Cortex: Neuroscience, Cell Mechanics & AI

• Cortex is a multifaceted term referring to structured peripheral domains in neuroscience, cell biology, and computing with distinct yet decisive organization.
• In neuroscience, the cerebral cortex features layered anatomy, specialized neurons, and conserved metabolic scaling laws that underpin efficient connectivity.
• In cell biology and computing, cortex denotes an active actomyosin shell and serves as a conceptual framework in AI, influencing imaging and computational models.

Cortex is a polysemous research term. In neuroscience it usually denotes the cerebral cortex, the layered outer tissue of the mammalian brain; in cell biology it denotes the thin actomyosin layer beneath the plasma membrane; and in contemporary computing literature it is also reused as an acronym for several methods and systems. The term therefore spans distinct biological and computational objects, but in each case it names a structured peripheral domain whose organization is treated as functionally decisive.

1. Cerebral cortex as a layered anatomical substrate

The neocortex is described as a large sheet of neural tissue, about 75% of brain volume, roughly 2–4 mm thick, and organized into six layers, with additional sublayers, cortical areas, columns, minicolumns, and cells (Giorgio, 2017). At the cellular level, the papers in this corpus repeatedly emphasize pyramidal neurons, dense dendritic arborization, and differentiated input streams to proximal, basal, and apical dendrites as key anatomical motifs (Giorgio, 2017).

Direct three-dimensional imaging confirms that cortical organization is not merely laminar in the abstract but volumetric. Hard x-ray microtomography of human frontal cortex, after Golgi silver impregnation, achieved $1.0\,\mu\mathrm{m}$ isotropic spatial resolution in reconstructed volumes and resolved gray and white matter, cortical layers II–VI, large pyramidal cells in the internal pyramidal layer, and apical dendrites extending toward the cortical surface (Mizutani et al., 2016). The same work showed that cortical microarchitecture can be inspected within intact tissue blocks rather than inferred from serial thin sections, making lamination, dendritic orientation, and vascular/luminal structures jointly visible in 3D (Mizutani et al., 2016).

A recurrent implication in current work is that the cerebral cortex is now studied simultaneously as tissue, as a graph of interacting cells, and as a high-dimensional image object. That shift underlies later attempts to model cytoarchitecture computationally, to relate it to metabolism and connectivity, and to use it as an explicit conditioning signal for generative models.

2. Conserved design, scaling laws, and neurovascular coupling

A central comparative claim is that mammalian cerebral cortex exhibits strong neuroanatomical and metabolic regularities despite spanning about four orders of magnitude in cortical volume (Karbowski, 2014). Adult synaptic density is reported to be about $5 \cdot 10^{11}\ \text{cm}^{-3}$, the excitatory synapse fraction is $0.83 \pm 0.03$, glial density is $\sim 2 \cdot 10^4\ \text{mm}^{-3}$, and capillaries occupy about $1\!-\!2\%$ of cortical volume (Karbowski, 2014). Intracortical axon diameters are about 0.2–0.3 $\mu m$, cortical thickness varies only weakly across species, and average path length between cortical areas is around 2.0 (Karbowski, 2014).

These anatomical regularities are paired with systematic metabolic scaling. Total brain metabolism scales approximately as $V^{5/6}$, whereas volume-specific cortical metabolism scales approximately as $V^{-1/6}$ (Karbowski, 2014). The same review argues that metabolic energy per cortical neuron, blood flow per neuron, the ratio $\text{CMR}/\text{CBF}$, and capillary length per neuron are all approximately invariant across mammals (Karbowski, 2014). Because capillary length density, neuron density, blood flow, and metabolism co-vary, cortical neuroanatomy and metabolic supply are treated as coupled systems rather than independent layers of description (Karbowski, 2014).

The same body of work also emphasizes trade-offs. Microscopic connection probability decreases sharply with cortical size, even while macroscopic area-to-area connectivity remains relatively high; larger brains therefore preserve global integration without sustaining dense local all-to-all coupling (Karbowski, 2014). Karbowski interprets the resulting architecture as a compromise among wiring economy, conduction delay, metabolic cost, robustness, and functional integration, rather than as the optimum of any single design principle (Karbowski, 2014).

3. Proposed operating principles of cortical computation and memory

Several papers in this corpus treat cortex not only as anatomy but as a candidate computational substrate with reusable principles. One proposal, due to Rvachev, casts the cerebral cortex as a reward-modulated classification architecture implemented in compartmentalized pyramidal neurons: apical tuft inputs initiate trial “guess” firing, basal dendritic synaptic clusters encode input subpatterns, and plasticity is revised to align with behavioral time scale synaptic plasticity (BTSP) (Rvachev, 2023). In that account, burst firing associated with dendritic $\mathrm{Ca}^{2+}$ spikes is linked to attentional, aware processing, whereas strengthened basal clusters later enable automatic classification without voluntary control (Rvachev, 2023).

