×

COMP 3340
Convolutional Neural Network
for Image Classification

COMP3340 Applied Deep Learning

2.1 Notations and Outline

• CV: Computer Vision

• NN: Neural Networks
• ReLU: Rectified Linear Unit
• SGD: Stochastic Gradient Descend
• FP: Forward Propagation
• BP: Backward Propagation
• CNN: Convolutional Neural Networks
• BN: Batch Normalization

COMP3340 Applied Deep Learning [1] Reference

2
2.1 Notations and Outline
2.1 Notations and Outline
2.2 Introduction
2.3 Boarder Impact
2.4 Benchmark
2.5 Basic Knowledge for Deep Learning
2.6 CNN Operators
2.7 CNN Architectures
2.8 Reference

Three questions for quiz and final.

1. L1 and L2 Regularization analysis

2. Shuffle conv implementation
3. Flops/Parameters computation

COMP3340 Applied Deep Learning [1] Reference

3
Where is CNN?

COMP3340 Applied Deep Learning

60% Of Human Brain Is “Visual”

The visual data domain is the most important data domain.

COMP3340 Applied Deep Learning
5
2.2.2 Background

COMP3340 Applied Deep Learning [1] [Link]

6
2.2.2 Background

• Representing an image in a computer involves storing it as a grid of pixels, where each pixel represents the
smallest unit of the image.
• RGB (Red, Green, Blue) color model, where each color is represented by a combination of these three primary
colors.
• Channel: each value ranges from 0 to 255, (255, 0, 0) red, while (0, 255, 0) green, and (0, 0, 255) blue. A
combination of these, like (255, 255, 0), yellow.

COMP3340 Applied Deep Learning [1] [Link]

7
2.2.2 Background

COMP3340 Applied Deep Learning [1] [Link]

8
2.2.2 Background

COMP3340 Applied Deep Learning [1] [Link]

9
2.2.2 Background

A picture is worth a thousand words.

COMP3340 Applied Deep Learning [1] [Link]

10
2.2.2 Background

COMP3340 Applied Deep Learning [1] [Link]

11
2.2.2 Problem definition: ConvNet for Image Classification

COMP3340 Applied Deep Learning [1] [Link]

12
2.2.2 Problem definition

• In 2012, Google Brain took an

artificial neural network and
spread the computation across
16,000 of our CPU cores, and
trained models with more than 1
billion connections.
• While there’s no accepted way
to compare artificial neural
networks to biological brains, as
a very rough comparison an
adult human brain has around
100 trillion connections. So we
still have lots of room to grow.

COMP3340 Applied Deep Learning [1] [Link]

13
2.2.3 ImageNet benchmark

COMP3340 Applied Deep Learning [1] [Link]

14
2.2.3 ImageNet benchmark

COMP3340 Applied Deep Learning [1] [Link]

15
2.2.3 ImageNet benchmark

COMP3340 Applied Deep Learning [1] [Link]

16
2.2.3 ImageNet benchmark

COMP3340 Applied Deep Learning [1] [Link]

17
2.2.3 ImageNet benchmark

COMP3340 Applied Deep Learning [1] [Link]

18
Teach Computer to See

COMP3340 Applied Deep Learning [Link]

19
2.3 Boarder impact

• Basis of various modern CV applications

• Image classification
• Semantic Segmentation
• Object Detection
• Video Understanding
• Caption
• Relation Prediction

• Variants:
• Low-shot learning
• Continual learning

• Applications:
• Face recognition
• Robotics
COMP3340 Applied Deep Learning [1] Reference
20
 2.3 Boarder Impact

Relation Prediction Caption

COMP3340 Applied Deep Learning [1] [Link]

21
2.4 Benchmark

• ImageNet (mentioned before)

• Cifar10. Cifar100

Process and visualize Cifar10 and Cifar100 dataset:

COMP3340 Applied Deep Learning [1] [Link]

22
23
PIL stands for Python Imaging Library, and it's the original library that
enabled Python to deal with images.

