May 14, 2021 – Andrew Nam

What underlies rapid learning and out-of-distribution transfer in humans?

Andrew Nam, PhD student. Stanford University.

Humans can sometimes learn a procedure from one or a small number of examples and then apply what they have learned to a much larger range of examples. Here, we explore this ability as it arises in learning a strategy from Sudoku. Participants naive to Sudoku went through a tutorial explaining a procedure for solving a specific instance of a general type of situation that arises in Sudoku without reference to a general rule. They then received practice with explanatory feedback on examples closely related to the one in the tutorial. Those who acquired the procedure from this experience did so within a small number of examples and, in a subsequent test, readily transferred what they had learned to examples outside the distribution of training examples. These participants’ learning was better characterized as a series of discrete transitions than as a continuous shift across strategies, and most of these participants described a reliably identifiable, valid strategy when asked to report how they solved one final puzzle. However, less than half of participants succeeded in acquiring the procedure, and success was associated with education, particularly basic mathematics education: no participants lacking both high-school algebra and geometry successfully acquired the procedure. We present these findings as constraints on computational principles guiding models of human intelligence for learning generalizable reasoning skills.