What if limits imposed by the laws of physics can guide the way we make decisions or design artificial intelligence (AI)? That’s the kind of groundbreaking question being explored by University of Hawaiʻi at Mānoa Department of Physics and Astronomy Professor Susanne Still and her research group.
Their research, recently featured in Quanta Magazine, studies how systems—whether they’re natural or man-made machines—process information to make decisions. This work could lead to smarter, more energy-efficient technologies and a better understanding of how to handle uncertainty in decision-making.
Still’s team studies “information engines,” versions of heat engines that use information as another resource. The strategies by which these engines collect and use information directly impact their energy efficiency. “Dissipation is related to how much useless information is kept,” Still explained.
Physics PhD student Dorian Daimer has worked with Still on models in which knowledge of the engine isn’t obvious but has to be inferred from observations. Together they analyzed canonical examples of such engines with intrinsic, observer-dependent, uncertainty. They’re now working with physics PhD student Thomas Redford to figure out how these engines can be used to better understand binary decisions based on imprecise measurements.
“The neat thing is how many applications these engines have,” said Redford. “And once we have strategies for how to run such an engine, then we know optimal, physically implementable solutions to decision making under uncertainty in the real world.”
Bringing together different fields
This research blends ideas from physics with ideas from machine learning and neuroscience, fields that study how systems learn, adapt and decide. Still’s interdisciplinary approach brings in students with diverse expertise. Still developed UH Mānoa’s first machine learning curriculum in the Department of Information and Computer Sciences (ICS). Daimer has a background in both physics and cognitive science. Christoph Haring, a UH Mānoa ICS grad student working with Still, has a background in mechanical engineering and machine learning.
“This work is fascinating because it touches on the basic fundamental assumptions made in AI. We try to understand them from a physical perspective. It is easy to relate to, yet deep and very useful,” Haring said.
Students in the spotlight
Helping students learn how to share their work beyond academic circles is something Still values. When she was interviewed for the Quanta article, she encouraged Daimer to take part. “It was a valuable learning experience,” Daimer said.
Big ideas, real-world impact
The research approach recognizes that the physical performance of any system that processes information is determined by its strategies—what information is stored and how it is used.
“The idea is to prefer a strategy that least hinders overall performance, unless there is a clear reason why a performance-limiting strategy must be used,” Still said. “We identify which strategies provide the possibility to come as close as possible to physical limits.”
Still has shown that this principle can be used to derive a well-known data compression method,
used in machine learning and AI, one that extracts relevant information and discards useless bits. In ongoing work, Daimer and Still investigate how the same principle might add insight into axioms employed in operational reconstructions of quantum mechanics.
Traditionally, the observer in physics is outside and independent of the physical reality that is described. But over the last few decades, there has been a shift towards giving the observer a more involved role, highlighting that information is central to our description of reality. The team’s approach studies observers that are embedded into the physical reality they observe.
