iwann 2019 transfer learning tutorialiwann.uma.es/wp-content/uploads/transfer_learning2.pdf · deep...

145
Transfer learning tutorial Dario Garcia Gasulla [email protected] Armand Vilalta Arias [email protected] IWANN 2019

Upload: others

Post on 29-May-2020

4 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

Transfer learning tutorial

Dario Garcia Gasulla [email protected] Vilalta Arias [email protected]

IWANN 2019

Page 2: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

Transfer Learning

“Don’t be a hero”Andrej Karpathy

Page 3: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

3

Your first driving lesson

Page 4: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

4

Your first driving lesson

Imagine learning to drive a car without knowing absolutely nothing about anything

Page 5: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

Your first driving lesson

5

Randomly initialized

Deep Neural Network

Page 6: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

6

Your first driving lesson Any previous learning can be useful

Knowing how to cook is better than knowing nothing at all

Page 7: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

7

Your first driving lesson Any previous learning can be useful

Knowing how to cook is better than knowing nothing at all

We naturally reuse what we previously learnt to be able to solve a new task.

Page 8: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

How transferlearning emerged

8

Page 9: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

9

Image classification

• 1998 LeNet-5

Gradient-based learning applied to document recognition.

Yann LeCun, Léon Bottou, Yoshua Bengio, Patrick Haffner

Page 10: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

Image classification

10

• 2012 AlexNet

ImageNet Classification with Deep Convolutional Neural Networks

Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton

Page 11: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

1998 LeNet-5

2012 AlexNet

2014 VGG19

2014 GoogLeNet

2015 Inception-V3

2015 ResNet-56

11

Page 12: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

ImageNet classification results

12From image-net.org

Page 13: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

13

At what price?

• Data available– 1,000 images per class

• Computational cost– Specific hardware

– Energy cost

• Human effort– Highly skilled professionals

– Architecture design

– Hyper-parameter fine tuning

Page 14: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

14

At what price?

• Data available– 1,000 images per class

• Computational cost– Specific hardware

– Energy cost

• Human effort– Highly skilled professionals

– Architecture design

– Hyper-parameter fine tuning

We can’t do that for every single problem!!

Page 15: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

15

At what price?

• Data available– 1,000 images per class

• Computational cost– Specific hardware

– Energy cost

• Human effort– Highly skilled professionals

– Architecture design

– Hyper-parameter fine tuning

We can’t do that for every single problem!!

Page 16: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

16

At what price?

• Data available– 1,000 images per class

• Computational cost– Specific hardware

– Energy cost

• Human effort– Highly skilled professionals

– Architecture design

– Hyper-parameter fine tuning

We can’t do that for every single problem!!

→ Transfer Learning to the rescue

Page 17: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

What is transfer learning?

17

Page 18: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

18

What is learning about?

Page 19: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

19

What is learning about?

Train set

Page 20: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

20

What is learning about?

Train set

Test set

Page 21: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

21

What is learning about?

Train set

Test setMODEL

Page 22: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

22

What is learning about?

Train set

Test setMODEL

Page 23: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

23

What is learning about?

Train set

Test setMODEL

Page 24: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

24

What is learning about?

Train set

Test set

Page 25: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

25

What is learning about?

Train set

Test set

DOG

Page 26: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

26

What is learning about?

Train set

Test set

DOG

Page 27: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

27

What is learning about?

Train set

Test set

DOG

Page 28: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

28

What is learning about?

Train set

Test set

Page 29: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

29

What is learning about?

Train set

Test set

Page 30: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

30

What is learning about?

Train set

Test set

GENERALIZATION

Page 31: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

31

What is transferlearning about?

Train set

Test set

Page 32: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

32

What is transferlearning about?

GENERALIZATIONTrain set Test set

Page 33: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

33

What is transferlearning about?

GENERALIZATIONTrain set Test set

Train a Machine Learning Model on a train set with the hope that what has been learnt will be useful to solve a different task.

Page 34: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

34

What is transferlearning about?

GENERALIZATIONTrain set Test set

Train a Machine Learning Model on a train set with the hope that what has been learnt will be useful to solve a different task.

Page 35: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

Formalizing transferlearning

35

Pan, Sinno Jialin, and Qiang Yang.

"A survey on transfer learning."

