Abstract
Some objects frequently co-occur in our environment, such as a mortar and pestle. While we quickly learn such statistical regularities, whether they confer any task advantage remains unclear. Here, we investigated this issue using parallel experiments in humans and monkeys. We created two sets of objects with spatially separated parts at three possible locations A/B/C. In Set 1, we selected 24 specific shape pairs at locations A-B to combine with 24 shapes at location C, to create 576 trigram objects. Likewise in Set 2, 24 shapes at location A were combined in all possible ways with 24 specific shape pairs at locations B-C to create 576 trigram objects. We therefore reasoned that repeated viewing of Set 1 or Set 2 should result in learning of specific part combinations (AB for Set 1, BC for Set 2, as they repeat 24 times more) over others. In Experiment 1, 16 human participants passively viewed all trigram objects from either Set 1 or Set 2 only once, and then were asked to choose the shape pair (AB vs BC) that was more familiar. We found that in both groups, participants identified the familiar pair with higher accuracy (average accuracy: 61.45%, p < 0.01, sign-rank test across 16 participants). In Experiment 2, two monkeys were familiarized during passive viewing to Set 1 or Set 2, and then tested on a same-different task with trigram objects in which the AB (e.g., A1B1C1 to A2B2C1) or BC (e.g., A1B1C1 to A1B2C2) locations changed. We found that each monkey detected a familiar pair breaking more accurately (average accuracy: 90%) than unfamiliar pair breaking (average accuracy: 73%, p < 0.05, unpaired t-test across pairs). Taken together, our results show that both humans and monkeys show rapid learning of statistical regularities, and that this results in faster object discrimination.