ECLAD: Extracting Concepts with Local Aggregated Descriptors

Andres Felipe Posada-Moreno, Nikita Surya, Sebastian Trimpe

Abstract: Highlights•we propose the concept extraction and localization method ECLAD.•Concepts extracted with ECLAD are closely related to the relevant features in models.•ECLAD outperforms state of the art (ACE and ConceptShap) in our experiments.•ECLAD explanations provide valuable insights in real-world use cases.•We propose a process for the quantitative evaluation and comparison of CE methods.