Connectomics

Building cellular-scale wiring diagrams of nervous systems: slice a brain into millions of thin sections, image each one with an electron microscope, and reconstruct the 3D shape of every neuron and every synapse between them. The output — a connectome — is a static map of what connects to what.

Held here as a machine-learning task, not as neuroscience. Reconstruction is image segmentation at a scale no one can do by hand, and the corpus's interest is in how the labels for it are produced.

The pipeline, as one project runs it

From male-fly-connectome (Google Research with HHMI Janelia, 2026-09-03):

1. Segment automatically. Flood-filling networks — convolutional networks that start from a single pixel and grow outward to claim every pixel of the same object — turn stacks of flat EM images into 3D neural shapes. PATHFINDER is the current reconstruction system.
2. Proofread by hand. Human experts verify and annotate the automated result. This takes "years of human effort" for a fly, and it is the step that sets the schedule.

The two halves make the field a useful case for training-data-quality: the automated pass and the human pass are not competing strategies here, they are consecutive, and the humans are correcting a model's output rather than labelling raw data.

Why the scale numbers matter

A fly's central nervous system is 166,000 neurons and 125 million synapses; a human brain is 86 billion neurons, which the post states plainly is not mappable with current techniques. The progression in one lab's own releases — a fully automated female fly brain in 2019, a human-verified half brain in 2020 (25,000 neurons, 21 million connections), the complete male CNS in 2026 — is a record of the verification cost, not of the segmentation cost.

The viewer is part of the release

A connectome is only useful if someone can look at it, and 166,000 neurons is not a table anyone reads. neuroglancer — Apache-2.0, WebGL, purely client-side — is how these datasets are actually published: point a browser at a URL and the volume streams in. Its precomputed format has become the interchange layer around it, with Julia and Python readers and meshing pipelines built by other labs.

What this page does not hold

No error rate, no segmentation accuracy, no proofreading-hours figure, no independent evaluation of any reconstruction system. One first-party blog post is the entire source, and the Cell paper behind it was verified as a citation and not read.

Related

male-fly-connectome · neuroglancer · training-data-quality · machine-learning · hhmi-janelia