---
title: "Neural Dynamics: Brain Function & Activity"
document_id: sobe-2025-main-ai
document_type: report_section
parent_document: ../report.md
section_id: main-section-11-neural-dynamics-brain-function-and-activity
section_order: 11
section_count: 17
language: en
license: CC-BY-4.0
---
# Neural Dynamics: Brain Function & Activity
> Selective-retrieval section 11 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-11-neural-dynamics-brain-function-and-activity`
- **Order:** 11 of 17
- **Words:** 8400
- **Previous:** [Humans](10-humans.md)
- **Next:** [Connectomics: Brain structure reconstruction](12-connectomics-brain-structure-reconstruction.md)
In the 1920s, English physiologist Edgar Adrian wondered if it might be possible to record the electrical activity of a single neuron as it fired, an idea that seemed almost impossible given the technology of the time. Most scientists believed neural signals were too fast and too small to measure. Using a primitive amplifier and an electrode thinner than a human hair, Adrian managed to record the electrical pulses from a single sensory nerve fiber. His 1928 paper The Basis of Sensation ([Adrian, 1928](https://digitalcollections.ucalgary.ca/archive/Basis-of-sensation---the-action-of-the-sense-organs--The-2R3408TII00D.html)) revealed for the first time that neurons communicate through discrete electrical impulses, or action potentials, and that the frequency of these impulses encodes the intensity of the stimulus. This discovery earned Adrian the Nobel Prize ([Adrian, 1932](https://www.nobelprize.org/prizes/medicine/1932/adrian/lecture/)).
Close to a century and many Nobel Prizes later, we now appreciate that neuronal activity is primarily shaped by fast synaptic inputs through chemical synapses and electrical coupling via gap junctions. These signals are modulated by slower-acting neuromodulators, hormones, and, to some extent, by surrounding glial cells. The structural features of neurons, particularly their dendritic/synaptic architecture, membrane properties, and ion channels, influence how these signals are integrated. Activity-dependent plasticity mechanisms allow neurons to adjust their properties based on experience. Note that the brain operates over time scales covering at least twelve orders of magnitude; electrical activity happens at the millisecond timescale, whereas the brain changes physically over decades.
Many modalities have been developed over the past century to measure neuronal activity. They differ substantially, by multiple orders of magnitude, in their temporal (sampling rate and recording duration) and spatial (neuron resolution and brain coverage) properties. The following figure illustrates this by comparing five different modalities across those axes. The figure highlights the benefits and drawbacks of each recording modalities, as well as the amount of information captured by the respective method. The ideal neuroscientific recording achieves single-neuron, single spike resolution across the whole brain and permits chronic recording.
Figure 17 - Simplified comparison of major neural recording modalities across key dimensions. Comparison across four dimensions. Resolution as number of individual cells recorded. Speed as temporal resolution in frames per second. Maximum recording duration per session and total volume recorded. An ideal recording method for whole-brain human recording would rank at the top of each bar.. Calculations are in the [data repository](https://docs.google.com/spreadsheets/d/1so1BojMiuSTadc6BvGAkQL-CM9ieRaUkmFYWh4iGmYc/edit?gid=1901438326#gid=1901438326).
### Machine-readable figure record: `main-fig-17`
```yaml
id: main-fig-17
document_id: sobe-2025-main-ai
document_path: report.md
anchor: main-fig-17
canonical_url: https://brainemulation.mxschons.com/ai/report.md#main-fig-17
figure: "Neural recording modalities comparison"
assets:
- "../assets/report/main-fig-17-publication.png"
data_status: final_circulation_export_raster_with_contextual_current_sources
generator: ../code/run_all_figures.py
generator_function: generate_recording_modalities
publication_export_members:
- "images/image41.png"
canonical_pdf_pages:
- 107
note: "Scores are literals in the generator, not loaded from the similarly named comparison CSVs."
```
#### Inline contextual data (not an exact publication input): hardcoded modality scores
```tsv
method resolution_score speed_score duration_score volume_score
fUS 3 3 4 3
Calcium 4 2 2 2
Voltage 4 3.25 1.5 1.75
MEA 4 4 4 1.5
EEG 1 4 3 3.5
fMRI 2 1 3 4
```
#### Figure-generation source map (valid Python; not standalone plotting code)
```python
GENERATOR_PATH = "../code/run_all_figures.py"
SOURCE_FUNCTION = 'generate_recording_modalities'
SELECTED_SOURCE_LINES = (
'categories = ["Resolution", "Speed", "Duration", "Volume"]',
'x = np.arange(len(categories))',
'ax.set_xlim(-0.2, len(categories) - 1 + 0.2)',
'ax.set_xticklabels(categories, fontsize=12, fontweight="500")',
'for i, cat in enumerate(categories):',
)
```
Generally speaking, non-invasive recording methods such as fMRI, EEG, fNIRS, or MEG are incapable of single-neuron resolution. While non-invasive methods tend to have a comparably high penetration depth, i.e., recording across deeper areas of the brain (which matters in particular for large brains) and can record over hours, they lack spatial resolution and only provide aggregate (indirect) information of the neuronal activity. Invasive methods, however, can provide spatial resolution of individual neurons and extremely high sampling rates, i.e., recordings at the millisecond level. Since the first single-cell recordings in the 1950s, these methods have evolved into sophisticated patch clamping, high-density microelectrode arrays (MEAs), and calcium or voltage imaging systems that measure electrical activity of individual neurons. The number of neurons being simultaneously recorded – with electrophysiological methods like MEA and patch clamping or imaging modalities such as calcium imaging – has roughly doubled every 5 years ([Stevenson & Kording, 2011](https://www.nature.com/articles/nn.2731); [Urai et al., 2022](https://www.nature.com/articles/s41593-021-00980-9); [Mineault et al., 2024](https://arxiv.org/abs/2411.18526)), but has rapidly accelerated with the rise of optical techniques. Parallel recording of individual neurons over time is now possible for up to one million cells simultaneously, though at low temporal resolutions (sampling rate of about 1 Hz). Invasive methods not only require surgery with direct access to the brain, but also often genetic engineering to express reporters of neural activity. Electrophysiological arrays are often limited to the more superficial parts of the brain.
Another dimension in which recording modalities differ is their ability to capture the full scope of an organism’s behavioral repertoire. Most modalities work exclusively or best in non-moving or fixated organisms, substantially limiting the range of behaviors. Many non-invasive methods are prone to movement artifacts and are almost unusable beyond resting or head-fixed organisms. Some methods, for example, calcium imaging, can work in both head-fixed and freely behaving animals; however, head-fixing animals typically reduces motion artifacts and permits higher spatial and temporal resolution. Miniaturization of microscopes permits calcium imaging in behaving mice, however, the full behavior repertoire is still substantially limited, given the equipment attached to the organism.
