Learn from Prior Knowledge

Illustrative figure by Shadi Albarqouni

Together with our clinical and industry partners, we realized that there is a need to incorporate domain-specific knowledge and let the model Learn from a Prior Knowledge. We first investigated modeling general priors, i.e., manifold assumptions, to learn powerful representations. Such representations achieved state-of-the-art on benchmark datasets, such as e IDRiD for Diabetic Retinopathy Early Detection (Sarhan et al. 2019), and 7 Scenes for Camera Relocalization (Bui et al. 2017). Then, we started looking into the laplacian graph, where prior knowledge can be modeled as a soft constraint, i.e., regularization, to learn feature representation that follows such manifold defined by graphs. We have shown in our ISBI (Kazi et al. 2019a), MICCAI (Kazi et al. 2019b), and IPMI (Kazi et al. 2019) papers that leveraging prior knowledge such as proximity of ages, gender, and a few lab results, are of high importance in Alzheimer classification.



  • Siemens AG
Shadi Albarqouni
Shadi Albarqouni
Professor of Computational Medical Imaging Research at University of Bonn | AI Young Investigator Group Leader at Helmholtz AI