---
title: "Humans"
document_id: sobe-2025-main-ai
document_type: report_section
parent_document: ../report.md
section_id: main-section-10-humans
section_order: 10
section_count: 17
language: en
license: CC-BY-4.0
---

<a id="main-section-10-humans"></a>

# Humans

> Selective-retrieval section 10 of 17 from the [complete AI-readable report](../report.md). Content is copied without editorial summarization; relative asset, data, and code paths are rebased for this directory.

- **Parent document:** [`report.md`](../report.md)
- **Section ID:** `main-section-10-humans`
- **Order:** 10 of 17
- **Words:** 3580
- **Previous:** [Mouse](09-mouse.md)
- **Next:** [Neural Dynamics: Brain Function & Activity](11-neural-dynamics-brain-function-and-activity.md)

### Anatomy & Behavior

The human brain reaches its peak volume in the third to fourth decade of life, typically measuring around 1000-1400 cm³ and weighing approximately 1.3-1.4 kg, or roughly 2% of total body weight ([Steen et al., 2007](https://doi.org/10.3174/ajnr.A0537), [Raichle et al., 2002](https://doi.org/10.1073/pnas.172399499)). The adult brain measures approximately 140 mm in width, 167 mm in length, and 93 mm in height, though considerable individual variation exists. The brain contains an estimated 86 billion neurons, with distinct regional distributions: approximately 16 billion neurons in the cerebral cortex, 69 billion in the cerebellum (which comprises only 10% of brain volume), and less than 1 billion distributed throughout other brain regions ([Azevedo et al., 2009](https://doi.org/10.1002/cne.21974)). While estimates for total brain synapses range between 100-1000 trillion, the most precise ones exist for the neocortex, containing an estimated 164 trillion synapses ±17% ([Tang et al., 2001](https://doi.org/10.1002/syn.1083)). Typical baseline firing rates observed in human cortical and hippocampal recordings (often derived from clinical patients) frequently average 0.5-4 Hz ([Aghajan et al., 2023](https://doi.org/10.1016/j.celrep.2023.113271)), with firing rates increasing in response to salient stimuli or during tasks, potentially reaching averages of 10-20 Hz or higher in some cortical neurons.

The breadth of human behavior patterns is not treated here, as we take reader familiarity as given.

### Neural Dynamics

Ethical considerations and the extensive genetic toolkit requirements make optical methods for single-neuron resolution recording currently non-viable in humans. Consequently, invasive electrophysiological techniques represent the only modality capable of achieving such resolution in the human brain. Typically employed within clinical research (e.g., as part of brain-computer interface trials) or during neurosurgical procedures, these methods provide valuable albeit highly localized data on individual neuron activity.

Pioneering efforts in chronic human intracortical recording have heavily relied on Utah Electrode Arrays (UEAs). These devices, often comprising a 96-channel grid of stiff silicon electrodes, are FDA-cleared for investigational BCI studies ([Sponheim et al., 2021](https://iopscience.iop.org/article/10.1088/1741-2552/ac3eaf)). UEAs have demonstrated impressive longevity in human participants, with successful recordings maintained for multiple years. While UEAs have enabled significant BCI achievement, they sample neurons primarily within a 2D cortical plane and typically yield a modest number of separable single units (often well below 150) ([Chung et al., 2022](https://doi.org/10.1016/j.neuron.2022.05.007); [Sponheim et al., 2021](https://iopscience.iop.org/article/10.1088/1741-2552/ac3eaf)).

