An illustrated argument · compiled from the public record · August 2026

The Unified Theory of
Will Gray-Roncal

On paper he is six different people: a brain-mapper, an AI researcher, a defense engineer, a learning engineer, a mentor of five hundred students, a public-health data scientist. This page argues the bibliography disagrees. Every project is the same machine: wring trustworthy, interpretable knowledge out of noisy, incomplete data — the graph is just his favorite container for it — calibrate whoever judges that knowledge, then hand the controls to someone new.

This is an interpretation, argued from sources — every claim links to one. His own one-line version waits at the end.

Fig. A — the same career, twice

Forty-four publications, filed two ways

Filed by field, the record looks like six unrelated careers. Filed by what each paper does — extract the knowledge, calibrate the judge, or widen the gate — it is one method, applied everywhere. Press the button; nothing moves between fields, only the filing changes.

Dot color = research thread (same colors everywhere on this page). Left state groups by publication venue's field; right state groups by the move the paper makes. Sources for every dot are in the plates below.

The theory

Three laws the bibliography obeys

A real theory has to be falsifiable against the record. Here are three claims; each one could have been broken by any of the ~90 entries in his publication list, and each survives contact with all of them. (One honest near-exception is flagged below — a theory with no exceptions should worry you.)

1. First, wring knowledge from noisy, incomplete data.

The raw material is never clean — electron micrographs with slicing artifacts, MRI scans that vary by scanner and day, clinical records with missing fields, unlabelled neural recordings, human behavior itself. The opening move is always to extract something interpretable from that mess: a structure you can question, test, and act on. His signature container for it is the graph — the dissertation is Images to Graphs for Discovery — but the graph is a representation, not the point: point processes, density maps, clusters and benchmarks appear wherever they serve the question better. Discovery is the point.

evidence: MRCAP (2012) · signal-subgraphs (TPAMI 2013) · images-to-graphs (2015) · X-ray cell-density maps (2017) · DotMotif (2021) · MYOW’s label-free representations (2021) · C. elegans motifs (2021) · COVID-19 patient clusters from messy hospital data (2023)

2. Then, calibrate the judge.

In every field he enters, he finds the same flaw: the instrument doing the judging has never itself been measured. Segmentation metrics that reward the wrong thing. "Superhuman AI" benchmarked against humans no one bothered to measure carefully. Pipelines whose results don't replicate. Admissions systems that score pedigree instead of ability. So he builds the calibration — a metric for the metric.

evidence: MRCAP repeatability (2012) · NRI (2018) · CONFIRMS (2021) · BENCHMARK R01 (2021–23) · human-performance framework (Sci Rep 2022) · aerial-search human benchmark (2021) · "Calibrate the Scientists, Not Just the Microscopes" (in press) · "Show Your Work" (2026 draft)

3. Then, hand over the controls.

A system is not finished when it works; it is finished when it runs without him — open source, cloud-hosted, fronted by a grammar or an API simple enough that a newcomer can drive it. And the newcomers are chosen deliberately: the students his field's gatekeepers would have filtered out are recruited, trained, and wired into the pipeline as working scientists — many of whom now appear above him on the author lists.

evidence: Open Connectome (2013) · Science in the Cloud (2017) · bossDB (2022) · intern toolkit (2021) · metadata standards (2022, 2024) · College Prep (2009–) · CIRCUIT (2017–) · NeuroTrailblazers · the student-to-first-author arcs traced in the story below

The promised near-exception. Stated as a sequence — knowledge, then judge, then gate — the theory implies the gate-widening was a conclusion the science led him to. It wasn't. The College Prep Program was co-founded in 2009: before MRCAP, before any graph existed, before the PhD. Law three ran first, in parallel, the whole time. The three moves survive as a kit he applies everywhere, but not as a derivation chain — and honestly, that's a better story. (A smaller misfit: the scale ladder isn't monotone in time — the mesoscale X-ray rung was filled after the nanometre floor, so Fig. 2's "down the ladder, then up" is an idealization the dates only roughly obey.)

The compressed form: extract the knowledge, calibrate the judge, widen the gate. The rest of this page is the evidence — a citation graph, a scale ladder, and every paper retold in plain language.

A fun supporting footnote, weighted accordingly: in 2016 — the hinge of the career — he and Eva Dyer posted an unreviewed, barely-cited arXiv preprint, “From sample to knowledge”, that sketches this very method: optimize the whole pipeline against the knowledge you want, not each stage’s surrogate metric. It is a preprint, so it anchors nothing here — the laws stand on the reviewed record — but it is lovely evidence that the through-line was articulated, not retrofitted. More below.

Fig. 1 — the related-work graph

One literature, wearing five disguises

His publications (filled disks, sized by citations), the prior work they build on (open rings), and the shared ideas that bridge threads (dashed boxes). Solid edges are real citations from reference lists; dashed edges are thematic links — my reading, not theirs. Click or tab to any node.

Select a node to see its story.

Fig. 2 — the spine

One ladder, from synapse to institution

Forget chronology; sort the bibliography by the scale of the thing being mapped. It tiles the whole ladder — from nanometre synapses to the institutions that decide who gets to do science — with no rung skipped, and the same three moves repeating at every rung. Dates show the career ran the ladder twice: down it (2010–2015, engineer to synapse), then back up (2015–2026, synapse to institution).

