S. Liu received CAREER award

Dr. Sijia Liu, Assistant Professor of Computer Science and Engineering, has been awarded the NSF CAREER Award entitled Zeroth-Order Machine Learning: Foundations and Emerging Applications.

Abstract:

The success of artificial intelligence (AI) methods in solving hard problems raises exciting challenges and opportunities to apply them to an even wider range of real-world scenarios. For some of these scenarios, the machine-learning (ML) approaches that automatically adjust the behavior of AI systems to match the training examples they are given are simply too computationally expensive to be practical. In other forward-looking applications, the information needed to carry out these calculations is completely unavailable. The overarching goal of this project is to foster technological breakthroughs in the applicability of ML methods using a learning paradigm termed zeroth-order machine learning (ZO-ML). ZO-ML provides a work-around to the core ML calculations, making them faster and more generally applicable. By integrating advanced ZO-ML techniques with applications, this project helps bridge the gap between foundational research and real-world AI challenges, developing practical and impactful solutions. The project?s deliverables include both research and educational outcomes and activities, and seek a lasting positive impact on the academic community and society at large.

This project includes both foundational and use-inspired research in ZO-ML. On the foundational side, ZO optimization methodologies are enhanced by incorporating deep learning priors and techniques, enabling more efficient and effective optimization methods, particularly in high-dimensional settings. Additionally, the project introduces innovative approaches beyond conventional ZO optimization algorithms, applicable to diverse domains such as hierarchical learning, federated learning, and distributed computing. On the use-inspired side, the project explores practical applications in trustworthy AI, foundation models, and AI for scientific research. Traditional ML methods often struggle in these areas due to their reliance on first-order learning and assumptions about the white-box nature of ML models. This project addresses these limitations by developing robust, efficient, steerable, and effective solutions using ZO-ML. A major mission of the project is to unify optimization and ML, bridging foundational research with practical applications. It also emphasizes developing multidisciplinary training and professional development programs that transcend traditional disciplinary boundaries. To maximize the impact, the project includes a series of education, outreach, and diversity programs designed to break down physical and cultural barriers through open online education. These programs foster cross-disciplinary training and actively attract and retain women and underrepresented minority students in STEM careers.

(Date Posted: 2024-07-15)