Speaker: Samer Nour Eddine (Tufts)
Title: Dynamic duo: Insights from a dual-unit predictive coding model of lexico-semantic processing
Abstract: What is the functional role of prediction in language processing? From one perspective, prediction is a passive consequence of the incremental, interactive nature of language processing. There is, however, an alternative perspective –– predictive coding –– that argues prediction is not a passive consequence, but rather a critical mechanism by which different levels of linguistic representation interact. Specifically, predictive coding proposes that neural dynamics aim to be as efficient as possible, only sending missing information up the hierarchy rather than redundant, already activated representations (i.e. prediction errors). To do this, the model relies on two kinds of units: 1) state units, whose activation levels correspond to the linguistic representations that are likely given the input, and 2) error units, which pass up information that is present in the bottom-up input but not yet active at higher levels. In this talk, I will show that this dual-unit architecture implements a unique mechanism of competition that captures asymmetries in bottom-up and top-down processing, and asymmetries in behavioral and neural measures of processing. Exploiting its unique properties, I will describe some counterintuitive patterns predicted by this framework and end with a discussion of our ongoing efforts to investigate them further.