The scaling study: our proofreading model against training data, and its transfer to expanded tissue

Brain Insights, measured 2026-09-28 (scaling) and 2026-09-30 (transfer). Every number on this page is the number the charts on the memo page are drawn from. The separate pre-registered test of a repair pipeline on nine held-out cubes of the public mouse cortex volume is on the research page, with its download.

What was measured

A small image model, ConvNeXt-Atto with 7 million parameters, was fine-tuned on ConnectomeBench2, the public benchmark of proofreading decisions built from the logged corrections of the fly and mouse electron-microscopy maps. Two tasks: split correction (the benchmark's false_split_correction: should these two pieces be joined?) and merge identification (false_merge_identification: is this one cell wrongly joined from two?). The model was trained on 1,000, 4,000, 16,000, 79,500 and all 531,734 training decisions, three seeds per point, and scored by balanced accuracy on the same 23,730 held-out decisions each time. The human-expert figures are the benchmark preprint's own, from a small sample of experts: 93.0 percent [87.2, 97.7] on split correction and 84.1 percent [75.8, 91.7] on merge identification.

Two models of ours appear on this site and they are different sizes of the same architecture. The 7 million parameter model is the one swept here, because a sweep of five sizes of training set times three seeds is cheap at that size. A 56 million parameter version (ConvNeXt-T), which the research page calls Proofreader, scores 97.75 and 95.52 percent on the benchmark's full 101,929-decision test split at full data (three seeds, seed sd 0.06 and 0.05); it is the model the pipeline on the research page is compared against, not part of that pipeline.

Accuracy against training data

Mean of three seeds per point. At full data the model scores 97.4 percent on split correction against the human 93.0, inside the human interval, so a match; and 94.4 against 84.1 on merge identification, above the interval, a clear lead. The benchmark authors' own models report 97.2 (ViT-L) and 93.0 (ViT-B) on their held-out split, so at full data ours is level with them, not ahead of them.

The fit. Error (one minus balanced accuracy) against training decisions is a power law over the sweep from 1,000 to 79,500, fitted on the seed means: exponent 0.201 for split correction and 0.158 for merge identification. Each doubling of the training data therefore removes about 13 and 10 percent of the remaining error; halving the error takes about 31 and 80 times the data. The curve had not leveled off at the last point.

Transfer to expanded tissue

The same model, trained only on electron-microscopy reconstructions, was then applied to E11 Bio's PRISM data: expansion microscopy with protein barcodes, from which the labels were derived. A held-out region of E11's evaluation volume supplied the test decisions (564 for split correction, 672 for merge identification). Two starting points were fine-tuned on 250 to 3,865 barcode-labeled decisions, three seeds per point: our electron-microscopy checkpoint, and the same architecture from generic ImageNet weights. The score is AUROC, where 0.5 is chance.

With no retraining the electron-microscopy model scores 0.695 on split correction and 0.758 on merge identification, against 0.5 for chance. Starting from it, 250 labels match what the generic start reaches with about 2,000 on split correction and 1,000 on merge identification, which is the “four to eight times the labels” on the memo page. Accuracy rose with every label added at every point measured.

What this does and does not show

It shows that the human proofreading decision can be learned from logged corrections, that accuracy rises predictably with data, and that what is learned on electron microscopy carries over to expanded light-microscopy tissue. It does not show a finished brain: scoring decisions on a benchmark is not running a pipeline over a volume, and the one end-to-end test we have published, on the research page, closed about a fifth of the breaks in nine held-out cubes while adding a few wrong joins. The exponents say how much data the next gains need; the memo's FAQ (“Your exponents, 0.20 and 0.16”) says which levers the plan expects to move the curve rather than ride it.

Sources: ConnectomeBench2 preprint (human figures, task definitions, the authors' model); E11 Bio, PRISM (the expansion data and barcodes); our runs, logged in the repository that produced this page (ConvNeXt-Atto, three seeds per point, balanced accuracy on 23,730 held-out decisions; PRISM AUROC on a held-out region, seed sd in parentheses).