Download presentation

Presentation is loading. Please wait.

Published byOdalys Alltop Modified over 2 years ago

1
Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

2

3
Concept Learning by Induction

4
Few Examples

5
Transfer Learning Target (New) Source (Original)

6
Define: Related Concept

7
Transfer Learning Approaches Common Inductive Bias Common Instances Common Features

8
Different Feature Space

9
Example

10
Example

11
Example

12
Common Inductive Bias

13
Common Inductive Bias

14
Common Instances

15
Common Features

16
New Approach to Transfer Learning

17
Our Solution: Metaphors

18
Metaphors Target (New) Source (Original)

19
Concept Learner Metaphor Learner Source Target +/-

20

21

22
Theorem

23
The Metaphor Theorem

24
Redefine Transfer Learning

25

26
Metaphor Learning Framework

27
Concept Learning Framework Search Algorithm Hypothesis Space Evaluation Function Data

28
Source Target Metaphor Learning Framework Search Algorithm Metaphor Space Evaluation Function

29
Metaphor Evaluation

30

31

32

33

34

35

36
Metaphor Spaces

37
General Few Degrees of Freedom Representation-Specific Bias

38
Geometric Transformations ЯR

39
Dictionary-Based Metaphors cheesequeso

40
Linear Transformations

41
Which metaphor space should I use?

42
Automatic Selection of Metaphor Spaces Which metaphor space should I use?

43
Occam’s Razor Automatic Selection of Metaphor Spaces Which metaphor space should I use?

44
Structural Risk Minimization Occam’s Razor Automatic Selection of Metaphor Spaces Which metaphor space should I use?

45
Automatic Selection of Metaphor Spaces

46

47

48
Empirical Evaluation

49
Reference Methods Baseline Target Only Identity Metaphor Merge State-of-the-Art Frustratingly Easy Domain Adaptation – Daumé, 2007 MultiTask Learning – Caruana, 1997; Silver et al, 2010 TrAdaBoost – Dai et al, 2007

50
Digits: Negative Image

51

52

53
Digits: Higher Resolution

54

55

56
Wine

57

58
Qualitative Results Transfer Learning Task Target Instance Target Sample Size Digits: Negative Image Digits: Higher Resolution

59
Discussion

60
Recap Problem: Concept learning with few examples Solution: Metaphors

61
Recap

62

63

64

65
What if the concepts are not related?

66

67
Metaphors are not a measure of relatedness

68
Metaphors are not a measure of relatedness Metaphors explain how concepts are related

69
Vision

70

71
Explaining how concepts are related since M E T A P H O R S

72
Concept Learning by Induction

73

74
Few Examples

75

76
Approaches Explanation-Based Learning Semi-Supervised Learning Transfer Learning

77
Explanation-Based Learning Axioms Data Logical Deduction

78
Semi-Supervised Learning

79
Transfer Learning

80
Target (New) Source (Original)

81
Transfer Learning

82
Target (New) Source (Original)

83
Define: Related Concept

84
Transfer Learning Approaches Common Inductive Bias Common Instances Common Features

85
Common Inductive Bias

86

87

88
Common Instances

89

90

91

92

93
Common Features 1.Perform feature selection on source 2.Use that selection on target

94
Which definition is better?

95
Different Feature Space

96
Example

97
Example

98
Example

99
Common Inductive Bias

100
Common Inductive Bias

101
Common Instances

102
Common Features

103
Our Solution: Metaphors

104
Performance with Automatic Selection of Metaphor Spaces Digits: Negative Image

105
Performance with Automatic Selection of Metaphor Spaces Digits: Negative Image Geometric Transformations Feature Reordering Orthogonal Linear Transformations Orthogonal Quadratic Transformations

106
Performance with Automatic Selection of Metaphor Spaces Digits: Negative Image Geometric Transformations Feature Reordering Orthogonal Linear Transformations Orthogonal Quadratic Transformations

107

108
What if I have more than one source?

109
Multiple Source Datasets B Я HRZ

110
B Я H R Z

111
Я R

112
Performance with Multiple Source Datasets Latin & Cyrillic

113
Performance with Multiple Source Datasets Latin & Cyrillic ABCDEFG HIJKLMN OPQRSTU VWXYZ

114
Performance with Multiple Source Datasets Latin & Cyrillic ABCDEFG HIJKLMN OPQRSTU VWXYZ

115
Performance with Multiple Source Datasets

116

117

Similar presentations

© 2017 SlidePlayer.com Inc.

All rights reserved.

Ads by Google