Artificial-intelligence systems have made remarkable progress in recent years, but they still struggle with the kind of flexible, relational reasoning that comes naturally to humans. Program Postdoctoral Fellow Shuhao Fu studies how to bridge that gap.
With a background in both machine learning and cognitive science, Fu designs models that replicate human-like abilities such as analogy-making and compositional understanding. During his Ph.D., he combined behavioral experiments with structural AI models to study how humans and machines perceive and reason about relationships. “I'm more into finding out why models fail,” he says, “not just making them score better on some metrics.”
His recent research explores how AI can move beyond simple object detection to represent relationships between elements, such as how objects are arranged, how they interact, or how a concept applies across different contexts and scenes. He has also applied AI tools in health-related domains, from collaborating on a project at UCLA that used large language models to assist in depression screening, to earlier work at Johns Hopkins University developing methods to detect tumors in CT scans.
At SFI, he will be working closely with Resident Professor Melanie Mitchell, starting with the development of a model associated with the Abstraction and Reasoning Corpus (ARC) challenge, a benchmark for artificial intelligence that tests a model’s ability to reason and generalize abstract concepts from a few visual grid-based input/output examples. Furthermore, he’s interested in using structural representations and new training approaches to help AI systems perform more human-like generalization.
Fu earned a B.S. in computer science and mathematics from the Hong Kong University of Science and Technology; from UCLA, he received an M.S. in statistics and an M.A. and Ph.D. in psychology.