Retinotopic vision and tonotopic hearing in BrainCore

In the single-seed acceptance runs, four letters are told apart in 15 of 16 test showings, the evoked response under each column follows the ink in its tile (Spearman ρ = 0.866), and a 1 kHz versus 2 kHz decoder is correct in 20 of 20 leave-one-out trials, with tonotopy ordered across the four columns. Shuffling the afferent order changes total spike production by at most 0.4 % while removing the retinotopic and tonotopic organisation; blank and silent inputs remove the stimulus information. Runs are bit-for-bit reproducible across processes.

Under a rule committed before 20 new seeds were run, letter discrimination passes (median accuracy 94 %, retinotopy on 20 of 20 seeds), tonotopy and the decoder hold on 20 of 20 seeds, and a presentation-level consistency criterion for hearing fails (13 of 20 seeds against the required 18). A geniculate relay built from published parameters fails its transfer-ratio criterion and is left out of the visual path. The results establish separately validated visual and auditory representations. They show neither learning nor reading.

The question

Data-driven spiking models of cortex are usually judged by their spontaneous activity: a microcircuit is accepted when its populations fire at the measured rates, irregularly and asynchronously. For a system meant to be assembled region by region the narrower question is different. When a physical stimulus passes through a peripheral model into cortical columns, does the spatial or spectral organisation of the stimulus survive, and does a readout depend on that organisation rather than on the mere amount of activity?

It is easy to answer badly. A decoder can succeed on activity level alone, a correlation can come from population sizes, a threshold chosen after the run can be set to whatever the run produced, and a single seed can be lucky. The protocol below is built against each of these.

The system

Regions, scheduler and digest

A region owns its neurons and every synapse that ends on them, advances by whole exchange intervals and talks to the rest of the brain only through batches of spikes. A scheduler runs all regions for one interval in parallel, waits at a barrier and routes each batch. Every inter-region delay is at least one interval, so the parallel phase is exact. The scheduler keeps a running SHA-256 over every batch; a region may run in the same process or in a child process behind gRPC, and each acceptance checks that this does not change the digest.

Neuron kernel

Populations are leaky integrate-and-fire neurons with exponential synaptic currents, integrated with the exact propagator of Rotter and Diesmann. Neurons are integrated in fixed chunks, each with its own random stream, and synaptic currents are summed as integer femtoamperes, so a step gives the same bits with one worker or twelve. Every physiological constant comes from the connectome file.

Columns and fields

A field is a sheet of layered columns, one per square millimetre, each with the eight populations of the Potjans–Diesmann microcircuit. Area 41 uses the original microcircuit, 77,169 neurons per column at full scale; area 17 uses the V1 column of Schmidt and colleagues’ multi-area model. Lateral connections fall off as exp(−r/β) with β of 0.232, 0.161, 0.125 and 0.120 mm for E→E, I→E, E→I and I→I. Fields are scaled to N = K = 0.1 with published compensation. Area 17 uses nine of the 4,767 columns its cited surface would hold; area 41 is a strip of four. Both extents are conventions forced by memory, not anatomical claims.

The two sensory paths

Vision. Four 7 × 9 bitmaps are centred on a 48 × 48 frame and shown to a primate parvo retina at 100 pixels per degree, which makes a letter 0.36°, inside the range of maximal reading rates. Each 24 × 24 ganglion mosaic is cut into 3 × 3 tiles of 64 cells, and tile (r, c) replaces the thalamic input of column (r, c). The network has 179,298 neurons.

Hearing. A cochlea of 32 gammatone channels between 125 Hz and 8 kHz drives Meddis synapses and three fibre classes. The high-spontaneous class forms four bands of eight channels; each band drives 96 bushy cells of the cochlear nucleus with one endbulb per fibre, and the bushy cells of band b replace the thalamic input of column b of area 41. The network has 31,896 neurons and runs at a 0.05 ms step.

Protocol

• Thresholds before runs. Each path has a connectome fixture in which every number carries a source and a label (measured, published, derived or convention), and an acceptance fixture with the criteria, thresholds and controls. The thresholds were in the fixtures before the runs and were not changed afterwards.
• Controls that must break. Shuffled retinotopy must break retinotopy; a blank frame must break contrast, evoked response, retinotopy and decoding; shuffled tonotopy must break tonotopy; silence must break discrimination and tonotopy. A control that leaves a required criterion intact fails the acceptance.
• A preregistered seed rule. The rule over 20 new seeds was committed together with the measured source tree and only then run on seeds 20261001–20261020, which no one had run before. A new seed redraws every synapse, delay and noise stream.
• Statistics. Showings within one run share a connectome and a noise stream, so binomial tails say only that one run did not succeed by chance. Robustness is judged across seeds.

Results

Letters: 15 of 16, and the shape survives

Over ink, OFF cells rise from 38.5 to 100.2 spikes/s and ON cells fall from 54.6 to 11.9 spikes/s. Summed over nine columns, layer 4 excitatory cells fire 3,017 spikes per showing above baseline. The evoked response of each column follows the ink in its tile, with negative excursions in tiles without ink. The decoder reads 15 of 16 test showings correctly; the one error is an O read as A.

