Unified machine learning: Classification with simultaneous observed and unobserved novelty detection

A unified approach of Positive and Unlabelled (PU)-learning, Semi-Supervised Learning (SSL), and Open-Set Recognition (OSR) would significantly enhance the development of cost-efficient application-grade classifiers. However, previous attempts have conflated the definitions of \mbox{\textit{observed}} and \mbox{\textit{unobserved}} novel categories. Observed novel categories are defined in PU-learning as those in unlabelled training data and exist due to an incomplete set of category labels for the training set. In contrast, unobserved novel categories are defined in OSR as those that only exist in the testing data and represent new and interesting patterns that emerge over time. To maintain safe and practical classifier development, models must generalise the difference between these novel category types. In this letter, we thoroughly review the relevant machine learning research fields to propose a new unified machine learning policy called Open-set Learning with Augmented Categories by exploiting Unlabelled data or Open-LACU. Specifically, Open-LACU requires models to accurately classify $K > 1$ number of labelled categories while simultaneously detecting and separating observed novel categories into the augmented background category ($K + 1$) and further detecting and separating unobserved novel categories into the augmented unknown category ($K + 2$). Open-LACU is the first machine learning policy to generalise observed and unobserved novel categories. The significance of Open-LACU is also highlighted by discussing its application in semantic segmentation of remote sensing images, object detection within medical radiology images and disease identification through cough sound analysis.