Optimal Synthesis of Fault-Tolerant IDK Cascades for Real-Time Classification

An IDK classifier is a computational element that classifies an input provided to it into one of a set of predefined categories provided that it can achieve the necessary confidence level to do so; otherwise, it outputs “I Don't Know” (IDK). The concept of IDK classifier cascades has emerged as a strategy for striking a balance between the requirements of rapid response and precise classification in machine perception. Effective algorithms for constructing IDK classifier cascades have recently been developed. Here we extend these prior approaches by incorporating fault-tolerance: enabling classification that is concurrently rapid and accurate even in the event of some of the IDK classifiers exhibiting faulty behavior.