High-density silicon probes, such as Neuropixels, have recently allowed for a significant increase in the scale and resolution of acute single-neuron recordings in humans during intraoperative settings ([Chung et al., 2022](https://doi.org/10.1016/j.neuron.2022.05.007)). For example, a recent study used Neuropixels probes (10 mm shank; 384 selectable channels out of 960 contacts/recording sites) in 8 neurosurgical participants, isolating 596 neurons in total and up to ~100 neurons simultaneously from a single insertion, with recordings typically lasting 10–20 minutes ([Chung et al., 2022](https://doi.org/10.1016/j.neuron.2022.05.007)). Another study employing the thicker Neuropixels 1.0-ST similarly obtained hundreds of spike-sorting clusters in three participants, with well-isolated neurons identified from those clusters after curation ([Paulk et al., 2022](https://www.nature.com/articles/s41593-021-00997-0)). An important challenge in these open craniotomy settings is the substantial brain motion due to cardiac and respiratory pulsations, which is negatively correlated with unit yield and necessitates sophisticated motion correction algorithms.

Further applications of Neuropixels have provided detailed accounts of single-neuron activity during human language processing. Recordings from 685 neurons across cortical layers in the superior temporal gyrus (STG) of 8 participants listening to speech ([Leonard et al., 2024](https://www.nature.com/articles/s41586-023-06839-2)) indicated that individual neurons encoded various speech sounds (e.g., consonants, vowels, pitch), with neurons at different cortical depths tuned to different speech features. In a study using Neuropixels in the language-dominant prefrontal cortex of 5 participants ([Khanna et al., 2024](https://www.nature.com/articles/s41586-023-06982-w)), activity from 272 neurons, even before speech, outlined the structure of upcoming words. This information about the word's structure was encoded in a specific, timed order. Similarly, Neuropixels recordings from 3 participants ([Jamali et al., 2024](https://www.nature.com/articles/s41586-024-07643-2)) identified single neurons in the prefrontal cortex that selectively represented word meanings, with activity also reflecting sentence context and semantic relationships. Such studies demonstrate the use of high-density probes to examine how individual human neurons process complex language elements, revealing specific encoding patterns and their organization.

Efforts towards next-generation, fully implantable, high-channel-count BCIs include Neuralink's N1 implant. This system uses 1,024 electrodes distributed across 64 flexible, robotically inserted "threads," designed for wireless data transmission and inductive charging ([Neuralink, 2024](https://neuralink.com/pdfs/PRIME-Study-Brochure.pdf)). As part of their PRIME clinical trial, three participants with quadriplegia have received the N1 implant. These participants have had their implants for over 670 days and used the BCI system for over 4,900 hours, with recent independent daily use averaging 6.5 hours ([Neuralink, 2025](https://neuralink.com/blog/a-year-of-telepathy/)). Initial BCI performance for cursor control has been reported at up to 8.0 bits-per-second (BPS). Wireless implants like Neuralink's N1 allow neural recording during more naturalistic daily activities than traditional wired systems. Challenges reported include the retraction of some electrode threads post-surgery in one participant, which required algorithmic adjustments to maintain performance. While still in early clinical stages, these systems aim to substantially increase the scale and practicality of human chronic neural recording.

While the aforementioned electrophysiological methods provide increasingly detailed recordings of neural activity, the ability to perturb specific human neurons at single-cell resolution in vivo to map function is severely limited. Optogenetic perturbation, a powerful tool in model organisms, is not currently viable for targeted modulation within the living human brain due to ethical barriers, challenges in precise gene delivery for opsin expression, and difficulty delivering light safely to deep brain structures.

Patch-clamp electrophysiology, however, serves as a robust method for perturbative studies on ex vivo human brain slices. Tissue resected during neurosurgery can be kept viable, allowing researchers to perform whole-cell patch-clamping. This enables precise current or voltage injections to characterize neuronal firing properties, pharmacological manipulations to block specific ion channels or activate receptors, and paired recordings with synaptic stimulation to investigate microcircuit connectivity ([Menéndez de la Prida et al., 2002](https://doi.org/10.1016/S0006-8993(02)02564-7); [Peng et al., 2019](https://doi.org/10.7554/eLife.48178)). Nevertheless, performing true whole-cell patch-clamp inside the intact, living human brain remains technically and ethically unfeasible. Other in-vivo human perturbation techniques, such as Transcranial Magnetic Stimulation (TMS) or Deep Brain Stimulation (DBS), affect larger neuronal populations and lack single-cell specificity.

