Abstract
The visual system can readily extract statistical relationships between objects over space and time. What is more remarkable is the ability to form new associations between objects that have never co-occurred before. What cognitive mechanisms support the formation of these new associations? Here we propose that statistical learning not only produces knowledge about objects that have previously co-occurred, but also generates new connections between objects that have never directly appeared together. Observers viewed a continuous sequence of colors while performing a cover task. Unbeknownst to the observers, the colors were grouped into base pairs (e.g., A-B, B-C). We found that observers learned not only the base pairs, but also a novel pair of objects (A-C) that had never co-occurred before and could only be associated through transitive relations between the base pairs (Experiment 1). We extended the chain of associations by adding one more base pair (e.g., A-B, B-C, C-D), but this time the observers did not learn the novel pair (A-D), despite having successfully learned all the base pairs (Experiment 2). This reveals the limit in the transitive associations afforded by statistical learning. We also explored how transitive associations are formed across the categorical hierarchy. In Experiment 3, after viewing a sequence of city pairs (e.g., New York-London), observers automatically learned subordinate park pairs (Central Park-Hyde Park), and also superordinate country pairs (USA-UK). Importantly, none of the park pairs or country pairs had been presented previously, and could only be associated through the city pairs. However in Experiment 4, after viewing a sequence of park pairs, observers only showed successful learning of city pairs but not country pairs, reflecting the limit in the transitive associations across the categorical hierarchy. In sum, the findings suggest that statistical regularities provide a scaffold through which new associations between objects are automatically formed.
Meeting abstract presented at VSS 2016