Paul Klee, "The Bounds of the Intellect," 1927 (detail)

This is the query driving a three-day workshop at SFI, which itself aims to understand how well scientific and mathematical reasoning can comprehend complex systems.

“When we went to school and we put down the answer to a question, we always had to explain how we’d gotten there,” says SFI President David Krakauer, one of the workshop’s co-organizers. “There was always a concern that we’d cheated or stumbled on the right answer.” With recent advances in machine learning, however, Krakauer says we have no easy way of evaluating the problem-solving process. “Machines are terrible at explaining things. We’re now living in an age where we’re confronted with significant limits in understanding. With artificial intelligence, the open question is ‘will the limits to our understanding ever be transcended by technologies or will technologies make understanding irrelevant?’”

The workshop was conceived by Krakauer, SFI and MIT Postdoctoral Fellow Brendan Tracey, and former SFI Omidyar Fellow Joshua Grochow (University of Colorado). It is designed to address a recurring challenge at the heart of complexity research: if complex systems are not amenable to the classic model of scientific explanation, on what basis does complexity research grant us scientific understanding?

The question arises because complexity researchers often rely on methods and computational tools that do not provide both predictive and explanatory power — two criteria for scientific understanding essential to the classic model of scientific explanation. A computational model that offers sound predictions and new insight into large-scale patterns, for example, may not help us arrive at an exacting phenomenological explanation.

According to Tracey, “We are in an age where we see much greater conflict between understanding and prediction.” Scientists and themselves asking: if a machine learning tool gives us predictive power but we don’t understand why it generates accurate predictions, do we have scientific understanding? Is research that relies on a black box properly called science?

A central goal of the workshop is to establish a clear definition of scientific understanding in light of the methods that scientists currently use. In Grochow’s words, “We will ask about the real foundations of the kinds of tools that we are using, and we will try to place these on firmer ground.”

Participants include philosophers, novelists, and scientific researchers, but the majority of invitees are practicing scientists who regularly confront the workshop’s questions in their research. The workshop takes place at SFI from Nov. 29-Dec.1.

Ultimately, the organizers hope to clarify not only where the limits to scientific understanding lie, but also how science can surpass them. 

Read more about the working group.