Quantitative and Ontology-Based Comparison of Explanations for Image Classification

Abstract

Deep Learning models have recently achieved incredible performances in the Computer Vision field and are being deployed in an ever-growing range of real-life scenarios. Since they do not intrinsically provide insights of their inner decision processes, the field of eXplainable Artificial Intelligence emerged. Different XAI techniques have already been proposed, but the existing literature lacks methods to quantitatively compare different explanations, and in particular the semantic component is systematically overlooked. In this paper we introduce quantitative and ontology-based techniques and metrics in order to enrich and compare different explanations and XAI algorithms.