[PS-1.3] A neurocomputational model of unsupervised associative learning predicts successful and impaired statistical learning

Tovar, &. ¹ , Westermann, G. ² , Torres, &. ¹ , Flores, M. ¹ , Molina, L. ¹ & Gámez, S. ¹

1 Universidad Nacional Autónoma de México
2 Lancaster University

We have developed a neurocomputational model that uses associative learning as a domain-general mechanism to learn from environmental regularities. The model incorporates long-term potentiation (LTP) and long-term depression (LTD), based on biological descriptions of synaptic adaptation, to capture the structure of the environment. This model highlights the interaction between stimulation regularities and learning constraints to describe properties of statistical learning. The model predicts learning interference effects when learning overlapping associations - different cues associated with the same target - in Serial Reaction Time Tasks (SRT). Moreover, it shows how this interference is overcome through more exposure to stimulation, but only when there is a typical balance between LTP and LTD. Altering this balance in the model, following neurobiological descriptions of intellectual disability, results in severe problems to learn overlapping associations. These predictions were supported in empirical studies evaluating human adults, typically developing children and children with Down syndrome in SRT. The model thus links reported atypicalities in low-level neural processing with specific expressions of cognitive disability in developmental disorders. Understanding how learning constraints interact with stimulation regularities will allow us to advance our knowledge of how humans learn from complex stimulation regularities, such as those underlying human language.