Managing the unknown: a survey on Open Set Recognition and tangential areas

Although this method has demonstrated its efficacy in numerous scenarios and remains relevant, there is an undeniable shift towards emphasizing autonomy and broader applicability in open scenarios. Consequently, there is a fervent quest for the emergence of a new era of Machine Learning (ML) models characterized by enhanced autonomy and generalization to perform a wide variety of tasks. But most formulations of such tasks still assume a so-called closed set scenario: all samples (or instances) at inference time belong to at least one of the classes existing in the training data from which the ML model was learned. Unfortunately, in many real-world circumstances, this closed set assumption may not necessarily hold, giving rise to open set environments where Unknown Classes (UC) can emerge at testing time. When this occurs, the model must detect the emergence of UC; otherwise, ML models designed under the open set assumption will incorrectly classify instances belonging to UC as any of the known classes (KC), often with a high confidence in their predictions. In this context, the Open Set Recognition (OSR) field has emerged [1] to address this problem by endowing ML models with the capacity to detect (and adapt) their knowledge to the appearance of new classes.