Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Bottom-up and top-down propagation of residual information across the left fronto-temporal cortex during sentence comprehension: Converging evidence from M/EEG and predictive coding simulations
Poster Session E, Sunday, September 14, 11:00 am - 12:30 pm, Field House
Thomas Hansen1,2, Anthony Yacovone1,2,3, Samer Nour Eddine1,4, Gina Kuperberg1,2; 1Tufts University, 2Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital, 3Boston University, 4University of Maryland
Introduction: Following the presentation of unexpected words, a bottom-up sweep of neural activity propagates from the left temporal to inferior frontal cortex between 300–500 ms, giving rise to the scalp-recorded N400. Within a predictive coding (PC) framework, this activity is posited to reflect a bottom-up residual error signal—lexical-semantic information within the bottom-up input that was not predicted by the prior context. In high-constraint sentences, the N400 to plausible unexpected words is often followed by a late frontal positivity (LFP) between 600–1000 ms. Recent MEG evidence has linked this later response to the reactivation of inferior frontal and temporal regions, but with a reversed dipole orientation, consistent with a reversal in information flow. Although the LFP is canonically evoked by plausible words that disconfirm top-down predictions, it can also be evoked by unexpected words in low-constraint contexts, provided they receive sufficiently prolonged processing. The present study asks (a) whether LFPs in low-constraint contexts similarly reflect late top-down reactivation of fronto-temporal regions, and (b) whether a single PC model can account for LFPs evoked by unexpected inputs in both high- and low-constraint contexts. Methods and Results: M/EEG. We recorded M/EEG from 31 native English speakers as they read 210 low-constraint sentences ending with unexpected words and 210 high-constraint sentences ending with expected words (600 ms SOA). ERPs revealed robust N400 and LFP effects to unexpected (vs. expected) sentence-final words. MEG source localization revealed an evoked response across left temporal and inferior frontal regions between 300–500 ms followed by a reactivation of the same regions from 600–1000 ms with the reversed dipole orientation. Simulations. To simulate these effects, we implemented a two-unit PC model of lexico-semantic processing, adapted from the visual system. The model received either top-down pre-activation or no pre-activation, followed by bottom-up inputs that corresponded to 1) expected words, 2) high-constraint unexpected words, or 3) low-constraint unexpected words. Relative to expected words, both types of unexpected words elicited larger bottom-up residual error whose rise-and-fall dynamics mirrored the timing of the N400. Additionally for the unexpected words, this was followed by top-down residual error whose dynamics mirrored the LFP. The larger top-down residual error to high-constraint unexpected vs. low-constraint unexpected words encoded mispredicted semantic information present within higher-level states that was absent from lower-level lexical states. In contrast, the larger top-down residual error to low-constraint unexpected vs. expected words encoded newly-activated, higher-level semantic information that was not already encoded in the lower-level lexical states (“postdiction”). In both cases, however, this top-down residual error was propagated down from higher to lower levels of the hierarchy as the PC algorithm converged over time. Conclusions: These findings support a PC account of language comprehension, in which early bottom-up residual errors propagate up the cortical hierarchy, followed by later top-down feedback to reconcile mismatches between levels.
Topic Areas: Meaning: Lexical Semantics, Computational Approaches