Autoassociative Learning of Structural Representations for Modeling and Classification in Medical Imaging

Annotation of medical imaging is notoriously time-consuming, prone to human biases, and hard to reconcile with the insatiable demands of contemporary machine learning. Deep Learning (DL) models trained on annotated data are often narrow in focusing on features that are specific to a given context (anomaly, pathology, etc.) rather than discovering and capturing general characteristics of observed structures and processes, which may make them susceptible to deceptive image features and lead to inferior generalization. We posit that one of the primary causes of this challenge is the unstructured character of DL architectures. Contemporary DL models are essentially intertwined compositions of dot products and nonlinearities, conglomerates of often billions of unsophisticated units that process data in a highly distributed and continuous, non-symbolic fashion. Their training requires large volumes of data, which are often hard to come by, and involves exorbitant amounts of compute and energy. If the task is posed within the supervised learning paradigm, those data need to be not only curated, but also annotated (labeled), which limits their availability even further. Last but not least, as each processing unit takes care only of a minuscule fraction of inference, it is very hard to explain the model and its decisions to a human in a transparent and succinct fashion. In this study, we argue for stronger involvement of unlabeled data in the construction of analytic and diagnostic ML models and propose ASR, a neurosymbolic architecture trained to form Auto-associative Structural Representations, in which a generative decoder synthesizes physically plausible structural models that explain the observed image.