Haraldur Hallgrimsson successfully defended his PhD on the topic of "Towards Interpretable Models of Health". He will join Apple as an Applied Research Scientist.

Committee: Ambuj Singh (Co-chair), Scott Grafton (Co-chair), Subhash Suri, Xifeng Yan

Abstract: Advances in machine learning have achieved promising results in recent years in predictive tasks, especially related to health data. However, the widespread use of such models is limited by the notion that such models are 'black box' algorithms whose performance is difficult to trust as they do not adequately explain their decisions.

During my degree, I developed methods to not just achieve high performance on singular metrics but also provide further descriptions of the health data itself. In this talk I will focus on my work towards characterizing regions of white matter connectivity, as imaged by diffusion-weighted MRI brain scans. In particular, I will discuss two dual problems: First, how informative are characteristics of an individual for a generative model to synthesize a brain region, including how that informativeness varies across brain regions. Secondly, discovering those brain regions that are most predictive of a given trait across a population. Developing methods to tackle these problems in a small-sample and high-dimensionality data set is challenging, but recent advances in both imaging fidelity as well as methodologies enable this study of white matter connectivity.