NeuroTrailblazers banner

Pick a problem, take the ramp

Each card below is an open problem that passes two tests: an undergraduate team can make real progress in one to two semesters, and progress matters to the field’s scaling effort — the NIH BRAIN CONNECTS program and projects like MouseConnects.

Every card has an on-ramp: five ordered steps from “never touched the data” to “scoped semester project with a named customer.” Steps 1–3 are each about an hour; step 4 is your first real measurement and may take a week. Every ramp crosses the data barrier the same way — through Getting Started with Data — and every ramp’s third step is reproducing a known number, because calibration comes before novelty.

The full analysis behind every card — state of the art with numbers, why CONNECTS needs it, scoped project menus, and a worked scoping example — is in the deep dive. Not sure which card is yours? The deep dive’s decision protocol is four questions; the first — what is your team’s strongest existing skill? — usually settles it, and each card’s “Team fit” pill is the shortcut.

1. The Proofreading Budget

Team fit: CS · data science · MLLaptop + Colab

Automated segmentation still needs human correction, and nobody knows how much proofreading a scientific claim actually requires. Measure it.

Customer: MouseConnects pipeline planners

On-ramp Deep dive

2. Synapse Detection That Generalizes

Team fit: Biology · ML · careful eyesLaptop + Colab

Detectors trained on one dataset degrade on the next, and the ground truth to measure that degradation mostly doesn't exist. Build some.

Customer: Every cross-region comparison in CONNECTS

On-ramp Deep dive

3. Cell Types Across Datasets

Team fit: Biology · data scienceLaptop only

The male and female fly connectomes are both public and the comparative literature is months old. Match types across sexes, individuals, and modalities.

Customer: Cross-dataset science in CONNECTS

On-ramp Deep dive

4. Graph Algorithms at Connectome Scale

Team fit: Math · CS · algorithmsLaptop only

An open optimization challenge with a public leaderboard, null models nobody agrees on, and stability analyses nobody runs. No biology background needed to start.

Customer: FlyWire leaderboard + every motif paper

On-ramp Deep dive

5. Structure → Function and Its Limits

Team fit: ML · computational neuroscienceColab GPU helpful

Wiring predicts orientation tuning well and receptive-field size badly. Map where the predictive boundary lies — the field's central open question.

Customer: The case for the next connectome

On-ramp Deep dive

6. Data Logistics at Exabyte Scale

Team fit: Systems · data engineeringLaptop + cloud egress patience

A whole mouse brain is ~an exabyte. How much can you compress EM before the science changes? Nobody has published the systematic answer.

Customer: CONNECTS acquisition budgets

On-ramp Deep dive

7. Annotation Science

Team fit: Education · psychology · HCILaptop + human participants

Connectomics runs some of the largest distributed annotation efforts in science, with almost no science about the annotators. Home turf for education, psych, and HCI teams.

Customer: Every distributed proofreading campaign

On-ramp Deep dive

The on-ramps

An on-ramp is not the project — it is the first two weeks, made explicit so that no team stalls at “we couldn’t get the data working.” Each step has a done when you can check without a mentor. If a step’s done when fails twice, that’s not a detour; it’s information — bring it to Ask an Expert with what you tried.

🔀 On-ramp 1: The Proofreading Budget

Automated segmentation still needs human correction, and nobody knows how much proofreading a scientific claim actually requires. Measure it. — full problem statement.

  1. Orient. Read the problem in full, then Error taxonomy and Metrics and QA.
    Done when You can explain to a teammate why an undetected merge error is worse than an undetected split error.
  2. Get data in hand. Work through Getting Started with Data to Level 2: a FlyWire snapshot loaded in pandas, plus a working CAVE token for versioned queries.
    Done when A notebook cell prints the edge count of a pinned materialization version.
  3. Calibrate. Reproduce the published FlyWire neuron count (139,255) from your downloaded snapshot.
    Done when Your count matches, and you can say exactly which filter (proofread flag, synapse threshold) the published number assumes.
  4. First measurement. Pull the same connectome at two public materialization versions and diff them: edges gained, edges lost, degree-distribution shift. This is one data point on a robustness-versus-proofreading curve.
    Done when A two-version diff table and one plot, with both version IDs recorded in the notebook.
  5. Scope the semester. Pick one of the three scoped projects in the deep dive (robustness curves, merge-detector benchmark, effort telemetry) and write a half-page abstract addressed to the customer.
    Done when The abstract names the minimum reportable result — the deliverable you get even if the ambitious version fails.

