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 -3 -2 0 2 3 Example -3 -2 0 0 2 3 4 9 Example -3 -2 0 𝑥𝑠 = 0 2 3 2 𝑥𝑡 4 9 Common Inductive Bias -3 -2 0 0 2 3 4 9 Common Inductive Bias -3 -2 0 0 2 3 4 9 Common Instances -3 -2 0 0 2 3 4 9 Common Features 3 2 4 -2 -3 𝑥𝑠 = 2 𝑦𝑡 9 New Approach to Transfer Learning Our Solution: Metaphors Metaphors Target (New) Source (Original) Source Concept Learner Target Metaphor Learner 𝑥𝑡 𝜇 𝑥𝑠 ℎ𝑠 +/- ℎ𝑡 𝑥𝑡 = ℎ𝑠 𝜇 𝑥𝑡 𝜇 is a perfect metaphor if: 1. 𝜇 is label preserving 𝑓𝑡 𝑥𝑡 = 𝑓𝑠 𝜇 𝑥𝑡 2. 𝜇 is distribution preserving 𝑥𝑡 ~𝑃𝑡 ⇒ 𝜇 𝑥𝑡 ~𝑃𝑠 Theorem If 𝜇 is a perfect metaphor - and ℎ𝑠 is a source hypothesis with 𝜀𝑠 error - then ℎ𝑡 𝑥𝑡 = ℎ𝑠 𝜇 𝑥𝑡 is a target hypothesis with 𝜀𝑠 error The Metaphor Theorem If 𝜇 is an 𝜀-perfect metaphor - and ℎ𝑠 is a source hypothesis with 𝜀𝑠 error - then ℎ𝑡 𝑥𝑡 = ℎ𝑠 𝜇 𝑥𝑡 is a target hypothesis with 𝜀 + 𝜀𝑠 error Redefine Transfer Learning Given source and target datasets, find a target hypothesis ℎ𝑡 such that 𝜀𝑡 is as small as possible. Redefine Transfer Learning Given source and target datasets, find an 𝜀-perfect metaphor 𝜇 such that 𝜀 is as small as possible. Metaphor Learning Framework Concept Learning Framework Search Algorithm Hypothesis Space ℎ Evaluation Function Data Metaphor Learning Framework Source Search Algorithm Metaphor Space 𝜇 Evaluation Function Target Metaphor Evaluation Metaphor Evaluation 1. 𝜇 is label preserving 𝑓𝑡 𝑥𝑡 = 𝑓𝑠 𝜇 𝑥𝑡 2. 𝜇 is distribution preserving 𝑥𝑡 ~𝑃𝑡 ⇒ 𝜇 𝑥𝑡 ~𝑃𝑠 Metaphor Evaluation 1. 𝜇 is label preserving Empirical error over target dataset 2. 𝜇 is distribution preserving Statistical distance between 𝜇 𝑥𝑡 and 𝑥𝑠 Metaphor Evaluation 𝑆𝐷 𝜇 𝑥𝑡 , 𝑥𝑠 Metaphor Evaluation 𝑆𝐷 𝜇 + 𝑥𝑡 + , 𝑥𝑠 Metaphor Evaluation 𝑆𝐷 𝜇 − 𝑥𝑡 − , 𝑥𝑠 Metaphor Evaluation 𝑆𝐷 𝜇 + 𝑥𝑡 + , 𝑥𝑠 + 𝑆𝐷 𝜇 − 𝑥𝑡 − , 𝑥𝑠 Metaphor Spaces Metaphor Spaces • General • Few Degrees of Freedom • Representation-Specific Bias Geometric Transformations Dictionary-Based Metaphors cheese queso Linear Transformations 𝜇 𝑥 =𝑤 ⋅𝑥 +𝑣 Which metaphor space should I use? Which metaphor space should I use? Automatic Selection of Metaphor Spaces Which metaphor space should I use? Automatic Selection of Metaphor Spaces Occam’s Razor Which metaphor space should I use? Automatic Selection of Metaphor Spaces Occam’s Razor Structural Risk Minimization Automatic Selection of Metaphor Spaces ℳ1 ℳ2 ℳ3 ℳ4 Automatic Selection of Metaphor Spaces ℳ1 𝜇1 ℳ2 𝜇2 ℳ3 𝜇3 ℳ4 𝜇4 Automatic Selection of Metaphor Spaces ℳ1 𝜇1 60% ℳ2 𝜇2 90% ℳ3 𝜇3 91% ℳ4 𝜇4 70% 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: Negative Image 𝜇 𝑥𝑡 = 1 − 𝑥𝑡 Digits: Negative Image Digits: Higher Resolution Digits: Higher Resolution → Digits: Higher Resolution Wine Wine Qualitative Results Transfer Learning Target Task Instance Digits: Negative Image Digits: Higher Resolution Target Sample Size 1 2 5 10 Discussion Recap • Problem: Concept learning with few examples • Solution: Metaphors Recap • Problem: Concept learning with few examples • Solution: Metaphors • Target → Source Recap • Problem: Concept learning with few examples • Solution: Metaphors • Target → Source • Generic framework Recap • Problem: Concept learning with few examples • Solution: Metaphors • Target → Source • Generic framework • Wide range of relations Recap • Problem: Concept learning with few examples • Solution: Metaphors • • • • Target → Source Generic framework Wide range of relations Learn the difference What if the concepts are not related? 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 M E TAPH O R S Explaining how concepts are related since 2012.