What functional organization does the model recover?
Seven reference systems · 7-network fit
- Network 2 pairs visual and default-mode organization. Its positive pole resembles the visual system (similarity 0.58; 44% of that pole’s weight), while its negative pole resembles default mode (similarity 0.62; 43% of that pole’s weight). This is the clearest two-system opposition in this fit; the signs name the learned poles, not activation or inhibition.
- Network 4 emphasizes somatomotor organization. Its positive pole has a strong somatomotor resemblance (similarity 0.74; 31% of that pole’s weight). Frontoparietal control is the strongest match on its negative pole (similarity 0.49; 34% of that pole’s weight), with a less selective resemblance.
- Visual organization also appears in Network 5. Its positive pole matches visual regions (similarity 0.57; 45% of that pole’s weight); its opposite pole has a more modest dorsal-attention match (similarity 0.39; 24% of that pole’s weight). The same reference system can therefore participate in several learned patterns.
These associations describe where a learned pattern is concentrated. Similarity and weight share answer different questions: a pole can resemble a system closely without placing most of its weight there. Weak or distributed matches should remain unnamed.
Findings use seven-system anatomy, regional Pearson similarity and all off-diagonal edges. They are descriptive associations, not validated functional labels. Network identities belong to this study and fitted rank.
These findings describe the seven-system reference; switch back to 7 systems to inspect the cited matches.
Explore the figure
Start with the networks described above. Click a score to select that pole and reference in the atlas. Click a system header to sort networks by their combined positive- and negative-pole similarity. A darker cell means greater resemblance; its bottom bar shows the share of that pole’s weight. In the atlas, click a ranked region to highlight its connections and inspect its strongest partners.
Studyforrest · Schaefer 100 cortical · min/max inputs · rank 7
Table help
Similarity compares each pole’s off-diagonal ROI weight totals with a reference system using regional Pearson correlation. Positive and negative poles are treated separately by magnitude. A bottom bar shows the fraction of that pole’s total regional weight assigned to the system. All edges contribute to matching; reducing visible atlas edges changes the display only.
Match thresholds are descriptive. A system can have several matches or none. The seven-system findings above name the clearest associations; a finer reference may split those associations. Dark outlines indicate the table’s similarity/share rule, not statistical significance.
How well does each pole match each system? · select a score to compare it in the atlas below
Read across a row to see whether a pole has one dominant anatomical association or several. Compare positive and negative columns to ask whether opposite coefficient patterns point toward different functional systems. Dark outlines flag candidates meeting the chosen similarity and weight-share thresholds; they are not statistical significance markers.
Highlight rule
Both thresholds must pass. Exploratory cutoffs, not statistical significance.
Scroll sideways for more references · use the scrollbar, trackpad, or mouse wheel over the table. Network labels stay pinned.
Where does a numerical match lie anatomically?
The atlas links a selected score to its underlying ROI pattern and the reference-system envelope. It asks whether a summary match is distributed across the system or driven by a few highly weighted regions. Select a node to inspect its connections; the strongest displayed edges are a visual subset, while scores use the complete selected edge scope.
Brain atlas comparison
Reading this view · size key · edge counts
Connectivity sizing follows the matching edge interpretation above. Absolute-magnitude mode draws unsigned edges and circular markers. Other modes retain original signed edges and regional projection signs; hover retains signed coefficients. Use Brain edge sign to inspect either polarity. ROI marker colors identify the learned network. In signed views, edges default to red for positive and blue for negative weights; all edges are solid. Choose By learned network to restore mode-colored edges. Circles/diamonds indicate positive/negative regional projections, independently of the selected size measure. Marker diameter is scaled within each network, so sizes do not compare absolute strength across networks. Reference layers: filled translucent surfaces in their own atlas colors. Colors and opacity are not similarity or probability values.
Drag to rotate · scroll to zoom · use the legend to hide individual layers. Overlapping layers can occlude one another; reduce opacity or deselect layers. The edge budget is a fraction of all unique off-diagonal pairs per mode. Strongest overall ranks absolute weight after the sign filter; it can show only one polarity even when both exist. Balanced selection reserves half the budget for each sign and fills any unused slots from the remaining strongest edges. It applies only when both signs are selected. Edge widths are uniform; hover for signed coefficient values. ROI markers are unchanged by edge sign/selection, so changing edges can leave the regional points looking similar. The node cutoff uses connectivity strength whether scaling is on or off; it hides values below the selected fraction of the network maximum absolute strength; zero values are hidden. It does not filter edges or scores. Connectivity strength excludes self-connections and uses every edge, including those hidden from the plot. These are learned-weight units, not probabilities. These display options do not change coefficients, model fitting or reference comparisons. Surfaces are atlas label-volume envelopes, not subject cortical reconstructions.
What the anatomy supports
The model recovers patterns that resemble familiar cortical systems, sometimes on opposing poles of the same network. The associations provide a functional vocabulary for interpreting expression over time. They do not assign cognitive states or establish excitation and inhibition.
Methods and provenance
Network 1 is omitted from this explorer and strength scaling; the separate constant bias is excluded. Different model ranks are fitted independently; network numbers and signs do not establish correspondence across fits. Switching seven/seventeen reference systems does not refit the model.
Regional scores compare off-diagonal edge sums with fractional overlap of Schaefer 100 cortical parcels in the selected Schaefer/Yeo cortical atlas. Matrix-edge scores compare covered ROI pairs with within-system templates, excluding self-pairs. The 10% cortical-overlap rule applies to scoring; subcortex and uncovered tissue remain unassigned. Reference surfaces show complete atlas systems; scores use covered common ROIs.
Matching magnitude uses |C|, positive-only uses max(C,0), and negative-only magnitude uses max(−C,0), applied to the shared fitted matrix before regional projection. This retains within-matrix strength despite opposing edge signs; it does not undo cancellation across subjects. These choices do not alter model fitting or the original signed matrices. Coordinate signs depend on the fit. Similarity is not a confidence probability. Bootstrap confidence, spatial-null tests and validated functional labeling are not implemented.
Exact membership and preprocessing/model provenance
When are the selected shared networks expressed?
The anatomical matrices above are fixed cohort coefficients. The participant selector below changes their event-aligned expression curves, using the same cohort basis.