Extracting Explainability from Multi-Layer Perceptrons for Binary Classification Problems through Łukasiewicz Logic

Multi-Layer Perceptrons are powerful tools for classification, but their lack of transparency poses a significant challenge for in-terpretability. In this paper, we propose a methodology for extracting human-readable explanations from neural networks using Łukasiewicz logic. Our architecture is designed to approximate logical minimum and maximum operations, allowing the model to be trained in a standard way. The final output of the network can be expressed as a disjunctive normal form formula, offering insight about when a class is predicted and what feature constraints apply. We validate our methodology on simple classification problems, focusing on explainability with a reasonable performance.