PhD Oral Exam - Inés Gonzalez Pepe, 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

Numerical variability, defined as the variability in computational results caused by the propagation of floating-point rounding errors, can alter statistical conclusions derived from neuroimaging-derived measures, and potentially affect downstream clinical studies. Convolutional neural networks (CNNs) are widely used in neuroimaging, yet their numerical behaviour during both training and inference remains poorly understood. This thesis investigates numerical variability in CNN for neuroimaging, haracterizing it experimentally and developing methods to leverage it to improve computational efficiency and predictive accuracy. During inference, we show that neuroimaging CNNs are substantially more stable than their non-deep-learning counterparts, and that many computations involve values that do not affect model outputs. Based on this observation, we introduce methods that skip these computations, avoiding up to a third of convolutions without measurable accuracy loss. To study CNN training efficiently, we developed Fuzzy PyTorch, a framework that achieves 5 − 60× faster execution than existing tools for numerical variability quantification. Using Fuzzy PyTorch, we provide the first quantification of numerical variability in CNN training. We find that training variability matches or exceeds that observed in non-deep-learning pipelines, and that retraining a CNN can produce distinct yet anatomically plausible brain measures. We further show that this variability can be reproduced at low cost through random seeds and leveraged through model ensembling to improve brain-age prediction accuracy. Numerical variability is thus low at inference and high in training, and accounting for it, by skipping computations in the former and augmenting data in the latter, can make CNN-based neuroimaging more reliable and efficient.