A more explicitly systems-level proposal argues that cognition may depend on a small repertoire of cortical primitives defined over sets of neurons rather than single synapses. The candidate set includes directed association $5 \cdot 10^{11}\ \text{cm}^{-3}$0, supervised memorization of conjunctions of the form $5 \cdot 10^{11}\ \text{cm}^{-3}$1, inductive learning of simple threshold functions, and hierarchical memory formation in which a new set $5 \cdot 10^{11}\ \text{cm}^{-3}$2 is allocated to represent a conjunction (Valiant, 2018). That paper’s emphasis is methodological: such primitives are proposed as experimentally testable by in-circuit stimulation and recording, rather than inferred only from behavior (Valiant, 2018).

A different theoretical line proposes that the physical form of long-term memory in cortex is a connected subgraph of a very large directed graph whose nodes are neurons and whose edges are effective synaptic connections (Wei et al., 2024). In that formulation, disparate sensory fragments of an event are linked by graph connectivity, and the abundance of possible connected subgraphs explains large storage capacity (Wei et al., 2024). The paper uses a distance-dependent connectivity model and random directed graph arguments; with an effective unidirectional connection saturation of about $5 \cdot 10^{11}\ \text{cm}^{-3}$3, the condition $5 \cdot 10^{11}\ \text{cm}^{-3}$4 is satisfied for $5 \cdot 10^{11}\ \text{cm}^{-3}$5, which the authors use to argue that Hamiltonian cycles, and therefore connected traces, are easy to construct in cortex-like graphs (Wei et al., 2024).

At a more engineering-oriented level, "Simple Cortex" abstracts neocortical ideas into stimuli, synapses, dendrites, forests, neurons, and areas, with sparse activation through inhibition, Hebbian-like permanence updates, and temporal prediction via feedback of previous neural states (Giorgio, 2017). These accounts are presented as models or hypotheses rather than as a single established doctrine, but together they show a common tendency: cortical explanation is shifted upward from isolated neurons toward dendritic subunits, neuronal sets, graph structure, and reusable computational operations.

4. Cell cortex as an active viscous shell

In cell biology, the cortex is a different object: a thin layer beneath the plasma membrane that gives animal cells mechanical resistance and drives shape change (Rocha et al., 2021). It is composed mainly of actin filaments, actin-binding proteins, and myosin molecular motors, and it is treated as a thin, curved, continuously renewed actomyosin layer whose large-scale dynamics are governed by viscous dissipation, active myosin stress, and active material turnover rather than long-lived elastic strain storage (Rocha et al., 2021).

A thin-shell reduction from a 3D incompressible active-gel model yields a 2D viscous active shell theory on the cortex midsurface (Rocha et al., 2021). In that theory, membrane tensions and bending moments depend on in-plane strain rate $5 \cdot 10^{11}\ \text{cm}^{-3}$6, curvature-rate $5 \cdot 10^{11}\ \text{cm}^{-3}$7, curvature $5 \cdot 10^{11}\ \text{cm}^{-3}$8, active contractility $5 \cdot 10^{11}\ \text{cm}^{-3}$9, and turnover via $0.83 \pm 0.03$0 and $0.83 \pm 0.03$1. Thickness is itself dynamic, with the leading-order evolution law

$0.83 \pm 0.03$2

The review’s distinctive claim is that turnover is mechanically active: depending on the sign of $0.83 \pm 0.03$3, it can contribute either contractile or extensile stress, and in simulated cytokinesis it can account for up to about 25% of the integrated active power input, whereas in rapid hyper-osmotic shocks passive bending can contribute up to about 60% of total viscous dissipation (Rocha et al., 2021).

Experimental work on mitotic HeLa cells adds a molecular mechanosensitivity layer. Short-lived peaks in active or passive cortical tension trigger a twofold mechanical reinforcement strategy: direct catch-bond mechanosensitivity of filamin and $0.83 \pm 0.03$4-actinin cross-linkers, and indirect reinforcement via enhanced actin polymerization (Ruffine et al., 2023). For example, filamin A FRAP recovery slowed from $0.83 \pm 0.03$5 s under reduced tension to $0.83 \pm 0.03$6 s in control cells, while active tension peaks increased cortical localization of FLNA, ACTN1, and MYH9 by more than 10%, with MYH9 increasing by about 30%, and cortical Lifeact by more than 35% (Ruffine et al., 2023). The authors describe this as a “molecular safety belt” that protects the actin cortex from mechanical injury (Ruffine et al., 2023). A common misconception is to equate this cortex with the cerebral cortex; the papers are explicit that it is instead a mechanically active cortical shell of the cell.

