Computer vision datasets frequently contain spurious correlations between task-relevant
labels and (easy to learn) latent task-irrelevant attributes (eg context). Models trained on …