Chip border = thread color. Rung heights are ordinal, not to scale — the real ladder spans about nine orders of magnitude in metres and then stops having units at all. Details for every chip are in the plates below.

The story

Six acts, in the order the ladder implies

Act I · 2003–2012

Two threads, started in parallel

He begins as an electrical engineer — Vanderbilt, then USC, then years at the Johns Hopkins Applied Physics Laboratory working across "space, submarines, and biomedicine." In 2009, two things start at once, and the order matters: he co-founds the all-volunteer College Prep Program with Karla Gray-Roncal, and soon after begins a PhD in computer science while working full-time. The first published move is MRCAP (2012): the brain becomes a graph, automatically. The gate-widening instinct was there before the science had a name.

Plates I, XXI · JHU CS seminar bio

Act II · 2013–2016

Down the ladder, to the floor

The same year the graphs exist, he starts distrusting them: MIGRAINE shows brain graphs are fingerprints (42/42), and signal-subgraphs shows they carry signal — with a self-warning about how much of the discovered signal is real. Then the descent to the nanometre floor: the saturated reconstruction proves the map is possible and that no shortcut exists — proximity does not predict wiring, and hand-tracing costs "two people-years of 24/7" per speck. So he automates: VESICLE finds the synapses, the images-to-graphs dissertation chains it all with zero humans and publishes its own unflattering score (F1 = 0.16), and SANTIAGO attacks the one error mode — spine necks — that the honest metric exposed. (A fun aside from the same year: an unreviewed arXiv preprint with Eva Dyer, “From sample to knowledge,” sketches exactly this integrated-pipeline philosophy — a period document worth a smile, not a landmark.)

Plates II–VII · marginalia: the 2016 preprint

Act III · 2013–2025

The archive is a promise

In parallel, the infrastructure thread: the Open Connectome Project serves 75 terabytes of brains to anyone with a browser in 2013, on the astronomy model. When the on-premise cluster hits its wall, the team rebuilds cloud-native as bossDB — before the petabytes existed, in anticipation of them. Standards follow (the card catalog for a petabyte, then the missing metadata standard for the whole modality), and in 2025 the promise is kept: MICrONS — 200,000 cells, half a billion synapses, one mouse — lands in the world's browsers from the archive his team runs.

Plates IX, XV, XVI, XX · field notes: SIC, substrate, intern, metadata standards · watch: his 2020 BRAIN Initiative symposium, “Toward Petascale and Exascale Connectomics”

Act IV · 2017–2023

Calibrate every judge in sight

The pattern discovered in Act II becomes a career-wide reflex. NRI catches the field's metrics rewarding the wrong thing; CONFIRMS industrializes the audit; ndmg catches whole scientific literatures disagreeing about magnitudes while agreeing about signs; discriminability turns the fix into a statistic. Then the same move outside neuroscience: the human-baseline framework catches "superhuman AI" claims resting on unmeasured humans (and finds the one-shot cliff), and the COVID-19 model is built to say "I don't know" rather than bluff.

Plates X–XII, XIV, XVIII, XIX · field notes: aerial-search benchmark, transfer learning, m2g

Act V · 2017–2024

The gate widens into the pipeline

The fusion. MICrONS needs proofreaders at impossible scale; the outreach thread has spent a decade learning how to find talent the gates miss. CIRCUIT connects them: trailblazing undergraduates — first-generation, low-income, no lab pedigree — do the real correction work, inside the real pipeline, and become authors. The loop closes in the bylines: the pilot paper is first-authored by its own students; a 2017 student-team annotator becomes CONFIRMS's first author; two CIRCUIT alumni first-author the standards paper (the portal greets one by name); the program's project manager co-authors the fly-compass robot and lead-authors the framework papers. The workforce program and the science program are one system — the casebook's name for it is "The Human Correction Layer at Petabyte Scale."

Plates XIII, XVII, XXI–XXIII · field notes: BOLT, Net-Hack, curriculum papers

Act VI · 2024–2026

Up the last rung

Finally the method turns on the institutions themselves. "Show Your Work" states the whole career as one claim — grades, transcripts and benchmarks certify proxies; only demonstration certifies capability — and the Capability Matters casebook assembles ~200 cases of systems that failed because "someone couldn't do what the system required." LENS, a new Johns Hopkins M.Ed. concentration launching Fall 2026, makes it a discipline. The synapse counter of 2015 is, by 2026, engineering the systems that decide who gets to count synapses.

Plate XXIV · LENS at IEEE ICICLE

The plates

Every paper, retold in plain language

Cajal published his arguments as numbered plates: a drawing, then the claim. Same format here — one bespoke illustration per publication, drawn from what the paper actually did, with its key findings and a link to the source. Citation counts are from Google Scholar, August 2026. Colors keep their meaning: maps trust archive machines people.

Plate Imaps2012 · 60 citations

The brain becomes a spreadsheet

Gray (Roncal), Bogovic, Vogelstein, Landman, Prince & Vogelstein · “Magnetic resonance connectome automated pipeline” · IEEE Pulse

The opening move of the career: an automated pipeline that takes raw MRI scans in one end and produces a brain graph out the other — 70 anatomical regions, 2,415 possible connections, no manual steps. Modular on purpose, so any stage could be swapped by someone else.

