Victor Almeida from LIM-27, Neuroscience lab, Institute of Psychiatry, Faculty of Medicine, University of São Paulo (USP) combines insights from animal studies, linguistics and neurophysiology to explain electrical potentials associated with predicton and prediction error.

The Predictive Coding (PC) framework was popularised by renowned neuroscientist Karl Friston following Rao and Ballard’s 1999 computational model of extra-classical receptive-field effects in the visual cortex. Since then, canonical tenets of PC theory have infiltrated various subfields of cognitive neuroscience (whether ipsis litteris or under adaptations). Arguably, an epitome of this phenomenon is none other than psycholinguistics, given its multitude of generative models of prediction and prediction error. Regrettably though, a critical aspect of PC theory has been largely overlooked in our field. Friston was originally inspired by a neurophysiological model which demonstrated how cortical feedback signals (predictions) and feedforward (residual, unpredicted error) could shape contextual interactions associated with peripheral receptive fields of lower visual cortex’s pyramidal neurons - namely, in such a way that it mimicked in vivo recordings. Hence, such mesoscopic neural operations constitute the cardinal pillar of the entire cognitive dimension of the PC framework - and much of its appeal in neuroscience. Yet, save for a few exceptions (to my knowledge), the same preoccupation with neural constraints of this nature appears to be lacking in language studies, which, in turn, might be problematic for a few reasons. Firstly, for example, associative and sensory cortices differ quite significantly in ways pertaining to microstructure, neurophysiology, and neural populations that behave differently, insofar as these differences should ideally be accounted for whenever one conjectures about prediction/prediction error in language, rather than perception. While it is infeasible to observe them via recordings of language processing in animals (for obvious reasons), they can still be safely inferred from neural behaviour during more basic cognitive processes in higher-order regions (e.g., working memory, selective attention, categorisation). Secondly, cognitive models can be extremely appealing even in spite of biological implausibility (as history itself teaches), and this poses a very real danger to the field - that is, it runs the risk of being misled into adopting questionable premises for empirical research, as well as non sequitur conclusions on the resulting data. In this seminar, I will thereby attempt to draw attention to these caveats. Namely, I will cover some of the transdisciplinary literature on the neural basis of prediction and prediction error - viz. as derived from in vivo studies and computational modelling of evoked-related potentials - and, by the end, I will make a case for a shift towards a more neurocentric approach in the study of language.