In particular, through use cases on medical and industrial inspection data, we demonstrate that CH effects systematically lead to significant performance loss of downstream models under plausible dataset shifts or reweighting of different data subgroups. Our empirical findings are enriched by theoretical insights, which point to inductive biases in the unsupervised learning machine as a primary source of CH effects. Overall, our work sheds light on unexplored risks associated with practical applications of unsupervised learning and suggests ways to systematically mitigate CH effects, thereby making unsupervised learning more robust.

For example, in image-based industrial inspection, which often relies on unsupervised anomaly detection9,10, we find that a CH decision strategy can systematically miss a wide range of manufacturing defects, resulting in potentially high costs. As another example, unsupervised foundation models of image data, advocated in the medical domain to provide robust features for various specialized diagnostic tasks, can potentially introduce CH effects into many of these tasks, with the prominent risk of large-scale misdiagnosis.

Read the full article at Nature.com.