Towards Better Interpretability and Controllability by Exploring Latent Space Geometry 113 views

Deep learning models have achieved remarkable performance in computer vision tasks, yet their complex architectures make them difficult to interpret and control. This lack of transparency limits their adoption in domains that demand reliability and accountability, such as medical imaging and image synthesis. This dissertation investigates geometric approaches for understanding and influencing neural network behavior. First, methods are developed to quantify continuous feature dependencies by aligning classifier gradients with human-interpretable features, intervening on target features within the data manifold, and decomposing prediction variance to obtain quantitative feature attribution scores. A diffeomorphism-based framework is then proposed for generating anatomically meaningful counterfactual explanations of medical image classifiers, identifying minimal, topology-preserving deformations that reverse classifier decisions. In a complementary line of work, controllability is examined through models that learn group actions in latent spaces, enabling structured transformations of learned representations, and through analyses of how topological assumptions in generative models affect their capacity to represent data with complex manifolds. Together, these studies advance geometric perspectives that enhance both the interpretability and controllability of deep neural networks.

PHD (Doctor of Philosophy)

English

2025-12-03