Workflow from Scientific Research

A benchmark analysis to see how well MA-MONET can identify the semantically meaningful concepts that lead to model error. To this end, we generated settings where we knew the ground truth (that is, concepts that lead to model errors); we created a training and a test dataset with spurious correlation (corr.). We used MA-MONET to identify which concepts led to model error for an AI model trained on this confounded dataset. MA-MONET returned a ranked list of concepts that explain model errors.

DOI