CAREER: New Algorithmic Framework to Enable Efficient Machine Learning

Modern artificial intelligence systems have become essential tools in areas such as healthcare, education, transportation, and scientific discovery. However, these systems are increasingly large and computationally expensive, making them difficult to train and deploy in many real-world settings. This creates a growing gap between what these technologies can achieve and who can access them. This project addresses the challenge of making advanced artificial intelligence systems more efficient so that they can run faster, use less energy, and operate on widely available hardware. Solving this problem is critical for enabling real-time applications such as autonomous vehicles, robotics, and mobile systems, where speed and reliability are essential, while also reducing the cost and environmental impact of large-scale computing. By improving efficiency, this work will expand access to artificial intelligence, allowing educators, researchers, and organizations with limited resources to benefit from these technologies. The project will support education by integrating research results into new courses, mentoring students at multiple levels, and engaging learners to build a skilled workforce that supports future technological innovation and economic growth. This project develops a unified and principled framework for improving the efficiency, adaptability, and scalability of modern deep learning systems through three interrelated research activities. The first task focuses on efficient fine-tuning of large models by enabling rapid adaptation to new tasks using minimal parameter updates, with emphasis on structured representations, improved initialization, and understanding how these design choices affect generalization. The second task designs efficient model architectures that scale to long inputs and multiple data types by reducing the computational cost of core operations, introducing approximate computation, and selectively activating the most relevant components during inference, while also exploring alternative sequence modeling approaches that reduce latency and memory usage in long-context and multimodal settings. The third task develops new optimization methods that leverage the structure of model parameters through matrix- and tensor-based updates, together with adaptive strategies that account for variations in data and training dynamics. The project integrates theoretical analysis with empirical evaluation on standard benchmarks across language, vision, and multimodal tasks, and will produce new algorithms, insights into efficient learning, and open-source tools that enable scalable, resource-efficient, and widely deployable artificial intelligence systems. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria. NSF Award ID: 2544337 | Program: 01002627DB NSF RESEARCH & RELATED ACTIVIT,01002930DB NSF RESEARCH & RELATED ACTIVIT,01003031DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: Kai Liu | Institution: Clemson University, CLEMSON, SC | Award Amount: $417,473 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2544337 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2544337.html