July 28 at 2:30 p.m.
Doctoral Programme | Computer Science - MAP joint programme
Defense | Statistical and Geometrical priors for Deep Learning with applications in Biomedical modalities
Student | Miguel Lopes Martins
Date: July 28
Time: 2:30 p.m.
Venue: Room FC6 0.29
President:
Pedro Gabriel Dias Ferreira
Full Professor
Department of Computer Science, Faculty of Sciences, University of Porto
Examiners:
Pietro Cerveri
Full Professor
Dipartimento di Ingegneria Industriale e dell'Informazione, Università di Pavia (Itália)
Giulio Rossolini
Assistant Professor
ReTis - Real Time Systems Laboratory, Scuola Universitaria Superiore Sant'Anna di Pisa (Itália)
Committee Members:
Jaime dos Santos Cardoso
Full Professor
Department of Electrical and Computer Engineering, Faculty of Engineering, University of Porto
Francesco Renna
Assistant Professor
Department of Computer Science, Faculty of Sciences, University of Porto
António Joaquim da Silva Teixeira
Full Professor
Department of Electronics, Telecommunications and Informatics, University of Aveiro
Abstract:
The advent of Deep Learning has fundamentally transformed the landscape of artificial intelligence, offering universal function approximation capabilities that solve complex tasks with unprecedented accuracy. However, this engineering revolution — driven largely by scale and gradient descent — faces critical limitations regarding interpretability, stability, and generalisation, particularly in high-stakes environments such as those involving biomedical data. This dissertation focuses on the role of domain knowledge as a necessary condition for overcoming the fragility of fully data-driven methods.
We make the learning process itself depend on explicit parameterisations that encode different statistical or geometric priors for the domain, thus defining the so-called hy brid or model-based deep learning methods. We first address the challenge of patholog ical detection in medical imaging by leveraging principles from (multi)fractal geometry. We articulate the self-seeding and cascading nature of pathological tissue abnormalities, showing that optimisation informed by these geometrical priors can guide networks to capture textural regularities that purely convolutional approaches may overlook. We also analyse regularities across time, and propose a statistical prior for sequential data. By jointly optimising a deep neural network with a Markovian prior, we demonstrate that this yields superior performance and robustness compared to unconstrained sequence modelling, highlighting its effectiveness in modelling the rhythmic nature of the cardiac cycle for the task of fundamental heart sound segmentation.
Finally, we investigate the fundamental trade-offs within contemporary Self-Supervised Learning (SSL), a paradigm often touted as agnostic to downstream tasks. We demonstrate that models designed to maximise invariance to subject transformations are subject to inherent bounds: the more degrees of variation a model captures, the more unstable the representation becomes regarding downstream generalisation. Ultimately, this work provides both methodological contributions from a Computer Science and Machine Learning perspective, illustrated by their impact on biomedical modalities.