Improving Generalisation and Robustness in Deep Learning Models

Throughout this thesis, we presented several methods for improving the generalization of deep neural networks from complementary perspectives. We begin with classical learning theory to establish the foundational notions of generalization, then review practical strategies used in modern deep learning, including architectural inductive biases, data-centric techniques, and optimization choices. Building on these foundations, we focused on a unifying geometric viewpoint: improving generalization by shaping the local geometry of the loss landscape, with particular emphasis on the role of flat minima. Within this theme, we introduced and developed approaches that seek solutions and model distributions whose training loss remains stable under perturbations—linking flatness to robustness, uncertainty estimation, and improved performance under distribution shift.