Lecture 3 of the connectomics lecture series. 58 slides in three parts, about 150 minutes. Openly licensed — CC BY-SA 4.0.

What this lecture covers

Segmentation and its error taxonomy, proofreading triage, graph construction and null models, and an honest account of what connectomics and machine learning give each other.

Learning objectives

By the end of this lecture, participants will be able to:

1. Describe how automated segmentation works and where it fails structurally.
2. Select quality metrics appropriate to a stated endpoint.
3. Construct a connectivity graph from a reconstruction, stating every consequential choice.
4. Justify a null model for a stated hypothesis and interpret a motif result against it.
5. Assess what connectomics and machine learning currently give each other.

Structure

Part A — Segmentation, error, and the labor problem

How automated segmentation works and where it fails structurally, the error taxonomy, quality metrics and their blind spots, and triage by endpoint change per annotator-minute.

Part B — From segmentation to a defensible graph

Six consequential construction choices, null models, the triad census, merge-error bias, and the error-sensitivity check.

Part C — Applications, NeuroAI, and what to claim

Comparative connectomics, cell typing, three results that actually landed, and the two symmetric errors about connectomics and machine learning.

What students produce

An analysis card: hypothesis, estimand, graph provenance block, null model with the “it would be uninteresting if…” sentence written out, success criterion set in advance, error band from measured merge and split rates, non-claim, and provenance.

The centrepiece

The reciprocity worked example in Part B, across three slides. The same data — 100 neurons, 1,200 edges, 210 reciprocal pairs — supports “2.9-fold enrichment, p < 10⁻⁶” under Erdős–Rényi, “1.4×, z = 5.0” under a degree-preserving null, and “no detectable effect” once distance is preserved too. Nothing about the data changed; only the question did.

Notes for whoever teaches it

Merge bias points toward the interesting answer. A merge fuses two neurons’ partner lists and manufactures triangles among partners never connected through one cell, so merges inflate dense motifs superlinearly while splits deflate everything proportionally. The errors do not cancel. Motif analysis on unproofread segmentation is not conservative — it is biased toward the result you were hoping for.

Parts A and B join at the error-sensitivity check. The Part A resample gives you measured merge and split rates; the Part B check turns them into an error bar on the motif claim. Neither half is useful alone, and students tend to treat them as separate topics until this is pointed out.

The NeuroAI section is written to prevent two errors, not one. Dismissing the connection and overselling it are both wrong. The accurate position — machine learning has given connectomics far more than the reverse — is specific, defensible, and slightly boring, which is usually the sign that it is right.

Licence and credit

CC BY-SA 4.0. Teach from this lecture anywhere, including commercially; copy and redistribute it in any medium; and re-cut, shorten, translate, or merge it into your own material. No permission needed. Two conditions: credit the original and say if you changed anything, and distribute your adapted version under the same licence.

Gray Roncal, W. (2026). Nanoscale Connectomics: Algorithms and Applications (EN.585.781 Frontiers in Neuroengineering, Module 9). NeuroTrailblazers. CC BY-SA 4.0. https://neurotrailblazers.org/teaching/lectures/

For an adaptation, prefix with “Adapted from” and note what you changed.

Found something wrong or out of date — or built a version worth sharing back? Open an issue.

Related

• Series overview
• Lecture 1: Introduction to Connectomics
• Lecture 2: Tools and Methods
• Technical training units — the long-form material behind these slides
• Journal club — papers and discussion prompts