Over 200 connectomes processed by publication — a large number, in 2012.
~8 hours per subject on a single CPU core — remember that number for Plate X.
Already validating itself: symmetry checks, repeatability analysis.

doi:10.1109/MPUL.2011.2181023

Plate IItrust2013 · 59 citations

Brain graphs are fingerprints

Gray Roncal, Koterba, Mhembere, … & Vogelstein · “MIGRAINE: MRI Graph Reliability Analysis and Inference for Connectomics” · IEEE GlobalSIP

One year after building the graph-maker, he builds the machine that asks whether its graphs can be trusted. MIGRAINE scales the pipeline up (to voxel-level graphs with ten million vertices) and adds a new rank-based reliability metric — the ancestor of “discriminability” (Plate XII).

For all 42 test–retest scans, the most similar graph in the whole collection was the same person, scanned on a different day — 42/42 identification.
Voxel graphs: ~10⁷ vertices cut to a 10⁵-vertex core keeping >99% of edges.
Over 1,500 subjects processed across five datasets.

arXiv:1312.4875

Plate IIImachines2013 · 89 citations

Reading sex from 70 brain regions

Vogelstein, Gray Roncal, Vogelstein & Priebe · “Graph classification using signal-subgraphs” · IEEE TPAMI

The payoff of making brains into graphs: classification becomes graph statistics. The “signal-subgraph” estimator finds the few edges that actually differ between classes — and classifies connectomes by sex at 0.16 error, versus 0.27 for lasso and 0.50 for chance, on graphs made by MRCAP (Plate I).

Best model: signal concentrated on 12 hub vertices and 360 edges.
The abstract warns about itself: with ~50 samples, only about half the discovered signal-edges are real — “taken with a grain of salt.”

doi:10.1109/TPAMI.2012.235

Plate IVmaps2015 · 1,287 citations

The speck where every wire got traced

Kasthuri, Hayworth, Berger, …, Gray Roncal, …, Lichtman · “Saturated Reconstruction of a Volume of Neocortex” · Cell

His most-cited work. Mouse cortex cut into 2,250 slices just 29 nm thick — collected on a conveyor belt of adhesive tape — and three tiny cylinders reconstructed saturated: every axon, dendrite, vesicle and synapse itemized. His role: the statistics and the open database serving it all. The result killed a comfortable shortcut.

~9 excitatory axons touch each dendritic spine; only one connects. An axon already connected to a dendrite is ~25× more likely to synapse on it again. Proximity does not predict wiring (Peters’ rule, refuted).
1,500 µm³ holds pieces of ~1,600 neurons — 5× the entire C. elegans nervous system, in “several billionths” of one brain.
Fully tracing the surrounding volume by hand: “two people-years of 24/7 tracing.” That number is the reason the next four plates exist.

doi:10.1016/j.cell.2015.06.054

Plate Vmaps2015 · 39 citations

Teaching a forest to spot synapses

Gray Roncal, Pekala, Kaynig-Fittkau, …, Hager · “VESICLE: Volumetric Evaluation of Synaptic Interfaces using Computer vision at Large Scale” · BMVC

The first synapse detector built for the hard, realistic case: anisotropic, non-post-stained tissue. Its trick is an anatomist’s trick — spot the cloud of vesicles first, then look for the cleft. A lightweight random forest using that biological context trains in 10 minutes and deploys ~200× faster than its own CNN twin.

≈11.6 million synapses detected across 20 teravoxels of mouse cortex — in 2015.
Philosophy stated outright: favor recall; humans can proofread false positives, but a missed connection is gone.

arXiv:1403.3724 · PDF

Plate VImaps2015 · 33 citations

The dissertation: images to graphs, no humans

Gray Roncal, Kleissas, Vogelstein, …, Hager · “An Automated Images-to-Graphs Framework for High Resolution Connectomics” · Frontiers in Neuroinformatics

The core of his PhD (“Enabling Scalable Neurocartography: Images to Graphs for Discovery”): the first pipeline that begins with imaged brain tissue and ends with a brain graph with zero human interaction. Just as important: it invents a metric that scores the graph, not the pixels — because the two disagree.

1,856 candidate pipelines searched (~8,000 cluster jobs); the best honest graph F1 was 0.16 — published as the community baseline to beat.
The best graph does not come from the best individual modules: tiny segmentation errors at spine necks destroy graph topology.

doi:10.3389/fninf.2015.00020 · PDF

Plate VIImaps2016

The graph lives in the spine necks

Gray Roncal, Lea, Baruah & Hager · “SANTIAGO: Spine Association for Neuron Topology Improvement and Graph Optimization” · arXiv

The follow-up attacks the single worst failure found in Plate VI: the wispy necks connecting dendritic spines to their shafts. They are a rounding error in voxels and almost all of the meaning in the graph. A classifier that re-attaches orphaned spines — using cues like branching angle — doubles graph quality without touching anything else.

