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.

Teach a 90-minute session

This selected-slide route is Session 4 in the short teaching sequence, but Lecture 3 / Module 9 in the original graduate course. Use the existing 58-slide deck and embedded notes; slide numbers include the cover. The full two-hour lab on slide 54 is an extension, not additional work required in this short session.

Outcome: build a directed graph under stated rules, compare reciprocity with an explicit null, and produce an analysis card that distinguishes enrichment, evidence and error sensitivity. Bring the study brief, audit and provenance record from the preceding sessions. Fractions are sufficient; no graph-library knowledge is assumed.

Preparation: open the deck and worksheet/key, check the definitions of reciprocal pairs versus reciprocated edges, and provide calculators. Optional Python 3 code enumerates a tiny synthetic null without packages or accounts. Test both thresholds before class; the same exact census is supplied on paper as a fallback.

• 0–10 min, slides 1–4, 12: trace the endpoint back to segmentation and synapses. Ask whether a split, merge or partner mistake would affect this endpoint in the same way. Do not assume error directions cancel or always have the same sign.
• 10–20 min, slides 16, 19–21: discuss proofreading priorities. Ask which correction would change the analysis most, and what evidence supports that ranking.
• 20–30 min, slides 25–28: state node, edge, threshold and boundary rules. On the worksheet, count {A,B} once even though it supplies two directed edges.
• 30–40 min, slides 29–30, 34: choose null constraints from the question. Explain the worksheet’s fixed-edge-count model and prespecified upper-tail rule. Distinguish it from an independent-edge model that fixes only the expected number of edges.
• 40–75 min: pairs complete the worksheet: 10 minutes constructing graphs, 10 comparing nulls, 5 checking the uncertain edge, and 10 writing the card. At minute 60, check whether anyone has confused enrichment ratio with tail probability.
• 75–85 min, slides 36–38, 40: exchange cards and compare with the model responses. A one-edge deletion is a scenario check, not a confidence interval. Ask what data would support a realistic error model.
• 85–90 min, slide 52: collect cards and exit tickets: “My result is conditional on __; the next measurement that could change it is __.” A justified negative result meets the learning objective.

Unselected slides are optional depth for this route; 57–58 carry references and credit. The full lecture adds segmentation algorithms, the larger hypothetical null comparison and applications/NeuroAI. Use the key’s four-dimension rubric for the short session rather than the original course exam requirements on slide 55.

Structure

The full, approximately 150-minute route uses all three parts.

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 sensitivity under a stated model, non-claim, and provenance. The full lab uses measured error rates; the short session uses a declared toy scenario.

The centrepiece

The full lecture compares hypothetical null summaries for 100 neurons, 1,200 edges and 210 reciprocal pairs. The interpretation depends on the null and test direction, not just the enrichment ratio. Those illustrative summaries are not outputs from supplied biological data. The short-session worksheet instead provides a four-node graph whose entire fixed-edge-count null can be checked exactly.

Notes for whoever teaches it

Error direction depends on the graph and endpoint. Merging nodes can combine partner lists, remove self-edges or collapse duplicate edges. Splitting can remove or redistribute connections. Neither a universal upward motif bias nor cancellation is guaranteed. Demonstrate a concrete perturbation and state what it changes.

Parts A and B join at the error-sensitivity check. The validation sample informs an error model; perturbing the graph then shows how the endpoint behaves under that model. Rates alone do not specify which nodes or edges are wrong. The resulting spread is conditional sensitivity, not automatically a confidence interval.

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.

The H01 cover image retains CC BY 4.0, with credit to Lichtman Lab / Harvard and Connectomics at Google, Shapson-Coe et al. (2024), doi:10.1126/science.adk4858. Preserve that separate attribution when adapting the deck.

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