702      阿里云  技术社区[云栖]

Top100论文导读：深入理解卷积神经网络CNN（Part Ⅰ）

https://yq.aliyun.com/cloud

Adrian Colyermorning paperAccel Partners

https://www.linkedin.com/in/adriancolyer/

Twitterhttps://twitter.com/adriancolyer

“ Top100” “”内容见下篇博客。首先是：

ImageNet classification with deep convolutional neural networks
Maxout networksGoodfellow
Network in network

OverFeat

Ujjwal KarnAn intuitive explanation of convolutional neural networks

ImageNet
Hinton成功ImageNet“”GPU
CNN
CNN
Krizhevsky8——531000softmax1000GPUGPU 

1. tanh使用ReLUtanh， ”
2. GPUGPUtop-1top-5%1.2%
3. “ ”
4. z x z之间top-1top-5%0.3% 

Dropout

ILSVRC2010top-1top-537.5%17.0%ILSVRC2012top-515.3%26.2%

Maxout

MaxoutDropoutDropoutMaxoutCNNMaxoutMaxout
kDropout
Maxout 

MaxoutDropoutMNISTCIFAR10CIFAR100
DropoutMaxoutDropoutDropoutMaxout
NIN

CNNGLMMaxout
NINGLM“”(MLP) 

第mlpconv种mlpconv 

CNNNINmlpconvsoftmax

CIFAR101% 

CIFAR1001%SVHN好 

OverFeat
OverFeatCNNILSVRC 2013

ConvNet
OverFeatConvNet

Krizhevsky265 

5 

——156

pdf

@

Convolutional neural networks, Part 1Adrian Colyer

最后更新：2017-05-01 08:01:17

  上一篇： spark源码分析之Executor启动与任务提交篇

  下一篇： spark源码分析Master与Worker启动流程篇