An Organizing Principle for Cerebral Function: The Unit Module and the Distributed System

TL;DR. In a 1978 essay, Vernon Mountcastle proposed that the neocortex is a tiling of one repeated microcircuit — the minicolumn — so that vision, audition, touch, language, and planning are all the same computation applied to different inputs. For nearly fifty years this was a framing device, not a testable claim: nobody could see the circuit, let alone compare it across areas. With the MICrONS cubic-millimeter connectome, whole-brain cell-type atlases, and multi-area Neuropixels recordings, it has finally become an empirical question — and it happens to be the biological bet underneath every architecture-uniform AI system you use, where one transformer block, weight-untied but structurally identical, is stacked and pointed at everything. (Note: I'm working from knowledge of the literature; the bibliographic details of the 1978 chapter were not verified against the source, and all numbers below are approximate.)

The idea as Mountcastle had it

Mountcastle's empirical starting point was his own 1957 discovery: when you drive a microelectrode vertically through cat somatosensory cortex, every neuron you meet responds to the same patch of skin and the same submodality (light touch, or deep pressure). Drive it obliquely, and the properties shift in discrete jumps. Cortex, he argued, is organized into narrow vertical columns — minicolumns of roughly 30–50 µm containing on the order of 100 neurons, grouped into macrocolumns of a few hundred microns. Hubel and Wiesel's orientation and ocular-dominance columns in visual cortex, found soon after, looked like the same story in a different area.

The 1978 essay (in The Mindful Brain, with Gerald Edelman) pushed this to its logical extreme. Two claims:

1. The unit module. The minicolumn is a canonical circuit — the same cell types, the same layered wiring, the same input–output transformation — replicated a few hundred million times across the human cortical sheet. What differs between areas is what is wired into them, not how they compute. Auditory cortex is visual cortex fed by the cochlea.

2. The distributed system. Function does not live in areas; it lives in re-entrant loops linking modules across many areas. An area is a node, not a faculty.

The supporting intuition was anatomical and evolutionary. Cortex looks strikingly monotonous under a Nissl stain — six layers everywhere, with local variations that seem like parameter tweaks (thick layer 4 where sensory input lands, thick layer 5 where motor output leaves). And cortical expansion in evolution happened by tiling more surface, not by inventing new local structure, which is exactly what you'd do if you had one good module and a way to stamp out copies.

The claim's radical content is often understated. It says there is one cortical algorithm. Find it in any patch — mouse V1 will do — and you've found the computation underlying language and planning too, up to inputs, outputs, and learned parameters.

Why it could not be tested in 1978

The claim quantifies over circuits — cell types plus their synaptic wiring plus their dynamics — compared across areas. Every term in that sentence was out of reach.

Recording. The state of the art was one tungsten electrode, one neuron at a time, in one area, in an anesthetized animal. Tetrodes and multi-electrode arrays were years to decades away. You could not measure the joint activity of a single minicolumn (~100 neurons), let alone compare population dynamics between V1 and A1 in the same brain.

Cell types. "Cell type" meant morphology under Golgi stains — a taxonomy of maybe a dozen categories, subjective and unquantified. There was no way to ask whether area X and area Y contained the same parts list in the same proportions.

Wiring. The only complete connectome then in progress was C. elegans — 302 neurons and roughly 7,000 synapses, hand-traced from electron micrographs over more than a decade (published 1986). A cortical minicolumn has ~100 neurons but its wiring is embedded in a volume containing millions of passing axons; a fair cross-area comparison needs on the order of a cubic millimeter per area, ~10⁵ neurons and ~10⁸–10⁹ synapses. Serial EM at that scale generates roughly a petabyte of imagery. In 1978 total data was measured in micrographs on film, segmented by human eyes; there was no digital pipeline, no ML segmentation, no storage. The gap is not an incremental one — it is roughly eight orders of magnitude in synapses mapped.

Perturbation. Testing "same computation" requires causal probes: silence cell type P in area X and area Y, compare the effect. Optogenetics arrived in 2005; cell-type-specific driver lines later still. In 1978 the tools were lesions and cooling — area-level sledgehammers.

So the hypothesis persisted as an organizing metaphor. The closest anyone came to a test was indirect: Douglas and Martin's "canonical microcircuit" (1991), inferred from intracellular recordings in cat V1 and proposed to generalize; and Mriganka Sur's rewired ferrets (late 1980s–2000), where retinal input routed into auditory thalamus made A1 develop orientation-tuned, visually responsive maps — strong evidence that a cortical area's function is set by its inputs, exactly as Mountcastle predicted. Suggestive, never decisive.

What changed

Three tool families, all maturing in the last decade, attack exactly the three missing measurements.