IEEE Transactions on knowledge

and data engineering (2010)

Page 36: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

36

Formalizing transfer learning

Domain:

Task:

Page 37: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

37

Formalizing transfer learning

Domain: 𝒟 = {𝒳, 𝑃 𝑋 }

• A feature space 𝒳

• A marginal probability distribution 𝑃 𝑋 ,𝑤ℎ𝑒𝑟𝑒 𝑋 = 𝑥1, … , 𝑥𝑛 ∈ 𝒳

Task:

➔ ≠“The Elgar Concert

Hall at the University

of Birmingham for

our third conference”

➔ Bag of words

➔ Content vector

Page 38: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

38

Formalizing transfer learning

Domain: 𝒟 = {𝒳, 𝑃 𝑋 }

• A feature space 𝒳

• A marginal probability distribution 𝑃 𝑋 ,𝑤ℎ𝑒𝑟𝑒 𝑋 = 𝑥1, … , 𝑥𝑛 ∈ 𝒳

Task: 𝒯 = {𝓎, 𝑓(·)}

• A label space 𝓎

CAT, DOG ≠ LION, WOLF

• An objective predictive function 𝑓 · ⟺ 𝑃(𝑦|𝑥)

➔ ≠“The Elgar Concert

Hall at the University

of Birmingham for

our third conference”

➔ Bag of words

➔ Content vector

Page 39: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

39

Formalizing transfer learning Source Target

Domain: 𝒟 = {𝒳, 𝑃 𝑋 }

Task: 𝒯 = {𝓎, 𝑓(·)}

Page 40: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

40

Formalizing transfer learning Source Target

Domain: 𝒟 = {𝒳, 𝑃 𝑋 }

• A feature space 𝒳– The Same (different)

• A marginal probability distribution 𝑃 𝑋– Different

– Similar

Task: 𝒯 = {𝓎, 𝑓(·)}

➔ ➔

Page 41: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

41

Formalizing transfer learning Source Target

Domain: 𝒟 = {𝒳, 𝑃 𝑋 }

• A feature space 𝒳– The Same (different)

• A marginal probability distribution 𝑃 𝑋– Different

– Similar

Task: 𝒯 = {𝓎, 𝑓(·)}

• A label space 𝓎– Different {CAT, DOG} {LION, WOLF}

– The same {FELINE, CANINE} {FELINE, CANINE}

• An objective predictive function 𝒇𝑺 · 𝒇𝑻 ·– Different (but similar?)

➔ ➔

Page 42: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

42

What is transferlearning about?

Train set

Test set

Page 43: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

43

What is transferlearning about? Source

domain

Target domain

Page 44: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

44

What is transferlearning about? Source

domain

Target domain

Page 45: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

45

Formalizing transfer learning Source Target

Domain: 𝒟 = {𝒳, 𝑃 𝑋 }

• A feature space 𝒳– The Same (different)

• A marginal probability distribution 𝑃 𝑋– Different

– Similar

Task: 𝒯 = {𝓎, 𝑓(·)}

• A label space 𝓎– Different {CAT, DOG} {LION, WOLF}

– The same {FELINE, CANINE} {FELINE, CANINE}

• An objective predictive function 𝒇𝑺 · 𝒇𝑻 ·– Different (but similar?)

➔ ➔

Page 46: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

46

Formalizing transfer learning Source Target

Domain: 𝒟 = {𝒳, 𝑃 𝑋 }

• A feature space 𝒳– The Same (different)

• A marginal probability distribution 𝑃 𝑋– Different

– Similar

Task: 𝒯 = {𝓎, 𝑓(·)}

• A label space 𝓎– Different {CAT, DOG} {LION, WOLF}

– The same {FELINE, CANINE} {FELINE, CANINE}

• An objective predictive function 𝒇𝑺 · 𝒇𝑻 ·– Different (but similar?)

➔ ➔

Page 47: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

47

Formalizing transferlearning

Source Target

𝒇𝑺 · 𝒇𝑻 ·

𝒇𝑺 · = 𝒇𝑻 · =

• Are they similar?

• Can we just use 𝒇𝑺 · to approximate 𝒇𝑻 · ?

• Can we reuse part of it?

Page 48: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

Representation learning

48

Page 49: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

49

Deep Neural Networks are representationlearning techniques

• Support Vector Machine (SVM) is just a classifier (a very good one).

• SVM find the best boundary separating the data instances into different classes in a given feature space.

