PhD Oral Exam - Mohammad Ali Salari, Computer Science

When studying for a doctoral degree (PhD), candidates submit a thesis that provides a critical review of the current state of knowledge of the thesis subject as well as the student’s own contributions to the subject. The distinguishing criterion of doctoral graduate research is a significant and original contribution to knowledge.

Once accepted, the candidate presents the thesis orally. This oral exam is open to the public.

Abstract

Variations of computational infrastructures, including operating systems, software versions, and hardware architectures, introduce variability in neuroimaging analyses that could affect the reproducibility of scientific conclusions. This is due to the creation, propagation, and amplification of numerical instabilities in analysis pipelines. In this regard, it is critical to identify numerical instabilities to make experiments computationally reproducible. In this thesis, we characterize the numerical stability of commonly-used complex pipelines in the context of neuroimaging analysis across the operating systems and provide accessible tools for developers and researchers to evaluate their pipelines and findings. First, we present the Spot tool that identifies the processes from which differences originate and the path along which they propagate in the pipelines. In the next step, to study the numerical instabilities more comprehensively, we introduce controlled numerical perturbations to the floating-point computations using the Monte-Carlo Arithmetic method. We propose an interposition technique to model the effect of operating system updates on analysis pipelines using the Monte-Carlo arithmetic. Finally, leveraging the interposition technique, we compare numerical variability with tool variability in an fMRI analysis. All the methods implemented in this thesis can be used to facilitate further investigations toward stabilizing pipelines.