Concept Bottleneck Models in the Wild: Assessing Interpretability under Image Corruptions

Concept Bottleneck Models (CBMs) route every prediction through a layer of named, human-auditable concepts, which makes their decisions inspectable. To date, CBMs have only been evaluated in the lab setting, however, when employed in the wild, new challenges may arise: We ask what happens to a deployed CBM when common image corruptions of real sensors alter its inputs. We analyse three CBMs and find that image corruptions reduce model accuracy by $3-67$%, depending on corruption type and severity. Two findings organize the paper. First, image corruption reorders concept activations. While the CLIP embedding moves by only a bounded amount even at the harshest severity (cosine $0.59$ or above), the standardized activations the classifier reads are re-ranked so thoroughly that the single top concept changes in $70$ to $100$% of images. Second, the reordering is uneven across concepts, so some concepts are more stable than others. We propose a per-concept stability score $d_k$. Our experiments show that pruning instable concepts improves accuracy on corrupted images, while having little impact on clean accuracy. Code will be released upon acceptance.