Snap the spine necks and the true graph collapses from 31,980 edges to 2,718.
Re-linking spines alone: graph F1 0.31 → 0.64. Show a human the top-5 candidates and the right shaft is there 89% of the time.

arXiv:1608.02307 · PDF

Plate VIIImaps2017 · 92 citations

Mapping a millimetre without a knife

Dyer, Gray Roncal, Prasad, …, Kasthuri · “Quantifying mesoscale neuroanatomy using X-ray microtomography” · eNeuro

Every map so far began by slicing the brain thousands of times. Here a millimetre-scale chunk is imaged intact in about six minutes, spinning in a synchrotron X-ray beam — filling the mesoscale rung between MRI and EM. And because the sample is prepared EM-compatibly, the same tissue can flow onward into nanoscale imaging: zoom-out and zoom-in on one specimen.

~16.6 gigavoxels in ~6 minutes; 48,689 cells auto-detected, analysis on a single workstation.
The automated segmentation agreed with the expert annotator better than a second human did.

doi:10.1523/ENEURO.0195-17.2017 · PDF

Plate IXarchive2013 · 66 citations

The astronomy playbook, applied to brains

Burns, Gray Roncal, Kleissas, …, Vogelstein · “The Open Connectome Project Data Cluster” · SSDBM

Law three gets its infrastructure: do for the brain what the Sloan Digital Sky Survey did for the sky. A hand-built university cluster serving EM volumes to anyone with a browser, with brain space carved into cuboids and indexed by a Morton space-filling curve — a design decision that survives, unchanged, into Plate XV a decade later.

75+ TB across 50+ datasets in 2013, including 20 trillion voxels of mouse cortex served as arbitrary 3D cutouts.
Companion result: ~20 million synapses detected in a 12 TB volume in 2 days on 256 cores (CAJAL3D).
The honest math in the limitations: a human brain has ~10¹⁵ synapses; manual anything can never scale.

arXiv:1306.3543

Plate Xtrust2018 · 51 citations

Robust connectomes, troublesome variability

Kiar*, Bridgeford*, Gray Roncal*, …, Vogelstein (*equal) · “A High-Throughput Pipeline Identifies Robust Connectomes But Troublesome Variability” · bioRxiv

The judge-calibration move at population scale: one pipeline, zero manual tuning, run over ~6,000 scans from dozens of studies — producing over 100,000 connectomes and a paradox. The direction of the brain’s wiring asymmetries is universal; the amount varies so much between studies that unmeasured experimental variables must be leaking in.

99.9% of diffusion connectomes: hemispheres talk to themselves; 99.4% of functional: left–right twins in sync — a perfect modality mirror.
Even studies of demographically near-identical students differed significantly in two-thirds of comparisons.
The sober coinage: “repeatable does not mean correct.”

doi:10.1101/188706

Plate XItrust2018 · 13 citations

Judge the wiring, not the paint job

Reilly, Garretson, Gray Roncal, …, Roos · “Neural Reconstruction Integrity” · Frontiers in Neuroinformatics

The keystone of law two. Classic metrics score reconstructions by voxel overlap — and stay near a perfect 1.0 while a reconstruction shatters every neuron, because the paint is mostly right even when the circuit is destroyed. NRI instead asks: for every pair of synapses that shared a neuron, is there still a path between them? Measure what the map is for.

Intuitive by design: split a neuron in two → 0.67; delete 20% of synapses → 0.78; a worked toy example → 0.333.
Precision vs recall diagnoses the error type: merges vs splits.
Tested on a synthetic cortex of 872 neurons and a million synapses — crash-test dummies for connectomics.

doi:10.3389/fninf.2018.00074 · PDF

Plate XIItrust2021 · 64 citations

A statistic for choosing honest pipelines

Bridgeford, Wang, …, Gray-Roncal, …, Vogelstein · “Eliminating accidental deviations to minimize generalization error and maximize replicability” · PLoS Computational Biology

MIGRAINE’s rank trick (Plate II), matured into a general statistic: discriminability — the probability that two measurements of the same person sit closer to each other than to anyone else’s. Then the mega-experiment: 192 analysis pipelines raced on >1,600 participants, with discriminability picking the winners — in brain imaging and genomics alike.

~80% of standard fMRI pipelines were significantly worse than the best (“FNNNCP” — pipeline settings as a six-letter genotype).
The classic statistic, ICC, goes negative with 2 outliers in 10 subjects; discriminability shrugs.
Their own honesty rule: discriminability is “neither necessary, nor sufficient, for a measurement to be practically useful.”

doi:10.1371/journal.pcbi.1009279 · PDF

Marginalia — a supporting curiosity

A 2016 preprint sketches the method early

Gray Roncal* & Dyer* (equal contribution), Gürsoy, Kording & Kasthuri · “From sample to knowledge: Towards an integrated approach for neuroscience discovery” · arXiv, 2016 — an unreviewed preprint

Filed at the hinge year between the career’s descent and ascent, this barely-cited note argues that acquisition, processing and analysis should be optimized together, against the knowledge objective — not against each stage’s surrogate metric — with graphs and point processes named explicitly as representations of the knowledge, not the knowledge itself, and “neurocartographers” as the audience. The three laws, sketched early, in the authors’ own hand.

It never went through peer review, so this page treats it as supporting color, not evidence: the argument stands on the reviewed record in the plates — the preprint is just a delightful footnote showing the through-line was articulated, not retrofitted.

arXiv:1604.03199 · PDF (self-hosted)

Plate XIIImaps2021 · 54 citations

A grammar for asking brains questions

Matelsky, Reilly, Johnson, Stiso, Bassett, Wester & Gray-Roncal · “DotMotif: an open-source tool for connectome subgraph isomorphism search and graph queries” · Scientific Reports

Once connectomes exist, most scientists still can’t query them — subgraph search is a computer-science problem. DotMotif is a tiny language where the code looks like the circuit: write A -> B, get every matching motif, on your laptop or a graph database, identically. The gate to graph analysis, widened.