While acknowledging the importance and briefly characterizing non-invasive recording modalities as well as electrophysiology studies in the following box, the method section as well as organism-specific reviews focus on modalities that – at least in theory – allow for whole brain recordings with single neuron resolution: Calcium imaging, voltage imaging and the nascent field of ultrasound currently in development.
While electric coupling via neurotransmitters is the primary way of communication for neurons, neuropeptides modulate neuronal activity and regulate long-term physiological processes, such as structural changes. The scale of signaling peptides is far from millions or billions of neurons; hundreds of neuropeptides could interact with their environment in different permutations. This adds an entirely new dimension to what needs to be captured to understand information processing in the brain. The totality of neuropeptides interacting with each other is sometimes called “chemical / peptidergic connectome” on top of the electric connectome, which is highly conserved across evolution. ([Jekley & Yuste 2024](https://www.sciencedirect.com/science/article/pii/S2352154624001074)). Unlike classical neurotransmitters, neuropeptides lack specific clearance mechanisms, allowing for sustained signaling that matches behaviorally-relevant timescales ([Guillaumin and Burdakov, 2021](https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2021.644313/full)).
At this point, it is not well understood what the relative contributions of neuropeptides to electrical activity and the overall information processing of the brain are. In C. elegans, for instance, where neuropeptide effects have been studied extensively ([Bhat et al., 2021](https://www.frontiersin.org/journals/molecular-neuroscience/articles/10.3389/fnmol.2021.786471/full)), a long list of behavioral defects can be demonstrated in mutants without the respective neuropeptide. A large variety of behaviors are regulated by neuropeptides in Drosophila,, including feeding, aggression, and sleep ([Nassel and Zandawala, 2022](https://royalsocietypublishing.org/doi/full/10.1098/rsob.220174)). One prominent example for humans is oxytocin, which has been convincingly demonstrated to influence complex bonding behaviors, e.g., in interactions of mothers and their newborn children ([Scatliffe et al., 2019](https://www.sciencedirect.com/science/article/pii/S2352013219303229?via%3Dihub)).
Lastly, combining recording modalities with methods that allow a deliberate and controlled manipulation or alteration of a specific part of a neural circuit (perturbation) is of exceptional value to understanding brain dynamics. By selectively activating or inhibiting specific neurons, researchers can directly test hypotheses and collect less correlated datasets for later computational modelling. Early studies focused on manipulating broad neural populations; modern approaches aim for increasingly precise control, from targeting anatomically defined regions to manipulating individual neurons.
In the following chapter, we survey major neural recording modalities for electrical activity and neuromodulators, including both invasive and non-invasive approaches, then cover perturbation methods, provide representative cost ranges, and discuss data management, standardization, and analysis.
## Electroencephalography (EEG)
When performed on humans, EEG records the brain’s electrical activity by placing electrodes on the scalp to detect voltage fluctuations produced by large populations of neurons, predominantly cortical pyramidal cells. The scalp-recorded signals in EEG are blurred by the skull and scalp, which limits spatial precision to roughly one centimeter or more. This resolution aggregates the activity of millions of neurons. EEG also predominantly captures signals from superficial cortical layers, as signals from deeper or subcortical structures attenuate significantly before reaching the scalp. Researchers can enhance spatial precision somewhat using high-density electrode arrays (64, 128, or more electrodes) and computational source localization methods. However, the reliance on scalp measurements makes it challenging to pinpoint activity at the resolution of individual cortical layers. Despite the poor spatial resolution, these signals allow researchers to sample data at rates of 250 Hz to over 1 kHz, i.e., well above the neuronal firing rates. In clinical or research settings, EEG can be collected continuously for hours or even days, enabling extended monitoring of conditions such as epilepsy.
Because EEG relies on electrodes attached to the scalp, participants can be seated upright or lying down, and some mobile systems allow moderate movement and study of real-world tasks. Despite this relative flexibility, excessive motion degrades data quality, and researchers must constrain movement. The most common disruptions come from eye movements and blinks (which can be 10-100 times larger than brain signals), muscle activity from jaw clenching or forehead movements, and electrical interference from nearby devices or power lines (showing up as a 50/60 Hz hum in the data). EEG systems are highly accessible, likely with tens of thousands of devices worldwide across research labs, hospitals, and increasingly, consumer applications. A research-grade system typically costs between \$20,000 and \$200,000 ([Ledwidge et al., 2018](https://pmc.ncbi.nlm.nih.gov/articles/PMC6312138/)). Operating costs are relatively modest, mainly involving electrode gel, cap maintenance, and technician time, typically ranging from \$50-200 per hour and relatively modest data sizes compared to imaging modalities, though high-density systems recording continuously can generate substantial datasets.
## Functional Magnetic Resonance Imaging (fMRI)
fMRIs use powerful magnets and radio waves to measure changes in blood oxygen levels (the BOLD signal) throughout the brain. This technique identifies regions with heightened neuronal activity by detecting differences in oxygen-rich and oxygen-poor blood: as neurons become more active, they consume more oxygen, and the subsequent increased blood flow to replenish this oxygen is what fMRI visualizes to map brain function. A key strength is non-invasive, whole-brain imaging. While typical fMRI studies often use voxel sizes on the order of a few millimeters, high-resolution protocols can achieve 1 mm³ whole-brain coverage ([de Martino et al., 2011](https://doi.org/10.1016/j.neuroimage.2011.05.008); [Heidemann et al., 2012](https://doi.org/10.1002/mrm.24156)), though this remains relatively uncommon. At 1 mm³, fMRI aggregates signals from tens of thousands of neurons, with the BOLD signal's spatial point-spread being a limiting factor at this scale ([Marblestone et al., 2013](https://doi.org/10.3389/fncom.2013.00137)). State-of-the-art scanners push spatial resolution further, with custom 7T systems achieving 0.56 mm isotropic whole-brain BOLD imaging ([Feinberg et al., 2023](https://www.nature.com/articles/s41592-023-02068-7)), and the 11.7T "Iseult" scanner reaching sub-0.5 mm "mesoscale resolution" for functional imaging ([Boulant et al., 2024](https://www.nature.com/articles/s41592-024-02472-7)). This approaches the spatial (though not temporal) resolution of functional Ultrasound Imaging (fUSI) at ~0.1 mm³. Even at these advanced fMRI resolutions, BOLD signals represent hundreds to thousands of aggregated neurons; single-neuron MRI remains far away without new contrast mechanisms ([Marblestone et al., 2013](https://doi.org/10.3389/fncom.2013.00137)). Temporally, the BOLD signal's intrinsic ~10-second rise and fall limits the effective sampling rate to ~0.1 Hz, despite scanners technically acquiring images every 1-3 seconds. As a consequence, to put this temporal resolution in context, a single brain image can be influenced by over 20 spoken words ([Tang et al., 2023](https://doi.org/10.1038/s41593-023-01304-9)).
