Convolutional Neural
Network Tutorial
Part1
Sungjoon Choi
(sungjoon.choi@cpslab.snu.ac.kr)
Overview
2
Part1: TensorFlow Tutorials
Handling images
Logistic regression
Multi-layer perceptron
Part2: Advances in convolutional neural networks
CNN basics
Four CNN architectures (AlexNet, VGG, GoogLeNet, ResNet)
Application1: Semantic segmentation
Application2: Object detection
Convolutional neural network
CNN
3
Convolutional Neural Network
CNNs are basically layers of convolutions followed by
subsampling and dense layers.
Intuitively speaking, convolutions and subsampling
layers works as feature extraction layers while a dense
layer classifies which category current input belongs to
using extracted features.
To understand CNN,
4
Zero-
paddingStride Channel
Convolution
5
http://deeplearning.stanford.edu/wiki/index.php/Feature_extraction_using_convolution
Zero-padding
6
What is the size of the input?
What is the size of the output?
What is the size of the filter?
What is the size of the zero-padding?
𝑛𝑖𝑛 = 5
𝑛 𝑜𝑢𝑡 = 5
𝑛 𝑓𝑖𝑙𝑡𝑒𝑟 = 3
𝑛 𝑝𝑎𝑑𝑑𝑖𝑛𝑔 = 1
Stride
7
(Left) Stride size: 1
(Right) Stride size: 2
If stride size equals the filter size, there will
be no overlapping.
8
9
10
11
12
13
14
CNN Architectures
15
AlexNet VGG
GoogLeNet ResNet
Top-5 Classification Error
16
AlexNet
AlexNet
18
ReLU
19
Rectified Linear Unit
tanhReLU
Faster Convergence!
VGG
VGG?
21
GoogLeNet
GoogLeNet
23
GoogLeNet
24
22 Layers Deep Network
Efficiently utilized computing resources, “Inception Module”
Significantly outperforms previous methods on ILSVRC 2014
Inception module
25
One by one convolution
26
One by one convolution
27
One by one convolution
28
GoogLeNet
29
Network in Network!
ResNet
Deep residual networks
31
152 layers network
1st place on ILSVRM 2015 classification task
1st place on ImageNet detection
1st place on ImageNet localization
1st place on COCO detection
1st place on COCO segmentation
Degeneration problem
32
CiFAR 100 Dataset
ImageNet
Residual learning building block
33
Residual mapping
34
Basic residual mapping (same dim.)
Basic residual mapping (different dim.)
“But we will show by
experiments that the
identity mapping is
sufficient for addressing
the degradation problem
and is economical, and
thus W is only used when
matching dimensions.”
Deeper bottle architecture
35
Dimension reduction
Convolution
Dimension increasement
Experimental results
36
Experimental results
37

CNN Tutorial