------------------------------------------------------------------------

### Connectomics

Despite the considerable technical challenges of imaging such large volumes of nervous tissue at synaptic resolution, electron microscopy efforts in primate and human brains have achieved several notable milestones. Wildenberg et al. imaged a 0.8 mm × 2.4 mm × 40 nm block of rhesus macaque cortex, revealing that primate neurons receive 2-5 times fewer synaptic inputs than their mouse counterparts ([Wildenberg et al., 2020](https://www.cell.com/cell-reports/fulltext/S2211-1247(21)01156-6)). Loomba et al. conducted comparative connectomic analysis across species by imaging multiple cortical samples, including ~175 × 220 × 100 μm³ volumes of layer 2/3 from macaque and human, and a larger 1.7 × 2.1 × 0.03 mm³ volume spanning all cortical layers in human temporal cortex ([Loomba et al., 2022](https://doi.org/10.1126/science.abo0924)). A landmark achievement came in 2024 when Shapson-Coe et al. demonstrated high-throughput serial section electron microscopy of a 170-μm-thick slab of human temporal cortex at nanoscale resolution. This volume, obtained during epilepsy surgery, spans approximately 1.05 mm³ of tissue (accounting for sectioning-induced compression) and was imaged at a resolution sufficient to resolve individual synapses and subcellular structures. Thus, it represents the most significant volume of human brain tissue imaged at a resolution sufficient for dense connectomics to date ([Shapson-Coe et al., 2024](https://doi.org/10.1126/science.adk4858)).

The most significant X-ray imaging effort for human connectomics is the SYNAPSE (Synchrotron for Neuroscience – an Asia-Pacific Strategic Enterprise) collaboration, which aims to map an entire human brain at 0.3 μm resolution ([Stampfl et al., 2023](https://doi.org/10.1016/j.physrep.2022.11.003)) in 2-3 mm-thick sections ([Chen et al., 2021](https://journals.iucr.org/s/issues/2021/05/00/ay5583/index.html)). The collaboration has made remarkable technical progress, achieving imaging speeds of 1 mm³ per minute and coordinating roughly 10 synchrotron facilities, including the first beamline entirely dedicated to connectomics at the Taiwan Photon Source ([Chen et al., 2021](https://journals.iucr.org/s/issues/2021/05/00/ay5583/index.html)). Their next phase, SYNAPSE 2.0, aims to achieve another 10-100x speed increase to 0.2 mm³/second ([Stampfl et al., 2025](https://doi.org/10.1080/08940886.2024.2414729)). Despite this, it is important to note that the project's target resolution is designed for mapping cellular distributions and long-range projections, not for resolving individual synapses, and the SYNAPSE roadmap does not currently include plans to reach synaptic-level detail. However, separate theoretical proposals suggest that a similar high-throughput synchrotron pipeline could potentially reach synaptic resolution by integrating expansion microscopy (ExxRM), thus bridging the current resolution gap ([Collins, 2023](https://logancollinsblog.com/2023/02/22/feasibility-of-mapping-the-human-brain-with-expansion-x-ray-microscopy/)).