Human fMRI across data repositories like OpenNeuro, DANDI, and Brain Image Library is typically on the order of 10-20h max with sessions of 1h each at ~1 mm³ resolution, with the most “intensive” datasets reaching 200 hours per subject ([Kupers et al., 2024](https://doi.org/10.1016/j.tins.2024.09.011), [Boyle et al., 2020](https://docs.cneuromod.ca/en/latest/_downloads/40f570bd627d600a2b26d7e6c56d331c/1939_BoylePinsard_OHBM2020.pdf)). Different groups generated multimodal datasets, such as fMRI and EEG ([Mayhew et al., 2013](https://www.sciencedirect.com/science/article/pii/S1053811913002255?via%3Dihub); [Pisauro et al., 2017](https://www.nature.com/articles/ncomms15808)), or fMRI-fNIRS-EEG ([Scarapicchia et al., 2017](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5563305/)). The key to good fMRI pictures is compliant participants. Motion correction and headcases compensate for subtle head movements during scanning, as even millimeter-scale shifts can contaminate the signal. Accordingly, fMRI is also limited in that it is restricted to analysis of supine, generally non-moving participants, meaning that movement and the complete behavior repertoire of organisms are challenging to study. Currently, thousands of these devices are used globally for research purposes. A state-of-the-art 7T machine can cost over \$10M ([Balchandani and Naidich, 2015](https://doi.org/10.3174/ajnr.A4180)). Operational costs are often quoted at approximately \$1000/hour ([Marek et al., 2022](https://www.nature.com/articles/s41586-022-04492-9)).
## Electrophysiology
Although invasive electrophysiological systems demonstrate remarkable and rapidly advancing capabilities for BCIs and fundamental neuroscience, their utility for whole-brain emulation is inherently limited. They provide exceptionally high-quality, high-temporal-resolution passive recordings of neuronal spiking activity. However, electrophysiological recordings are, and will likely remain, highly localized to the small brain volumes immediately surrounding the implanted electrodes. Indeed, scaling current electrophysiological approaches to achieve significant coverage of the human brain faces fundamental physical, biological and ethical hurdles. The sheer number of discrete implants required would be infeasible, and the associated tissue response to such widespread penetration would be prohibitive. Thus, these methods primarily offer sparse, high-fidelity data points rather than a pathway to comprehensive functional mapping of the entire human brain.
### Patch Clamp
Patch clamp recordings involve attaching a glass micropipette to a neuron's membrane to measure electrical currents directly through individual ion channels. This technique achieves the highest possible spatial resolution by recording from single neurons or even isolated membrane patches containing individual ion channels. The temporal resolution is exceptional, capturing signals at microsecond timescales (\>10 kHz sampling rates), allowing researchers to observe individual ion channels' opening and closing kinetics. However, patch clamp recordings, depending on the exact technique, are typically limited to minutes or hours for a single cell due to membrane destabilization and cell deterioration. In slice preparations or cell cultures, experienced researchers might maintain stable recordings from multiple sequential cells over an 8-12 hour experiment ([Mayer Jr and Brown, 1998](https://doi.org/10.1016/S0165-0270(98)00143-5)).
Patch clamp is highly invasive and requires precise micromanipulation under microscopic visualization, limiting its use mostly to in vitro preparations (brain slices, cell cultures). Patch clamp experiments have been done in vivo, but involve highly laborious and relatively low-throughput procedures. We estimate that a basic setup costs between \$50,000 and \$150,000 (e.g., ScienceDirect Topics, Patch-Clamp Technique [reports](https://www.sciencedirect.com/topics/immunology-and-microbiology/patch-clamp-technique?utm_source=chatgpt.com) ‘almost \$100 000 at current list prices’ ), with operational costs of approximately \$200-500 per day (see e.g. [the Ohio State University whole cell patch clamping service](https://medicine.osu.edu/departments/neuroscience/core-services/electrophysiology-core) at \$375/day) in consumables (micropipettes, solutions) and requiring highly skilled operators.
### Microelectrode Arrays (MEAs)
Microelectrode Arrays (MEAs) consist of recording electrodes embedded in a substrate to simultaneously record from multiple individual neurons. MEAs bridge the gap between single-cell and population-level recordings, typically capturing activity from dozens to thousands of neurons, depending on electrode density. Modern high-density MEAs can contain thousands of electrodes within a 1mm² area, approaching single-neuron resolution within local networks. MEAs sample neural activity at 10-50 kHz, capturing action potentials and local field potentials with millisecond precision. Unlike patch clamp, MEAs can maintain stable recordings for days to weeks in vitro, and implanted chronic arrays can sometimes function for months to years in vivo, enabling long-term studies of neural network dynamics and plasticity. Common challenges include signal quality decreases with motion, electrode impedance changes over time, and tissue inflammatory responses around electrodes. MEA systems are increasingly available in research settings, with perhaps thousands of systems globally. We estimate that equipment costs range from \$50,000 for basic in vitro systems to over \$250,000 for high-density or wireless in vivo systems. Operational costs include replacement arrays (\$500-2,000 each), surgical procedures for implantation (\$1,000-3,000 per animal), and data storage for the substantial volumes generated.
## Ultrasound
Ultrasound‐based neuroimaging is distinct from electromagnetic approaches in that it uses mechanical waves, brief “pings” of sound, to probe tissues. These waves travel through soft tissue at around 1,540 m/s and get partially scattered back, carrying information about local density and compressibility. Frequencies typically range from about 3 MHz (wavelength ~500 µm) to 25 MHz (wavelength ~60 µm), with attenuation in water-rich tissue scaling linearly ([Marblestone et al., 2013](https://doi.org/10.3389/fncom.2013.00137)). Bone attenuates ultrasound more severely, though this can be addressed through minimally invasive preparations such as acoustically transparent cranial windows.
