reconstruction

Reconstruction is the step between a mountain of electron microscope imagery and a usable wiring diagram: tracing every neuron through the volume and marking every synapse. For 139,255 neurons and roughly 50 million connections, this is not a task a team of humans can do by hand in a lifetime. The fly connectome exists because machine segmentation got good enough to do the bulk of it, and because a large community of scientists and volunteers spent years proofreading the machine's output.

segmentation

A segmentation model assigns every pixel of the volume to a neuron. Errors come in two kinds: splits, where one neuron is broken into fragments, and mergers, where two neurons are fused into one false object. A merger is worse than a split. A split loses a branch; a merger invents a connection pathway that never existed. Proofreading is largely the work of hunting mergers.

proofreading

The fly brain reconstruction was proofread openly, by a distributed community, neuron by neuron, over years. This matters for the archive's provenance: the wiring diagram our partitions are cut from is public science, checked by thousands of eyes, not a private artifact. What Eon Systems built on top of it, and what we built beside them, are separate things resting on the same shared map.

from diagram to dynamics

A traced diagram still does not run. To emulate it, each neuron becomes a computational unit, each synapse a coupling with a sign and a strength, and the whole system is integrated forward in time. The modeling choices here are surprisingly spare. The published emulation work used simple neuron models, uniform assumptions within neurotransmitter classes, and no per-synapse tuning against behavioral data. The point was precisely to see how far structure alone would go. How far it went is the subject of the 91% result.

reconstruction errors and the archive

No reconstruction is perfect. Residual split and merger errors exist in the diagram at some low rate, and therefore exist in every partition we run. We note this because the errata reference it once, in connection with LA-05, and we want the reader to have the mundane explanation available before encountering the less mundane one. A merger error is a wire that should not be there. Most of the time, that is all it is.