What is it about?

Most theories of how people learn new concepts start with fixed descriptions of the objects to be categorized. By contrast, people often create new object descriptions to help them learn needed categories. To study how people generate new perceptual interpretations while they are learning new concepts, we built a computational model that solves Bongard Problems in a human-like way. In a Bongard Problem, 6 or so scenes belong to one category and another 6 scenes belong to another category, and a creative problem solver must come up with a rule that perfectly sorts the scenes into the two categories.

Featured Image

Why is it important?

Our computational model presents a rigorous, publicly testable theory for an often neglected capacity of human creativity – our ability to create new descriptions for situations. This mode of creative thought characterizes scientific reasoning such as John Snow figuring out in 1854 that whether a Londoner contracted Cholera was determined by whether they got their water from a particular, contaminated pump, or James Maxwell developing a conception of gas pressure grounded in invisible gas molecules colliding against a container wall. Major innovations in science, mathematics, music, and art often involve constructing radically new descriptions. Our model shows the cognitive importance of building descriptions for a scene that depend on: the descriptions being simultaneously developed for other scenes, our ability to mentally simulate physical events, and emerging rules that we construct to structure our world.


This article has been over 16 years in the making. The idea for it was hatched from a dissatisfaction with Artificial Intelligence approaches to creative concept learning. Many of these systems were billed as creating new descriptions, but only in a trivial sense of creating descriptions like “blue or square” if the system started with “blue” and “square” as descriptions. For a creative problem solver to create truly novel, open-ended, and interesting descriptions, we felt it would necessarily have a perceptual system capable of grouping elements from a scene, discovering new relations between parts of a scene, constructing graded and fuzzy descriptions, and imagining what would happen in a scene if physical processes like gravity, collision, and support applied to it. We knew about Bongard Problems from reading Douglas Hoftadter’s 1979 classic “Gödel, Escher, Bach” and felt that these open-ended and fun problems would be a perfect domain for exploring how perceptual interpretations both shape and are shaped by the concepts we acquire.

Robert Goldstone
Indiana University Bloomington

Read the Original

This page is a summary of: Perception and simulation during concept learning., Psychological Review, July 2023, American Psychological Association (APA),
DOI: 10.1037/rev0000433.
You can read the full text:




The following have contributed to this page