Currently, most neuroscientific research involving ultrasound revolves around hemodynamic functional ultrasound imaging (fUSI), which measures changes in blood flow rather than neural activity ([Rabut et al., 2020](https://www.sciencedirect.com/science/article/pii/S0896627320307030)). Hemodynamic fUSI involves transmitting plane waves at multiple angles and detecting Doppler shifts from moving red blood cells, producing images that quantify local cerebral blood volume with spatial resolution around 100μm and temporal resolution of 400 ms, all while enjoying significant coverage. State-of-the-art probes are reportedly capable of imaging over 30% of the human brain by volume ([Forest Neurotech, 2025](https://forestneurotech.org/forest-1)). However, as discussed above with fMRI, the hemodynamic signal is ultimately a slow, indirect correlate of neural activity: the blood‐flow response emerges with a latency of around one second and is integrated over multiple neural events ([Aydin et al., 2020](https://www.nature.com/articles/s41467-020-16774-9)). Norman et al. achieved a 0.4 mm³ resolution with fUS for an 8 cm³ volume in monkeys ([Norman et al., 2021](https://www.cell.com/neuron/fulltext/S0896-6273(21)00151-3)).
A key development toward more direct ultrasonic measures of neural activity involves gas vesicles (GVs): air‐filled protein nanostructures that scatter ultrasound waves. These 2-nm-thick protein shells enclosing gas compartments can be genetically introduced into mammalian cells and engineered to respond to biological signals, making them promising “reporter genes” for functional ultrasound ([Shapiro et al., 2014](https://www.nature.com/articles/nnano.2014.32)). Recent developments include the engineering of GVs to respond dynamically to calcium transients, resulting in ultrasonic reporters of calcium (URoCs) – the first genetically encodable calcium indicators for ultrasound imaging ([Jin et al., 2023](https://doi.org/10.1101/2023.11.09.566364)). Significant challenges remain, including slow sensor kinetics (requiring 3.5 seconds to reach half-maximum signal upon calcium binding) and achieving robust expression in neurons. Further, research has only demonstrated resolutions of about 100 μm, insufficient to capture single neuron activity.
While the approaches described above leverage ultrasound as an imaging modality, another proposed ultrasound-based technique, "neural dust", uses ultrasound for power delivery and communication with implanted electrodes for direct neural recording ([Seo, 2018](https://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-146.html)). These miniature devices (“motes”) leverage piezoelectric materials so that when ultrasound waves from an external transducer reach these implants, some of the ultrasound’s energy powers the device while changes in electrical impedance caused by neural activity modulate how ultrasound is reflected back, creating a backscatter communication channel. This approach has been successfully demonstrated in peripheral nerves and muscles in rodents, achieving wireless recordings comparable to conventional wired systems ([Seo et al., 2016](https://doi.org/10.1016/j.neuron.2016.06.034)). In general, although still in its early stages, neural dust offers significant potential: its ultrasonic powering can enable extremely small (50-100 μm) implants, promising high-density, minimally invasive neural recording. However, many significant challenges remain, including delivering these microscopic motes to target tissues, achieving sufficient backscatter sensitivity for reliable signal detection, developing robust biocompatible encapsulation, and employing the sophisticated beamforming required to isolate signals from large numbers of motes.
In conclusion, ultrasound represents a modality with significant potential. For imaging approaches like fUSI and gas vesicles, at least theoretically, higher frequencies (60-100 MHz) could achieve single-neuron resolution (~15-25 μm) with significantly higher attenuation. Meanwhile, neural dust offers a complementary approach through distributed, miniature implantable sensors that can directly record neural activity with potentially less tissue displacement than traditional electrodes. Together, these ultrasound-based technologies could advance the frontier of neural recording by potentially enabling whole-brain recordings at cellular resolution in organisms as large as mice, while offering broader coverage in larger brains than is currently achievable with other techniques. However, substantial engineering challenges remain across all ultrasound approaches, and realizing the full potential of ultrasonic neural interfaces will require significant further research and development.
## Optical methods
Before we dive into two of the most prominent recording modalities (calcium and voltage imaging) we want to provide a quick primer on photon microscopy. Readers familiar with the systems can skip this section.
Traditional single-photon fluorescence microscopy relies on a fundamental quantum mechanical process. When a fluorescent molecule absorbs a photon of the appropriate energy (typically in the visible light range), an electron in this molecule is excited from its ground state to a higher energy level. After a brief period (nanoseconds), the electron returns to its ground state, releasing a photon with slightly less energy (and thus longer wavelength) than the excitation photon. This difference between excitation and emission wavelengths, known as the Stokes shift, allows us to separate the emitted fluorescence signal from the excitation light using optical filters. In the context of brain imaging, this process faces significant limitations. When visible light enters brain tissue, it encounters numerous cell membranes, protein structures, and other cellular components that can either absorb the light or change its direction (scattering). Both effects reduce the number of photons that reach the focal point and make it harder to collect the emitted fluorescence. Moreover, because single-photon excitation occurs wherever an excitation photon encounters a fluorescent molecule, out-of-focus fluorescence creates a background signal that reduces contrast and spatial resolution. While techniques like confocal microscopy can reduce this out-of-focus fluorescence using a pinhole, they cannot overcome the fundamental depth limitation imposed by tissue scattering, typically restricting 1P imaging to depths of 200-300 micrometers of tissue ([Xu et al., 2024](https://doi.org/10.1016/j.cell.2024.07.036)).
Two-photon microscopy, first demonstrated by Denk and colleagues in 1990 ([Denk et al., 1990](https://doi.org/10.1126/science.2321027)), revolutionized deep tissue imaging through an elegant quantum mechanical principle: instead of using one high-energy photon to excite a fluorescent molecule, it uses two lower-energy photons that arrive nearly simultaneously (within about 10⁻¹⁶ seconds). These lower-energy photons, typically in the infrared range (700-1040 nm), can penetrate tissue much more deeply than visible light because they experience less scattering. Critically, the probability of two photons arriving simultaneously is only high at the focal point of the microscope, creating natural optical sectioning and reducing photobleaching and photodamage in surrounding tissue. This allows imaging up to about 600-800 micrometers deep ([Xu et al., 2024](https://doi.org/10.1016/j.cell.2024.07.036)) in the cortex.
Three-photon microscopy extends these principles further by using three even lower-energy photons (typically \>1300 nm) to achieve excitation. This technique offers several advantages for deep imaging: the longer wavelengths experience even less scattering, and the requirement for three coincident photons provides better background rejection. Three-photon microscopy can reach depths of 1-1.3 millimeters ([Xu et al., 2024](https://doi.org/10.1016/j.cell.2024.07.036)), accessing structures like the hippocampus in mice that are largely inaccessible to two-photon imaging. However, three-photon systems require more expensive laser sources and typically operate at slower speeds due to the need for higher pulse energies; further, another consideration with three-photon systems is that three-photon imaging heats brain tissue more readily than two-photon systems.
