法 鲁棒性
关键词由CSDN通过智能技术生成

摘要

深度神经网络在图像分类上表现的很好，但是测试数据稍有不同的时候，分类会有很大的下降。为了解决这个问题，通过高斯对抗训练，文章提出了鞍点方法来提高模型的健壮性。通过对抗的样本来训练，这些对抗样本的扰动呈高斯分布
pass

视频编码小菜鸟
关注
  • 0
    点赞
  • 踩
  • 0
    收藏
    觉得还不错? 一键收藏
  • 2
    评论
专栏目录
【论文阅读笔记】NeurIPS2020文章列表Part1
zincrain的博客
12-09 2万+
A graph similarity for deep learning An Unsupervised Information-Theoretic Perceptual Quality Metric Self-Supervised MultiModal Versatile Networks Benchmarking Deep Inverse Models over time, and the Neural-Adjoint method Off-Policy Evaluation and Learning.
论文阅读笔记： 2019 ICCV 《Improving Adversarial Robustness via Guided Complement Entropy》
m0_37932351的博客
12-06 1092
2019 ICCV 《Improving Adversarial Robustness via Guided Complement Entropy》IntroductionGuided Complement EntropySynthetic Data Analysis Introduction 作者提出以往的防御方法有两点不足：1. 需要额外的流程来或者额外的模型来使目标模型更加鲁棒，这样会导致不...
【论文泛读】An Unsupervised Information-Theoretic Perceptual Quality Metric
⛄
03-25 197
解决问题 监督成本太高 解决对策 将信息理论目标函数的最新进展与由人类视觉系统生理学和对视频帧的无监督训练提供的计算架构相结合，得出感知信息度量(PIM)。 创新点： PIM在最近的和具有挑战性的BAPPS图像质量评估数据集上与监督度量具有竞争力，并且在预测CLIC 2020中图像压缩方法的排名方面优于它们 我们还使用ImageNet-C数据集进行定性实验，并确定PIM在架构细节方面是稳健的。 手工设计的模型，如MS-SSIM，由于其简单和快速，仍然是流行的人类图像质量判断的预测器。 ...
论文解读：Improving the robustness of machine reading comprehension model with hierarchical knowledge and
夏栀的博客
12-21 496
论文解读：Improving the robustness of machine reading comprehension model with hierarchical knowledge and auxiliary unanswerability prediction 摘要： 先前深度学习方法在MRC任务上均成功应用，但他们普遍脆弱且在给定一些对抗噪声时不鲁棒。为了提升MRC，我们通过引入额外知识库的信息增强模型的表征能力，其次引入辅助不可回答预测模块并使用监督类的多任务学习来进行抽取式答案区间预测。
Robust Facial Landmark Detection via a Fully-Convolutional Local-Global Context Network
Cody的博客
01-07 360
目录摘要简介 摘要 全卷积神经网络在提取局部特征上很棒，但是感受野受限制，对于全局特征表现不好。现在的一些方法通过引入级联网络、池化、或者一个数据驱动的模型来解决这个问题。 本文中，我们提出了一种新的方法，直接将全局语义特征引入到全卷积神经网络中。一个关键的概念是implicit核，这个卷积核可以将局部信息模糊化，然后通过一个dilate卷积加强成一个全局信息网络。卷积核对于网络的收敛很关键，因为...
Improving Neural Language Modeling via Adversarial Training 引入噪声
08-03
【神经语言模型对抗训练简介】 在机器学习领域，统计语言建模是一项基础任务，它广泛应用于自动语音识别、机器翻译和计算机视觉等多个领域。近年来，深度神经网络模型，特别是基于循环神经网络（RNN）的模型，已经...
Feature Denoising for Improving Adversarial Robustness.pdf
06-18
我在阅读论文《Feature Denoising for Improving Adversarial Robustness》的时候，对论文做了标注，其中加入了一些论文中提到的知识概念的补充，也包括我个人对论文中某一部分的理解。
论文研究-Spherical projection for improving the accuracy and robustness of ICA or BSS algorithms against outliers.pdf
08-16
Spherical projection for improving the accuracy and robustness of ICA or BSS algorithms against outliers，Liu Xianhua ，Robert B. Randall，Outliers are a few erroneous points or spiky noise in ...
Improving Adversarial Transferability via Neuron Attribution-Based Attacks.pdf
最新发布
10-13
Improving Adversarial Transferability via Neuron Attribution-Based Attacks 深度神经网络（DNNs）已经被部署在许多安全关键的实际应用中，如自动驾驶和医疗诊断等。然而，DNNs易受对抗样本的攻击，这些攻击可能...
A New Adaptive High-Order Unscented Kalman Filter for Improving the Accuracy and Robustness of Target Tracking
02-07
A New Adaptive High-Order Unscented Kalman Filter for Improving the Accuracy and Robustness of Target Tracking
Dependence of Eigenvalues on the Regular Fourth-Order Sturm$-$Liouville Problem
12-26
四阶正则S-L问题的特征值的依赖性，索建青，师志洁，Sturm-Liouville 问题的特征值不仅连续而且光滑依赖于该问题,本文给出了特征值关于给定系数：区间端点、边界条件、系数、和权函数的导�
Graph Theoretic Methods in Multiagent Networks
06-19