Tones: each lands in its own column

Each tone drives its band’s column by 349 to 565 spikes above baseline while the other columns change by −39 to 84. The loudest column is the band’s own in 90 % of the 300 Hz showings and in all showings of the other tones. In silence all 32 populations stay within 25 % of the rates of the column alone, and inhibition exceeds excitation in all 16 layers.

Organisation, not activity

Shuffling the retinotopic tiling changes the total spike count by less than 0.01 % (8,976,028 against 8,975,700) but drops the retinotopic correlation from 0.866 to 0.184 and decoding from 94 % to 56 %. Shuffling the tonotopic afferents changes the count by 0.4 % but moves every tone off its column and drops the decoder from 100 % to 60 %, where chance is 50 %. Blank and silent controls remove the stimulus information: letters fall to 19 % (chance 25 %). Every control broke every criterion it was required to break.

Robustness across seeds

For letters every confirmatory seed passes every per-run criterion: median accuracy 93.8 % (interquartile range 92.2–100 %, minimum 81.2 %) and retinotopy between 0.812 and 0.923. The rule passes. For hearing, tonotopy and the decoder hold on all 20 seeds, but the requirement that a single showing reproduce the loudest column in nine of ten cases holds on only 13 of 20 seeds against the required 18. On each failing seed one or two tones agreed in 8 of 10 showings instead of 9. The rule fails, and is reported as failed.

A geniculate relay that does not transmit enough

A lateral geniculate relay was built between a cat X-cell retina and one column: 49 ON and 49 OFF relay cells, 54 reticular cells and corticothalamic feedback. It passes six of seven criteria, among them a spontaneous rate of 4.49 Hz and retinotopic maps with a mean correlation of 0.784. Its transfer ratio, however, is 0.124 relay spikes per ganglion spike under a drifting grating, against the measured 0.46 ± 0.16 of Sincich and colleagues. Moving one sourced number inside its stated uncertainty raises it at most to 0.243; removing reticular inhibition overshoots to 2.31. This composition of published parameters is incompatible with the measured ratio, so the relay is reported as a negative result and kept out of the visual path.

Reproducibility

Within each acceptance, two runs give one digest, and moving the eye, or the cochlea and cochlear nucleus, into a child process over gRPC gives the same digest. The acceptances were rerun on a clean extraction of the committed tree: all 16 digests of the original runs matched. All runs used one Apple M2 Max with 12 cores and 64 GB.

go test ./cmd/regionlab -count=1 -run 'TestSeedSweepAcceptanceIsCompleteAndSourced$|TestReadingFixturesAreCompleteAndSourced$|TestHearingFixturesAreDerivedFromTheParts$'
BRAINCORE_READING_ACCEPTANCE=1 go test ./cmd/regionlab -run 'TestReadingAcceptance$' -count=1 -v -timeout 3600s
BRAINCORE_HEARING_ACCEPTANCE=1 go test ./cmd/regionlab -run 'TestHearingAcceptance$' -count=1 -v -timeout 3600s
BRAINCORE_LGN_ACCEPTANCE=1 go test ./cmd/regionlab -run 'TestLGNAcceptance$' -count=1 -v -timeout 3600s
BRAINCORE_HEARING_PREREG=1 go test ./cmd/regionlab -run 'TestHearingPreregisteredSweep$' -count=1 -v -timeout 7200s
BRAINCORE_GLYPH_PREREG=1 go test ./cmd/regionlab -run 'TestGlyphPreregisteredSweep$' -count=1 -v -timeout 14400s

Limitations

• The fields are small, nine and four columns, and scaled to N = K = 0.1; the full-scale sheet synchronises in our hands and is not used.
• Retinotopy is a 3 × 3 tiling without cortical magnification. The auditory pathway collapses the superior olive, inferior colliculus and medial geniculate body into one relay, and there is one ear.
• No plasticity is enabled and the decoders are external: nothing here is learned.
• The visual path bypasses the thalamus, so the cortex under a grey frame is driven about ten times above its isolated rate.
• The seed sweeps cover the model schedules only; controls, silence and transport runs use the fixture seed. Several full texts behind thresholds could not be read and only their abstracts are cited in the source audits. All runs used one machine.

What comes next

The two paths are the baseline for Paper 2: learning an arbitrary association between a letter and a tone through a hippocampal loop, with plasticity on, and with the unlearned pairing, a shuffled pairing and plasticity switched off as controls. Because each modality has been validated on its own here, a failure of that experiment cannot be blamed on either sensory path alone.