Because two-photon and three-photon microscopy rely on point-scanning, increasing the number of recorded neurons requires distributing the available laser power across more points and reducing the dwell time per neuron to maintain temporal resolution. Analysis of publicly released datasets demonstrates this relationship clearly - as the number of simultaneously recorded neurons increases from hundreds to thousands, there is a systematic increase in noise levels unless temporal resolution is sacrificed ([Rupprecht, 2021](https://gcamp6f.com/2021/10/04/large-scale-calcium-imaging-noise-levels/)). This relationship holds across different imaging configurations and preparations, reflecting the inherent physical limits of current optical recording approaches.
### Calcium imaging
Calcium imaging uses genetically encoded proteins that light up when they bind to calcium. These fluorescent proteins, most commonly variants of GCaMP, are introduced into neurons through genetic engineering. When neurons fire, calcium rushes into the cell; GCaMP binds to calcium, resulting in a conformational change that permits fluorescence.
The preparation approach for introducing the GCaMP transgene varies significantly. Viral injection methods generally result in higher and more rapid expression, though this expression can be quite variable. For viral approaches, researchers perform surgery to inject viral vectors carrying the calcium indicator genes. After this injection surgery, there is a waiting period of days to multiple weeks for proper protein expression, with expression levels changing over time. In contrast, generating transgenic lines provides reliable, heritable expression but may require months to years to develop and validate. For in vivo imaging with either method, a surgical procedure to implant a clear window in the skull is also necessary to provide optical access to the brain.
The technique requires substantial infrastructure – a typical multiphoton microscope costs more than \$500,000 ([Holmes et al., 2022](https://doi.org/10.1017/S1551929522000657)) and comes with significant service and maintenance costs. Globally, we estimate the number of two-photon microscopes in active use for neuroscience research to likely be between a few hundreds and a few thousands. Additional costs include surgical supplies, viral vectors, and specialized habituation equipment. The price per hour falls into the range of \$20-\$100 ([NYU Langone Health, n.d.](https://med.nyu.edu/research/scientific-cores-shared-resources/microscopy-laboratory/fees), [Sunnybrook Research Institute, n.d.](https://sunnybrook.ca/research/content/?page=sri-core-multiphoton-laser-fees))
Calcium imaging operates in an interesting middle ground between fMRI and electrophysiology. The calcium indicator GCaMP7f, for example, has a half-rise time of about 60 milliseconds and half-decay time of about 150 milliseconds ([Dana et al., 2019](https://www.nature.com/articles/s41592-019-0435-6)), while the newer jGCaMP8f reaches half-rise times of about 2–7 ms and half-decay times of about 40 ms ([Zhang et al., 2023](https://www.nature.com/articles/s41586-023-05828-9)). Typical two-photon systems can image at 30 Hz for a single plane, though this drops to 1-5 Hz when imaging multiple planes to capture a volume. Recordings can last from hours in acute preparations to months in chronic experiments. Key factors limiting continuous imaging are photobleaching and photodamage (phototoxicity). Photobleaching, the irreversible loss of indicator fluorescence due to light exposure, is counteracted by cellular synthesis of new indicator proteins. Given typical fluorescent protein half-lives of approximately 24 hours ([Snapp, 2009](https://doi.org/10.1016/j.tcb.2009.08.002)), significant fluorescence recovery through this replenishment can occur over hours to days. Separate from signal loss, photodamage refers to light-induced cellular injury and physiological disruption, often mediated by reactive oxygen species or thermal effects, particularly with two-photon microscopy ([Grienberger et al., 2022](https://doi.org/10.1038/s43586-022-00147-1); [Icha et al., 2017](https://doi.org/10.1002/bies.201700003)). This damage can be subtle, affecting cellular processes before morphological changes are evident ([Icha et al., 2017](https://doi.org/10.1002/bies.201700003)). Since both photobleaching and photodamage restrict imaging duration and can compromise data integrity, careful optimization of light exposure is critical.
The spatial resolution of calcium imaging is sufficient for individual neurons and even their dendrites at approximately 0.5-1 μm resolution. A typical field of view might be 500x500 μm, containing hundreds of neurons. Standard two-photon microscopes can image up to 600-800 μm deep in the cortex, while three-photon systems can reach 1-1.3 mm ([Xu et al., 2024](https://doi.org/10.1016/j.cell.2024.07.036)). For advanced systems, this results in up to 1 mm³ volumes that can be imaged at single-cell resolution. While these capabilities allow for recording large neuronal populations, it's important to recall the inherent trade-offs in point-scanning systems between neuron count, signal quality, and temporal resolution, as discussed previously.
### Voltage Imaging
While calcium imaging indirectly measures neural activity through calcium transients, voltage imaging directly detects changes in membrane potential. This allows voltage imaging to capture the rapid dynamics of action potentials, spike timing, and subthreshold synaptic events, all of which remain invisible to calcium imaging ([Peterka et al., 2011](http://dx.doi.org/10.1016/j.neuron.2010.12.010)). These capabilities are also valuable for mapping dendritic computation, axonal propagation, and inhibitory/excitatory balance ([Kulkarni & Miller, 2017](https://doi.org/10.1021/acs.biochem.7b00490)), and position voltage imaging as complementary to both calcium imaging and electrophysiology. Similar to calcium imaging, both viral and transgenic approaches exist.
Voltage imaging relies on indicators that transduce membrane potential changes into optical signals through two primary approaches: voltage-sensitive dyes (VSDs) and genetically-encoded voltage indicators (GEVIs). VSDs are small organic molecules that often employ mechanisms like electrochromism, where the electric field directly shifts the dye's absorption or emission spectrum, and that exhibit fast kinetics (\<1 ms) and high sensitivity but lack cell-type specificity ([Aseyev et al., 2023](https://doi.org/10.3390/bios13060648)). GEVIs, in contrast, are fluorescent proteins engineered to be voltage-sensitive, typically using voltage-sensing domains from proteins like Ci-VSP or microbial rhodopsins, enabling genetic targeting to specific neuronal populations. However, GEVIs typically trade speed for sensitivity: most have response times of 2–10 ms, limiting single-trial action potential detection in vivo ([Kulkarni & Miller, 2017](https://doi.org/10.1021/acs.biochem.7b00490)). Both methods face challenges in signal-to-noise ratio (SNR) due to low photon counts at high acquisition rates (\>1 kHz) and phototoxicity.
