Which network patterns are shared across people?
Spatial agreement generally falls across the learned network ordering. The earlier patterns are more consistently represented within the shared basis; later patterns show more individual variation. The figure cannot separate genuine individual differences from instability in short recordings.
Explore the figures
Each dot represents one participant. Higher agreement means their pattern of ROI relationships more closely resembles the other participants’ average. Click an unusual point to inspect that person; change the fitted rank to explore a different model. Red inspection flags identify unusual values without excluding anyone.
How consistently are the learned networks represented across people?
A group network is useful only if its spatial pattern describes more than an average that conceals individual differences. This section asks which learned networks are represented similarly across the 96 NYU control participants, and where individual recordings depart from that common description. Resting-state data provide a view of spontaneous BOLD organization, without task events that would explain a particular departure.
The analysis first estimates one common predictive basis from all 96 recordings, with equal weight per subject. It then holds that basis fixed while calibrating each person’s spatial coefficient matrices. Network k therefore has the same coordinate definition across people. The comparison measures individual variation within a shared representation; it does not test whether separate subject fits would discover the same networks.
For each network, the subject’s unique off-diagonal ROI weights form a vector wₛ. Its reference is the arithmetic average of the other 95 subjects, w₋ₛ. Signed spatial agreement is their cosine, (wₛ · w₋ₛ) / (||wₛ||₂ ||w₋ₛ||₂). This asks whether the same connections carry similar relative signs and weights, independently of a positive overall rescaling. A separate strength ratio, ||wₛ||₂ / ||w₋ₛ||₂, asks whether the coefficient pattern is larger or smaller.
What this comparison can—and cannot—tell us
Each network is a signed coefficient matrix over the same 100 cortical ROIs, contributing to predicted regional signal second moments rather than conventional centered covariance. We compare the unique off-diagonal connection weights using signed cosine similarity: +1 means the same pattern up to a positive scale, 0 means orthogonal patterns, and −1 means reversed signs. This is not correlation between subjects’ BOLD time courses, simultaneous network activation, or a measure of neural excitation/inhibition.
These are deviations within the shared representation, not a test of individual dynamics outside that representation. Because every subject contributed to learning the basis, excluding them from the reference average does not make this an independent validation. A low-rank shared basis can miss individual structure. Resting-state recordings provide no task blocks here with which to explain differences.
Red outlines mark agreement more than 3.5 scaled MAD units below the network’s median, or absolute log-strength departures beyond 3.5 scaled MAD units. These are descriptive screening rules, not significance tests, diagnoses, or exclusion criteria. Near-zero MAD yields no defined flag. All subjects and learned networks are retained; the explicit constant bias is separate.
Which networks have the most consistent spatial patterns?
Each dot compares one person’s network with the other subjects’ average for the same network. Agreement near +1 means proportional connection patterns with the same signs; 0 means orthogonal patterns; −1 means sign reversal. High, compact distributions support a reproducible description within this fitted basis. Broad distributions or low tails identify patterns that need individual inspection. Red outlines apply the descriptive screening rules described above.
Show the subject-by-mode overview
Ordered by median agreement, not clustered. Click a cell to select the subject and mode.
Could recording quality help explain low agreement?
Differences between people may reflect their neural organization, but also motion, incomplete ROI coverage and finite recording length. The scatterplot asks whether subjects with higher average head motion also tend to have lower median agreement across networks. It is a quality-context check, not an explanation of any single network.
What can be concluded?
Most subject–network comparisons resemble the cohort pattern within the shared predictive basis, with a smaller tail of unusual patterns or strengths. Later networks are generally less consistent, but this cannot distinguish weaker estimation from individual structure. These findings support a common description of the fitted cohort while identifying cases that merit closer inspection; they do not establish independent reproducibility, clinical abnormality or a task-related cause.
Additional diagnostics and provenance
The detailed diagnostic explorer ↗ contains deviation maps, cross-network correspondence and mixture comparisons for closer inspection.
· ·
Rank 20 · Schaefer 100 · existing min–max inputs · original fits unchanged.
More about the measures and interpretation
Spatial agreement generally decreases across the learned network ordering. Leading dimensions are retained first in the predictive spectral decomposition; later coordinates show greater variation across subjects. One plausible explanation is that later coordinates capture less stable predictive structure in these short recordings; genuine individual differences can contribute too. The trend alone cannot distinguish those explanations. Network order is not a ranking of explained BOLD variance or biological importance, and smaller coefficient magnitudes do not directly lower the agreement score. Mean-like network 1 should be interpreted separately.
Fixed −1 to +1 scale · boxes show the middle half of subjects · network 1 is mean-like · click a dot to inspect its numbers.
Amber = recorded ROI coverage warning; blue = no recorded coverage warning. Diamonds indicate a metric or coverage flag. Click a point to select that subject above.
A motion–agreement association would motivate closer quality review; its absence would not establish that motion has no effect. This unadjusted subject-level comparison cannot isolate physiology, motion or individual neural differences. Visual QC and verified framewise censoring remain incomplete. All subjects remain in the analysis.