A brain assembled region by region from published models, and graded against its papers.
Website · Paper 1 · Progress, October 2026 · DOI 10.5281/zenodo.23119475
BrainCore is a spiking neural network built the way a brain is described in the literature: leaky integrate-and-fire neurons in Potjans–Diesmann cortical columns, a model retina, a cochlea, a cochlear nucleus, a hippocampal region. Each part comes from a published model, every number in a fixture carries its source, and each region has to pass an acceptance written before the run.
There are no conventional neural networks inside. Learning, where it is switched on, is local synaptic plasticity.
This repository is the public page of the project: what it is, what has been measured, and what has not worked yet. The source code is private.
How it is built, Paper 1 with its controls and its two failed criteria, letter–tone learning, the eye on a live camera, the ear on 81 recordings, and where it does not work yet: youtu.be/DCkoTII_WII
flowchart LR
subgraph S[Scheduler]
direction TB
B[barrier every exchange interval]
D[running SHA-256 over every spike batch]
end
R1[Region: retina] -- spike batches --> S
R2[Region: cochlea + cochlear nucleus] -- spike batches --> S
S -- routed batches --> C1[Region: area 17 columns]
S -- routed batches --> C2[Region: area 41 columns]
S -- routed batches --> H[Region: hippocampus CA1]
- A region owns its neurons and every synapse that ends on them. It talks to the rest of the brain only through batches of spikes.
- All regions run one exchange interval in parallel, wait at a barrier, and the scheduler routes the batches. Every delay between regions is at least one interval, so the parallel phase is exact.
- A region can run in the same process or in a child process behind gRPC. The acceptance checks that this does not change the digest.
- Neurons are integrated with the exact propagator of Rotter and Diesmann, in fixed chunks with their own random streams, and synaptic currents are summed as integer femtoamperes. One worker or twelve give the same bits.
Retinotopic vision and tonotopic hearing in BrainCore, a spiking neural network assembled from published cortical models, technical report, 11 September 2026. DOI 10.5281/zenodo.23119475. Full text: getbraincore.com/research.html.
flowchart LR
L[Letter on a 48×48 frame] --> RET[Primate parvo retina<br/>24×24 ganglion mosaic]
RET -- 3×3 tiles, 64 cells each --> V1[Area 17<br/>9 layered columns<br/>179,298 neurons]
T[Tone, sound pressure] --> COC[Cochlea, 32 gammatone channels<br/>Meddis synapses, 3 fibre classes]
COC -- 4 bands × 8 channels --> CN[Cochlear nucleus<br/>96 bushy cells per band]
CN --> A1[Area 41<br/>4 layered columns<br/>31,896 neurons]
V1 --> DEC1[External decoder<br/>reads spike counts]
A1 --> DEC2[External decoder<br/>reads spike counts]
The network receives only luminance and sound pressure. No plasticity is switched on.
| Result | Value |
|---|---|
| Letters told apart, acceptance run | 15 of 16 showings |
| Evoked response follows the ink in each tile | Spearman ρ = 0.866 |
| 1 kHz vs 2 kHz decoder | 20 of 20 |
| Letters over 20 new seeds | median 93.8 %, retinotopy on 20 of 20 seeds |
| Tonotopy and decoder over 20 new seeds | 20 of 20 seeds |
| Hearing consistency over 20 new seeds | fails: 13 of 20 against the required 18 |
| Geniculate relay from published parameters | fails its transfer ratio: 0.124 against the measured 0.46 ± 0.16, left out of the visual path |
| Reproducibility | two runs give one digest; all 16 digests matched on a clean extraction of the committed tree |
Organisation, not activity. Shuffling the retinotopic tiling changes the total spike count by less than 0.01 % but drops the correlation from 0.866 to 0.184 and decoding from 94 % to 56 %. Shuffling the tonotopic afferents changes the count by 0.4 % and drops the decoder from 100 % to 60 %, where chance is 50 %. Every control broke every criterion it was required to break.
How the rules were kept. Thresholds were in the fixtures before the runs and were not changed afterwards. The rule over 20 new seeds was committed with the measured source tree, and only then run on seeds no one had run before. Failed criteria are reported as failed.
Progress report, 9 October 2026: getbraincore.com/progress-2026-10.html. This is a status report, not a result. The stage is not accepted.
flowchart LR
V[Letter] --> V1[Area 17]
A[Tone] --> EAR[Human ear model<br/>Verhulst et al. 2018] --> CN[Cochlear nucleus] --> A1[Area 41]
V1 --> CA1[Hippocampus CA1<br/>plasticity on]
A1 --> CA1
CA1 -- recall --> OUT[Tone from letter<br/>letter from tone]
| What | Status |
|---|---|
| Letter–tone association, fixed course | 9 of 9 in both directions, 169,925 weight updates |
| Same course with plasticity frozen | 0 of 9, no false recalls |
| Training interrupted and restored from a checkpoint | same output as the uninterrupted run |
| Unknown words from new voices | not rejected: the separation margin is negative with learning on and off |
Six candidate mechanisms were rejected on the way; each is listed with its score in the report.
- The model cannot yet say "I do not know this word": unfamiliar voices are not separated from familiar ones.
- Paper 2 is not published. Nothing about learning is claimed beyond the progress report.
- The fields are small (nine and four columns) and scaled to N = K = 0.1. The full-scale sheet synchronises and is not used.
- The auditory path collapses the superior olive, inferior colliculus and medial geniculate body into one relay, and there is one ear.
Everything above runs on one MacBook Pro, Apple M2 Max, 12 cores, 64 GB. Memory is the limit, and the papers say so:
- Area 17 uses 9 of the 4,767 columns its cited surface would hold. Both field sizes are conventions forced by memory, not anatomical claims.
- A full Potjans–Diesmann column is 77,169 neurons and about 300 million synapses, 6 GB at 20 bytes per synapse.
- One pass over the 54-recording speech corpus is 80.9 million spike events. The cochlear nucleus cell alone needs two minutes to initialise.
Potjans and Diesmann (2014), Schmidt et al. (2018), Rotter and Diesmann (1999), Wohrer and Kornprobst (2009), Meddis (1986, 1988), Sumner et al. (2002), Glasberg and Moore (1990), Kuenzel et al. (2011), Sincich et al. (2007), Merzenich et al. (1975). The full list with DOIs is in the reference section of Paper 1.
Later stages also use the human ear model of Verhulst et al. (2018, Ghent University), a hippocampal CA1 model from ModelDB, and other published models. They keep their own licenses and authorship.
Cherepanov V. (2026). Retinotopic vision and tonotopic hearing in BrainCore, a spiking neural network
assembled from published cortical models. Technical report. doi:10.5281/zenodo.23119475
Vitalii Cherepanov · ORCID 0009-0009-8289-8435 · LinkedIn · getbraincore.com