Page 50: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

50

Deep Neural Networks are representationlearning techniques

• Support Vector Machine (SVM) is just a classifier (a very good one).

• SVM find the best boundary separating the data instances into different classes in a given feature space.

Page 51: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

51

Deep Neural Networks are representationlearning techniques

• SVMs using the kernel trick can overcome the linear limitation through an implicit mapping to a higher dimensional feature space

Page 52: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

52

Deep Neural Networks are representation learning techniques

OutputInput

Page 53: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

53

Deep Neural Networks are representation learning techniques

Linear

Classifier

OutputInput

Page 54: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

54

Deep Neural Networks are representation learning techniques

Linear

Classifier

Intermediate

representations

OutputInput

Page 55: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

Reusing DNNs knowledge

55DNNs

Page 56: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

Reusing DNNs knowledge

56

Feature Extraction

Fine-tuning

Page 57: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

Reusing DNNs knowledge

57

Feature Extraction

Fine-tuning

Page 58: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

58

Feature extraction

Linear

Classifier

Intermediate

representations

OutputInput

Page 59: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

59

Feature extraction

Linear

Classifier

Intermediate

representations

OutputInput

Pre-trained CNN, frozen weights

Page 60: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

60

Feature extraction

Linear

Classifier

Intermediate

representations

OutputInput

Page 61: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

61

Feature extraction

Intermediate

representations

Input

Page 62: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

62

Feature extraction

Intermediate

representations

Input

S

V

M

Target

Task

Labels

Page 63: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

63

Simple solutions

• DNN last layer features + SVM(Feature extraction)

Page 64: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

64

Simple solutions

• DNN last layer features + SVM(Feature extraction)

We need: Similar task and domain

Page 65: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

Reusing DNNs knowledge

65

Feature Extraction

Fine-tuning

Page 66: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

66

Fine tuning

Linear

Classifier

Intermediate

representations

OutputInput

Page 67: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

67

Fine tuning

Linear

Classifier

Intermediate

representations

OutputInput

What if the tasks

are quite different?

Page 68: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

68

Fine tuning

Linear

Classifier

Intermediate

representations

OutputInput

Source

Task

What if the tasks

are quite different?

Page 69: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

69

Fine tuning

Intermediate

representations

Input

Features learned

for the Source

Task

Page 70: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

70

Fine tuning

Intermediate

representations

Input

Features learned

for the Source

Task

Can we make

them better?

Page 71: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

71

Fine tuning

Intermediate

representations

Input

D

N

N

Target

Task

Labels

D

N

N

Page 72: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

72

Fine tuning

Intermediate

representations

Input

D

N

N

Target

Task

Labels

D

N

N

Error back-propagation

Page 73: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

73

Fine tuning

Intermediate

representations

Input

D

N

N

Target

Task

Labels

D

N

N

Error back-propagation

Page 74: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

74

Fine tuning

Intermediate

representations

Input

D

N

N

Target

Task

Labels

D

N

N

Error back-propagation

Page 75: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

75

Fine tuning

Intermediate

representations

Input

D

N

N

Target

Task

Labels

D

N

N

Error back-propagation

Page 76: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

76

Fine tuning

Intermediate

representations

Input

D

N

N

Target

Task

Labels

D

N

N

Error back-propagation

Page 77: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

77

Fine tuning

Intermediate

representations

Input

D

N

N

Target

Task

Labels

D

N

N

Error back-propagation

• Difficult to control how much do we re-train.

→ Reduced learning rate (1/10)

→ Early stopping

→ Alternate source/target sampling

Page 78: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

78

Simple solutions

• DNN last layer features + SVM(Feature extraction)

We need: Similar task and domain

• Add one or several NN layers +

Fine-tuning pre-trained layers

Page 79: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

79

Simple solutions

• DNN last layer features + SVM(Feature extraction)