Connectomics at the right scale. The MICrONS project (published across a set of papers in 2025) reconstructed roughly a cubic millimeter of mouse visual cortex: on the order of 200,000 cells, ~half a billion synapses, kilometers of axonal wiring — co-registered with calcium-imaged activity from tens of thousands of the same neurons. For the first time, structure and function of a cortical volume exist in one dataset. Human and non-visual-cortex volumes are following.

Cell-type atlases. Single-cell transcriptomics (Allen Institute and others) has produced quantitative taxonomies of thousands of cell types across the whole mouse brain. The headline result is directly relevant: the cortical parts list is largely shared across areas, but the proportions and expression gradients vary systematically, especially along the sensory-to-associative hierarchy. That is neither "identical modules" nor "different machines" — it looks like one architecture with per-area hyperparameters.

Multi-area recording. Neuropixels probes record hundreds of units per shank; multi-probe rigs and the International Brain Laboratory's brain-wide surveys yield simultaneous population recordings across many areas during behavior. Comparing the dynamics of putatively identical circuits across areas is now routine engineering rather than fantasy.

And on the AI side, the hypothesis quietly won a proxy war. The transformer is Mountcastle's architecture in spirit: one block, structurally identical, repeated, with all specialization living in learned weights and in what you feed it. The same block handles text, images, audio, proteins, and actions. That doesn't prove cortex works this way — but it proves a uniform repeated module is sufficient for general competence, which removes the main a priori objection ("surely language needs different machinery than vision").

What a serious 2026 revival looks like

The hypothesis finally admits a clean falsifiable form: fit one module, test it everywhere.

1. Extract the module. From MICrONS-style joint connectome-plus-activity data in area X, fit a mechanistic circuit model — cell-type-resolved, connectivity-constrained, with free per-synapse-class weights — or a learned neural surrogate distilled from it. This is the modern replacement for Douglas–Martin's hand-drawn canonical circuit.

2. Freeze the architecture, test transfer. Take the same module — same cell types, same wiring statistics, ideally same fitted dynamical parameters — and ask whether it predicts stimulus responses and perturbation effects in areas Y and Z (S1, A1, a higher associative area) given only their measured inputs. Mountcastle predicts high transfer with only input/gain adjustments; the null predicts you need to refit the circuit itself. Cell-type-specific optogenetic perturbations, applied identically in two areas, give the sharpest test: the module predicts the perturbation response cross-area or it doesn't.

3. The ML mirror. Train a multimodal system where a single module (optionally weight-tied) is replicated across modalities, versus per-modality specialized modules, and compare which better predicts real cross-area neural data. Keep Mountcastle's re-entrant "distributed system" too: modules communicating through long-range loops, not a feedforward stack.

Reuse from the paper: the module/system decomposition, the input-defines-function principle, the tiling-by-replication evolutionary argument. Replace: the anatomical minicolumn as the unit (its status as a functional unit is genuinely contested — rodents mostly lack orientation columns yet see fine) with a statistically defined canonical circuit at the ~100 µm scale.

Has it been vindicated?

Partially, and the honest answer is contested. In favor: cross-area conservation of cell types, Sur's rewiring, cross-modal plasticity in blind humans (V1 recruited for Braille and language), the Douglas–Martin lineage, Bastos et al.'s mapping of predictive coding onto the canonical circuit (2012), Hawkins' Thousand Brains theory, and the transformer's existence proof. Against: systematic cell-type gradients across the hierarchy, primate V1's peculiar specializations, agranular motor cortex lacking a classic L4, and pointed critiques — Marcus, Marblestone, and Dean's "The atoms of neural computation" (2014) argues cortex is more plausibly a family of related circuits than one algorithm. My read: the field is converging on "one architecture, many parameterizations," which preserves Mountcastle's deep claim (a shared computational motif) while abandoning the literal one of byte-identical modules. What remains open is the part that matters: nobody has yet extracted the motif's computation from data and shown it transfers. That experiment is now merely hard, not impossible.

Where to read it

The essay appears in The Mindful Brain (Mountcastle & Edelman, MIT Press, 1978); a related and more accessible statement is Mountcastle's 1997 Brain review "The columnar organization of the neocortex." No link was provided and I have not verified the chapter details against the source. Read alongside: Douglas & Martin (1991, and their 2004 review) for the canonical microcircuit; Sur's ferret rewiring papers; Horton & Adams' skeptical "The cortical column: a structure without a function" (2005); Marcus et al. (2014) for the strongest modern counter-position; the MICrONS 2025 paper collection for the dataset that makes the test possible; and Hawkins' A Thousand Brains for the maximalist modern descendant.