Funded Grants

Prediction as a mechanism of infant language acquisition

Despite extraordinary progress over the past decade, remarkably little is yet understood about how infants acquire language. This is an issue of pressing societal importance, given that approximately 5-10% of school-aged children lack developmentally appropriate language skills. A growing body of evidence demonstrates that infants are remarkably skilled at laboratory language learning tasks, including tracking myriad statistical properties of linguistic input. However, all of this work involves off-line measures of learning, in which infants are first exposed to some set of stimuli (often from an artificial miniature language), and subsequently tested on discrimination of familiar versus novel patterns. This essay describes new thinking that has the potential to significantly extend the empirical and theoretical reach of infant language research by focusing on the role of on-line prediction as a mechanism of language processing and learning.

There are several sources of indirect evidence to suggest that prediction may play a role in infant language learning. However, this hypothesis has not been experimentally tractable given the standard methods in infant language research. Motivated by the theoretical and practical importance of understanding how infants learn language, we have recently developed new methods that will allow us to directly investigate the prediction hypothesis. We propose to interrogate infant language processing and learning using measures of eye-gaze, focusing on the critically important question of how infants learn syntax. On-line measures of infants’ predictions will be used to test important theoretical questions concerning the degree to which statistical regularities are actually employed during online processing. In addition, these methods afford the opportunity to test an assumption of computational models of learning in actual infant learners. One of the key insights of modern computational neuroscience is that if a learner at time-step t predicts what will occur in the input at time-step t+1, discrepancies between the predicted input and the actual input provide a learning signal. Can infants also detect discrepancies between their predictions and actual language input in real time, as simulated in these models? If so, predictions could facilitate learning via covert self-correction. The prediction framework also motivates tests of theoretical claims concerning domain-specificity and critical periods, can be applied to both typically and atypically developing infants, and will be integrated with subsequent studies concerning the neural bases of infant learning.