2026-05-20T19:55:02Zhttps://keep.lib.asu.edu/oai/request

oai:keep.lib.asu.edu:node-2014442025-05-12T19:35:22Zoai_pmh:alloai_pmh:repo_items

201444 https://hdl.handle.net/2286/R.2.N.201444 http://rightsstatements.org/vocab/InC/1.0/ All Rights Reserved 2025 143 pages Doctoral Dissertation Academic theses en Tahir, Anique Rogers Liu, Huan Cheng, Lu Shakarian, Paulo Choi, YooJung Arizona State University Partial requirement for: Ph.D., Arizona State University, 2025 Field of study: Computer Science The recent advancements in Generative Artificial Intelligence (AI) have facilitated the widespread integration of Machine Learning (ML) systems into a wide range of daily activities. The abilities of ML models to generate informed text and produce images is astounding. However, these abilities come with several caveats. Generative models are large, resource intensive, prone to learning spurious correlations and produce responses with hallucinations. These cases can range from being mildly offensive to hazardous. Safety solutions to assuage unintended consequences tend to be resource intensive, discouraging potential adopters, wasting energy, and negatively affecting the environment. This dissertation aims to address these issues by exploring two dimensions of the problem: safety and efficiency. In practice, it is important to have AI systems which are not just accurate, but also safe and use-able especially in terms of hardware resources. AI safety and efficiency are vast and developing areas. This dissertation highlights various aspects of safety and efficiency, the relation between them, and how ML systems can optimize in both dimensions simultaneously. Safety in AI systems consists of several aspects. Deploying and developing AI systems could create several risks. The primary objective of research in AI safety is to identify, evaluate, and mitigate these risks. These risks can be categorized based on their origin: malicious use, malfunctions, and systematic risks. In this dissertation, we focus on malfunctions. When safety is incorporated in ML systems' pipeline, it usually comes at the cost of efficiency. For instance, if we divide the pipeline into the data, training, and inference stages, then safety may, for instance, be achieved by sanitizing the data, changing the training, or controlling the output, respectively. The goal of my research is to search for the Pareto dominant approach towards safety and efficiency. However, since there are several dimensions for both safety and efficiency, this dissertation focuses on specific combinations of aspects. Artificial Intelligence Computer Science AI Safety Large Language Models Uncertainty Quantification Toward Safe and Efficient AI