The reconstruction of primate connectomes at synaptic resolution remains a massive undertaking. Storage requirements alone would be significant: assuming one byte per voxel, a typical marmoset brain would require 5.8 exabytes at 10 nm isotropic resolution (0.37 exabytes at 25 nm); a typical macaque brain 104 exabytes (6.65 exabytes at 25nm); and a human brain roughly 1-1.4 zettabytes (64-90 exabytes at 25nm). The volume imaged by Shapson-Coe et al., representing an extremely minute part of a whole human brain, or approximately 0.00007% of a whole human brain, produced over 1.4 petabytes of data ([Shapson-Coe et al., 2024](https://doi.org/10.1126/science.adk4858)), motivating the development of dedicated infrastructure ([Maitin-Shepard and Leavitt, 2022](https://research.google/blog/tensorstore-for-high-performance-scalable-array-storage/)). Given the scale of this challenge, some whole-brain mapping efforts opt for lower-resolution imaging approaches. Even so, storage requirements remain substantial - the SYNAPSE collaboration, which aims to map the human brain at 0.3 μm isotropic resolution using X-ray imaging, estimates they will need exabyte-level storage for a single brain dataset and approximately 100 exabytes for their complete connectome mapping goals. Reconstruction would also represent a daunting challenge. The H01 dataset released by Shapson-Coe et al., even with thousands of Google-developed Tensor Processing Units (TPUs) for automated neuron segmentation and synapse detection methods ([Blakely and Januszewski, 2021](https://research.google/blog/a-browsable-petascale-reconstruction-of-the-human-cortex/)), still required extensive manual proofreading. While the computational pipeline successfully identified about 16,000 cells and 150 million synapses, ensuring accuracy still demanded intensive human effort, with expert reviewers spending over 3.5 hours per neuron to correct remaining errors. Proofreading efforts continue to this day, with the original release including only 104 proofread cells.

### Computational Modeling

Unlike organisms like Drosophila or larval zebrafish, computational neuroscience of the human brain operates under severe data constraints. There is no synapse-level connectome, and functional recordings either cover only hundreds of cells or provide only indirect measures of neural activity at low spatial and temporal resolution (in the case of non-invasive methods like fMRI). This fundamental limitation has shaped a landscape dominated by feasibility studies - efforts focused primarily on demonstrating the possibility of simulating human brain-scale networks on current supercomputing infrastructure, rather than attempting to replicate specific circuits or behaviors with high biological fidelity.

#### Yamazaki et al., 2021

Yamazaki and colleagues developed one of the first human-scale simulations of the cerebellum using their MONET simulator on Japan's K supercomputer, capable of 11.3 PFLOPS ([Yamaura et al., 2021](https://doi.org/10.3389/fninf.2020.00016)). Their model leveraged anatomical studies providing cerebellar layer thicknesses and cell density measurements across different regions. This structural data informed the spatial organization and the number of neurons of each type in their model.

They constructed a network of 68 billion neurons and 5.4 trillion synapses from this anatomical data. Neurons were modeled as conductance-based leaky integrate-and-fire units with α-function synapses implementing four receptor types (AMPA, NMDA, GABAA, GABAB). Connection patterns between neurons were defined using two-dimensional Gaussian distributions, with parameters like connection probabilities and spatial extents derived from anatomical studies.

Model validation occurred in two stages. First, they examined resting state activity, comparing baseline firing rates across different cell types with experimental data. Second, they tested the model's ability to reproduce the optokinetic response (OKR), a reflexive eye movement controlled by the cerebellum. For OKR, they found Purkinje cells modulated their firing rates between 50-80 Hz, within the range observed across various animal studies (20-100 Hz). Running on 82,944 nodes of the K computer, the simulation achieved speeds 578 times slower than real-time.

#### Lu et al., 2023

Lu and colleagues developed the "Digital Twin Brain" (DTB) platform, implementing a whole-brain simulation of 86 billion neurons and 47.8 trillion synapses on a GPU-based supercomputing cluster ([Lu et al., 2024](https://www.nature.com/articles/s43588-024-00731-3)). Their model's structure was informed by three types of macroscopic brain imaging data: structural MRI provided regional neuron densities at 3x3x3mm³ voxel resolution, diffusion tensor imaging (DTI) determined voxel-to-voxel connection probabilities, and PET data established what proportion of connections remained local vs projecting to other regions (varying from ~29% external connections in cortex to ~13% in cerebellum). Of all possible connections between voxels, only 0.72% were realized, reflecting the brain's sparse connectivity.

