Machine-learning tools have powerfully accelerated the process of doing science. They can sort through and analyze vast sums of data, revealing insights and connections about the world never before possible. But could we ever fully automate the scientific process? Could we make an AI physicist? It’s a question that captured Seungwoong Ha, an incoming Applied Complexity Fellow, as a student at Korea Advanced Institute of Science and Technology (KAIST) where he completed his B.S. and integrated M.S. and Ph.D., all in physics.
The basic roadblock is an AI’s ability to comprehend. “You can’t command a computer, ‘find something interesting.’ But humans can find it. We understand what ‘interesting’ means,” says Ha. However, if we ask an AI a more specific question — for instance, does a dataset shows internal symmetry or a conserved quantity? — our machines could point toward something “interesting.”
Ha now wants to apply the powerful pattern-finding abilities of machine-learning systems on complex-systems questions. And while many of his previous research questions have been rooted in physics, Ha is turning his attention toward social systems, “the hardest complex system,” as he says. During his two-year fellowship, Ha will work closely with SFI Professor Mirta Galesic and External Professor Henrik Olsson on their belief dynamics project, using machine-learning tools and natural language models to explore how people behave and influence one another in online spaces.