Optimally ordering IDK classifiers subject to deadlines

Abstract A classifier is a software component, often based on Deep Learning, that categorizes each input provided to it into one of fixed set classes. An IDK may additionally output “I Don’t Know” (IDK) for certain inputs. Multiple distinct classifiers be available the same classification problem, offering different trade-offs between effectiveness, i.e. probability successful classification, and efficiency, execution time. Optimal offline algorithms are proposed sequentially ordering such expected duration successfully classify an minimized, optionally subject hard deadline maximum time permitted classification. Solutions considering independent dependent relationships pairs classifiers, as well mix two.