Predictive Coding explains the dynamics of neural activity within the left ventromedial temporal lobe during reading comprehension

Abstract

Abstract Predictive coding has been proposed as a computational theory of brain function. However, there is little conclusive evidence linking it to language comprehension. We measured brain activity with magnetoencephalography during reading comprehension and simulated this activity using a predictive coding model of lexico-semantic processing. Between 300-500ms, the left ventromedial temporal lobe produced a larger N400 to unexpected than expected inputs. Our simulations showed that this could be explained by the production of lexico-semantic prediction error. To distinguish predictive coding from other frameworks that can also account for the univariate N400 effect, we carried out two multivariate analyses. First, we showed that between 300-500ms, expected words produced neural patterns that matched those that were pre-activated before the bottom-up input appeared. Our simulations showed that this could be explained by a reinstatement of item-specific lexico-semantic predictions within state units. Second, expected inputs produced consistent patterns that were distinct from those produced by unexpected inputs. Our simulations showed that this emerged from the differential activation of functionally distinct state and error units during the predictive coding algorithm. Together, these findings provide strong evidence that the left ventromedial temporal lobe employs predictive coding to infer meaning from orthographic form during reading comprehension.