Prediction error, processing elements, and the development of early linguistic skills: An integrated EDL-PRIMs model

Abstract

Computational theories in the early linguistic acquisition, especially statistical learning and cognitive processing perspectives, have independently acknowledged the crucial aspects of learning mechanisms in lexical or syntactic acquisition. In this research, the prediction error-driven learning mechanism (EDL) has been integrated into the primitive information processing element architecture (PRIMs) in modeling the simultaneous learning of lexicons and procedures in multiple early linguistic tasks. In the first proof-of-concept simulation study, the integrated EDL-PRIMs model demonstrates procedural learning in an algebraic task, extending the statistical learning perspective in a syntactic phenomenon. In the second simulation study, the model incorporates a set of information processing elements constrained by the empirically-identified skills; and moderates a global efficiency parameter to simulate age-related differences in processing algebraic, statistical learning, and non-adjacent dependency patterns. The integrated model's learned lexical and procedural activation profiles are adaptive when the task is syntactic, lexical, and hybrid, respectively. Moreover, the simulated efficiencies of processing elements are linked with age-related performance changes revealed by empirical findings. The simulated results are discussed with references to the empirical literature and computational theories of early linguistic development.

Simulating Repetition Suppression and Enhancement in Young Children: A Prediction-Based Approach

“…Nevertheless, it is possible to consider the model acquires cue-outcome associations initially based on top-down reinforcements (feedback), which allows bottom-up predictions of future outcomes based on learned cue-outcome associations (feed-forward). Thus, only feed-forward vs. feedback signals are issued when the model is still processing the stimulus without reaching an outcome prediction (for a processing-based prediction error model, see Ji, 2022). In this way, the Rescorla-Wagner learning equation does not assume continuous top-down influence during sensory perception, consistent with the recent view that bottom-up feed-forward signals also encode prediction error (Sederberg, MacLean, & Palmer, 2018).…”