Statistical learning mechanisms are flexible and can adapt to structural input properties

Abstract

Abstract Evidence has been found for two different learning mechanisms that might underlie statistical learning, computation of transitional probabilities and chunking. It is not clear though whether humans use either of the two learning mechanisms or whether they flexibly implement both. Mixed results have also been taken to indicate individual differences in the employment of learning mechanisms. In our study, we examined whether learning mechanisms are exploited differentially depending on the structure of the input to be learned. Participants were presented with three different input structures. We measured reaction times in a self-paced task and created Bayesian models that formalised different learning mechanisms. There were chunking model, transitional probabilities model and three other models were a hybrid combination of these two models. We compared the reaction times with the models’ predictions to determine which model best described learning of each input structure. The results show that the employment of the learning mechanisms indeed depends on the input structure. Additionally, hybrid models had a better fit to the data than the traditional models which might lead to the reconsideration of the traditional chunking and transitional probability models. Lastly, our findings reveal only a minor role of individual differences in learning mechanisms.