Extracting boolean and probabilistic rules from trained neural networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Neural networks composed of linear threshold functions are foundational in AI.
• Extracting interpretable rules from these networks remains a challenge.
• Existing methods may produce overly complex or limited rule sets.

Purpose of the Study:

• To develop and evaluate two distinct algorithms for rule extraction from trained neural networks.
• To enhance the conciseness and interpretability of extracted rules.
• To explore the extraction of both deterministic (Boolean) and probabilistic rules.

Main Methods:

• Algorithm 1: Rule extraction into Boolean functions, optimized for specific network structures (1-decision lists, majority functions).
• Algorithm 2: Rule extraction into probabilistic rules using dynamic programming, with efficiency considerations for networks with constant-width hidden layers.
• Computational experiments to validate the effectiveness of both approaches.

Main Results:

• The first approach generates significantly more concise Boolean rules for certain network architectures compared to existing methods.
• The second approach successfully extracts probabilistic rules, revealing input-output relationships.
• Both methods demonstrated effectiveness in computational evaluations.

Conclusions:

• The presented algorithms offer improved methods for rule extraction from linear threshold function neural networks.
• These approaches contribute to making complex neural network models more transparent and understandable.
• The findings suggest potential for enhanced interpretability in AI systems.