We need: Similar task and domain

• Add one or several NN layers +

Fine-tuning pre-trained layersWe need: Enough data

Page 80: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

80

Beyond the last layer

Linear

Classifier

Intermediate

representations

OutputInput

Page 81: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

81

Knowledge inside DNN

Page 82: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

82

Knowledge inside DNN

Page 83: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

83

Knowledge inside DNN

Convolutional

filter

Page 84: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

84

Knowledge inside DNN

Image width

Image

height

# Filters

Depth

Page 85: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

85

Knowledge inside DNN

Image width

Image

height

# Filters

Depth

Spatial

Average

Pooling

Page 86: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

86

Knowledge inside DNN

Image width

Image

height

# Filters

Depth

Spatial

Average

Poolingfi

fifi

f1

Page 87: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

87

Knowledge inside DNNConv3_3 FC7

Spatial

Average

Pooling

Page 88: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

88

Knowledge inside DNN

Spatial

Average

Pooling

Conv3_3 FC7

Page 89: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

89

Knowledge inside DNNConv3_3 FC7

Page 90: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

90

Knowledge inside DNNConv3_3 FC7

Task

Labels

Input

Domain More

influenced by

domain

More

influenced by

task

Page 91: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

91

Knowledge inside DNNConv3_3 FC7

Task

Labels

Input

Domain More

influenced by

domain

Simple

features

More

influenced by

task

Complex

features

Visualizations from: Yosinski, Jason, et al. "Understanding neural networks through deep visualization."

arXiv preprint arXiv:1506.06579 (2015).

Page 92: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

92

Fine tuning inside DNN?

Fine-tuning

Conv3_3 FC7

Task

Labels

Input

Domain

Page 93: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

93

Fine tuning inside DNN?

Fine-tuning

Conv3_3 FC7

Task

Labels

Input

Domain

Page 94: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

94

Fine tuning inside DNN?

Fine-tuning

Conv3_3 FC7

Task

Labels

Input

Domain

Page 95: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

95

Feature extraction inside DNN?Conv3_3 FC7

Task

Labels

Input

Domain

Feature extraction

Page 96: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

96

Feature extraction inside DNN?Conv3_3 FC7

Task

Labels

Input

Domain

Feature extraction

Page 97: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

97

Feature extraction inside DNN?Conv3_3 FC7

Task

Labels

Input

Domain

Feature extraction

Page 98: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

98

Feature extraction inside DNN?Conv3_3 FC7

Task

Labels

Input

Domain

Feature extraction

Page 99: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

99

Feature extraction inside DNN?Conv3_3 FC7

Task

Labels

Input

Domain

Feature extraction

VGG16 dim: 12,416

Page 100: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

100

Feature extraction inside DNN?Conv3_3 FC7

Task

Labels

Input

Domain

Feature extraction Different range

of values

VGG16 dim: 12,416

Page 101: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

101

Feature behavior in Transfer LearningConv3_3 FC7

0 50 100 150 200 250 300 350

Page 102: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

102

Standardization: features in same range

0 50 100 150 200 250 300 350

L2 n

orm

Typically, it is used

a L2-normalization

0,00 0,13 0,27 0,40 0,54 0,67 0,80 0,94

Page 103: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

103

Standardization: features in same range

Featu

re

Sta

nd

ard

isati

on

Each feature independently

-5 -4 -3 -2 -1 0 1 2 3 4 50 50 100 150 200 250 300 350

L2 n

orm

Typically, it is used

a L2-normalization

0,00 0,13 0,27 0,40 0,54 0,67 0,80 0,94

Page 104: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

104

Standardization: features in same range

Featu

re

Sta

nd

ard

isati

on

Each feature independently

-5 -4 -3 -2 -1 0 1 2 3 4 50 50 100 150 200 250 300 350

L2 n

orm

Typically, it is used

a L2-normalization

0,00 0,13 0,27 0,40 0,54 0,67 0,80 0,94

Page 105: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

105

Standardization: features in contextConv3_3 FC7

• Distinct feature behavior for

a subset of data wrt rest

Page 106: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

106

Stardardization: features in context

CUB-200 - birds

Garcia-Gasulla et al.

On the behavior of

convolutional nets for

feature extraction. 2017

Page 107: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

107

Curse of dimensionalityConv3_3 FC7

Task

Labels

Input

Domain

VGG16 dim: 12,416

Page 108: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

108

Curse of dimensionalityConv3_3 FC7

Task

Labels

Input

Domain

VGG16 dim: 12,416

PCA ?

Page 109: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

109

Curse of dimensionalityConv3_3 FC7

Task

Labels

Input

Domain

VGG16 dim: 12,416

PCA

Page 110: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

110

Curse of dimensionalityConv3_3 FC7

Task

Labels

Input

Domain

VGG16 dim: 12,416

PCA

• Unsupervised:

Only takes into account

the domain

Page 111: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

111

Curse of dimensionalityConv3_3 FC7

Task

Labels

Input

Domain

VGG16 dim: 12,416

Supervised Feature Selection

?

