Canonical computation of cortical columns
Do the repeated columns of the neocortex all perform the same basic computation?
Summary
Why it is open
The idea began with the column. Recording from the somatosensory cortex of cats, Vernon Mountcastle and colleagues found that all cells along a vertical electrode track responded either to superficial stimulation of skin and hair or to deep stimulation of joints and fascia, and in 1957 Mountcastle proposed "an elementary unit of organization in the somatic cortex made up of a vertical group of cells extending through all the cellular layers" [5]. In 1997 he described the classical areas of the neocortex as made of local circuits "repeated iteratively within each area"; these modules may differ between large areas in cell types, connections and mode of processing, but within one area they "have a basic similarity of internal design and operation" [3].
The problem bundles three questions: whether areas share a common circuit, whether they share a common computation, and what that computation is. The first two can come apart. Matteo Carandini and David Heeger write that "a canonical neural computation can rely on diverse circuits and mechanisms", which different regions or species may implement with different components [7]. The unit itself is disputed: half a century after Mountcastle's discovery, Jonathan Horton and Daniel Adams found that neuroscience "has no commonly accepted definition of 'column'" and concluded that the column "has failed as a unifying principle for understanding cortical function" [5]. How neural activity carries information is a related but separate problem, the neural code.
No body sets a bar for a resolution. A proposal becomes testable when it says what every area should show. Georg Keller and Thomas Mrsic-Flogel, for example, write that if predictive processing "describes a canonical cortical computation, we should find similar prediction error signals in other cortical areas", and note that most work on its circuits has focused on primary sensory areas [12].
What is known
Columns (1957–2005). After Mountcastle, columns were described in many cortical areas, and the term came to mean any vertical cluster of cells that share tuning for some receptive-field attribute [5]. Ocular dominance columns, which group cells in primary visual cortex by their preferred eye, are present in some species but not others, and vary among normal squirrel monkeys; Horton and Adams found no visual ability that appears to require them [5].
Neuron counts (1980–2013). A. J. Rockel, R. W. Hiorns and T. P. S. Powell counted about 110 neurons in a 30-micrometre strip through the depth of the cortex in several areas of mouse, rat, cat, monkey and human brains; the exception was primate primary visual cortex (area 17), whose binocular part had about 2.5 times as many [1]. They suggested that differences in cytoarchitecture and function "reflect differences in connections" [1]. Suzana Herculano-Houzel and colleagues found that the number of neurons under 1 mm² of cortical surface varies three-fold across nine primate species and the tree shrew, and concluded that models cannot assume that columns in different primates contain invariant numbers of neurons [6]. C. Nikoosh Carlo and Charles Stevens repeated Rockel and colleagues' counts with stereological methods in mouse, rat, cat and macaque, and confirmed them across areas and species, leaving primate primary visual cortex aside; they argue that the 2008 study, which used average cortical volume and surface area, is open to their objection that averages across areas cannot be assumed to be correct [9].
Circuits (1989–2015). Rodney Douglas, Kevan Martin and David Whitteridge combined the anatomy of single neurons with intracellular recordings in visual cortex into a simplified circuit of three neuronal populations, in which intracortical connections rather than the thalamic input supply most of the excitation each neuron receives [2]. The recordings came from cat visual cortex, and the circuit was proposed as a basic theme repeated, with minor variations, throughout the cortical sheet [8]. Reviewing mainly rodent data, Kenneth Harris and Gordon Shepherd describe a small number of neuron classes whose properties are "remarkably similar across areas", arranged in a basic pattern repeated with area- and species-specific modifications, an organisation they call serially homologous [10]. Some differences are qualitative: spiny stellate cells occur in primary visual cortex of cats and monkeys but not of rodents, and are generally absent from primary auditory cortex [10].
Cell-type atlases (2018–2023). In mouse, Bosiljka Tasic and colleagues sequenced 23,822 cells from primary visual cortex and anterior lateral motor cortex and defined 133 transcriptomic types; nearly all inhibitory (GABA-releasing) types were shared by both areas, whereas most excitatory (glutamatergic) types were found in only one [13]. In human tissue from 5 postmortem donors, Nikolas Jorstad and colleagues found all 24 cell subclasses in each of 8 areas, with large differences in their proportions: layer 4 intratelencephalic neurons ranged from 3% to 30% of excitatory neurons, and the ratio of excitatory to inhibitory neurons was 2:1 in most areas but 4.5:1 in primary visual cortex [15]. They write that "both a canonical and a non-canonical architecture were apparent, depending on the granularity of cellular detail analyzed" [15].