This accessible book provides an introduction to the analysis and design of dynamic multiagent networks. Such networks are of great interest in a wide range of areas in science and engineering, including: mobile sensor networks, distributed robotics such as formation flying and swarming, quantum networks, networked economics, biological synchronization, and social networks. Focusing on graph theoretic methods for the analysis and synthesis of dynamic multiagent networks, the book presents a powerful new formalism and set of tools for networked systems. The book's three sections look at foundations, multiagent networks, and networks as systems. The authors give an overview of important ideas from graph theory, followed by a detailed account of the agreement protocol and its various extensions, including the behavior of the protocol over undirected, directed, switching, and random networks. They cover topics such as formation control, coverage, distributed estimation, social networks, and games over networks. And they explore intriguing aspects of viewing networks as systems, by making these networks amenable to control-theoretic analysis and automatic synthesis, by monitoring their dynamic evolution, and by examining higher-order interaction models in terms of simplicial complexes and their applications. The book will interest graduate students working in systems and control, as well as in computer science and robotics. It will be a standard reference for researchers seeking a self-contained account of system-theoretic aspects of multiagent networks and their wide-ranging applications. This book has been adopted as a textbook at the following universities: University of Stuttgart, Germany Royal Institute of Technology, Sweden Georgia Tech, USA University of Washington, USA Ohio University, USA
使用imagecorruptions处理图片
cc__cc__的博客
04-26 2802
0.安装imagecorruptions pip install imagecorruptions 源码地址：https://github.com/bethgelab/imagecorruptions 1.函数介绍 def corrupt(image, severity=1, corruption_name=None, corruption_number=-1) """This fu...
科研篇二：对抗样本（Adversarial Example）综述
weixin_38316806的博客
10-10 6727
文章目录一、写作动机与文献来源二、术语定义2.1.对抗样本/图片（Adversarial Example/Image）2.2.对抗干扰（Adversarial perturbation）2.3.对抗训练（Adversarial Training）2.4.对抗方（Adversary）2.5.黑盒攻击（Black-box attacks）2.6.探测器（Detector）2.7. 愚弄率（Foolin...
论文阅读笔记-----Single-frame Regularization for Temporally Stable CNNs
qq_25283239的博客
02-21 1145
Eilertsen G, Mantiuk R K, Unger J. Single-frame Regularization for Temporally Stable CNNs[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2019: 11176-11185. 摘要： ...
科研篇一：NeurIPS2019 分类整理-对抗样本&Meta-Learning
weixin_38316806的博客
09-08 5491
文末获取paper整理资源，NIPS2019对抗样本&Meta-Learning
[cvpr2015]Improving training of deep neural networks via Singular Value Bounding
MataFela的博客
03-11 641
introduction 【training method】 Singular Value Bounding (SVB)：在网络训练过程中，通过将权重矩阵的奇异值限制在1附近，保证权值矩阵的正交性。 Bounded Batch Normalization (BBN)：用SVB的思想对BN的改进，去除了BN的ill-conditioning（ill-conditioning参考）的风险 算法...
10_week
qq_44157835的博客
03-09 1595
第10教学周（11.1-11.7） Papers [A Graph Similarity for Deep Learning](./A Graph Similarity for Deep Learning.md) Learning to See in the Dark（2018 CVPR） [An Unsupervised Information-Theoretic Perceptual Quality Metric](./A Graph Similarity for Deep Learning.md) P
自适应学习率算法
qq_28234239的博客
11-17 2233
神经网络中的学习率是难以设置的参数之一，如果学习率太小,则梯度很大的参数会有一个很慢的收敛速度;如果学习率太大,则己经优化得差不多的参数可能会出现不稳定的情况 。为了更有效地训练模型,比较合理的一种做法是,对每个参与训练的参数设置不同的学习率,在整个学习的过程中通过一些算法自动适应这些参数的学习率。 ...
Improving Language Understanding by Generative Pre-Training
05-13
ChitGPT is a language model developed by CSDN, which is similar to OpenAI's ChatGPT in that it is a generative pre-training model for natural language processing. However, ChitGPT is a separate ...
评论 2
添加红包

请填写红包祝福语或标题

个

红包个数最小为10个

元

红包金额最低5元

当前余额3.43元 前往充值 >
需支付：10.00元
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付元
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明：

1.余额是钱包充值的虚拟货币，按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载，可以购买VIP、付费专栏及课程。

余额充值