Traditional neural networks exhibit excellent classification performance, but suffer from limitations in verifying and extracting their learning content. This paper proposes a neural logic network (Neural Logic Network) with an interpretable structure that learns the logical mechanism between inputs and outputs through AND and OR operations. We generalize the network by adding a NOT operation and a bias that accounts for unobserved data, and develop rigorous logical and probabilistic modeling in terms of concept combinations to facilitate its use. Furthermore, we propose a novel factorized IF-THEN rule structure and a modified learning algorithm. The proposed method improves on the state-of-the-art in Boolean network discovery and can learn relevant and interpretable rules, particularly in tabular classification in medical and industrial fields where interpretability is crucial.
