The functional localization of neural networks using genetic algorithms.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Neural networks offer powerful predictive capabilities but often function as "black boxes."
• Extracting Boolean functions from neural network units aids interpretability.
• Complex extracted Boolean functions can hinder understanding and indicate poor functional localization.

Purpose of the Study:

• To develop and evaluate an algorithm for the functional localization of hidden units in neural networks.
• To improve the understandability of neural network components by approximating hidden units with low-order Boolean functions.

Main Methods:

• An algorithm was developed to extract Boolean functions from trained neural network units.
• Functional localization is achieved by approximating hidden units with low-order Boolean functions.
• Genetic algorithms were employed for optimization, evaluating localization via approximation error.

Main Results:

• The algorithm successfully performs functional localization on neural network hidden units.
• Application to vote, mushroom, and chess datasets demonstrated the algorithm's effectiveness.
• The method enhances the understandability of complex neural network models.

Conclusions:

• The proposed algorithm effectively localizes functions within neural network hidden units.
• This approach contributes to making complex neural networks more interpretable.
• Functional localization is a viable strategy for understanding neural network behavior.