Page 112: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

112

Curse of dimensionalityConv3_3 FC7

Task

Labels

Input

Domain

VGG16 dim: 12,416

Supervised Feature Selection

?• High computational cost!

Page 113: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

113

Curse of dimensionalityConv3_3 FC7

Task

Labels

Input

Domain

VGG16 dim: 12,416

Page 114: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

114

Curse of dimensionalityConv3_3 FC7

Task

Labels

Input

Domain

VGG16 dim: 12,416

• Which is the real problem?

Page 115: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

115

Curse of dimensionalityConv3_3 FC7

Task

Labels

Input

Domain

VGG16 dim: 12,416

• Which is the real problem?

• Too many features?

• Too few images?

Page 116: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

116

Curse of dimensionalityConv3_3 FC7

Task

Labels

Input

Domain

VGG16 dim: 12,416

• Which is the real problem?

• Too many features?

• Too few images? A requirement

Page 117: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

117

Curse of dimensionalityConv3_3 FC7

Task

Labels

Input

Domain

VGG16 dim: 12,416

• Which is the real problem?

• Too many features?

• Too much information!

• Too few images?

Page 118: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

118

Meaningful discretizationConv3_3 FC7

Task

Labels

Input

Domain

FC7

Real

values

1 -1 0 1 1 0 0 0 1 -1 0 1 1 0 0 0 1 -1 0 1 1 0 0 0 1 -1 0 1 1 0 0 0 1 0 0 0 1 -1 0 1 1 0 0 0 1 -1 0 1 1

Quantitization to {-1,0,1} based on

feature values

Conv3_3

Discrete

values

Page 119: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

119

Full-Network Embedding Recipe

1. Spatial Average Pooling

2. Standardisation

3. Discretization

Garcia-Gasulla, et al.

An out-of-the-box full-

network embedding for

convolutional neural

networks. 2018.

Page 120: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

120

Full Network embedding

Page 121: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

121

FNE - Results: Similar source task

Network pre-trained on Places2 for mit67 and on ImageNet for the rest.

Page 122: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

122

FNE - Results: Similar source task

Network pre-trained on Places2 for mit67 and on ImageNet for the rest.

Page 123: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

123

FNE - Results: Similar source task

Network pre-trained on Places2 for mit67 and on ImageNet for the rest.

+2.9

+4.2

Page 124: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

124

FNE - Results: Similar source task

Network pre-trained on Places2 for mit67 and on ImageNet for the rest.

-0.3 -0.4 -0.5

+2.9

+4.2

Page 125: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

125

FNE - Results: Similar source task

Network pre-trained on Places2 for mit67 and on ImageNet for the rest.

-0.3 -0.4 -0.5

+2.9

+4.2

Page 126: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

126

FNE - Results: Dissimilar source task

Network pre-trained on ImageNet for mit67 and on Places2 for the rest.

Page 127: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

127

FNE - Results: Dissimilar source task

Network pre-trained on ImageNet for mit67 and on Places2 for the rest.

Page 128: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

128

FNE - Results: Dissimilar source task

Network pre-trained on ImageNet for mit67 and on Places2 for the rest.

+3.3 +11.9 +15.4 +17.5 +8.0 +9.3 +10.6

Page 129: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

129

Simple solutions

• DNN last layer features + SVM(Feature extraction)

We need: Similar task and domain

• Add one or several NN layers +

Fine-tuning pre-trained layersWe need: Enough data

• Full Network Embedding– Robust to different task and domain

– Works with little data

Page 130: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

Practicaltips

130

Page 131: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

131

Practicaltips

Fine-tuning

• Whenever possible don’t start from scratch.

Page 132: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

132

Practicaltips

Fine-tuning

• Whenever possible don’t start from scratch.

• External data can help prevent overfitting (even from a different problem).

Page 133: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

133

Practicaltips

Fine-tuning

• Whenever possible don’t start from scratch.

• External data can help prevent overfitting (even from a different problem).

• Begin freezing as much as possible and proceed with

caution (particularly for large models)

Page 134: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

134

Practicaltips

Feature

extraction

• Easy baseline for every problem.

Page 135: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

135

Practicaltips

Feature

extraction

• Easy baseline for every problem.

