On the coherency of quantitative evaluation of visual explanations

Benjamin Vandersmissen, José Oramas

Abstract: Highlights•A comparative study of different visual explanations methods on natural data using quantitative evaluation metrics.•A correlation study of multiple quantitative evaluation protocols used in Explainable AI.•We discover that there are different competing ideas hidden within the evaluation protocols.•We study the impact of sparsity in explanations on the evaluation protocols.•We find that sparsity in explanations has an impact on some of the evaluation methods.