09 Connectome Analysis and NeuroAI

Technical Training: Nanoscale Connectomics

Session outcomes (60 minutes)

  • Formulate a motif/graph hypothesis with explicit estimand.
  • Choose and justify a null model.
  • Report bounded claims with uncertainty and reproducibility metadata.

Pedagogical arc

  • Hook: why graph/motif results are often overclaimed.
  • Model: hypothesis -> query -> null -> interpretation.
  • Practice: design and critique analysis plans.
  • Check: one bounded claim + non-claim pair.

Motivation and framing

  • Structure can constrain models; it does not automatically explain intelligence.

Representation framing

  • Define representation before inference.

Limits of reverse engineering claims

  • Teach boundary statements as required output.

Analysis workflow overview


Hypothesis -> Query -> Search -> Null comparison -> Interpretation -> Reproducibility package

Motif search context

  • Distinguish candidate motifs from validated mechanisms.

Query language and reproducibility

  • Human-readable queries reduce hidden assumptions.

Complexity constraints and feasibility

  • Computational limits are part of methodological validity.

Historical benchmark caution

  • Use old benchmark values as context, not current truth.

Comparative analysis caveats

  • Cross-dataset claims require aligned preprocessing and null assumptions.

Misconceptions to correct

  • "Significant motif enrichment implies mechanism."
  • "One null model is enough for any claim."
  • "Query scripts without provenance are acceptable."

Activity

Design one analysis card with:

  • hypothesis,
  • estimand,
  • null model,
  • success criterion,
  • non-claim,
  • provenance fields.

Rubric checkpoint

  • Pass: coherent hypothesis-null-estimand chain.
  • Strong: includes sensitivity analysis and boundary statement.
  • Flag: result-first narrative without methodological controls.

External paper figure integration

  • Bassett, Zurn, Gold 2018 (model taxonomy and claim type framing).
  • Large-scale connectome motif-analysis papers with null-model details.
  • Graph-method papers showing sensitivity to preprocessing choices.

External inserted figure (open license)

References and attribution