Interpretability Is in the Mind of the Beholder: A Causal Framework for Human-Interpretable Representation Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Human-Computer Interaction

Background:

• Explainable AI (XAI) research is shifting towards concept-based explanations.
• Current methods for acquiring interpretable concepts lack standardization and neglect human understanding.
• A key challenge is modeling the human element in representation learning.

Purpose of the Study:

• To propose a mathematical framework for acquiring interpretable representations.
• To bridge the gap between human and algorithmic interpretability in AI.
• To establish a foundation for future research in human-interpretable representation learning.

Main Methods:

• Developed a formalization of human-interpretable representation learning (hrl).
• Integrated causal representation learning principles.
• Modeled a human stakeholder as an external observer to define alignment.

Main Results:

• Derived a principled notion of alignment between machine representations and human concept vocabularies.
• Linked alignment and interpretability via a name transfer game.
• Clarified relationships between alignment, disentanglement, concept leakage, and content-style separation.

Conclusions:

• The proposed framework offers a principled approach to human-interpretable representation learning.
• Alignment is a crucial factor for creating AI representations understandable by humans.
• This work provides a stepping stone for advancing AI interpretability research.