OOD Detectors Are Best Used Runtime Verifiers, Not Semantic Shift Classifiers

Abstract Out-of-distribution (OOD) detection is a widely studied problem in machine learn- ing, and involves identifying inputs that are drawn from a different distribution than what a trained network is intended to model. Conventionally, OOD detectors are evaluated in terms of their capabilities for detecting instances where the inputs are semantically distinct from the training data, such as when a network trained on nu- meric digits encounters letters. In this position paper, we contend that this problem setting significantly undersells the true potential of OOD detectors have, namely as runtime verifiers that detect instances of subtle, semantics-preserving shifts in the covariates of the data that nevertheless adversely impact network accuracy. We base this argument on the fact that OOD detectors effectively measure the degree to which a datum has support in the training distribution, and that this is a necessary condition for a neural network to reliably predict correctly. We support our position empirically through a cost-benefit analysis in a polyp segmentation case study, where we compare the expected lifetime costs per-patient in a system utilizing OOD detectors as runtime verifiers, to a conventionally implemented system. Our results show that implementing OOD detectors as runtime verifiers reduces the expected costs per patient by upwards of 40%. Overall, we position OOD detection as a promising candidate towards endowing deep learning systems with the necessary resilience for responsible deployment in high-stakes applications, and encourage a shift in the focus of OOD detection research to this end.