Learning under privileged information: a survey

Real-world learning scenarios often exhibit asymmetries between training and inference, where additional or higher-quality information is available only during training. The Learning Under Privileged Information (LUPI) paradigm provides a formal framework to study such settings. Despite research interest over the years, the extensive literature on LUPI remains highly fragmented. In this survey, we present a comprehensive and structured overview of research on LUPI and its related formulations. We review 107 papers published between 2009 and 2025 and propose a novel unified taxonomy based on two core dimensions: the underlying machine learning models, spanning both shallow and deep architectures, and the methodological strategies used to integrate privileged information. Additionally, we provide an extensive overview of the various application domains and the particular types of privileged modalities used. We also discuss key insights and open challenges, focusing on how and when privileged information can be effectively leveraged in modern machine learning systems.