What is Interpretability?

Area of Science:

• Artificial Intelligence
• Machine Learning
• Philosophy of Science
• Medical Artificial Intelligence

Background:

• Theoretical engagements with artificial neural networks (ANNs) raise key questions about their explainability and interpretability.
• Existing accounts of 'explanation' for ANNs often lack novelty, potentially reinventing established philosophical concepts.
• Confusion exists in machine learning literature regarding the distinct meanings of 'explainability,' 'understandability,' and 'interpretability.'

Purpose of the Study:

• To clarify the concepts of explainability and interpretability in the context of artificial neural networks.
• To propose a novel theory and typology for interpretation in machine learning.
• To address the specific challenges and applications within medical artificial intelligence.

Main Methods:

• Applied four established philosophical accounts of explanation to artificial neural networks.
• Diagnosed the equivocation of terms like 'explainability,' 'understandability,' and 'interpretability' in machine learning literature.
• Developed a theory and typology of interpretation, defining it as a process to create more understandable explanations.

Main Results:

• Demonstrated that familiar philosophical accounts of explanation are applicable to ANNs.
• Distinguished between explainability, understandability, and interpretability, resolving conceptual confusion.
• Proposed a framework for interpretation categorizing methods as Total/Partial, Global/Local, and Approximative/Isomorphic.

Conclusions:

• Artificial neural networks are fundamentally explainable using established scientific explanation frameworks.
• A clear distinction between explainability, understandability, and interpretability is crucial for advancing machine learning.
• The proposed theory of interpretation offers a structured approach consistent with both machine learning practices and philosophy of science, with implications for medical AI.