Ontology-Based Post-Hoc Explanations via Simultaneous Concept Extraction*

Abstract

Ontology-based explanation techniques allow one to get explanation why a neural network arrived to some conclusion using human-understandable terms and their formal definitions. The paper proposes a method to build post-hoc ontology-based explanations by training a multi-label neural network mapping the activations of the specified "black box" network to ontology concepts. In order to simplify training of such network we employ semantic loss, taking into account relationships between concepts. The experiment with a synthetic dataset shows that the proposed method can generate accurate ontology-based explanations of a given network.