The worm has cliques the mouse cortex lacks: K6 (six neurons all mutually connected) appears thousands of times in C. elegans and exactly zero times in the MICrONS cortex graph.
Worm brains fan in; mouse cortex fans out — the most common 3-neuron pattern flips between species.
Random-graph stand-ins for brains always undercount motifs — only degree-preserving shuffles come close.

doi:10.1038/s41598-021-91025-5 · PDF

Plate XIVtrust2021 · 10 citations

Auditing a brain map with ten-second questions

Bishop, Matelsky, Wilt, …, Plaza, Wester & Gray-Roncal · “CONFIRMS: a toolkit for scalable, black box connectome assessment” · IEEE EMBC

NRI (Plate XI) needs ground truth; CONFIRMS manufactures it, cheaply. Any reconstruction — no source code needed — is sampled into atomic micro-questions a person can answer in seconds, and the answers are fused into quality estimates. Quality assurance as a swarm of yes/no/maybe decisions, engineered so newcomers can contribute on day one.

~60 users, >90,000 tasks over three years; 244,663 synapses tagged; most forced-choice answers arrive within ten seconds.
Fusion rule: five intermediate annotators with three agreeing ≈ one expert — laws two and three in a single equation.
First author Caitlyn Bishop first appears in this bibliography on a 2017 student-team abstract.

doi:10.1109/EMBC46164.2021.9630109

Plate XVarchive2022 · 43 citations

The archive, rebuilt for petabytes

Hider, Kleissas, Gion, …, Johnson, Gray-Roncal & Wester · “The Brain Observatory Storage Service and Database (BossDB)” · Frontiers in Neuroinformatics

The Open Connectome cluster (Plate IX) hit its wall, so the team rebuilt it cloud-native and serverless — same Morton-curve DNA, new physics: storage that scales to zero when idle and fans out to thousands of parallel functions under load. Built before the data existed, in anticipation of MICrONS — the archive as a promise.

Sustained ingest of 230 GB per minute; autoscaling tested to 5,000+ concurrent functions.
~2 petabytes of lossless EM imagery stored for MICrONS; 30+ public collections at bossdb.org.
Every dataset one Neuroglancer link away from any browser.

doi:10.3389/fninf.2022.828787 · PDF

Plate XVIarchive2022

Ask a petabyte a question

Sanchez, Moore, Johnson, Wester, Lichtman & Gray-Roncal · “Connectomics Annotation Metadata Standardization” · Frontiers in Neuroinformatics

A petabyte you can’t query is a warehouse, not a library. This paper writes the card catalog: a community metadata standard plus a portal where dropdown menus ask questions of the H01 human-cortex volume — 150 million synapses — and answers come back as clickable 3D neurons. The first authors, Morgan Sanchez and Dymon Moore, are CIRCUIT alumni; the demo portal literally greets its own builder: “Welcome Morgan!”

~27 GB of metadata makes ~1.4 petabytes of imagery queryable — 119 query combinations from seven question templates.
One schema spans two brains: H01 human cortex and the Kasthuri mouse volume (Plate IV).

doi:10.3389/fninf.2022.828458 · PDF

Plate XVIImachines2022 · 13 citations

The map runs backwards

Robinson, Norman-Tenazas, Cervantes, …, Zhang & Gray-Roncal · “Online learning for orientation estimation during translation in an insect ring attractor network” · Scientific Reports

Why map brains at all? Here is the answer running on wheels: the fly’s head-direction circuit — 141 neurons, wired per the connectome — driving a TurtleBot. The circuit’s heading “bump” teaches itself which visual landmarks to trust, with no labels, in a single pass. The mapping thread and the machines thread are one project.

Online learning cut heading error 28% (rotation) and 38% (translation) in simulation, 14% on the physical robot.
Estimated to run in ~18.6 microwatts on neuromorphic hardware — insect-grade power.
The optimal solution rediscovers what animals do: trust distant landmarks, because nearby ones lie when you move.

doi:10.1038/s41598-022-05798-4 · PDF

Plate XVIIItrust2022 · 45 citations

Measure the human before crowning the machine

Cowley, Natter, K. Gray-Roncal, …, Wester & Gray-Roncal · “A framework for rigorous evaluation of human performance in human and machine learning comparison studies” · Scientific Reports

“Superhuman AI” claims rest on human baselines that were never measured carefully. This framework imports psychometric rigor — matched trials, naive participants, IRB, attention checks, deliberately alien stimuli, even thumbnails blurred on purpose to equalize memory — and then shows why it matters, with a one-shot learning study of 134 people.