🎯 On-ramp 2: Synapse Detection That Generalizes

Detectors trained on one dataset degrade on the next, and the ground truth to measure that degradation mostly doesn't exist. Build some. — full problem statement.

  1. Orient. Read the problem in full, then Synapse classification and skim Unit 05.
    Done when You can state the three EM criteria for calling a synapse without looking them up.
  2. Get data in hand. Work through Getting Started with Data to Level 3-lite: one small EM imagery cutout from a public volume, plus the CREMI ground-truth volumes (plain HTTP download, no account).
    Done when You can open the cutout in Neuroglancer and scroll through sections without it feeling alien.
  3. Calibrate. Find five synapses by eye in your cutout using the three criteria, then check yourself against the dataset’s released synapse annotations.
    Done when At least four of your five calls appear in the released table — or you can explain each disagreement.
  4. First measurement. Load a CREMI volume with its annotations and compute synapse density per unit volume; compare against the same measurement in a MICrONS subvolume’s synapse table.
    Done when A density comparison with an honest paragraph on why the numbers differ (tissue? annotation protocol? detector?).
  5. Scope the semester. Pick cross-dataset degradation study or ground-truth annotation (with agreement statistics) and write the half-page abstract.
    Done when If annotating: your protocol names the annotator count, the region, and the agreement statistic before anyone starts clicking.

🧬 On-ramp 3: Cell Types Across Datasets

The male and female fly connectomes are both public and the comparative literature is months old. Match types across sexes, individuals, and modalities. — full problem statement.

  1. Orient. Read the problem in full and Neuron type identification.
    Done when You can explain the difference between morphology-based and connectivity-based typing, with one strength and one failure mode each.
  2. Get data in hand. Work through Getting Started with Data: Level 1 gets you FlyWire’s annotation and classification tables (no account); Level 2’s neuPrint token gets you the male CNS.
    Done when Two DataFrames — female types, male types — loaded side by side.
  3. Calibrate. Reproduce the per-superclass neuron counts that Codex displays for the female brain from the raw annotation table.
    Done when Your groupby matches what the Codex interface shows for the same release.
  4. First measurement. Pick one cell type reported in both sexes, and compare: neuron count, mean synaptic partners, top-5 partner types. You are now doing comparative connectomics.
    Done when A one-page comparison; flag every difference and label it stereotypy, dimorphism candidate, or probably-methodological.
  5. Scope the semester. Pick the female-versus-male recount, the typing-method shootout, or the transfer-learning classifier, and write the abstract.
    Done when The abstract states which dataset releases and which type annotations are being treated as ground truth — and why that's defensible.

🕸️ On-ramp 4: Graph Algorithms at Connectome Scale

An open optimization challenge with a public leaderboard, null models nobody agrees on, and stability analyses nobody runs. No biology background needed to start. — full problem statement.

  1. Orient. Read the problem in full, then Network analysis methods and Motif analysis.
    Done when You can explain why motif significance depends on the null model, with the reciprocity-under-distance-null example.
  2. Get data in hand. Work through Getting Started with Data Level 1: the FlyWire connections table. Then download the Minimum Feedback Challenge graph and baseline code.
    Done when The connectome loads as a directed weighted graph in igraph or graph-tool (not NetworkX — see the guide for why).
  3. Calibrate. Recompute the challenge graph’s total synapse weight and check it against the number stated on the challenge page.
    Done when Your total matches the challenge's stated denominator exactly.
  4. First measurement. Run the provided baseline ordering and reproduce its score; or compute reciprocity in one neuropil under an Erdős–Rényi null and a configuration-model null.
    Done when Either the baseline score reproduces, or you have two z-scores for the same statistic that disagree — and understand why that's the finding.
  5. Scope the semester. Pick the challenge attack, the null-model sensitivity atlas (a full worked scoping example is in the deep dive), or approximate-algorithm benchmarking.
    Done when For the challenge: your abstract names the algorithm family you'll try and the compute budget. For the atlas: it names the three published claims you'll re-test.

