UC San Diego

In the dynamically progressing domain of Artificial Intelligence (AI), Machine Learning (ML) stands out as a pivotal innovation, acclaimed for its proficiency in enabling autonomous learning from data without direct human supervision. This advancement has positioned AI systems at the vanguard of technological progress, endeavoring to amplify human capabilities across diverse tasks through heightened efficiency, precision, and the reduction of manual intervention. The drive towards automation within these systems has given rise to Automated Machine Learning (AutoML), an advanced, AI-facilitated methodology that automates the process of model selection, dataset optimization, weight parameterization, and hyperparameter tuning, significantly reducing the need for extensive human expertise and intervention.Central to AutoML research is the area of Neural Architecture Search (NAS), which has shown exceptional performance, rivaling human capabilities across various applications. However, the deployment of ML solutions, especially in critical areas such as healthcare diagnostics and autonomous vehicle guidance, is hampered by concerns about their trustworthiness. To alleviate these apprehensions, it is imperative to rigorously investigate and incorporate principles of trustworthy AI, such as interpretability, fairness, robustness, and reliability, thereby encouraging their wider acceptance and utilization. This dissertation investigates the creation and evaluation of multi-level optimization (MLO) frameworks aimed at augmenting the trustworthiness and effectiveness of NAS methodologies. By weaving together elements of interpretability, fairness, robustness, and generalization, this research endeavors to devise NAS frameworks that are distinguished not only by their performance but also by their ethical and dependable operation. We systematically apply our innovative frameworks to diverse tasks, including image classification, natural language processing, and image captioning, to empirically verify their impact and efficacy. Expanding on this foundation, the study examines the implementation of these trustworthy NAS frameworks in crucial healthcare contexts, undertaking thorough experiments to assess their accuracy and reliability. Additionally, this dissertation thoughtfully explores potential future research directions and the limitations of the proposed approaches. Through this examination we highlight the ongoing necessity for research to address the complexities and ethical issues surrounding the broad implementation of AI specially NAS.