Purpose The quantitative detection of failure modes is important for making deep neural networks reliable and usable at scale. We …