Data and code

Measured source tree: commit b0e46be4e71b9f2af02b1492274c70ced4d0eab6. Report, seed sweeps, clean-copy check and a SHA-256 manifest of every fixture and artifact: commit fc9a23671dc447cfcc3240f07f8a41494ff2111a, tag sensory-report-2026-09-11. Archive of the tag: SHA-256 578068fc8b555b392a63a0c8c7c00a0e5c5f4fd22c9aaa852b8fa7aeccf0033e. The repository is private; reviewers and partners get access on request at [email protected].

Key references

1. Potjans TC, Diesmann M (2014). The cell-type specific cortical microcircuit. Cerebral Cortex 24:785–806. doi:10.1093/cercor/bhs358
2. Schmidt M, Bakker R, Shen K, Bezgin G, Diesmann M, van Albada SJ (2018). A multi-scale layer-resolved spiking network model of resting-state dynamics in macaque visual cortical areas. PLoS Comput Biol 14:e1006359. doi:10.1371/journal.pcbi.1006359
3. van Albada SJ, Helias M, Diesmann M (2015). Scalability of asynchronous networks is limited by one-to-one mapping between effective connectivity and correlations. PLoS Comput Biol 11:e1004490. doi:10.1371/journal.pcbi.1004490
4. Senk J, Hagen E, van Albada SJ, Diesmann M (2024). Reconciliation of weak pairwise spike-train correlations and highly coherent local field potentials across space. Cerebral Cortex 34:bhae405. doi:10.1093/cercor/bhae405
5. Reimann MW, Horlemann A-L, Ramaswamy S, Muller EB, Markram H (2017). Morphological diversity strongly constrains synaptic connectivity and plasticity. Cerebral Cortex 27:4570–4585. doi:10.1093/cercor/bhx150
6. Rotter S, Diesmann M (1999). Exact digital simulation of time-invariant linear systems with applications to neuronal modeling. Biol Cybern 81:381–402. doi:10.1007/s004220050570
7. Wohrer A, Kornprobst P (2009). Virtual Retina: a biological retina model and simulator, with contrast gain control. J Comput Neurosci 26:219–249. doi:10.1007/s10827-008-0108-4
8. Legge GE, Pelli DG, Rubin GS, Schleske MM (1985). Psychophysics of reading — I. Normal vision. Vision Res 25:239–252. doi:10.1016/0042-6989(85)90117-8
9. Glasberg BR, Moore BC (1990). Derivation of auditory filter shapes from notched-noise data. Hear Res 47:103–138. doi:10.1016/0378-5955(90)90170-t
10. Meddis R (1986; 1988). Simulation of mechanical to neural transduction in the auditory receptor; Simulation of auditory-neural transduction: further studies. J Acoust Soc Am 79:702–711; 83:1056–1063. doi:10.1121/1.393460
11. Meddis R, Hewitt MJ, Shackleton TM (1990). Implementation details of a computation model of the inner hair-cell auditory-nerve synapse. J Acoust Soc Am 87:1813–1816. doi:10.1121/1.399379
12. Sumner CJ, Lopez-Poveda EA, O’Mard LP, Meddis R (2002). A revised model of the inner-hair cell and auditory-nerve complex. J Acoust Soc Am 111:2178–2188. doi:10.1121/1.1453451
13. Liberman MC (1978). Auditory-nerve response from cats raised in a low-noise chamber. J Acoust Soc Am 63:442–455. doi:10.1121/1.381736
14. Kuenzel T, Borst JG, van der Heijden M (2011). Factors controlling the input-output relationship of spherical bushy cells in the gerbil cochlear nucleus. J Neurosci 31:4260–4273. doi:10.1523/JNEUROSCI.5433-10.2011
15. Sincich LC, Adams DL, Economides JR, Horton JC (2007). Transmission of spike trains at the retinogeniculate synapse. J Neurosci 27:2683–2692. doi:10.1523/JNEUROSCI.5077-06.2007
16. Merzenich MM, Knight PL, Roth GL (1975). Representation of cochlea within primary auditory cortex in the cat. J Neurophysiol 38:231–249. doi:10.1152/jn.1975.38.2.231
17. Formisano E, Kim DS, Di Salle F, van de Moortele PF, Ugurbil K, Goebel R (2003). Mirror-symmetric tonotopic maps in human primary auditory cortex. Neuron 40:859–869. doi:10.1016/s0896-6273(03)00669-x
18. Andrews TJ, Halpern SD, Purves D (1997). Correlated size variations in human visual cortex, lateral geniculate nucleus, and optic tract. J Neurosci 17:2859–2868. doi:10.1523/JNEUROSCI.17-08-02859.1997
19. Pauli R, Weidel P, Kunkel S, Morrison A (2018). Reproducing polychronization: a guide to maximizing the reproducibility of spiking network models. Front Neuroinform 12:46. doi:10.3389/fninf.2018.00046
20. Nosek BA, Ebersole CR, DeHaven AC, Mellor DT (2018). The preregistration revolution. PNAS 115:2600–2606. doi:10.1073/pnas.1708274114

The full preprint cites 53 sources, each checked against PubMed or Crossref on 11 September 2026.