The one-shot cliff: a ResNet falls from 92.7% to 54.4% on a class it saw once; humans go 94.6% → 95.8% — no penalty at all.
The pitfalls list doubles as a field checklist: unmatched trials, careless annotators, no confidence intervals.

doi:10.1038/s41598-022-08078-3 · PDF

Plate XIXtrust2023 · 19 citations

A model designed to say “I don’t know”

Cowley, Robinette, Matelsky, …, Zeger, Garibaldi & Gray-Roncal · “Using machine learning on clinical data to identify unexpected patterns in groups of COVID-19 patients” · Scientific Reports

The same measurement discipline, turned on a pandemic: 1,175 Johns Hopkins patients, first 48 hours in six-hour epochs. The classifier’s thresholds are deliberately asymmetric — telling someone “you’ll be fine” is made the hardest thing for the model to say — and clustering surfaces the phenotypes where the model gets fooled.

The symptom paradox: clusters with worse outcomes reported fewer symptoms — flagged honestly as documentation artifact or immunology, untested.
One epoch scored 90% accuracy with precisely zero predictive skill (MCC 0.00) — accuracy theater, exposed.
Two specific patient clusters generate most false “you’ll be fine” calls — a map of when not to trust the algorithm.

doi:10.1038/s41598-022-26294-9 · PDF

Plate XXmaps2025 · 456 citations

The poppy seed that weighed 1.6 petabytes

The MICrONS Consortium (Gray-Roncal among; JHU/APL led archiving) · “Functional connectomics spanning multiple areas of mouse visual cortex” · Nature

The landmark the previous nineteen plates were building toward: ~1 mm³ of mouse visual cortex — more than 200,000 cells and half a billion synapses — co-registered with the recorded activity of ~75,000 of those same neurons while the mouse watched movies. A relay of institutions made it; his team’s archive (Plate XV) is where the world reads it.

~28,000 serial sections at 40 nm; ~2 PB of raw imagery; over a million proofreading edits released.
~4 km of axons in a volume the size of a poppy seed (figure from the project’s public materials).
The New York Times called it “an advance … once considered impossible.”

doi:10.1038/s41586-025-08790-w · bossdb.org · PDF

Plate XXIpeople2018 (program: 2009–)

The thread that predates the PhD

K. Gray-Roncal, Huynh, Kolarik, Roncal, Saunders & Gray-Roncal · “The college prep program at APL” · IEEE ISEC

Before the dissertation, before the archive — in 2009 — he co-founded an all-volunteer college-prep program with Karla Gray-Roncal for underserved high-schoolers, and ran it like an engineered system: competency-based curriculum, capstone portfolio, longitudinal follow-up, published metrics. The “people” thread isn’t a late-career pivot; it ran in parallel the whole time.

168 students over nine summers; 98% of surveyed alumni on track to four-year degrees (self-reported survey — the paper says so plainly).
Zero paid staff. The companion paper (ISEC 2019) documents the low-cost technology stack that lets volunteers run it like an institution.

doi:10.1109/ISECon.2018.8340511

Plate XXIIpeople2018 (program: 2017–)

The fusion point: outreach wired into the pipeline

Encarnacíon, Bishop, Downs, …, Rivlin, Wester & Gray-Roncal · “CIRCUIT summer program” · IEEE ISEC

The moment the two parallel threads fuse. MICrONS needed proofreaders “at a scale not possible with conventional approaches” — so he built a program where underserved undergraduates do that real, sponsor-grade work while being trained as researchers. The paper describing the pilot is first-authored by the students themselves. The casebook later names it “The Human Correction Layer at Petabyte Scale.”

191 fellows, 58 cohort research projects, 46 challenge teams across four cycles (2023 count); ~220 undergrads, 15 grads, 23 high-schoolers through 2022; >90% first-generation, low-income, and/or underrepresented.
Students matched to projects with the Gale–Shapley algorithm — the medical-residency match, repurposed for interns.
His site: cohorts “contributed to reconstructions published in Nature.”

doi:10.1109/ISECon.2018.8340503

Plate XXIIIpeople2023–24

The program becomes a framework

Cervantes, Floryanzia, Sharp, Gray-Roncal & Johnson (ASEE 2023) · Floryanzia, Jayabharathi, Carr, K. Gray-Roncal & Gray-Roncal (IEEE ISEC 2024)

Law three applied to the program itself: CIRCUIT is abstracted into an eight-pillar model anyone can run, written up with the program’s own people as lead authors — its project manager (Cervantes, who also co-authored the fly-compass paper of Plate XVII) and a graduate fellow (Floryanzia, first author of the 2024 framework and of three first-author bioengineering papers of her own).

The thesis, verbatim: “capability, or talent, is distributed equally among individuals, whereas opportunity is not.”
Alumni offers at Google, Apple, Microsoft, Merck; graduate offers at MIT, Harvard, Berkeley, Chicago; 85% intend a Masters or PhD.
Selection is essays and interviews, not GPA cutoffs — judge the work, not the pedigree.

peer.asee.org/43271 · doi:10.1109/ISEC61299.2024.10664684

Plate XXIVpeople2026

The synthesis: capability is a system parameter

Gray-Roncal & Diamond · “Show Your Work” (working draft) · and the Capability Matters casebook, JHU Learning Design & Technology

The newest work states law two as a philosophy and law three as a curriculum. “Show Your Work” argues that grades, transcripts and benchmarks certify proxies — only a demonstration certifies capability: “The patient, the bridge, the cockpit, and the reaction vessel do not grade on a curve.” The ~200-case casebook builds the discipline around it, and LENS — a new Johns Hopkins M.Ed. concentration launching Fall 2026 — teaches it.

