当前位置:网站首页>深度学习分类网络 -- ZFNet
深度学习分类网络 -- ZFNet
2022-07-06 12:13:00 【有时候。】
前言
ZFNet这篇论文中的motivation有两点:
- 深度神经网络为什么表现好?
- 如何改进网络?
作者通过Deconvnet来可视化深度神经网络各层的特征图,并据此来针对性地对AlexNet进行改进,取得了2013年ImageNet分类挑战赛的冠军。
1. 网络结构
在可视化特征图的基础上,作者对ALexNet做了两处修改:
- 第一个卷积的尺寸由11×11改为7×7
- 第一个卷积的stride由4改为2

当然,作者这样修改不是凭空臆造,而是有据可依的,而这个依据就是特征图的可视化结果。
2. 要点
Deconvnet可视化方法

网络各层特征特点


网络各层特征学习过程

根据特征可视化改进网络结构

特征不变性

遮挡敏感性


参考资料
边栏推荐
- [infrastructure] deployment and configuration of Flink / Flink CDC (MySQL / es)
- 信息系统项目管理师---第八章 项目质量管理
- 【计网】第三章 数据链路层(4)局域网、以太网、无线局域网、VLAN
- 5. 无线体内纳米网:十大“可行吗?”问题
- Wonderful coding [hexadecimal conversion]
- Groovy基础语法整理
- Learning and Exploration - function anti shake
- POJ 3207 Ikki' s Story IV – Panda' s Trick (2-SAT)
- Analysis of rainwater connection
- 枚举根据参数获取值
猜你喜欢

Pay attention to the partners on the recruitment website of fishing! The monitoring system may have set you as "high risk of leaving"

Standardized QCI characteristics

Understand yolov1 Part II non maximum suppression (NMS) in prediction stage
腾讯字节等大厂面试真题汇总,网易架构师深入讲解Android开发

Teach you to learn JS prototype and prototype chain hand in hand, a tutorial that monkeys can understand
![[translation] linkerd's adoption rate in Europe and North America exceeded istio, with an increase of 118% in 2021.](/img/09/106adc222c06cbd2f4f66cf475cce2.jpg)
[translation] linkerd's adoption rate in Europe and North America exceeded istio, with an increase of 118% in 2021.

Tencent Android interview must ask, 10 years of Android development experience

Monthly report of speech synthesis (TTS) and speech recognition (ASR) papers in June 2022

夏志刚介绍

Interview assault 63: how to remove duplication in MySQL?
随机推荐
腾讯Android面试必问,10年Android开发经验
The "white paper on the panorama of the digital economy" has been released with great emphasis on the digitalization of insurance
leetcode先刷_Maximum Subarray
Cesium 两点之间的直线距离
信息系统项目管理师---第八章 项目质量管理
121. 买卖股票的最佳时机
【基础架构】Flink/Flink-CDC的部署和配置(MySQL / ES)
MySQL information schema learning (II) -- InnoDB table
Interpretation of Dagan paper
Transformer model (pytorch code explanation)
POJ3617 Best Cow Line 馋
Poj3617 best cow line
广州首个数据安全峰会将在白云区开幕
技术分享 | 抓包分析 TCP 协议
Tensorflow2.0 self defined training method to solve function coefficients
After solving 2961 user feedback, I made such a change
Finally, there is no need to change a line of code! Shardingsphere native driver comes out
Tencent Android interview must ask, 10 years of Android development experience
AsyncHandler
Classic 100 questions of algorithm interview, the latest career planning of Android programmers