A connectomics milestone: mapping the complete male fruit fly brain

Michał Januszewski and Viren Jain, Google Research, 2026-09-03. The announcement that a complete wiring diagram of the male Drosophila melanogaster brain and central nervous system — over 166,000 neurons and 125 million synaptic connections — has been released, from a project led by HHMI Janelia and described as "a decade-long partnership."

T3, on the same reading this spoke applied to timesfm-3: a first-party post about the poster's own work. What it describes about the pipeline can be taken at face value; what it claims about how well that pipeline works cannot be checked from here, and this post publishes no accuracy figure at all.

The paper underneath is peer-reviewed and was verified, not read. Crossref returns "Sexual dimorphism in the complete Drosophila male central nervous system connectome," Cell 189(18), September 2026, doi 10.1016/j.cell.2026.08.015, 111 authors, first author Stuart Berg, last author Gregory S.X.E. Jefferis. (The blog renders the title in a different word order.) Its methods, error rates and validation are where the real evidence sits, and this page does not hold them.

Why a fly brain is in a machine-learning wiki

The subject is neuroscience and there is no spoke for it. What routes here is the pipeline, and it lands on this spoke's most-worked question — where labels come from — with both of the answers the corpus has been separating, running in one system.

The pipeline. Slice a brain into millions of thin sections, image each with an electron microscope, then reconstruct 3D neural shapes from flat images. The segmentation is done by flood-filling networks: convolutional networks that start at a single pixel and grow outward to claim every pixel belonging to the same object. Then a team of human experts at Janelia proofreads the result — the post says the connectome "has been annotated and verified, or proofread" by people.

Scale, over three releases. 2019: a fully automated reconstruction of a female fly brain, unverified. 2020: a human-verified half brain — 25,000 neurons, 21 million connections, "a record at the time." 2026: the full male CNS at 166,000 neurons and 125 million connections. The automated step arrived first; six years of the difference is the verification.

The two label strategies, in one system

training-data-quality has spent two months separating collect-and-label from manufacture-the-labels (code-as-world-report). This project runs both, and says which one binds.

• Manufactured supervision, and it worked. "A recent effort incorporated synthetic neurons into the training data, successfully improving the speed and accuracy of our state-of-the-art reconstruction system, PATHFINDER." Synthetic structures are generated into the training set for a segmenter whose real targets are hand-traced neurons — the same move as reading labels off a simulator, in a domain where the ground truth is a physical object rather than a physics engine.
• And the human layer is still the constraint. "Currently, mapping the fruit fly brain requires years of human effort just to verify and annotate the neural shapes." The stated goal is "reducing this need for manual error correction," so that "research groups can tackle even larger brain mapping projects within reasonable budgets and timelines."

That is the corpus's cleanest instance of the boundary the synthesis drew on 2026-08-31: manufacture where the labels can be generated, annotate where they cannot, and the annotation half sets the schedule. Here the two are not alternatives at all — the synthetic data trains the model, and the humans correct what the model gets wrong, which makes proofreading a post-hoc label-error correction pass on the model's output rather than a pre-hoc labelling step. pervasive-label-errors measured wrong labels sitting undetected in finished test sets; this is a field that assumes them and budgets years for removing them.

What the post does not say

No accuracy number, no error rate, no proofreading-hours figure, no before/after for the synthetic-data change beyond "speed and accuracy," and no comparison against any other reconstruction system. This is the same shape as timesfm-3 — a first-party post asserting an improvement while publishing no score — and it is worth recording twice, because the pattern is the publisher's rather than the project's.

Also unmeasured here: how much of the 166,000-neuron result is automated output versus corrected output, which is the one number that would say what the decade actually bought.

What else the post announces

Vertebrates are next, and three items are named: a Nature paper led by Columbia on the elephantnose fish hindbrain ("Connectome analysis of a cerebellum-like circuit for sensory prediction"), which the post says is the first work to combine a connectome with other data to study plasticity and learning; an upcoming Harvard collaboration on the larval zebrafish brain; and two released datasets, ZAPBench and a preliminary Fish Fire&Wire dataset combining whole-brain activity and structure in the same larva. The viewer, Neuroglancer, is open source — and is paged separately, because the release of a connectome is in practice the release of a viewer URL.

Cross-spoke, noted not paged: the neuroscience content — sexual dimorphism, courtship and aggression circuits, the ventral nerve cord as the brain-to-body link — has no owner in this hub; ../psychology-wiki holds behavioural neuroscience but not neuroanatomy. If two more sources like this arrive, that gap is worth reopening rather than settling by precedent from this ingest.

Related

training-data-quality · connectomics · code-as-world-report · pervasive-label-errors · timesfm-3 · neuroglancer · viren-jain · michal-januszewski · hhmi-janelia · google · synthesis