Attribution Explanations for Deep Neural Networks: A Theoretical Perspective.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Explainable AI (XAI)

Background:

• Attribution explanation methods aim to quantify input variable contributions to deep neural network (DNN) predictions.
• A fundamental challenge is the lack of faithfulness, questioning if these methods truly reflect DNN decision-making logic.

Purpose of the Study:

• To provide a comprehensive review of recent theoretical advances in attribution methods for DNNs.
• To address the core challenges of theoretical unification, rationale, and principled evaluation of attribution methods.

Main Methods:

• Reviewing theoretical studies on attribution methods.
• Focusing on three directions: theoretical unification, theoretical rationale, and theoretical evaluation.
• Analyzing commonalities, differences, mathematical justifications, and faithfulness principles.

Main Results:

• Identified key advances in theoretically unifying attribution methods.
• Clarified mathematical and conceptual justifications for existing methods.
• Established principled methods for evaluating attribution faithfulness.

Conclusions:

• Theoretical studies are crucial for understanding and improving attribution methods.
• Practical recommendations are provided for method design, selection, and usage based on theoretical findings.
• Open problems and future research directions in attribution explanation are discussed.