Facilitator Guide

Read this first: the constraint that shapes everything

Connectomics teaching has a structural problem that most technical curricula do not. The gap between what a learner can be told and what they can do is unusually wide, because the core skills — reading an EM image, judging a segmentation, choosing a null model — are perceptual and judgmental, not procedural. You cannot transmit them by explanation. A learner who can recite the three criteria for calling a synapse will still, on their first real patch, call a tangentially cut membrane a synapse.

Everything below follows from that. The design principle throughout is:

Time spent making and comparing judgments beats time spent receiving explanations. Aim for at least half of contact time on learner judgment with immediate comparison against a reference or a peer.

The corollary is uncomfortable and worth stating plainly: a lecture-only delivery of this material will feel good, review well, and produce very little transferable capability. If you have 60 minutes and can either explain thoroughly or have learners make 20 scored judgments, choose the judgments.

Session design

The standard four-phase structure

Every module and technical unit is built to run in this shape:

1. Learn — capability target and the concepts needed to attempt it. Short.
2. Practice — a studio activity producing an artifact with evidence.
3. Check — a rubric-based competency check.
4. Teach — slides and worksheet, so the learner can transfer it onward.

The 60-minute template

The modeling phase is where most facilitators lose the session. The temptation is to present clean examples that make the method look reliable. Do the opposite: choose an example where you genuinely have to weigh conflicting evidence, and narrate the weighing. Learners calibrate their own standard for “how sure should I be?” almost entirely from watching an expert be uncertain in public.

Question stems that work

Replace “is that right?” — which teaches learners to seek your approval — with:

• “Show me your evidence chain before your label.”
• “Which cue family is that from?” (Units 05–06)
• “Which cue would you drop first if the staining got worse?”
• “What would change your mind?”
• “What is the cheapest observation that would settle this?”
• “If you had to be wrong in one direction, which would you choose, and why?”

The last one is the highest-value stem in the set. It surfaces whether the learner understands the asymmetric cost of merge versus split errors, which is the conceptual core of the whole technical track.

Differentiation across the four personas

The site’s learner personas are not decoration — they describe genuinely different failure modes, and a session that works for one can fail another. In a mixed cohort you will usually have all four.

Julian — first-generation undergraduate

Predictable friction: hidden curriculum. Not the content — the norms. Whether it is acceptable to say “I don’t know”, how to ask a question without appearing under-prepared, what “read the paper” actually means in practice, whether their uncertainty is normal.

What to do:

• State norms explicitly, in the first five minutes, every time. “In this session, ‘uncertain’ is a passing answer and I will say so out loud when someone uses it well.”
• Give a worked example of the process, not just the content — including how you decided which paper to read first and what you skipped.
• Scaffold the first judgment heavily, then remove scaffolding fast. Under-challenge is as damaging as over-challenge here.
• Make office hours scheduled and normal, not available on request. “Optional” reads as “for people who are struggling” to a learner watching for signals.

Maya — graduate student bridging computation and biology

Predictable friction: depth is uneven and she knows it. Strong on methods, less secure on the biology, or the reverse — and reluctant to expose the gap in front of a cohort.

What to do:

• Pair her with someone whose gap is complementary and make the pairing’s purpose explicit, so exposure of a gap reads as the assignment rather than as a deficiency.
• Give her teaching responsibility for the part she is strong in. Explaining is the fastest route to consolidation, and this cohort member usually wants it.
• Push her toward the assumption-naming habit (Bin B claims, Unit 01). Learners who are strong at methods tend to be the ones who over-claim from them.

Amir — industry researcher entering the field

Predictable friction: speed and vocabulary. Used to shipping; frustrated by the pace and by biology terminology that seems arbitrary. Prone to assuming the biology is a detail he can pick up later.

What to do:

• Front-load the dictionary — much of the apparent difficulty is vocabulary, and it is fixable in a week.
• Let him start at Unit 04 or Unit 08 where his systems intuitions transfer directly, then send him back to Units 05–07. Entry point order is negotiable; skipping 05–07 entirely is not.
• Be direct about why the biology matters. “A model trained on annotations from people who cannot distinguish an astrocytic process from a dendrite will learn to make that mistake at scale” lands better than an appeal to thoroughness.

Dr. Nguyen — faculty mentor

Predictable friction: wants material she can hand to a trainee tomorrow, and has no time to adapt it.

