Unifying Fourteen Post-Hoc Attribution Methods With Taylor Interactions

Area of Science:

• Artificial Intelligence
• Machine Learning
• Explainable AI

Background:

• Deep neural networks (DNNs) are powerful but lack transparency.
• Numerous attribution methods exist to explain DNNs, but they lack a unified theoretical foundation.
• Comparing the effectiveness and relationships between different attribution methods is challenging.

Purpose of the Study:

• To provide a unified theoretical understanding of existing deep neural network attribution methods.
• To reveal the core mechanism shared by fourteen distinct attribution methods.
• To propose new principles for evaluating and comparing attribution methods.

Main Methods:

• Mathematical reformulation of attribution scores using Taylor interactions.
• Analysis of attribution methods as weighted allocations of independent and interaction effects.
• Development of three principles for fair effect allocation.

Main Results:

• Demonstrated that fourteen diverse attribution methods share a common underlying mechanism.
• Showed that attribution scores are mathematically equivalent to weighted allocations of independent and interaction effects.
• Identified method differences based on the weights assigned to these effects.

Conclusions:

• A unified perspective on fourteen deep neural network attribution methods has been established.
• Essential similarities and differences among these methods are theoretically clarified.
• Proposed principles offer a fair and direct comparison framework for attribution methods.