Teaching Machines to Learn by Metaphors Omer Levy & Shaul Markovitch Technion – Israel Institute of Technology
Concept Learning by Induction
Few Examples
Transfer Learning Target (New) Source (Original)
Define: Related Concept
Transfer Learning Approaches Common Inductive Bias Common Instances Common Features
Different Feature Space
Example
Example
Example
Common Inductive Bias
Common Inductive Bias
Common Instances
Common Features
New Approach to Transfer Learning
Our Solution: Metaphors
Metaphors Target (New) Source (Original)
Concept Learner Metaphor Learner Source Target +/-
Theorem
The Metaphor Theorem
Redefine Transfer Learning
Metaphor Learning Framework
Concept Learning Framework Search Algorithm Hypothesis Space Evaluation Function Data
Source Target Metaphor Learning Framework Search Algorithm Metaphor Space Evaluation Function
Metaphor Evaluation
Metaphor Spaces
General Few Degrees of Freedom Representation-Specific Bias
Geometric Transformations ЯR
Dictionary-Based Metaphors cheesequeso
Linear Transformations
Which metaphor space should I use?
Automatic Selection of Metaphor Spaces Which metaphor space should I use?
Occam’s Razor Automatic Selection of Metaphor Spaces Which metaphor space should I use?
Structural Risk Minimization Occam’s Razor Automatic Selection of Metaphor Spaces Which metaphor space should I use?
Automatic Selection of Metaphor Spaces
Empirical Evaluation
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
Digits: Negative Image
Digits: Higher Resolution
Wine
Qualitative Results Transfer Learning Task Target Instance Target Sample Size Digits: Negative Image Digits: Higher Resolution
Discussion
Recap Problem: Concept learning with few examples Solution: Metaphors
Recap
What if the concepts are not related?
Metaphors are not a measure of relatedness
Metaphors are not a measure of relatedness Metaphors explain how concepts are related
Vision
Explaining how concepts are related since M E T A P H O R S
Concept Learning by Induction
Few Examples
Approaches Explanation-Based Learning Semi-Supervised Learning Transfer Learning
Explanation-Based Learning Axioms Data Logical Deduction
Semi-Supervised Learning
Transfer Learning
Target (New) Source (Original)
Transfer Learning
Target (New) Source (Original)
Define: Related Concept
Transfer Learning Approaches Common Inductive Bias Common Instances Common Features
Common Inductive Bias
Common Instances
Common Features 1.Perform feature selection on source 2.Use that selection on target
Which definition is better?
Different Feature Space
Example
Example
Example
Common Inductive Bias
Common Inductive Bias
Common Instances
Common Features
Our Solution: Metaphors
Performance with Automatic Selection of Metaphor Spaces Digits: Negative Image
Performance with Automatic Selection of Metaphor Spaces Digits: Negative Image Geometric Transformations Feature Reordering Orthogonal Linear Transformations Orthogonal Quadratic Transformations
Performance with Automatic Selection of Metaphor Spaces Digits: Negative Image Geometric Transformations Feature Reordering Orthogonal Linear Transformations Orthogonal Quadratic Transformations
What if I have more than one source?
Multiple Source Datasets B Я HRZ
B Я H R Z
Я R
Performance with Multiple Source Datasets Latin & Cyrillic
Performance with Multiple Source Datasets Latin & Cyrillic ABCDEFG HIJKLMN OPQRSTU VWXYZ
Performance with Multiple Source Datasets Latin & Cyrillic ABCDEFG HIJKLMN OPQRSTU VWXYZ
Performance with Multiple Source Datasets