CNN Tutorial

Preview:

Citation preview

Convolutional Neural Network Tutorial

Part1

Sungjoon Choi(sungjoon.choi@cpslab.snu.ac.kr)

2

OverviewPart1: TensorFlow Tutorials

Handling imagesLogistic regressionMulti-layer perceptron

Part2: Advances in convolutional neural networksCNN basics

Four CNN architectures (AlexNet, VGG, GoogLeNet, ResNet)Application1: Semantic segmentationApplication2: Object detection

Convolutional neural network

3

CNNConvolutional Neural Network

CNNs are basically layers of convolutions followed by subsampling and dense layers.

Intuitively speaking, convolutions and subsampling lay-ers works as feature extraction layers while a dense layer classifies which category current input belongs to using ex-tracted features.

4

To understand CNN,

Zero-paddingStride Channel

5

Convolution

http://deeplearning.stanford.edu/wiki/index.php/Feature_extraction_using_convolution

6

Zero-padding

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?

7

Stride

(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

15

CNN ArchitecturesAlexNet VGG

GoogLeNet ResNet

16

Top-5 Classification Error

AlexNet

18

AlexNet

19

ReLURectified Linear Unit

tanhReLU

Faster Convergence!

VGG

21

VGG?

GoogLeNet

23

GoogLeNet

24

GoogLeNet

22 Layers Deep Network

Efficiently utilized computing resources, “Inception Module”

Significantly outperforms previous methods on ILSVRC 2014

25

Inception module

26

One by one convolution

27

One by one convolution

28

One by one convolution

29

GoogLeNet

Network in Network!

ResNet

31

Deep residual networks

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

32

Degeneration problemCiFAR 100 Dataset

ImageNet

33

Residual learning building block

34

Residual mappingBasic residual mapping (same dim.)

Basic residual mapping (different dim.)

“But we will show by experiments that the identity mapping is sufficient for address-ing the degradation problem and is eco-nomical, and thus W is only used when match-ing dimensions.”

35

Deeper bottle architecture

Dimension reduction

Convolution

Dimension increasement

36

Experimental results

37

Experimental results

Recommended