Instructor script: "Hold this motivating question against an asymmetry. Machine learning has given connectomics far more than the reverse so far — dense segmentation, synapse detection, error detection and proofreading candidate generation are all learned systems, and none of the petascale datasets would exist without them. Saying that plainly is part of writing about NeuroAI credibly."
Before any statistic, six construction choices (unit §1): what is a node (cell, type, compartment); what is an edge (and its synapse threshold); what is the weight; is it directed; which cells are included; where is the boundary. Each changes the answer. A graph is a versioned artifact with recorded parameters, not a script someone ran once. Inclusion criteria are the highest-leverage and least-reported of the six.
What a connectome does not give machine learning (unit §5): not weights — synapse count is a proxy for strength, and sign, short-term dynamics, plasticity state and neuromodulatory context are absent; not dynamics — a static diagram from one animal at one moment; not a runnable brain; and not, so far, a competitive advantage in mainstream deep learning.
The image is the techtalk's framing of an embodied agent; the teaching point is the pipeline beneath it. The strongest current result type is a connectome-constrained model: fix connectivity from the measurement, fit what remains, predict responses that are then tested. This has been done in the fly visual system. That is the version where the connectome removes free parameters — which is exactly what makes the model falsifiable — rather than decorating a figure.
A motif is enriched only relative to a null. Ask what would count as the uninteresting explanation — degree heterogeneity, spatial proximity or cell-type composition — and require the null to preserve it. Worked example (unit §2): 100 neurons, 1,200 edges, 210 reciprocal pairs. Erdős–Rényi expects 72.7 -> 2.9x enrichment. Degree-preserving expects ~150 -> 1.4x, z = 5.0. Degree- and distance-preserving expects ~185 -> 1.14x, z = 1.8, p ~ 0.07. Same data: "2.9-fold enrichment, p < 10^-6" or "no detectable effect", depending on a choice made before any test ran. Pre-register the null, or report under all three.
The real value of a declarative motif query is that the motif definition becomes an explicit, reviewable artifact instead of a description in prose — which turns "pre-register the motif and the null" into a practical instruction. Note what it does not do: it returns counts against whatever graph you built, carrying every construction choice, including a synapse threshold that may have quietly removed most of your edges.
Subgraph isomorphism is the computational core, but in practice the hard part is statistical rather than algorithmic. There are 16 directed triad classes; testing all 16 at alpha = 0.05 expects about one false positive, and triad counts are strongly correlated with one another — so treating them as independent overstates confidence. That argues for permutation-based inference, which respects the dependence. Report how many tests you ran, including the ones you did not report.
Read benchmarks the way the atlas advises: scores on small, clean volumes systematically overstate performance on production data with artifacts, rare morphologies and volume boundaries. The question for a project is not leaderboard position but error rate on your tissue — a pilot sub-volume of your own data.
Comparisons are the most durable result type in this field — across development, as in the C. elegans developmental series, or across species — because a shared reconstruction bias partly cancels between the two sides. The condition is that both sides were built with the same construction parameters and, where relevant, the same materialization version; otherwise the comparison measures the pipeline rather than the biology.
Why this matters (unit §3): merges fuse two partner lists and manufacture triangles among partners never connected through one cell — they inflate dense motifs superlinearly in the error rate. Splits mostly remove edges, deflating all motifs roughly proportionally. The two do not cancel; the residual bias points toward more dense motifs, the direction of the interesting result. Instructor script: "If the band crosses the null expectation, your result is not robust to your own measured error rate. Say so before a reviewer does. It is a few dozen lines of code."
On the second: someone objects that a distance-preserving null "throws away the biology". It depends on the hypothesis. Testing for specific reciprocal wiring beyond generic spatial structure: control distance. Comparing reciprocity between two circuits: a within-circuit distance null may over-control. Asking what generates the reciprocity: compare generative models rather than pick one null. Write out in words what would count as the uninteresting explanation before choosing.