Several efforts have been made to synthesize semi-supervised learning (SSL) and open set recognition (OSR) within a single training policy. However, each attempt violated the definition of an open set by incorporating novel categories within the unlabeled training set. Although such \textit{observed} novel categories are undoubtedly prevalent in application-grade datasets, they should not be conflated with the OSR-defined \textit{unobserved} novel categories, which only emerge during testing. This study proposes a new learning policy wherein classifiers generalize between observed and unobserved novel categories. Specifically, our open-set learning with augmented category by exploiting unlabeled data (Open-LACU) policy defines a background category for observed novel categories and an unknown category for unobserved novel categories. By separating these novel category types, Open-LACU promotes cost-efficient training by eliminating the need to label every category and ensures safe classification by completely separating unobserved novel categories that appear over time. Finally, we present a unified approach to establish benchmark results for this emerging and more application-grade learning policy.

翻译：已有若干研究尝试将半监督学习（SSL）与开放集识别（OSR）整合到单一训练策略中。然而，这些方法均因在未标记训练集中纳入新类别而违背了开放集的定义。尽管此类*可观测*新类别在应用级数据集中无疑普遍存在，但不应将其与OSR定义的仅在测试阶段出现的*未观测*新类别混为一谈。本研究提出一种新的学习策略，使分类器能够在可观测与未观测新类别之间进行泛化。具体而言，我们的利用未标记数据扩展类别的开放集学习（Open-LACU）策略为可观测新类别定义了背景类别，为未观测新类别定义了未知类别。通过分离这两类新类别，Open-LACU无需对每个类别进行标注即可实现经济高效的训练，同时通过完全分离随时间出现的未观测新类别来确保安全分类。最后，我们提出一种统一方法，为这一新兴且更具应用价值的学习策略建立基准结果。