⚡ On-ramp 5: Structure → Function and Its Limits

Wiring predicts orientation tuning well and receptive-field size badly. Map where the predictive boundary lies — the field's central open question. — full problem statement.

  1. Orient. Read the problem in full and NeuroAI bridge.
    Done when You can summarize what connectome-constrained models got right in the fly visual system and where connectome-only prediction breaks down.
  2. Get data in hand. Per Getting Started with Data Level 1: install the released fly visual system model (flyvis) and pull the ZAPBench loaders — both are pip/git installs with no credentials.
    Done when The pretrained model runs end to end on the demo stimulus in Colab.
  3. Calibrate. Reproduce one published response property from the released model’s own demo (e.g., a cell type’s direction selectivity).
    Done when Your reproduction matches the released figure qualitatively, and you know which random seed and checkpoint produced it.
  4. First measurement. Perturb one thing — shuffle synapse signs within a cell type, or delete the weakest decile of edges — and measure how the reproduced property degrades.
    Done when A one-plot sensitivity result: property X survives perturbation A but not B.
  5. Scope the semester. Pick a ZAPBench benchmark entry, a prediction-versus-physiology audit, or a full lesion study, and write the abstract.
    Done when For the audit: the abstract lists the specific published predictions you will check and the physiology sources you will check them against.

💾 On-ramp 6: Data Logistics at Exabyte Scale

A whole mouse brain is ~an exabyte. How much can you compress EM before the science changes? Nobody has published the systematic answer. — full problem statement.

  1. Orient. Read the problem in full, then Data formats and Reconstruction pipeline.
    Done when You can sketch the path from microscope to Neuroglancer tab and name the storage format at each hop.
  2. Get data in hand. Work through Getting Started with Data Level 3: CloudVolume cutouts from a public EM volume at two different mip levels.
    Done when You predicted a cutout's size in bytes from its dimensions and dtype before downloading, and were right.
  3. Calibrate. Measure actual transfer time and effective bandwidth for the same cutout at mip 0 vs a downsampled mip.
    Done when A small table: mip, voxels, bytes, seconds — and the ratio makes sense.
  4. First measurement. Recompress a cutout at increasing lossy ratios and plot a fidelity metric (SSIM to start) against compression ratio. This is the seed of the compression-safety benchmark.
    Done when A rate-distortion curve on real EM data, with the code clean enough to swap in a downstream model as the metric.
  5. Scope the semester. Pick the compression-safety benchmark (upgrade the metric from SSIM to ‘does segmentation output change’), access-pattern profiling, or the artifact QA screener.
    Done when The abstract names the downstream model whose output defines 'scientifically safe', not just an image metric.

🧑‍🤝‍🧑 On-ramp 7: Annotation Science

Connectomics runs some of the largest distributed annotation efforts in science, with almost no science about the annotators. Home turf for education, psych, and HCI teams. — full problem statement.

  1. Orient. Read the problem in full, then the proofreading side quest overview and Error taxonomy.
    Done when You can name the error classes a proofreading study would score, and what 'expert consensus gold standard' means operationally.
  2. Get data in hand. Per Getting Started with Data Level 0: open a public segmentation in the browser (no install, no account). Your ‘data’ is also this site’s own worksheets — the standardized training instrument.
    Done when You can drive Neuroglancer well enough to demonstrate a merge error and a split error to someone else.
  3. Calibrate. Do the proofreading side-quest scenario yourself, timing every correction by error type.
    Done when Your own time-per-correction table, by error class — the pilot datum for any study design.
  4. First measurement. Run 2–3 volunteers through the same exercise informally and compute inter-annotator agreement and per-error-type times. (A real study needs ethics review — start that paperwork now, it takes longer than the pilot.)
    Done when An agreement statistic plus a draft IRB/ethics application if you're going the human-subjects route.
  5. Scope the semester. Pick the learning-curve study, the consensus-aggregation benchmark, or retrospective analysis of public edit logs (which sidesteps most human-subjects overhead).
    Done when The abstract states whether the design needs ethics approval and the timeline accounts for it.

After the ramp

The last step of every ramp is the same on purpose: a half-page abstract addressed to the problem’s customer, naming the minimum reportable result — the deliverable you get even if the ambitious version fails. Write it in week two, not week ten. Then work the project menu in the deep dive, and when your team needs the skills a step assumes, the technical course and modules are the reference layer behind all of it.