Accelerating Fault-Tolerant Real-Time Classification with IDK Classifiers

We consider the problem of rapid, fault-tolerant classification using IDK classifiers—components that return a class label or 'I Don't Know' if confidence is insufficient. Multiple such classifiers may be available, each with different speed-accuracy trade-offs. Prior work has developed scheduling algorithms to minimize expected classification duration under classifier faults. We present improved schemes that achieve lower expected classification durations while maintaining fault tolerance.