Logic Explained Networks

The large and still increasing popularity of deep learning clashes with a major limit neural network architectures, that consists in their lack capability providing human-understandable motivations decisions. In situations which the machine is expected to support decision human experts, comprehensible explanation feature crucial importance. language used communicate explanations must be formal enough implementable friendly understandable by wide audience. this paper, we propose general approach Explainable Artificial Intelligence case showing how mindful design networks leads family interpretable models called Logic Explained Networks (LENs). LENs only require inputs predicates, they provide terms simple First-Order (FOL) formulas involving such predicates. are cover number scenarios. Amongst them, consider directly as special classifiers being explainable, or when act additional role creating conditions for making black-box classifier explainable FOL formulas. Despite supervised problems mostly emphasized, also show can learn unsupervised settings. Experimental results on several datasets tasks may yield better classifications than established white-box models, trees Bayesian rule lists, while more compact meaningful explanations.