The equity turn, verbatim: “a demonstrated output is harder to inherit than a proxy.”
The casebook’s first maxim: “Systems don’t fail because the technology breaks. They fail because someone couldn’t do what the system required.”
CIRCUIT appears inside as a case — with an explicit conflict-of-interest disclosure. Calibrating the judge of his own program.

Show Your Work (PDF) · capabilitymatters.org

Field notes

The rest of the record, accounted for

The plates cover the spine; the method shows up everywhere else too. Briefly, with sources:

machines Glocal invariants (2013) — graph statistics for 16-million-vertex brains on one core and 6.5 GB of RAM; coined “glocal.”
maps Gradient-domain fusion (2015) (PDF) — color-correcting five-teravoxel EM stacks in a 600 MB memory footprint; the unglamorous plumbing under Plates V–VI.
archive Science in the Cloud (2017) — a full connectome pipeline for ~1.4 cents per scan; a live, clickable paper hosted for ~$100/year.
archive substrate (2017) — composable browser 3D visualization; intern (2021) — one Python interface over every volumetric store; EM/XRM metadata standards (2024) (PDF) — the missing BIDS/NWB equivalent for nanoscale data.
maps Thalamocortical dataset (2020) (PDF) — a cortex-to-hypothalamus “core sample” at micron scale, published for others; “Agents” (MICCAI 2020) — virtual rovers borrowed from autonomous navigation, path-finding through brains; multi-scale X-ray segmentation (2021); C. elegans development (2021) — one motif grows 353-fold from birth to adulthood while triangles grow 9×.
machines MYOW (2021) (PDF) — self-supervised learning that mines a dataset for its own teaching examples; beat the supervised baseline on all four monkey-reach datasets. Transfer learning for EM workflows (2019) — measuring whether models survive new brains instead of assuming it.
trust Aerial-search benchmark (2021) (PDF) — 666 people searched satellite images: targets under ~100×100 px slip past, and on nearly-empty images 57% of “finds” were phantoms. m2g (2021) — the democratized pipeline: any laptop with 16 GB of RAM, ~6,000 scans, every one yielding a plausible connectome.
people Outreach tech stack (2019), trailblazer training (2020), the immersive curriculum (2023) — Learner Blueprints, a git-repo Knowledge Base, stoplight assessments; Net-Hack (2023) — a hackathon judged on proposals, not code; BOLT (2023) — swarm robotics as laser tag.
the rest ~30 conference abstracts and invited talks, catalogued in the source bibliography — and no longer all dead ends: several now link to the actual artifacts, including the SfN posters behind the archive thread (Open Connectome & NeuroData, 2015; scalable synapse detection, 2015; whole-brain CLARITY statistics, 2015), the NeurIPS BigNeuro 2015 and 2017 workshops he co-organized, his AAAI 2021 workshop on trailblazing STEM students, and a recording of his 2020 BRAIN Initiative symposium, “Toward Petascale and Exascale Connectomics.”

The real thing

Three figures from the papers themselves

Illustration is interpretation; these are the originals, reproduced from the three CC BY–licensed papers below with full attribution. Most of the other papers on this page are under restrictive licences and are redrawn, never copied.

MICrONS pipeline figure: a mouse undergoes in-vivo imaging, its cortex is prepared and sectioned, imaged by multiple electron microscopes, then assembled, segmented, synapses detected, and proofread. — **The relay, end to end.** From a living mouse watching movies to a proofread wiring diagram. Figure 2 of The MICrONS Consortium, “Functional connectomics spanning multiple areas of mouse visual cortex,” *Nature* 640:435–447 (2025), doi:10.1038/s41586-025-08790-w. Licence: CC BY 4.0. Modified: resized and recompressed only.

BossDB spatial indexing figure: image slices from a microscope are chunked into small 3D cuboids, assigned Morton-order indices, and stored in a cuboid store with an index table. — **How a petabyte becomes addressable.** Brain volumes chunked into cuboids and flattened onto a Morton space-filling curve — the design decision that dates to 2013 (Plate IX). Figure 4 of Hider R Jr, Kleissas D, Gion T, et al., “The Brain Observatory Storage Service and Database (BossDB),” *Frontiers in Neuroinformatics* 16:828787 (2022), doi:10.3389/fninf.2022.828787. Licence: CC BY. Modified: resized and recompressed only.

Screenshot of the Connectomics Query Engine portal greeting a user named Morgan, with dropdown-menu questions about synapses and layers over a backdrop of colorful reconstructed neurons, and a table of answered queries. — **“Welcome Morgan!”** The H01 query portal greeting its own builder — Morgan Sanchez, CIRCUIT alumna and first author. Figure 1 of Sanchez M, Moore D, Johnson EC, Wester B, Lichtman JW, Gray-Roncal W, “Connectomics Annotation Metadata Standardization for Increased Accessibility and Queryability,” *Frontiers in Neuroinformatics* 16:828458 (2022), doi:10.3389/fninf.2022.828458. Licence: CC BY. Modified: resized and recompressed only.

Handouts — take the story with you

The same theory, at three reading levels

In the spirit of law three, this argument ships with its controls handed over: three printable one-pagers, each with a group activity. Print this page and these survive; the animations don't.