Despite significant advances, voltage imaging still faces substantial challenges. The signal-to-noise ratio of voltage indicators generally remains problematic, with fluorescence changes of only 2-50% per 100 mV compared to calcium indicators' signals that can exceed 1000%. Moreover, voltage indicators are restricted to the membrane (since they necessarily must detect the membrane potential), which is a much smaller volume than the cytosol, and consequently, voltage imaging is often much dimmer than calcium imaging. These challenges create a tradeoff where faster indicators often require substantial excitation light, causing photodamage that restricts recording sessions to typically just minutes rather than the hours possible with calcium imaging. Furthermore, the membrane localization requirement makes cell segmentation difficult, as adjacent labeled neurons create overlapping "chicken wire" patterns instead of easily distinguishable volumes ([Kulkarni & Miller, 2017](https://doi.org/10.1021/acs.biochem.7b00490)). Generally, voltage imaging to whole brain coverage in mammals would require orders-of-magnitude improvements in sensor brightness, photostability, and multiplexed imaging systems ([Aseyev et al., 2023](https://doi.org/10.3390/bios13060648)). Nonetheless, voltage imaging has already proven invaluable for studying aspects of neural computation inaccessible to other techniques, and its continued development holds significant promise, particularly in organisms with relatively optically accessible brains.
### Neurotransmitter imaging
While the methods discussed above capture electrical activity, understanding brain function also requires monitoring the dynamics of chemical signaling. Neurotransmitters and neuromodulators orchestrate neural communication, from rapid synaptic transmission to slower, widespread modulatory effects. Genetically encoded fluorescent biosensors have emerged as powerful tools for visualizing these chemical signals in real-time, offering insights that complement neural recordings. These biosensors generally operate by fusing a specific ligand-binding domain – which recognizes a particular neurotransmitter or neuromodulator – to a fluorescent protein, often a circularly permuted fluorescent protein (cpFP). The binding of the target molecule induces a conformational change in the sensor, which in turn alters the fluorescence properties of the cpFP. This change can then be detected using standard microscopy techniques similar to those employed for calcium or voltage imaging. A key advantage of this approach is the ability for targeted expression in specific cell types or even subcellular compartments, typically achieved via viral vectors (such as AAVs) or through the creation of transgenic lines.
The detection of fast excitatory neurotransmission, particularly glutamate signaling, has been a significant focus. Early FRET-based sensors for glutamate, such as FLIPE, had limitations in signal-to-noise ratio (SNR) and kinetic performance, which restricted their application for studying rapid synaptic events ([Hao and Plested, 2022](https://doi.org/10.1016/j.jneumeth.2022.109531)). An important development in this area was the introduction of iGluSnFR (intensity-based glutamate-sensing fluorescent reporter) by Marvin and colleagues ([Marvin et al., 2013](https://doi.org/10.1038/nmeth.2333)). This sensor, which utilizes a glutamate-binding protein from E. coli (GltI) fused with a cpGFP, offered improved dynamic range and SNR, enabling the detection of glutamate transients with a temporal resolution of approximately 100 ms in various experimental settings. iGluSnFR has since been widely adopted for studying glutamate dynamics in contexts ranging from sensory processing and synaptic plasticity to investigations of pathological states ([Hao and Plested, 2022](https://doi.org/10.1016/j.jneumeth.2022.109531)).
The initial utility of iGluSnFR prompted the engineering of numerous variants with refined characteristics. These include versions with faster off-kinetics (e.g., iGluf, SF-iGluSnFR.S72A) to better resolve successive release events, improved brightness and stability through the use of superfolder fluorescent proteins (e.g., SF-iGluSnFR), and altered emission spectra (e.g., R-iGluSnFR1) to facilitate multicolor imaging ([Helassa et al., 2018](https://doi.org/10.7554/eLife.35174);[ Marvin et al., 2018](https://www.nature.com/articles/s41592-018-0110-1)). To enhance spatial precision, iGluSnFR has also been fused to synaptic proteins like Neurexin 1 for presynaptic targeting or Stargazin for postsynaptic localization ([Kim et al., 2020](https://doi.org/10.1016/j.neuron.2020.04.026)). These ongoing improvements have broadened the applicability of glutamate sensors, for example, in quantal analysis and disease modeling. Similar design principles have also been applied to create sensors for other fast neurotransmitters, including GABA (e.g., iGABASnFR) and acetylcholine (e.g., iAchSnFR) ([Marvin et al., 2019](http://doi.org/10.1038/s41592-019-0471-2);[ Jing et al., 2018](https://doi.org/10.1038/nbt.4184)).
For monitoring neuromodulators and neuropeptides, many of which signal through G-Protein Coupled Receptors (GPCRs), a common sensor design strategy involves engineering the native GPCRs themselves. This is typically achieved by inserting a cpFP into an intracellular loop of the target GPCR. The design aims to preserve the receptor's natural ligand affinity while blocking endogenous G-protein coupling, thereby preventing downstream signaling but allowing ligand binding to be transduced into a fluorescence change ([Girven et al., 2022](https://www.cell.com/trends/neurosciences/abstract/S0166-2236(22)00184-9?_returnURL=https://linkinghub.elsevier.com/retrieve/pii/S0166223622001849?showall=true)). This approach has yielded sensor families such as "GRAB" (GPCR Activation-Based) sensors and others like dLight for dopamine. Currently, validated sensors are available for approximately 12-15 different neuromodulators and neuropeptides, including dopamine (e.g., dLight, GRABDA), norepinephrine (e.g., GRABNE), serotonin (e.g., GRAB5-HT), acetylcholine (acting on muscarinic receptors, e.g., GRABACh), and various opioid peptides ([Muir et al., 2024](https://doi.org/10.1126/science.adn6671);[ Sun et al., 2018](https://doi.org/10.1016/j.cell.2018.06.042);[ Feng et al., 2019](https://doi.org/10.1016/j.neuron.2019.02.037)).
These GPCR-based sensors generally provide temporal resolution (1-10 Hz frame rates) and allow recording durations (hours to months) comparable to calcium imaging systems. For example, the dopamine sensor dLight1.2 has reported response times of approximately 9.5 ms for binding and 90 ms for unbinding ([Patriarchi et al., 2018](https://www.science.org/doi/10.1126/science.aat4422)). Spectrally distinct sensors permit simultaneous monitoring of multiple neuromodulators, such as green dLight1.3b for dopamine alongside red GRABNE2m for norepinephrine ([Muir et al., 2024](https://doi.org/10.1126/science.adn6671)). Such tools have enabled new observations, for instance, tracking dopamine transients with ~100 ms precision during reward-related behaviors ([Mohebi et al., 2019](https://www.nature.com/articles/s41586-019-1405-8)) and identifying selective opioid peptide release during specific physiological states ([Castro et al., 2022](https://pmc.ncbi.nlm.nih.gov/articles/PMC8858443/)). The imaging setups, data handling, and physical limitations for these biosensors are largely analogous to those for calcium imaging. While the range of available sensors continues to grow, expanding this toolkit to cover a broader array of neurochemicals and further enhancing sensor performance remain active areas of research.