• ImageNet, a model to pre-train them all. (but not always)

Page 136: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

136

Practicaltips

Feature

extraction

• Easy baseline for every problem.

• ImageNet, a model to pre-train them all. (but not always)

• Always normalize features.

Page 137: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

137

Practicaltips

Feature

extraction

• Easy baseline for every problem.

• ImageNet, a model to pre-train them all. (but not always)

• Always normalize features.

• If source and target task are closely related:

Last two layers are your best chance.

Page 138: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

138

Practicaltips

Feature

extraction

• Easy baseline for every problem.

• ImageNet, a model to pre-train them all. (but not always)

• Always normalize features.

• If source and target task are closely related:

Last two layers are your best chance.

• If source and target task are quite different:

Try everything

Use FNE

Page 139: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

139

UsefulReferences • Zhe Xu, Shaoli Huang, Ya Zhang, and Dacheng Tao. Augmenting

strong supervision using webdata for fine-grained categorization.

2015.

• Steve Branson, Grant Van Horn, Serge Belongie, and Pietro

Perona. Bird species categorizationusing pose normalized deep

convolutional nets. 2014.

• Chang Liu, Yu Cao, Yan Luo, Guanling Chen, Vinod Vokkarane, and

Yunsheng Ma. Deep-food: Deep learning-based food image

recognition for computer-aided dietary assessment. 2016.

Don’t start training

from scratch

Page 140: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

140

UsefulReferences • Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How

transferable are features in deep neural networks? 2014.

• Dario Garcia-Gasulla, Ferran Parés, Armand Vilalta, Jonatan

Moreno, Eduard Ayguadé, Jesús Labarta, Ulises Cortés, and

Toyotaro Suzumura. On the behavior of convolutional nets for

feature extraction. 2017Transfer learning 101

Page 141: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

141

UsefulReferences

• Hossein Azizpour, Ali Sharif Razavian, Josephine Sullivan, Atsuto

Maki, and Stefan Carlsson. Factors of transferability for a generic

convnet. 2016.

• Ali Sharif Razavian, Hossein Azizpour, Josephine Sullivan, and

Stefan Carlsson. CNN features off-the-shelf: an astounding baseline

for recognition. 2014.

• Yunchao Gong, Liwei Wang, Ruiqi Guo, and Svetlana Lazebnik.

Multi-scale orderless pooling of deep convolutional activation

features. 2014.

• Jeff Donahue, Yangqing Jia, Oriol Vinyals, Judy Hoffman, Ning

Zhang, Eric Tzeng, and Trevor Darrell. Decaf: A deep convolutional

activation feature for generic visual recognition. 2014.

• Arsalan Mousavian and Jana Kosecka. Deep convolutional features

for image based retrieval and scene categorization. 2015.

Basic feature

extraction

Page 142: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

142

UsefulReferences • Garcia-Gasulla, Dario, Armand Vilalta, Ferran Parés, Eduard

Ayguadé, Jesus Labarta, Ulises Cortés, and Toyotaro Suzumura. An

out-of-the-box full-network embedding for convolutional neural

networks. 2018.

• Vilalta, Armand, Dario Garcia-Gasulla, Ferran Parés, Jonathan

Moreno, Eduard Ayguadé Jesús Labarta, Ulises Cortés, and

Toyotaro Suzumura. Full-Network Embedding in a Multimodal

Embedding Pipeline. 2017.

Multi-layer feature

extraction

Page 143: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

143

UsefulReferences • Weifeng Ge and Yizhou Yu. Borrowing treasures from the wealthy:

Deep transfer learning through selective joint fine-tuning. 2017.

• Marcel Simon and Erik Rodner. Neural activation constellations:

Unsupervised part model discovery with convolutional networks.

2015.Advanced fine tuning

to solve small tasks

Page 144: IWANN 2019 Transfer learning tutorialiwann.uma.es/wp-content/uploads/Transfer_learning2.pdf · Deep Neural Networks are representation learning techniques •Support Vector Machine

CTE-Power9

144

Hands on session

16:00

2 GPUs per account

52 computing nodes. Each one:

• 2 x IBM Power9 8335-GTH @

2.4GHz(total 160 threads)

• 512GB of main memory

• 4 x GPU NVIDIA V100 (Volta)

with 16GB each

Available through PRACE and RES