For kids & classrooms (ages ~8–12)

The Brain Mappers

Your brain is made of billions of tiny cells called neurons, joined by even tinier connections called synapses. Scientists like Dr. Will Gray-Roncal take super-zoomed-in pictures of brains and trace every wire, like following one noodle through a bowl of spaghetti — without pulling any noodles out. In 2025, his team helped share a map of a piece of mouse brain the size of a poppy seed that held about 200,000 cells and more than 500 million connections (the real paper). The map was so big it wouldn't fit on 1,000 laptops, so his team built a giant online library where any scientist — or any curious kid — can look at it for free. And here's the best part: a lot of the map-checking was done by students, some not much older than you, who were invited onto the team before anyone said they were "ready." They learned by doing the real thing. That's the whole method: draw the map, check the map, share the map — and invite somebody new.

Group activity — be a connectome (20 min) Everyone stands in a circle holding a ball of yarn. The first "neuron" holds the end and tosses the ball to someone across the circle; keep tossing until everyone is connected at least twice. Now you're a brain map! Then pick one "proofreader": they must check three connections by following the yarn hand-over-hand and asking each holder "who sent you this?" If a connection is wrong or remembered differently — that's exactly what brain-map proofreading finds. Talk about it: what made checking hard? What would help? (Real answer from the lab: better tools, clear rules, and more proofreaders.)

For peers — researchers, educators, program builders

A checklist stolen from a bibliography

Thirty-odd papers by one JHU/APL engineer ( Scholar) keep making the same three moves, in order. 1 · Extract the knowledge: decide what you actually need to learn from your noisy, incomplete data, then hold it in a representation you can query — his is usually a graph (DotMotif); yours might be a cluster map or a spreadsheet. The representation serves the question, never the reverse. 2 · Calibrate the judge: list every metric, review, or gate your system uses to score people or results, and ask what evidence shows the metric measures what it claims (NRI and the human-baseline framework are the worked examples). 3 · Widen the gate: your system is finished when someone new can run it — which means docs, open interfaces, and deliberately recruiting people the current gate filters out (CIRCUIT wired undergraduates into a petabyte pipeline as working scientists).

Group activity — calibrate your judge (45 min) In pairs: each person names one gatekeeping instrument their org uses (an admissions score, a code-review rubric, a model benchmark, an annual review). Swap. Your partner designs the cheapest possible experiment that could show the instrument failing — scoring the wrong thing, or scoring pedigree instead of performance. Reconvene and vote on the one experiment the group should actually run. The uncomfortable rule, from "Show Your Work": if you can't imagine evidence that would fail your instrument, it isn't a measurement — it's a habit.

For everyone else — the two-minute version

Why map a brain the size of a poppy seed?

Because nobody has ever seen how a brain is actually wired, at the level of individual connections, at scale — and you cannot fix, mimic, or marvel at what you cannot see. One team's answer took a relay of institutions: image a speck of mouse brain into ~28,000 slices, each far thinner than paper; let AI trace the wiring; store the result — about 1.6 petabytes, a small national library's worth of data — in an open cloud archive (bossdb.org) that anyone can browse free. The engineer behind that archive spent twenty years on a second, quieter project: proving that the students usually filtered out of science — first-generation, low-income, no lab connections — do publishable work when you hand them real controls. Over 400 students later, some of those "interns" are first authors on the field's standards. The map and the mapmakers were built the same way: carefully, openly, and on purpose.

Kitchen-table activity — judge the work (10 min) Two volunteers each fold the same paper airplane design. One narrates their credentials first ("I've watched every paper-airplane video"); the other says nothing. Fly both. Ask the room which information predicted the flight — the résumé or the throw. Then the question behind this whole page: how many throws does your school, team, or company never get to see, because the résumé went first?

Coda — his version

He already said it, in one line

A theory like this should worry you until you learn the subject got there first. The tagline at the top of his own homepage reads:

“Learning is the through-line — patterns in brains, in machines, and in minds. Giving everyone a chance to participate, and to thrive.”

Same claim, better compression. The first sentence is laws one and two — learning as the object being mapped and measured, in every substrate. The second sentence is law three. This page is just the long-form proof, with the reference list attached.

And the machine keeps running: the newest work — Capability Matters, the “Show Your Work” paper, the LENS concentration launching at Johns Hopkins in Fall 2026 — is the three laws applied to the judging systems themselves. “The patient, the bridge, the cockpit, and the reaction vessel do not grade on a curve.”

Sources & honesty

What this page is, and is not

This page is an interpretation. The “three laws,” the thread names and colors, the scale-ladder ordering, and the act structure are this page's reading of the record, not the subject's claims — except where his own words are quoted and linked. Every publication fact, number, and quotation traces to a linked source: his Google Scholar profile (citation counts, retrieved August 2026), his personal site and its public CV data, and the individual papers linked throughout. In the related-work graph, solid edges come from actual reference lists; dashed edges are thematic links drawn by this page and labelled as such.

Where we couldn't verify, we said so. Claims that rest only on self-reported sources are attributed to those sources. Reproduced figures appear only from CC BY–licensed papers, with author, source, year, licence, and modifications stated under each; papers under restrictive licences (Cell, Nature Methods, ACM, IEEE, ASEE) are described and redrawn, never reproduced.

Compiled August 2026 · built from the public record · corrections welcome: the source of truth for the career facts is the subject's own site repository.