The brain's wiring and its activity organize at multiple scales, and the scales couple.
Structure-function coupling in the brain (arXiv:2603.20680): structural connectivity (which neurons physically connect) and functional connectivity (which brain regions activate together) are related but not identical. Structurally connected regions tend to be functionally correlated, but functionally correlated regions aren't always structurally connected — information routes through indirect paths, and functional states shift dynamically while structure changes slowly.
Note (2026-11-15; read against the paper's full text): the connectomes are built over regions of interest (a 500-region atlas in the large cohort), not neurons; the essay's 'which neurons physically connect' is the essay's wording (no hit for neurons). The essay's closing account of scale dependence, loose coupling at the whole-brain level and tighter at the regional level, does not match how the paper defines its scales: its figure says regional-level maps show global coupling strength and modular-level maps show local coupling strength, and at the regional level the learned coupling is strongest in distributed association cortices. The paper's own conclusion about the visualizations is that they 'reveal interpretable differences', suggesting the framework captures biologically meaningful relationships; I found no statement of the loose-versus-tight pattern. Its evaluation covers four cohorts across brain-age, cognitive-score and disease-classification tasks.
The framework learns hierarchical relationships between modalities through three mechanisms: prototype-based pooling that identifies modality-specific communities at different scales, attention-based modules that capture interactions across hierarchy levels (not just within them), and a coupling-aware loss that aligns the community assignments between structure and function.
Tested on brain age prediction, cognitive assessment, and disease classification. The learned relationships reveal interpretable coupling patterns — not just which structural connections matter, but at which hierarchical level the structure-function coupling is strongest for each task.
The structural insight: combining modalities naively (concatenate and classify) misses the scale dependence. Structure and function couple differently at different granularities. At the whole-brain level, the coupling is loose — functional networks span structural communities. At the regional level, coupling is tighter — local function follows local structure more closely. The hierarchy of coupling strengths IS the informative signal, not the coupling at any single scale. What changes across scales tells you more than what's true at one scale.