Optimal Synthesis of IDK-Cascades
Sanjoy Baruah, Alan Burns, Yue Wu · 2021
A classifier is a software component, often based upon deep learning (DL), that categorizes each input provided to it into one of a fixed set of “classes”. An IDK classifier may additionally output an “I don’t know” (IDK) on certain input. Given several different IDK classifiers for the same operation, the problem is considered of using them in concert in such a manner that the average duration to successfully classify any input is minimized. Optimal algorithms are proposed for solving this problem, both as is and under an additional constraint that the operation must be completed within a specified hard deadline).