Testing one proposal (2018–2021). In layer 2/3 of mouse primary visual cortex, a subset of neurons responds when expected visual flow is absent, a "mismatch" response that Keller and Mrsic-Flogel read as a negative prediction error [12]. Selina Solomon and colleagues recorded in primary visual cortex and area V4 of macaque monkeys, and with scalp EEG in humans, while occasionally violating a fixed sequence of stimuli; responses to expected and pattern-violating stimuli were "nearly identical" [14]. They say this challenges the assertion that signalling prediction errors is a fundamental computational motif in sensory cortex, at least for predictions derived from temporal patterns [14].
Main approaches
Predictive coding. Feedback from higher areas carries predictions, and feedforward connections carry the residual errors between prediction and input [4]. Andre Bastos and colleagues found "a remarkable correspondence" between the microcircuitry of the cortical column and the connectivity that predictive coding implies, which they treat as a form of Bayesian inference [8]; see the Bayesian brain. Strongest evidence: in Rajesh Rao and Dana Ballard's model, trained on natural images, units developed receptive fields like those of simple cells and error units showed endstopping and other effects from outside the classical receptive field [4]; mismatch responses in mouse visual cortex [12]. Main difficulty: nearly identical responses to expected and unexpected stimuli in macaque and human visual cortex [14], and little work outside primary sensory areas [12].
Normalisation. A neuron's response is divided by a common factor that typically includes the summed activity of a pool of neurons [7]. Strongest evidence: first proposed for primary visual cortex, it is now reported throughout the visual system, in other modalities and brain regions, and in species from invertebrates to mammals [7]. Main difficulty: it also operates outside the cortex, for example in the fruit fly's olfactory system, and may rely on different mechanisms in different places [7], so on its own it does not say what is specific to cortical columns.
Columns as models of objects. Jeff Hawkins, Subutai Ahmad and Yuwei Cui propose that each column learns predictive models of the objects it senses by combining sensory input with a signal for location relative to the object, computed in the layers below layer 4 [11]. Strongest evidence: in simulations, small single-column networks learned to recognise hundreds of objects, and networks of several columns needed fewer movements of the sensors [11]. Main difficulty: the evidence comes from simulations, and the location signal is a proposal [11].
A shared plan with local variations. Areas share cell classes and a basic circuit, and quantitative differences may let them serve different functions [10]. Strongest evidence: area-specific excitatory types in mouse [13] and large differences in subclass proportions across human areas [15]. Main difficulty: the same atlases find nearly all inhibitory types shared between the two mouse areas and all 24 subclasses in every human area sampled [13, 15], a result Jorstad and colleagues call both canonical and non-canonical, depending on the level of detail [15].
Open sub-questions
- Prediction errors beyond sensory cortex. Do neurons in mouse prefrontal or motor cortex show feature-specific mismatch responses like those in layer 2/3 of primary visual cortex, as Keller and Mrsic-Flogel expect if predictive processing is canonical [12]?
- Self-generated predictions in primates. Does macaque primary visual cortex show prediction-error responses when the prediction comes from the animal's own movement rather than from a stimulus sequence, the case Solomon et al. did not test [12, 14]?
- Proportions and computation. Across the human areas sampled by Jorstad et al., does the ratio of excitatory to inhibitory neurons predict the strength of response normalisation measured in the same areas [7, 15]?
- Neuron counts. Do stereological counts in the primate species studied by Herculano-Houzel et al. reproduce their three-fold variation in neurons per square millimetre, or the constancy Carlo and Stevens found [6, 9]?
- Area-specific excitatory types. Do the glutamatergic types that Tasic et al. found in only one of two mouse areas also differ in their local connections, or only in their long-range projections [13]?
- Location in columns. Do neurons below layer 4 of somatosensory cortex encode where a touched point lies on an object, relative to the object, as Hawkins et al. propose [11]?
Work on this in Papex
If you have a testable answer to one of the sub-questions above, you can draft it as a private hypothesis in Papex, or prepare the draft from Claude or ChatGPT with the Papex connector. You review the exact version, sign it and decide whether to publish it; a public record that names this problem is listed on this page. A Papex record shows who signed which version and when. It is not peer review and does not mean the hypothesis is correct.
