Figure 2.

Prediction performance of encoding models based on five language feature spaces in cortex and cerebellum. Encoding models fit with 5.4 h of BOLD data were tested against a held-out story (10 min). A, Correlation (r²) between predicted and actual BOLD response is plotted on flattened cortical and cerebellar surfaces for 1 subject (UT-S-02; other subjects are shown in Extended Data Fig. 2-1). Significance testing for each model in each voxel was done using a one-sided FDR-corrected permutation test with a threshold of p < 0.05. The higher-level models have better prediction performance in both cerebellum and cortex. To confirm this, we also tested a low-level spectrotemporal modulation model, which was not substantially more predictive than the spectral model (see Extended Data Fig. 2-3). In cortex, the areas best predicted by each of the three feature categories are spatially distinct. However, in the cerebellum, the areas best predicted by each feature space are highly overlapping. B, To compare across subjects, we plotted average signed r² across all voxels in the cerebellum and cortex for each subject and each feature space. The context-level semantic feature space has the highest predictive performance in both the cerebellum and cortex for all subjects. Performance scales roughly linearly in both cerebellum and cortex across the hierarchy of language representations, albeit with higher r² in cortex than cerebellum. C, Because cortical and cerebellar BOLD responses might have different levels of noise, which could obscure differences in representation, we also computed noise ceiling-corrected correlations (Schoppe et al., 2016). This correction caused the average r² to be less biased in favor of cortex (for corrected correlation flatmaps, see Extended Data Fig. 2-2) and suggests that each feature space might be represented to a similar extent in cerebellum and cortex. However, overlapping prediction performance between different feature spaces in the cerebellum suggests that the cerebellum may not be separately representing each stage of language processing.