Four researchers at UC Santa Barbara's Robert Mehrabian College of Engineering have received a $1 million National Science Foundation award to advance artificial intelligence (AI) methods, focusing on systems where multiple components interact. This grant addresses a significant limitation in current AI, which often models interactions primarily in pairs, such as between two atoms in a chemical reaction or two regions in the brain. The project aims to enable AI to understand “higher-order interactions” within groups of components, recognizing that these multi-component systems can create outcomes that individual pairs cannot.
TheThe team includes Nina Miolane, an assistant professor of electrical and computer engineering (ECE); Haewon Jeong, also an ECE assistant professor; Sanjukta Krishnagopal, an assistant professor of computer science; and Yao Qin, an ECE associate professor. Their work seeks to uncover scientific laws that govern complex systems, particularly in fields like neuroscience where such laws are not yet fully understood. This award marks the first federal grant for the Real AI initiative, launched earlier this year at UCSB, which Miolane, Jeong, and Qin co-direct.
Rather than confining AI models to simple pairwise relationships, the researchers intend to develop systems capable of accounting for overlapping interactions among numerous components, across varying scales, and over time. This approach is critical for building AI that can not only process human language but also comprehend and predict phenomena in the physical world, like fluid flow, where current AI often struggles. The collaboration is notable for bringing together four female early-career faculty members from different departments, reflecting UCSB’s strength in interdisciplinary research.
Each co-principal investigator brings distinct expertise to the project, titled “Foundations for Discovering Higher-Order Scientific Laws with Deep Learning.” Nina Miolane, who directs UCSB’s Geometric Intelligence Lab and co-directs the AI core at the Ann S. Bowers Women’s Brain Health Initiative, specializes in using novel mathematics, including topological deep learning, to help AI identify and compute group interactions in data. This method allows models to perceive how multiple parts function collectively, potentially revealing the governing rules of an entire system.
Haewon Jeong’s research group focuses on machine learning for complex scientific systems. Her role involves exploring new AI architectures designed to uncover hidden information and move beyond mere prediction to interpret what the models have learned. The goal is to develop AI capable of accurately predicting complex physical systems and understanding the underlying principles of the physical world, utilizing techniques such as Kolmogorov-Arnold networks and topological neural networks.
Yao Qin’s expertise in the robustness and trustworthiness of machine learning is crucial for this project. She addresses the reliability of new AI architectures, which may lack guarantees of their accuracy. Her work on certified robustness is particularly vital in fields like molecular biology and drug discovery, where minor prediction errors can have significant consequences. Qin aims to establish a theoretical foundation for these models, ensuring scientists can trust their predictions, especially when dealing with the inherent noise in real experimental data.
Sanjukta Krishnagopal, with a background spanning physics and computer science, studies interconnected social and biological systems. Her contribution ensures the project considers the behavior of systems over time, seeking explanations for how entire systems evolve rather than just capturing snapshots.
The team's overarching objective is to create a mathematical framework that is robust, mathematically consistent, and applicable across diverse scales, systems, and disciplines. They will begin by establishing these mathematical foundations, then rigorously test the framework using both synthetic and real-world datasets, including those from neuroscience, molecular biology, and cosmology. Using such varied scientific contexts will allow the researchers to verify the approach’s broad applicability beyond a single field or data type. Ultimately, the team plans to make their tools open source, enabling other scientists to utilize these advanced methods without needing deep expertise in robustness. Qin views the project’s success partly in its ability to shift AI for science toward models that are not only accurate but also demonstrably trustworthy.
