37
Convolutional Neural Network Tutorial Part1 Sungjoon Choi ([email protected])

CNN Tutorial

Embed Size (px)

Citation preview

Page 1: CNN Tutorial

Convolutional Neural Network Tutorial

Part1

Sungjoon Choi([email protected])

Page 2: CNN Tutorial

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

Page 3: CNN Tutorial

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.

Page 4: CNN Tutorial

4

To understand CNN,

Zero-paddingStride Channel

Page 5: CNN Tutorial

5

Convolution

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

Page 6: CNN Tutorial

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?

Page 7: CNN Tutorial

7

Stride

(Left) Stride size: 1

(Right) Stride size: 2

If stride size equals the filter size, there will be no overlapping.

Page 8: CNN Tutorial

8

Page 9: CNN Tutorial

9

Page 10: CNN Tutorial

10

Page 11: CNN Tutorial

11

Page 12: CNN Tutorial

12

Page 13: CNN Tutorial

13

Page 14: CNN Tutorial

14

Page 15: CNN Tutorial

15

CNN ArchitecturesAlexNet VGG

GoogLeNet ResNet

Page 16: CNN Tutorial

16

Top-5 Classification Error

Page 17: CNN Tutorial

AlexNet

Page 18: CNN Tutorial

18

AlexNet

Page 19: CNN Tutorial

19

ReLURectified Linear Unit

tanhReLU

Faster Convergence!

Page 20: CNN Tutorial

VGG

Page 21: CNN Tutorial

21

VGG?

Page 22: CNN Tutorial

GoogLeNet

Page 23: CNN Tutorial

23

GoogLeNet

Page 24: CNN Tutorial

24

GoogLeNet

22 Layers Deep Network

Efficiently utilized computing resources, “Inception Module”

Significantly outperforms previous methods on ILSVRC 2014

Page 25: CNN Tutorial

25

Inception module

Page 26: CNN Tutorial

26

One by one convolution

Page 27: CNN Tutorial

27

One by one convolution

Page 28: CNN Tutorial

28

One by one convolution

Page 29: CNN Tutorial

29

GoogLeNet

Network in Network!

Page 30: CNN Tutorial

ResNet

Page 31: CNN Tutorial

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

Page 32: CNN Tutorial

32

Degeneration problemCiFAR 100 Dataset

ImageNet

Page 33: CNN Tutorial

33

Residual learning building block

Page 34: CNN Tutorial

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.”

Page 35: CNN Tutorial

35

Deeper bottle architecture

Dimension reduction

Convolution

Dimension increasement

Page 36: CNN Tutorial

36

Experimental results

Page 37: CNN Tutorial

37

Experimental results