Inferences about Uniqueness in Statistical Learning

AbstractThe mind adeptly registers statistical regularities in experience, often incidentally. We use a visual statistical learning paradigm to study incidental learning of predictive relations among animated events. We ask what kinds of statistics participants automatically compute, even when tracking such statistics is task-irrelevant and largely implicit. We find that participants are sensitive to a quantity governing associative learning, P, independently of conditional probabilities and chunk frequencies, as previously considered. P specifically reflects the uniqueness, as well as strength, of conditional probabilities; we find that uniqueness is equally affected by a single strong alternative predictor as by several weak predictors. Performance is well captured with an adapted version of the Rescorla-Wagner delta learning rule (Rescorla & Wagner, 1972). We conclude that incidental predictive learning is governed by considerations of uniqueness, and that this is computed by normalizing conditional probabilities by events' base-rates. This opens the possibility of common mechanisms between statistical learning, associative learning, and causal inference.

Return to previous page