Entropy-Based Logic Explanations of Neural Networks

Explainable artificial intelligence has rapidly emerged since lawmakers have started requiring interpretable models for safety-critical domains. Concept-based neural networks arisen as explainable-by-design methods they leverage human-understandable symbols (i.e. concepts) to predict class memberships. However, most of these approaches focus on the identification relevant concepts but do not provide concise, formal explanations how such are leveraged by classifier make predictions. In this paper, we propose a novel end-to-end differentiable approach enabling extraction logic from using formalism First-Order Logic. The method relies an entropy-based criterion which automatically identifies concepts. We consider four different case studies demonstrate that: (i) enables distillation concise in domains clinical data computer vision; (ii) proposed outperforms state-of-the-art white-box terms classification accuracy.