71
Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

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

Embed Size (px)

Citation preview

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

Teaching Machines to Learn by Metaphors

Omer Levy & Shaul MarkovitchTechnion – Israel Institute of Technology

Page 2: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology
Page 3: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Concept Learning by Induction

Page 4: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Few Examples

Page 5: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Transfer Learning

Target (New)

Source (Original)

Page 6: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Define: Related Concept

Page 7: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Transfer Learning Approaches

• Common Inductive Bias

• Common Instances

• Common Features

Page 8: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Different Feature Space

Page 9: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Example

0 2 3-3 -2

Page 10: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Example

0 2 3-3 -2

0 4 9

Page 11: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Example

0 2 3-3 -2

0 4 9

𝑥𝑠=𝑥𝑡2

Page 12: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Common Inductive Bias

0 2 3-3 -2

0 4 9

Page 13: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Common Inductive Bias

0 2 3-3 -2

0 4 9

Page 14: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Common Instances

0 2 3-3 -2

0 4 9

Page 15: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Common Features

23

-3-2

4 9

𝑥𝑠=𝑦𝑡2

Page 16: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

New Approach to Transfer Learning

Page 17: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Our Solution: Metaphors

Page 18: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Metaphors

Target (New)

Source (Original)

Page 19: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Concept Learner

Metaphor Learner

𝜇 h𝑠

Source

Target

+/-𝑥𝑡 𝑥𝑠

Page 20: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

h𝑡 (𝑥𝑡 )=h𝑠 (𝜇 (𝑥𝑡 ))

Page 21: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

is a perfect metaphor if:

1. is label preserving

2. is distribution preserving

Page 22: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Theorem

If is a perfect metaphor- and -

is a source hypothesis with error- then -

is a target hypothesis with error

Page 23: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

The Metaphor Theorem

If is an -perfect metaphor- and -

is a source hypothesis with error- then -

is a target hypothesis with error

Page 24: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Redefine Transfer Learning

Given source and target datasets, find a target hypothesis such that is as small as possible.

Page 25: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Redefine Transfer Learning

Given source and target datasets, find an -perfect metaphor such that is as small as possible.

Page 26: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Metaphor Learning Framework

Page 27: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

h

Concept Learning Framework

Search Algorithm

Hypothesis Space

Evaluation Function

Data

Page 28: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

𝜇

Source

Target

Metaphor Learning Framework

Search Algorithm

Metaphor Space

Evaluation Function

Page 29: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Metaphor Evaluation

Page 30: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Metaphor Evaluation

1. is label preserving

2. is distribution preserving

Page 31: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Metaphor Evaluation

1. is label preservingEmpirical error over target dataset

2. is distribution preservingStatistical distance between and

Page 32: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Metaphor Evaluation

𝑆𝐷 (𝜇 (𝑥𝑡 ) ,𝑥𝑠 )

Page 33: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Metaphor Evaluation

𝑆𝐷 ¿

Page 34: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Metaphor Evaluation

𝑆𝐷 (𝜇 (𝑥𝑡− ) , 𝑥𝑠−)

Page 35: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Metaphor Evaluation

𝑆𝐷 ¿

Page 36: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Metaphor Spaces

Page 37: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Metaphor Spaces

• General

• Few Degrees of Freedom

• Representation-Specific Bias

Page 38: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Geometric Transformations

Я R

Page 39: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Dictionary-Based Metaphors

cheese queso

Page 40: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Linear Transformations

Page 41: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Which metaphor space should I use?

Page 42: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Automatic Selection of Metaphor Spaces

Which metaphor space should I use?

Page 43: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Occam’s Razor

Automatic Selection of Metaphor Spaces

Which metaphor space should I use?

Page 44: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Structural Risk Minimization

Occam’s Razor

Automatic Selection of Metaphor Spaces

Which metaphor space should I use?

Page 45: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Automatic Selection of Metaphor Spaces

ℳ1

ℳ2

ℳ3

ℳ 4

Page 46: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Automatic Selection of Metaphor Spaces

ℳ1

ℳ2

ℳ3

ℳ 4

𝜇1

𝜇2

𝜇3

𝜇4

Page 47: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Automatic Selection of Metaphor Spaces

ℳ1

ℳ2

ℳ3

ℳ 4

𝜇1

𝜇2

𝜇3

𝜇4

60 %

9 0 %

91 %

7 0 %

Page 48: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Empirical Evaluation

Page 49: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

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

Page 50: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Digits: Negative Image

Page 51: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Digits: Negative Image

𝜇 (𝑥𝑡 )=1−𝑥𝑡

Page 52: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Digits: Negative Image

Page 53: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Digits: Higher Resolution

Page 54: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Digits: Higher Resolution

Page 55: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Digits: Higher Resolution

Page 56: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Wine

Page 57: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Wine

Page 58: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Qualitative ResultsTransfer Learning

TaskTarget

InstanceTarget Sample Size

1 2 5 10

Digits: Negative Image

Digits: Higher Resolution

Page 59: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Discussion

Page 60: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Recap

• Problem: Concept learning with few examples• Solution: Metaphors

Page 61: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Recap

• Problem: Concept learning with few examples• Solution: Metaphors

• Target Source

Page 62: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Recap

• Problem: Concept learning with few examples• Solution: Metaphors

• Target Source• Generic framework

Page 63: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Recap

• Problem: Concept learning with few examples• Solution: Metaphors

• Target Source• Generic framework• Wide range of relations

Page 64: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Recap

• Problem: Concept learning with few examples• Solution: Metaphors

• Target Source• Generic framework• Wide range of relations• Learn the difference

Page 65: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

What if the concepts are not related?

Page 66: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

What if the concepts are not related?

Page 67: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Metaphors are not a measure of relatedness

Page 68: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Metaphors are not a measure of relatedness

Metaphors explain how concepts are related

Page 69: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Vision

Page 70: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology
Page 71: Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology

Explaining how concepts are related since 2012.M E T A P H O R S