What to do:

• Point at the teaching kits and the rubrics rather than the reading. What she needs is the assessment instrument and the run-of-show.
• The technical units are written so that a trainee can work through one unsupervised and produce a reviewable artifact. That is the property she is looking for.

Mixed-cohort rule

Do not differentiate by content. Differentiate by scaffolding and entry point, holding the capability target constant. Everyone reaches the same competency check; they arrive with different amounts of support and by different routes. Splitting the target itself produces a two-tier cohort and is visible to everyone in the room.

Assessment that scales

Grade the reasoning, not the answer

For every judgment task, the artifact is label + confidence + evidence chain + one alternative considered. Grade the last three. A correct label with no evidence chain should not outscore a well-reasoned incorrect one, and saying so publicly changes behavior within one session.

This also solves the scaling problem: evidence chains can be peer-reviewed reliably against a rubric, whereas correctness often cannot be judged by a peer at all.

Calibration is the metric that matters

Track, per learner: accuracy overall, and accuracy within their high-confidence calls. A learner at 84% overall with 100% accuracy on high-confidence calls and a non-zero uncertain rate is performing better than one at 90% flat with no uncertain calls — because the first learner’s confidence carries information a team can act on.

Teach this explicitly. It is counter-intuitive to learners trained by conventional exams, and it is the single most transferable habit in the track.

Peer review, structured

Peer assessment scales and works, provided:

• The rubric is specific enough that two reviewers agree.
• Reviewers assess evidence quality, not correctness.
• You spot-check ~10% and publish the agreement rate between your marks and theirs. If peer marks diverge from yours, that is a rubric problem to fix, not a learner problem to complain about.

Run a calibration round before the cohort matters

Have everyone score the same three items, then compare publicly before proceeding. Inter-rater spread is typically large on the first attempt and shrinks sharply after one round of discussion. That shrinkage is the learning, and it doubles as a live demonstration of why annotation protocols need calibration sessions — the same mechanism the field uses in production.

Running without live instruction

Much of the intended audience will work through this material alone. Design for it.

A self-study path that works:

1. Read the unit’s Before you start and What you’ll be able to do.
2. Read the numbered sections, attempting each Check yourself before opening the answer. Opening it first converts a retrieval exercise into re-reading, which feels productive and is not.
3. Do the lab. Produce the artifact.
4. Self-grade against the rubric, honestly, then find one person to review the artifact — a peer, a mentor, a lab-mate. One external review is worth more than three self-reviews, because the failure mode of self-assessment is not laziness but blindness to the thing you did not know to check.
5. Log which rubric row you scored lowest on and target it in the next unit.

What a lone learner cannot get from the page, and should seek deliberately: calibration against other people’s judgments, and exposure to genuinely ambiguous cases someone has curated. Both are reasons to join a journal club or a community proofreading effort, and it is worth telling learners that directly rather than implying the reading is sufficient.

Failure modes specific to this material

Lecturing the perceptual units. Units 05–07 taught as lectures produce learners who can describe cues and cannot apply them. Convert at least half the time to scored drills on short z-stacks.

Single-image practice. Any ultrastructure or glia exercise built from single images teaches a habit — the single-plane call — that the units explicitly tell learners to break. Always use short z-stacks, even when it is more work to prepare.

Clean examples only. A curated set of unambiguous patches produces overconfidence that collapses on real data. Every drill needs ambiguous cases, and at least one case where the correct answer is “this is a segmentation error, not a biological structure”.

Rewarding speed. If throughput is what you praise, throughput is what you get, with accuracy quietly falling. Praise well-justified uncertainty at least once, early, and visibly.

Skipping the null-model discussion in Unit 09. It is the least visual and most skippable section and it is where most published errors in this field live. If you must cut something from Unit 09, cut the survey of methods in §4, not the worked reciprocity example in §2.

Treating the labs as homework. The labs are the curriculum. The reading exists to make the labs possible. A cohort that does the reading and skips the labs has not done the course.

Preparation checklist

Before a session, confirm:

• You can state the capability target in one sentence without reading it.
• You have one worked example you will narrate, including where you are unsure.
• Your practice set includes ambiguous cases and at least one trap.
• Practice items are z-stacks, not single images, for any perceptual unit.
• The rubric is visible to learners before they start, not after.
• You know the two or three misconceptions this unit names, and how you will surface them.
• You have decided what “uncertain” earns, and you will say so out loud.
• Data access works — accounts, viewer, notebook — verified today, not last week.

Where to find materials