Combining an explainable model based on ontologies with an explanation interface to classify images
Matthieu Bellucci, Nicolas Delestre, Nicolas Malandain, Cécilia Zanni-Merk · Procedia Computer Science · 2022
Numerous explainability methods have appeared thanks to the surge of popularity of the explainable AI (XAI) domain. The DARPA depicted an explainable system, where an explainable model interacts with an explanation interface to generate explanations adapted to a user. We propose an explainable image classification system that follows this combination of explainable model and explanation interface as described by the DARPA. It takes advantage of the explainability of ontologies as well as the performance of machine learning models. This system is able to predict the class and properties of an object in an image. The results of this classification system are displayed in an explanation interface to help users understand and analyze the predictions of the proposed system. Our system exploits an ontology to build models that classify an object and its properties. The class and properties predicted by these models are instantiated in the ontology and added as assertions to an individual in order to verify the consistency of these predictions. Therefore, the system is able to warn the user when a prediction is uncertain and explain why, by using the ontology. These capacities will help users trust this system and better understand the predictions.