[-1,+1]

24
PIL stands for Python Imaging Library, and it's the original library that
enabled Python to deal with images.

25
26
27
2.5.1 Preliminaries for Image Classification

COMP3340 Applied Deep Learning [1] [Link]

28
[Link] K-Nearest Neighbor

COMP3340 Applied Deep Learning [1] [Link]

29
[Link] K-Nearest Neighbor

COMP3340 Applied Deep Learning [1] [Link]

30
…

31
…

32
33
34
…

Accuracy: 35.39%
35
[Link] K-Nearest Neighbor

COMP3340 Applied Deep Learning [1] [Link]

36
[Link] K-Nearest Neighbor

COMP3340 Applied Deep Learning [1] [Link]

37
[Link] K-Nearest Neighbor
Try it by yourself [demo]

COMP3340 Applied Deep Learning [1] [Link]

38
[Link] Cross-validation

COMP3340 Applied Deep Learning [1] Reference

39
[Link] Cross-validation

COMP3340 Applied Deep Learning [1] [Link]

40
[Link] Cross-validation

COMP3340 Applied Deep Learning [1] [Link]

41
[Link] Disadvantages of k-Nearest Neighbor

COMP3340 Applied Deep Learning [1] [Link]

42
[Link] Linear Classifier 10x3072

Weight matrix Bias vector

COMP3340 Applied Deep Learning [1] [Link]

43
[Link] Linear Classifier and Softmax

Wx+b
COMP3340 Applied Deep Learning [1] [Link]
44
[Link] Neural Network

1*0.55+2*0.54+0.45=2.08
W11
b1 = 0.45

1 W12

W21

W22
b2 = 0.89

b3 = 0.96

COMP3340 Applied Deep Learning [1] [Link]

45
[Link] Neural Network

COMP3340 Applied Deep Learning [1] [Link]

46
[Link] Neural Network

COMP3340 Applied Deep Learning [1] [Link]

47
[Link] Neural Network

COMP3340 Applied Deep Learning [1] [Link]

48
[Link] Neural Network
More Activation functions

COMP3340 Applied Deep Learning [1] [Link]

49
[Link] Neural Network
Forward Propagation

write the equations above more compactly as:

COMP3340 Applied Deep Learning [1] [Link]

50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
[Link] Linear Classifier and Softmax

COMP3340 Applied Deep Learning [1] [Link]

78
[Link] Linear Classifier and Softmax

COMP3340 Applied Deep Learning [1] [Link]

79
Visualization

80
[Link] Linear Classifier and Softmax

COMP3340 Applied Deep Learning [1] [Link]

81
[Link] Linear Classifier and Softmax

COMP3340 Applied Deep Learning [1] [Link]

82
[Link] Linear Classifier and Softmax

COMP3340 Applied Deep Learning [1] [Link]

83
[Link] Regularization Regularization

COMP3340 Applied Deep Learning [1] [Link]

84
[Link] Regularization

COMP3340 Applied Deep Learning [1] [Link]

85
[Link] Regularization

COMP3340 Applied Deep Learning [1] [Link]

86
[Link] Neural Network
Backward Propagation

cost function with respect to that single example

cost function with respect to m examples

COMP3340 Applied Deep Learning [1] [Link]

87
[Link] Neural Network
Backward Propagation

gradient descent

where

COMP3340 Applied Deep Learning [1] [Link]

88
[Link] Neural Network (example)

COMP3340 Applied Deep Learning [1] [Link]

89
[Link] Neural Network

COMP3340 Applied Deep Learning [1] [Link]

90
[Link] Neural Network

COMP3340 Applied Deep Learning [1] [Link]

91
COMP3340 Applied Deep Learning [Link]
92
[Link] Neural Network
Backward Propagation

COMP3340 Applied Deep Learning [1] [Link]

93