Fig. 4: Results of predictive coding (PC) neural modeling.

a Two computational models were constructed based on the continuous and sparse updating hypothesis. Left panel: the continuous updating hypothesis assumes that the higher-level representations are updated continuously as inputs change over time. Right panel: the sparse updating hypothesis assumes that the higher-level only predicts and updates at its preferred timescales (i.e., sentence boundaries). b, c Simulated data generated by the PC models were converted into the putative BOLD signals via the hemodynamic model, and further compared with the real fMRI responses to evaluate model performance. d Model performance in the forward and backward conditions. e The sparse PC model outperforms the continuous PC model only in the forward condition. f Model performance without word-level prediction error. g Comparison of MSE values for sparse and continuous models by leveraging PE for all subjects across stories. h, i Examples of simulated signals from the PC models at the word (Subject 10, story 2) and sentence levels (Subject 04, story 2), shown alongside the corresponding real BOLD signals. Significant levels are indicated as p ≤ 0.001 (***), p ≤ 0.01 (**), p ≤ 0.05 (*), and p > 0.05 (n.s.).