Both Default and Multiple-Demand Regions Represent Semantic Goal Information

Abstract

We used a semantic feature-matching task combined with multivoxel pattern decoding to test contrasting accounts of the role of the default mode network (DMN) in cognitive flexibility. By one view, DMN and multiple-demand cortex have opposing roles in cognition, with DMN and multiple-demand regions within the dorsal attention network (DAN) supporting internal and external cognition, respectively. Consequently, while multiple-demand regions can decode current goal information, semantically relevant DMN regions might decode conceptual similarity regardless of task demands. Alternatively, DMN regions, like multiple-demand cortex, might show sensitivity to changing task demands, since both networks dynamically alter their patterns of connectivity depending on the context. Our task required human participants (any sex) to integrate conceptual knowledge with changing task goals, such that successive decisions were based on different features of the items (color, shape, and size). This allowed us to simultaneously decode semantic category and current goal information using whole-brain searchlight decoding. As expected, multiple-demand cortex, including DAN and frontoparietal control network, represented information about currently relevant conceptual features. Similar decoding results were found in DMN, including in angular gyrus and posterior cingulate cortex, indicating that DMN and multiple-demand regions can support the same function rather than being strictly competitive. Semantic category could be decoded in lateral occipital cortex independently of task demands, but not in most regions of DMN. Conceptual information related to the current goal dominates the multivariate response within DMN, which supports flexible retrieval by modulating its response to suit the task demands, alongside regions of multiple-demand cortex.SIGNIFICANCE STATEMENT We tested contrasting accounts of default mode network (DMN) function using multivoxel pattern analysis. By one view, semantically relevant parts of DMN represent conceptual similarity, regardless of task context. By an alternative view, DMN tracks changing task demands. Our semantic feature-matching task required participants to integrate conceptual knowledge with task goals, such that successive decisions were based on different features of the items. We demonstrate that DMN regions can decode the current goal, as it is applied, alongside multiple-demand regions traditionally associated with cognitive control, speaking to how DMN supports flexible cognition.

Keywords: control; decoding; default mode network; feature representation; multiple demand; semantic.

Figure 1.

A–C, Illustration of semantic feature-matching task, spatial working memory task, and math task. D, The behavioral performance of semantic feature matching task. **p < 0.01; ***p < 0.001.

Figure 2.

Regions involved in the semantic feature matching task that represent goal in each period and across periods. A, Univariate response to the semantic feature matching task. B, Regions that represent goal information for probe period (no smoothing). C, Schematic summarizing the decoding accuracy of each set of regions found to represent goal information during the goal cue period (green), probe (red), and target (blue) periods. D–F, Regions that represent goal information for each period (smoothed). G, H, Temporal generalization decoding between goal cue and probe periods and between probe and target periods. I, The regions where decoding accuracy between probe and target periods are significantly higher than decoding accuracy between goal cue and probe periods. All the maps are thresholded at FWE-corrected p < 0.05.

Figure 3.

Regions representing goal information overlap with DMN and networks contributing to MD cortex (DAN and FPCN) defined by Yeo et al. (2011). A, The seven networks identified by Yeo et al. (2011). B, C, Overlap between regions representing goal information and large-scale networks during the probe and target word periods, respectively. D, E, Pie chart shows the percentage of voxels within each decoding map falling within each network defined by Yeo et al. (2011) during the probe and the target word periods, respectively. Values <3% are not shown.

Figure 4.

Top panels, Regions representing goal information overlap with DMN and MD cortex defined by the localizer tasks. A, DMN and MD cortex defined using the localizer tasks. B, C, The overlap between regions representing goal information and DMN (blue) and MD cortex (red) during the probe and the target period, respectively. Regions that overlap with neither DMN nor MD cortex are in green. Bottom panels, ROIs defined for the ROI-based goal-decoding analysis. D, E, ROIs within the DAN (16∼23) and FPCN (34∼35) and within the DMN (38∼50), as defined by Schaefer et al. (2018). F, ROIs that could decode goal information (FWE corrected, p < 0.05). The numbers refer to the ROI index provided by Schaefer et al. (2018).

Figure 5.

A, Region that represents category information of probe words (FWE corrected, p < 0.05). B, Overlap of category classifier with regions that represent goal information of probe words. C, The two ROIs defined for the ROI-based category classification analysis.

Figure 6.

A, The regions in which there was a greater than zero (FWE corrected, p < 0.05) correlation between the neural RDM of color trials and the color RDM, plus significant partial correlations between these two RDMs after controlling for semantic distance. B, Overlap between regions representing color feature information and large-scale networks identified by Yeo et al. (2011). C, Pie chart shows the percentage of voxels in B falling within each network defined by Yeo et al. (2011). D, E, The regions in A overlapped with the regions that represented categorical goal information during the probe and target periods, respectively.