Trust and Transparency in Deep Learning, from Medical Imaging to Language Models
Skip to main content
eScholarship
Open Access Publications from the University of California

UC Davis

UC Davis Electronic Theses and Dissertations bannerUC Davis

Trust and Transparency in Deep Learning, from Medical Imaging to Language Models

Abstract

As deep learning models have grown deeper and more complex, and as they have been deployed in high stakes decision-making settings, their lack of transparency has become an increasingly urgent problem. This dissertation addresses that problem across two domains, medical artificial intelligence (AI) and large language models (LLMs), with the shared goal of improving our ability to understand, characterize, and ultimately trust the outputs of black box models. The first part of this work lies within a broader effort to build AI-driven, patient-specific synthetic data capable of supporting tasks like diagnosis. This dissertation lays the groundwork for that goal in two ways. First, it shows that combining physiologically heterogeneous data sources in a shared latent space can improve the diagnostic performance of a deep learning model on acute respiratory distress syndrome (ARDS). Second, it demonstrates the ability of optimal transport-based models to generate synthetic chest x-rays by selectively and reliably manipulating disease-relevant features of images, and introduces a new, more effective training method to that end. The second part of this dissertation turns to LLMs. First, it introduces a method for improving the reliability of agentic tool use by quantifying the uncertainty of LLM responses and learning when models should say ``I don't know." It then presents a mathematically rigorous and verifiable approach to machine unlearning, addressing the growing need to remove information from trained models for legal, regulatory, and data management reasons. Finally, this dissertation explores a novel method for quantifying the temperature used during LLM text generation and the lessons that can be learned from where the method fails. Together, these contributions span medical imaging and natural language processing, but they share a common throughline that trust in deep learning systems is based on an understanding of how they behave and where their limitations lie.