## Invasive perturbation experiments
Under one of the most common evaluation criteria for brain emulations, the emulation should match the internal dynamics of the target brain, likely at least at the level of neural activity. Meeting this benchmark requires accurately modeling each neuron's input-output function – how its outputs depend on the various inputs it receives. While passive recordings constrain possible input-output functions, standard inputs and the fact that different circuit configurations can produce identical activity patterns make it challenging to gather enough data to uniquely identify each neuron's parameters ([Haspel et al., 2023](https://doi.org/10.48550/arXiv.2308.06578)). This fundamental limitation persists even with extensive recordings under diverse conditions, as the system's inherent complexity and feedback loops mean that many different parameter sets remain consistent with observed activity patterns ([Pospisil et al., 2024](https://www.nature.com/articles/s41586-024-07982-0)). This challenge of parameter identification from passive recordings alone, which exists even when the connectome is known, means that experimental, perturbation methods will also be key to generating accurate brain emulations.
In this discussion, we focus on perturbation techniques like optogenetic activation and silencing, as well as patch-clamp electrophysiology, since these permit temporally and spatially precise changes to neural activity. Other forms of perturbing neural activation, such as ablation experiments, are generally not discussed, since these techniques are often less useful for downstream computational modeling, as their exact influence on circuit activity is typically less precise and often irreversible, complicating the modeling of dynamic input-output functions.
Invasive perturbation methods provide a solution by breaking natural correlations between neurons and allowing direct manipulation of specific circuit elements. As established in causal inference theory, determining genuine causal relationships generally requires perturbations ([Woodward, 2004](https://doi.org/10.1093/0195155270.001.0001); [Pearl, 2009](https://doi.org/10.1017/CBO9780511803161)). In neural circuits specifically, establishing the causal influence of one neuron on another requires stimulating the former while recording from the latter ([Haspel et al., 2023](https://doi.org/10.48550/arXiv.2308.06578)). By precisely controlling individual neurons or groups of neurons while recording from others, experimenters can drive circuits into novel states that would not arise naturally. Data from such causal perturbations enables dramatically more sample-efficient fitting of neuronal input-output functions than passive recordings alone ([LaFosse et al., 2024](https://doi.org/10.1073/pnas.2318837121); [Wagenmaker et al., 2024](https://arxiv.org/abs/2412.02529)). These targeted interventions thus provide crucial constraints for reverse engineering neural computation.
Patch clamp recording, one of the oldest of these interventional techniques, provides exact control and measurement of individual neurons' electrical activity. By forming a tight seal with the cell membrane it enables direct manipulation of membrane potential while simultaneously recording with sufficient temporal resolution to detect individual action potentials and minute features of the membrane voltage dynamics. Dual patch recordings are required to establish causal relationships between neurons, stimulating one neuron while recording from another. This method produces the highest-quality data for determining input-output functions, with minimal off-target effects and exceptional signal fidelity. However, patch clamp recordings are technically demanding, especially in vivo, labor-intensive, and typically limited to short durations. Combined with the need for dual recordings, this makes it impractical for mapping large circuits.
Optogenetics enables precise temporal control of genetically defined neuron populations using light-sensitive ion channels or pumps ([Rost et al., 2022](https://www.nature.com/articles/s41593-022-01113-6)). This approach allows simultaneous manipulation of many neurons and can be combined with large-scale recording methods. Optogenetic approaches can be used to both activate and silence sets of neurons. Standard optogenetic approaches using fiber optics suffer from light scattering in tissue, limiting spatial resolution. However, methods like two-photon holography combined with soma-targeted opsins can achieve single-cell resolution in vivo, though spatial precision varies with the specific tools and light delivery systems used ([Adesnik and Abdeladim, 2021](https://www.nature.com/articles/s41593-021-00902-9)). While optogenetic stimulation offers millisecond-precise control, recording neural activity in response to these perturbations presents additional challenges. The most common recording method, calcium imaging, struggles to reliably detect individual spikes, particularly in fast-spiking neurons, despite ongoing improvements in indicators. Furthermore, optical crosstalk between stimulation and imaging wavelengths can interfere with simultaneous recording and perturbation.
Beyond detailed circuit mapping, it is important to note that optogenetic perturbation experiments linking neural population activity to behavior have the potential to contribute significantly towards embodied emulations, possibly more so than whole-brain imaging alone for certain objectives ([Cowley et al., 2024](https://www.nature.com/articles/s41586-024-07451-8)). Optogenetic activation behavioral data can be relatively cheap and fast to collect compared to some functional recording datasets, and can generate very useful, quantitative data ([Cande et al., 2018](https://elifesciences.org/articles/34275)). Furthermore, such perturbation experiments directly test causal relationships, avoiding the interpretational problems of messy correlations often found in neural recording data, and directly link neural activity to behavior.
Chemogenetics complements these approaches by offering sustained but less temporally precise control through engineered receptors (DREADDs) that respond to specific synthetic compounds ([Roth, 2016](https://doi.org/10.1016/j.neuron.2016.01.040); [Minamimoto et al., 2024](https://doi.org/10.2183/pjab.100.030)). After viral delivery of these receptors to targeted neuron populations, systemic drug administration can modulate neural activity over hours to days. While this approach enables manipulation of specific cell types and even specific neural projections through retrograde viral vectors, its utility for reverse engineering precise input-output functions is more limited. The slow kinetics of drug action and intracellular signaling cascades, combined with the sustained nature of the manipulation, make it impossible to probe the precise temporal interactions needed to characterize neuronal computation. Instead, chemogenetics is most valuable for behavioral studies and understanding the role of broader cell populations over longer timescales.
## Data management: storage, standardization, analysis
The advent of high-density neural recording technologies routinely generates terabyte datasets that were unimaginable a decade ago. This scale presents both opportunities and fundamental challenges for data management and analysis.
To illustrate – high-resolution scanning at 1 mm³ can generate about 60 GB of raw image data per hour of continuous scanning (brain dimensions: ~150mm × 180mm × 150mm = 4,050,000 voxels at 4 bytes over 1 hour at 1Hz is 16.2 MB × 3600 = 58.32 GB). High-density arrays recording at high sampling rates can produce 10-100 GB per hour of continuous recording. A typical two-hour calcium imaging session at 30 Hz can generate 500 GB to 2 TB of raw data, depending on the field of view and resolution ([Stringer et al, 2024](https://www.science.org/doi/10.1126/science.adp7429)). Raw data includes both the brain data and associated behavioral measurements.
Nowadays, functional recordings in neuroscience are often shared with the community via dedicated data repositories. The appendix provides a list of over 50 such repositories.
Because large scale neuroscience projects can generate huge data volumes, individual academic labs often do not make their raw data easily accessible, and instead state that the data is “available upon request” ([Tedersoo et al., 2021](https://www.nature.com/articles/s41597-021-00981-0)). Public, easily accessible data sharing has not been the default over the past two decades (although it has long been pioneered by dedicated organizations like the [Allen Institute](https://portal.brain-map.org/)) and is only slowly finding adoption, as some funders make it a requirement. Only recently have efforts like Neurodata Without Borders ([Rübel et al., 2022](https://doi.org/10.7554/eLife.78362)) started to standardize neurophysiology data, making them more comparable and interoperable. From private conversations with multiple experts, they assume that less than 5% of all the existing data is publicly available (in any form) at this point. In conversations with computational neuroscientists, the anecdotally reported access to multiple consistent and well-cleaned datasets is one of the main limiting factors to computational models making use of said data. This data barrier also prevents non-neuro-specialist engineers from other fields from getting involved and making meaningful contributions.
We can think about data repositories in 4 categories.
1. Preferred repositories: these comprise the majority of the well-formatted, easily accessible data: OpenNeuro, DANDI, Brain Image Library, EBrains, FigShare, Allen Institute, or Zenodo.
2. Single datasets: Study Forrest and FlyWire are examples of large datasets hosted on their dedicated websites.
3. Other repositories, such as CRCNS, Neurovault, and others, are often much less well-maintained and/or easily accessible.
4. Meta-repositories, registries, etc.: neuinfo does not host data and just has metadata scraped from elsewhere. OpenEphys predominantly lists other repositories, and the data it does link might just be repeats of previously mentioned repositories.
While some data repositories offer APIs facilitating mass data download, others may be impossible to scrape due to absent or scattered metadata; despite this, a repository's user-friendliness correlates positively with data volume, allowing access to significant portions of data across various modalities. It is unclear how much duplication exists across repositories. Heterogeneity of the data quality is another issue, with image data particularly vulnerable to quality issues like poor resolution.
The standardization of neurodynamic datasets presents unique challenges due to the extraordinary complexity of neural recordings across different temporal and spatial scales across different organisms. These technical challenges are compounded by sometimes significant methodological variations across laboratories and institutions. Different research groups often employ distinct preprocessing pipelines, quality control procedures, and analysis methods, making direct comparisons between datasets difficult. Even within the same recording modality, variations in experimental protocols, equipment configurations, and environmental conditions can introduce systematic differences that complicate data integration. The situation is further complicated by inconsistent documentation practices, with many datasets lacking crucial metadata about preprocessing steps, quality control measures, or experimental conditions.
The raw data distributed needs various post-processing steps. For calcium imaging, for instance following motion correction, cell identification and segmentation algorithms can process hours of recording in parallel. Real-time processing of 10,000 neurons requires a modern workstation and a high-end GPU. Scaling to 100,000 neurons demands at least 128GB RAM and multiple GPUs, with processing taking 3-5x real-time. At the extreme end, analyzing 1,000,000 neurons requires distributed computing systems with at least 1TB of RAM, processing at 10-20x real-time. ([Stringer et al 2024](https://www.science.org/doi/10.1126/science.adp7429)) The primary bottlenecks emerge in population analyses that require pairwise computations between neurons – for instance, a correlation matrix for 100,000 neurons requires 80GB of RAM just for storage. Scaling these analyses to even larger populations (millions to billions of neurons) would require fundamental algorithmic innovations. Current approaches often scale quadratically with neuron count, making them impractical for large populations.
Recently, Mineault et al. published an analysis of the “contents of DANDI, OpenNeuro, iEEG.org, as well as large-scale individual datasets” as part of their NeuroAI for AI Safety roadmap ([Mineault et al., 2024](https://doi.org/10.48550/arXiv.2411.18526)). We will quote them verbatim here.
Figure 18 - Availability of neural data to train a large-scale model (Reproduced from [Mineault et al, 2024](https://doi.org/10.48550/arXiv.2411.18526))
“The past decade has seen an explosion in the quantity of neural data freely available online. These public datasets represent a unique opportunity to learn good representations of neural data for a variety of downstream tasks, including brain-computer interfaces, clinical diagnoses for computational psychiatry and sleep disorders, and basic neuroscience.
Here we present a breakdown of the available data sources from an analysis of the contents of DANDI, OpenNeuro, iEEG.org, as well as large-scale individual datasets. Some of the highlights from this analysis include:
- There are around 100,000 hours of neural data available in freely accessible archives.
- There are roughly 3.3 million neuron-hours of single-neuron recordings from animals.
- The most abundant data type in terms of number of hours is intracortical EEG in humans–an invasive modality generated from the typically continuous, week-long recordings performed during epilepsy monitoring.
- Single neuron data is concentrated in a few datasets; the top 10 largest datasets in terms of neuron-hours account for more than 94% of total neuron-hours across all of DANDI. These come mostly from zebrafish and mouse, with one dataset from macaques.
- Large fMRI recordings are split into two categories: broad neuroimaging surveys, including HCP and UK Biobank, which scan many people for a short time; and intensive neuroimaging datasets, including Courtois Neuromod and the Natural Scenes Dataset, which scan few people for a very long time.”
### Machine-readable figure record: `main-fig-18`
```yaml
id: main-fig-18
document_id: sobe-2025-main-ai
document_path: report.md
anchor: main-fig-18
canonical_url: https://brainemulation.mxschons.com/ai/report.md#main-fig-18
figure: "Neural data availability"
assets:
- "../assets/report/main-fig-18a.png"
- "../assets/report/main-fig-18b.png"
- "../assets/report/main-fig-18c.png"
- "../assets/report/main-fig-18d.png"
data_status: unavailable_in_repository
publication_export_members:
- "images/image24.png"
- "images/image30.png"
- "images/image13.png"
- "images/image21.png"
canonical_pdf_pages:
- 124
note: "Four-panel external reproduction after Mineault et al. (2024)."
```
No local numeric plot input was found. The figure is a conceptual diagram, video still, or reproduction from an external publication; its caption and source remain the machine-readable provenance.