They developed a "hierarchical mesoscopic data assimilation" (HMDA) approach to tuning model parameters. This involved first training a smaller network (0.2 billion neurons) to estimate hyperparameters for synaptic conductances and then using these to initialize the full model. Parameters were iteratively refined by comparing simulated BOLD signals to real fMRI data during a visual evaluation task. Each neuron was modeled as a conductance-based LIF unit with four receptor types (AMPA, NMDA, GABAa, GABAb). Average input synapses per neuron were set to 1000 for cortical/subcortical regions and 100 for brainstem/cerebellum. Long-range connections were restricted to excitatory neurons only. To manage computational demands on their 14,012 GPU cluster (each node contained four GPUs with 16GB of memory each), they developed a partitioning algorithm that balanced computational load and minimized inter-GPU communication. The fitted model achieved correlations above 0.98 with experimental BOLD signals in input regions (with a 2-timepoint lag) and 0.75 across the cortex. When trained to predict participants' emotional image ratings, the model significantly correlated with actual ratings (r=0.655, p=0.006). Running at different firing rates (7Hz, 15Hz, and 30Hz), the simulation achieved speeds between 65-118.8 times slower than real-time.

### Gap Analysis

The human brain sits in a league of its own, being more than three orders of magnitude larger than the mouse brain. This extraordinary scale creates a fundamental divide in feasibility: while whole-brain recording at cellular resolution is achievable in zebrafish and likely feasible in Drosophila, and whole-cortex recording remains at least imaginable for the mouse, comprehensive neural recording in humans faces insurmountable physical barriers with current or near-future technologies, even setting aside regulatory and ethical challenges. Most models are based either on low-resolution functional data (fMRI) or anatomical observations, requiring significant extrapolation from these sparse constraints. Unlike more approachable organisms where neural recordings can constrain brain simulations, human-scale models will likely need to find different approaches, potentially relying on infering parameters primarily from structural datasets. Given current technology, human brain connectome efforts would require industrial-scale operations that dwarf any existing neuroscience facility. Facilities the size of modern semiconductor manufacturing plants: thousands of electron or optical microscopes running in parallel, or entire synchrotron facilities with multiple beamlines dedicated to brain mapping. Such facilities would demand large-scale storage and computing infrastructure to process the multi-exabyte datasets generated daily.

<a id="table-12"></a>

### Table 12 — Model Organism Overview: Human

| Model Organism Overview: Human | Pros | Cons |
|---|---|---|
| Anticipated Scientific Insights | • Fundamental insights into Neuroscience: It seems plausible that advanced brain models will transform our understanding of how the brain works and how consciousness and personality traits arise.<br>• Applications for human health<br>• Feasibility of personality-preserving brain emulations | • Miscellaneous ethical consideratioins: Potential risks to individuals or even whole society have to be carefully evaluated. |
| Experimental Tractability | • Computational models approaching human scale: Recent modelling on high-performance computing datacenters is getting close to managing the loads necessary to model whole human brains. Additionally, progress towards more efficient neuromorphic hardware is ongoing.<br>• First connectome reconstructions of human tissue: First reconstructions of human brain tissue can rely on the same methods as smaller organisms. | • Methodological limitations in humans: Many technologies require genetic engineering or highly invasive brain surgery.<br>• Resolution vs Coverage Trade-off. Non-invasive methods provide whole-brain coverage but at the cost of spatial and temporal resolution. fMRI measures slow hemodynamic responses averaged across hundreds of thousands of neurons, while EEG and MEG provide faster temporal sampling but even coarser spatial resolution. Conversely, invasive recordings can resolve individual neurons but are restricted to tiny fractions of brain tissue, capturing thousands of neurons at best compared to the brain's ~86 billion. This trade-off between resolution and coverage appears fundamental rather than technological, making comprehensive recordings with full human brain coverage at single-neuron resolution permanently out of reach.<br>• Scale of human brain: At 1000x the scale relative to the mouse brain, the operational challenges mentioned in the section about mouse are magnified substantially. For instance, achieving synaptic resolution in a volume of ~1,200 cm³ (vs. ~0.5 cm³ for the mouse) demands capacity and throughput far beyond the capabilities of current laboratories. Meeting this challenge will require near-industrial levels of coordination, equipment scale-up, and data infrastructure, elevating human connectomics from a traditional scientific project to a massive engineering and logistics undertaking.<br>• Likely strong interindividual and intercultural differences: Given human diversity differences in brain connectivity are expected. |

------------------------------------------------------------------------

<a id="table-13"></a>

### Table 13 — Gaps and Opportunities: Human

| Gaps and opportunities: Human | Gaps (non-exhaustive selection) | Illustrative Project Opportunities |
|---|---|---|
| Neural dynamics | • Structure-to-Function Dependency. A human connectome's ultimate value for brain emulation depends critically on bridging the structure-to-function gap. Even with molecular annotations, a fully reconstructed wiring diagram is only the first step. | • Extensive “Human in a dish recording”: Instead of gathering data in vivo, human neuronal cell lines could be used to collect large-scale data sets and facilitate modelling. |
| Connectomics | • Electron Microscopy. At current effective scan speeds, achieving a complete human EM dataset within a decade would require approximately 30,000 parallel electron microscopes (Jefferis et al., 2023). Beyond microscope availability, sample preparation, automated image segmentation, and proofreading each introduce substantial computational and manual labor demands, though recent advances in machine learning have improved automation capabilities (Januszewski et al., 2020; Schmidt et al., 2021).<br>• More scalable methodologies: Expansion Microscopy (ExM) and X-ray Microscopy (XRM): ExM has never been attempted at anything close to the scale of a human brain. Tissue anisotropies, distortion, and batch-to-batch consistency remain unresolved, particularly at the extremely high expansion factors needed to achieve synapse-level resolution in large volumes. XRM promises high-speed volumetric imaging of thick specimens, thanks to high-brightness synchrotron beams, and is not constrained by the diffraction limit. However, XRM's practical viability for a complete human connectome remains untested, with significant challenges in achieving dense reconstructions at synapse-level resolution.<br>• Lack of baselines: In humans we have a surprisingly poor understanding of variables such as number of synapses per neuron in different areas of the brain or the ratio of local and distant connections for neurons. | • High-throughput EMs: Substantial optimization of beamlines, automated sample handling, etc., could increase effective data acquisition speeds by an order of magnitude, potentially more.<br>• Prototypes of advanced imaging modalities: Establishing protocols and scalability tests for methodologies like ExM and XRM.<br>• Small connectomics studies across many areas of the brain: Sample tiny areas from many areas of the brain from multiple individuals to determine variables for estimating the total computational demands. |
| Computational Neuroscience | • Data Scarcity : The most fundamental limitation is the lack of adequate functional and structural data to constrain human brain models. While future technologies may eventually provide detailed structural data through connectomics, functional data at cellular resolution will likely remain permanently out of reach due to physical and ethical constraints. This forces models to rely on massive extrapolations from animal studies or indirect measurements, severely limiting their biological validity.<br>• Extreme Resource Demands : Even with optimal implementation, biophysically detailed models at accurate human scale (86B neurons with realistic synaptic connectivity) would require exascale computing systems for real-time simulation. The "memory wall" and interconnect bandwidth limitations pose particular challenges for efficiently simulating such massive, densely connected networks.<br>• Model Validation Challenges : A key issue is the limited direct data on human brain activity. In animal models (e.g., mice), scientists can gather extensive and controlled neural recordings. However, data from human brains is significantly less comprehensive, and researchers have less control over recording locations. Further, relying solely on behavioral validation is insufficient to confirm that the emulation truly replicates the brain's underlying neural mechanisms. | • Human-scale neuromorphic computing experiments: Stress testing the currently existing human-scale neuromorphic computing systems and identifying and iterating on their strengths and weaknesses. |