Year: 2025
Type: preprint
Abstract: People learn languages, music, games, mathematics, and a seemingly limitless assortment ofother structures across domains. How do we learn such a large variety of richly structuredrepresentations efficiently? One possibility is that people "learn by programming," synthesizingdata-generating algorithms to explain what they observe. We examine the nature andorigins of structure learning mechanisms in human adults, human children, and nonhumanprimates (rhesus macaques), using a highly unconstrained sequence prediction task. Humanadults and older children (4-7 y.o.) learned many richly structured sequences, while monkeysand younger children (3 y.o.) succeeded mostly on simple, continuously-varying sequences(e.g. linear or approximately linear patterns). We test multiple learning models and findthat adults and older children are best explained by an inference model that generates programsin a "Language of Thought" with motor and geometry primitives, while monkeys arebest explained by local linear extrapolation strategies. Younger children exhibit variationin strategies but pattern more closely with monkeys than adults. By age 4, children showstrong program-like inductive biases similar to adults and are best fit in aggregate by theLoT model.
Authors: Tracey Mills, Nicole Hope Coates, Alessandra Acadia Silva, Kaylee Ji, Stephen Ferrigno, Laura Schulz, Joshua Tenenbaum, Samuel Cheyette
Language: English
Subfield: Developmental and Educational Psychology
Field: Psychology
Domain: Social Sciences
SDG: Quality education
Open Access status: gold
OpenAlex ID: W4414622669
Cites:
Cited by:
Related to: