Position Paper - Bringing Classifiers into Critical Systems: Are We Barking up the Wrong Tree?

Domain experts are desperately looking to solve classification problems by designing and training Machine Learning algorithms with the highest possible accuracy. No matter how hard they try, classifiers will always be prone to misclassifications due to a variety of reasons that make the decision boundary unclear. This complicates the integration of classifiers into critical systems and infrastructures, where misclassifications could directly impact the health of people, infrastructures, or the environment. Differently, a classifier should be considered as a component to be deployed into a system, and never in isolation. This provides more flexibility to the classifier, which can even afford to reject those outputs that are likely to be misclassifications, triggering system-level mitigation strategies instead. The resulting fail-controlled classifier will output a noticeably lower amount of misclassifications, making this system-level conceptualization and design of ML classifiers an actual step toward the deployment of classifiers in real, critical systems. Evaluation metrics should be adapted to cope with this paradigm change, scoring rejections differently from predictions, which are either correct or incorrect.