5. Cortical morphology in organoids, representation learning, and image generation

In organoid morphogenesis, the term again shifts meaning. In the quasi-2D organoids modeled by the cortex-core paper, the cortex is the surrounding region of radially stretched or extended cells outside a dense compact core, and the authors explicitly note that this is not a six-layered cerebral cortex in the mature in vivo sense (Borzou et al., 2021). Their theory has two stages: an initial hydrodynamic instability in which cortex-core structures are seeded when an effective attraction exceeds a pressure-like term, $0.83 \pm 0.03$7, and a later extended buckling-without-bending morphogenesis model in which a growing cortical annulus around an incompressible core develops asymmetric sulci and gyri when active strain regulation is included (Borzou et al., 2021).

Recent computational work turns cortical microarchitecture itself into a learned representation. CytoNet is a foundation model trained on ten postmortem human brains, 4,654 histological sections, and microscopy patches of size $0.83 \pm 0.03$8 at $0.83 \pm 0.03$9, using a proximity-weighted contrastive loss called SpatialNCE (Schiffer et al., 21 Oct 2025). The model’s features support 113-area cortical classification with macro-F1 0.69 under linear probing and 0.71 with finetuning on seen brains, layer segmentation with macro-F1 0.63 from only 1% of the training labels, and unsupervised separation of frontal pole areas Fp1 and Fp2 with 94.75% accuracy in the two-cluster setting (Schiffer et al., 21 Oct 2025). Attention maps highlight histological landmarks such as the stripe of Gennari, Betz giant cells, and prominent layer IV patterns, indicating that the learned space captures classical cytoarchitecture rather than only gross geometry (Schiffer et al., 21 Oct 2025).

A complementary development, Cor2Vox, treats the cortex as an explicit generative prior for 3D brain MRI synthesis (Bongratz et al., 27 Jan 2026). It converts high-resolution pial and white matter surfaces into signed distance fields, fits a cortical statistical shape model from 33,403 UK Biobank scans, and uses a 3D shape-to-image Brownian bridge diffusion process to generate MRI from a cortex representation $\sim 2 \cdot 10^4\ \text{mm}^{-3}$0 (Bongratz et al., 27 Jan 2026). On ADNI test data, Cor2Vox reports SSIM 0.906 ± 0.018, left white-matter ASSD 0.283 ± 0.029 mm, and left pial ASSD 0.251 ± 0.019 mm, outperforming several diffusion and translation baselines on cortical surface recovery (Bongratz et al., 27 Jan 2026). It also supports controlled simulation of cortical thinning, with mean absolute error about 0.14 mm in recovered relative cortical thickness change, and harmonization of external frontotemporal dementia scans without retraining (Bongratz et al., 27 Jan 2026). This suggests a new use of cortex in imaging research: not only as the object of analysis, but as the conditioning variable that organizes image synthesis.

6. “CORTEX” as a recurring acronym in AI and systems research

Outside biological usage, CORTEX is repeatedly adopted as an acronym for unrelated computational systems. Examples include a token-level hallucination detector for retrieval-augmented generation that compares final-layer hidden states with and without retrieved references (Furumai et al., 30 Jun 2026), a large-scale spiking brain simulator for Fugaku built around Indegree Sub-Graph Decomposition (Lyu et al., 2024), a cost-sensitive surrogate tree and rule extractor for multiclass explainable AI (Kopanja et al., 5 Feb 2025), a correlation-aware database indexing system that extends the reach of a primary index to additional attributes (Nathan et al., 2020), a workflow-aware serving platform for agentic workloads based on stage isolation (Pagonas et al., 15 Oct 2025), and a structured reasoning benchmark for 3D chest CT MLLMs whose name expands to Clinically Organized Reasoning and sTructured EXplanation (Malik et al., 25 Jun 2026).

The same naming tendency appears in Cortex Neural Network, which treats “cortex” as an upper architecture over ordinary neural networks, with a sensory cortex, an association cortex area, and a reflection mechanism that creates specialist subnetworks for error-prone cases; the paper reports 98.32% on MNIST and 62% on CIFAR10 in a mixed setting, alongside a claim of about 40% loss reduction (Gao, 2018). The reuse of the name across such different systems does not imply conceptual identity. In some cases it signals layered control, in others structure-aware verification or large-scale coordination; in all cases the acronym is local to the paper rather than a shared technical standard.

One further caution follows from this multiplicity. Even when a paper uses “CORTEX” in a reliability or reasoning context, the target of verification is often narrow. The RAG hallucination detector, for example, is explicit that it detects lack of grounding in the provided references, not general factual incorrectness (Furumai et al., 30 Jun 2026). The same terminological precision is needed across the broader literature: “cortex” may designate cerebral tissue, cell mechanics, organoid morphology, or a named computational method, and the correct interpretation depends entirely on domain.