High-capacity deep learning models have achieved state-of-the-art predictive accuracy across diverse classification tasks, yet they frequently operate as black-box models, lacking the transparency necessary for high-stakes decision making. Such opacity creates a persistent trade-off between performance and explainability. This paper proposes a solution to address this gap: a knowledge distillation framework that results in globally explainable Boolean rule-sets from neural network models. Adapting the EVOTER symbolic infrastructure, this research treats neural networks as targets for the evolution process, distilling their performance into minimal sets of logical expressions. There are three primary contributions: (1) an evolutionary method for distilling black-box neural network models into explicit rule-set models; (2) a method for making rule sets more explainable by including a brevity objective to evolution; and (3) a demonstration that the distillation is viable even without access to ground truth data. The paper thus establishes that black-box neural network models can be made explainable and therefore useful in real-world applications where trustworthiness is paramount.

PDF

Video