concept bottleneck models (CBMs) assume that training examples (e.g., x-ray
images) are annotated with high-level concepts (e.g., types of abnormalities),
and perform classification by first predicting the concepts, followed by
predicting the label relying on these concepts. The main d