Key references
- Rockel, A. J., Hiorns, R. W. & Powell, T. P. S. (1980). The basic uniformity in structure of the neocortex. Brain 103(2), 221–244. 10.1093/brain/103.2.221 · Checked 15 Oct 2026
- Douglas, R. J., Martin, K. A. C. & Whitteridge, D. (1989). A canonical microcircuit for neocortex. Neural Computation 1(4), 480–488. 10.1162/neco.1989.1.4.480 · Checked 15 Oct 2026
- Mountcastle, V. B. (1997). The columnar organization of the neocortex. Brain 120(4), 701–722. 10.1093/brain/120.4.701 · Checked 15 Oct 2026
- Rao, R. P. N. & Ballard, D. H. (1999). Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects. Nature Neuroscience 2(1), 79–87. 10.1038/4580 · Checked 15 Oct 2026
- Horton, J. C. & Adams, D. L. (2005). The cortical column: a structure without a function. Philosophical Transactions of the Royal Society B 360(1456), 837–862. 10.1098/rstb.2005.1623 · Checked 15 Oct 2026
- Herculano-Houzel, S., Collins, C. E., Wong, P., Kaas, J. H. & Lent, R. (2008). The basic nonuniformity of the cerebral cortex. Proceedings of the National Academy of Sciences 105(34), 12593–12598. 10.1073/pnas.0805417105 · Checked 15 Oct 2026
- Carandini, M. & Heeger, D. J. (2012). Normalization as a canonical neural computation. Nature Reviews Neuroscience 13(1), 51–62 (published online 2011); erratum in 14(2), 152 (2013), DOI 10.1038/nrn3424. 10.1038/nrn3136 · Checked 15 Oct 2026
- Bastos, A. M., Usrey, W. M., Adams, R. A., Mangun, G. R., Fries, P. & Friston, K. J. (2012). Canonical microcircuits for predictive coding. Neuron 76(4), 695–711. 10.1016/j.neuron.2012.10.038 · Checked 15 Oct 2026
- Carlo, C. N. & Stevens, C. F. (2013). Structural uniformity of neocortex, revisited. Proceedings of the National Academy of Sciences 110(4), 1488–1493. 10.1073/pnas.1221398110 · Checked 15 Oct 2026
- Harris, K. D. & Shepherd, G. M. G. (2015). The neocortical circuit: themes and variations. Nature Neuroscience 18(2), 170–181. 10.1038/nn.3917 · Checked 15 Oct 2026
- Hawkins, J., Ahmad, S. & Cui, Y. (2017). A theory of how columns in the neocortex enable learning the structure of the world. Frontiers in Neural Circuits 11, 81. 10.3389/fncir.2017.00081 · Checked 15 Oct 2026
- Keller, G. B. & Mrsic-Flogel, T. D. (2018). Predictive processing: a canonical cortical computation. Neuron 100(2), 424–435. 10.1016/j.neuron.2018.10.003 · Checked 15 Oct 2026
- Tasic, B., Yao, Z., Graybuck, L. T., Smith, K. A., Nguyen, T. N., Bertagnolli, D. et al. (2018). Shared and distinct transcriptomic cell types across neocortical areas. Nature 563(7729), 72–78. 10.1038/s41586-018-0654-5 · Checked 15 Oct 2026
- Solomon, S. S., Tang, H., Sussman, E. & Kohn, A. (2021). Limited evidence for sensory prediction error responses in visual cortex of macaques and humans. Cerebral Cortex 31(6), 3136–3152. 10.1093/cercor/bhab014 · Checked 15 Oct 2026
- Jorstad, N. L., Close, J., Johansen, N., Yanny, A. M., Barkan, E. R., Travaglini, K. J. et al. (2023). Transcriptomic cytoarchitecture reveals principles of human neocortex organization. Science 382(6667), eadf6812. 10.1126/science.adf6812 · Checked 15 Oct 2026
Source
Open, as listed by Wikipedia · Unsolved problems in neuroscience. Checked 08 Oct 2026.
The status is the source’s. Papex does not decide whether a problem is solved.
Records on this problem
No public record addresses this problem yet.