A Call to the Lexicon: Predictive Coding Offers a Biologically Plausible Computational Framework for Implementing Language Comprehension in the Brain, Gina Kuperberg (Tufts University)
It is widely accepted that the brain generates predictions during language comprehension. Yet, capturing these predictions has been challenging. I will present studies that combine multivariate methods with MEG, EEG, and intracranial recordings to show direct neural evidence of predictive pre-activation at multiple levels of representation, from coarse-grained semantic features, to individual words, to orthographic forms. These findings align with the Parallel Architecture's proposal that the brain pre-activates upcoming information via stored interface links across an extended lexicon. However, they force us to revisit the fundamental question of how the brain prevents these predictions from fully overriding bottom-up inputs. I will argue that the computational principles of a particular neural network and algorithm, known as predictive coding, provide key insights into this question. I will present the results of simulations using an implemented predictive coding model of lexico-semantic processing, demonstrating how this theory resolves several longstanding puzzles in the neurobiology of language: Why the confirmation of prior predictions facilitates behavior but reduces neural activity; why we don’t incur neural or behavioral costs when our predictions are disconfirmed; and why we don’t hallucinate our predictions. Together, these findings suggest that predictive coding might provide a biologically plausible computational framework for implementing a Parallel Architecture of language comprehension in the brain.