Open set classification using tolerance intervals

This paper presents a classification methodology that addresses the problem of open set classification. For our purposes here, we define open set classification as classification of data from signal classes that were not part of the training set data (the closed set). Classifiers that are not designed to account for this eventuality will often attempt to assign the received signal to one of the training set classes, potentially resulting in misclassification. We propose a method for overcoming this problem, based on tolerance interval analysis, that may be efficiently implemented in ASICs or FPGAs. In addition, our method leads naturally to a "probabilistic maximum level of significance" interpretation of type I errors for those signals that are falsely rejected from one of the training set classes.