Toward global interpretability of neural networks via Boolean transformation

Area of Science:

• Artificial Intelligence
• Computer Science
• Machine Learning

Background:

• Deep neural networks (DNNs) present interpretability challenges, especially in critical applications.
• Existing methods often struggle to provide global, transparent views of complex network architectures.

Purpose of the Study:

• To develop a unified framework for post-hoc global interpretability of general neural network architectures.
• To transform neural networks into equivalent decision diagrams for transparent analysis.

Main Methods:

• Proposed a novel framework to convert neural networks (e.g., Residual Networks, Transformers) into equivalent decision diagrams.
• Introduced a modified node merging process for efficient diagram construction, reducing size while preserving equivalence.
• Utilized decision diagrams for exploring logical properties like decision boundary tracing and robustness analysis.

Main Results:

• Successfully transformed various neural network architectures into interpretable decision diagrams.
• Demonstrated significant reduction in diagram size through an optimized merging process, enhancing efficiency.
• Validated the framework's effectiveness and scalability for neural network analysis and verification.

Conclusions:

• The proposed decision diagram framework offers a globally interpretable surrogate for deep neural networks.
• This approach facilitates reliable analysis, verification, and understanding of complex AI models.
• The method holds significant potential for safety-critical and decision-sensitive AI applications.