Towards Explaining Image-Based Distribution Shifts

Abstract

Distribution shift can have fundamental consequences such as signaling a change in the operating environment or significantly reducing the accuracy of downstream models. Thus, understanding such distribution shifts is critical for examining and hopefully mitigating the effect of such a shift. Most prior work has focused on either natively handling distribution shift (e.g., Domain Generalization) or merely detecting a shift while assuming any detected shift can be understood and handled appropriately by a human operator. For the latter, we hope to aid in these manual mitigation tasks by explaining the distribution shift to an operator. To this end, we suggest two methods: providing a set of interpretable mappings from the original distribution to the shifted one or providing a set of distributional counterfactual examples. We provide preliminary experiments on these two methods, and discuss important concepts and challenges for moving towards a better understanding of image-based distribution shifts.