Work-in-Progress: Utilizing Probabilistic Analysis to Fine-Tune Optimal IDK Cascades

In an effort to reduce the runtime of classification algorithms, IDK (I Don’t Know) cascades have been presented as an alternative to current classification models. These structures comprise of a “cascade” of classifiers that only categorizes an input if a classifier outputs a confidence level that exceeds a predetermined threshold. If it does not, it outputs the class “I don’t know” and moves on to the next classifier in the cascade. However, these IDK cascades often reach their worst-case execution time, where classification is only completed by the final classifier. This paper aims to improve the static structure employed by these cascades, deploying a dynamic IDK framework that skips certain classifiers upon meeting specific conditions. A probabilistic analysis run on a previously validated optimal IDK cascade helped us identify these specific conditions, using the outputted confidence level of the first classifier to dictate if a cascade should run its middle classifiers. Ultimately, we generated a dynamic IDK cascade that ran up to 17% faster than its static counterpart.