---
title: "Part I: Foundations"
document_id: sobe-2025-main-ai
document_type: report_section
parent_document: ../report.md
section_id: main-section-05-part-i-foundations
section_order: 5
section_count: 17
language: en
license: CC-BY-4.0
---
# Part I: Foundations
> Selective-retrieval section 5 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-05-part-i-foundations`
- **Order:** 5 of 17
- **Words:** 2324
- **Previous:** [Technical Overview](04-technical-overview.md)
- **Next:** [Part 2: State of Brain Emulation across Organisms — C. elegans](06-part-2-state-of-brain-emulation-across-organisms-c-elegans.md)
## Introduction
In 2008, seventeen years prior to the publication of this report, Sandberg and Bostrom laid out an ambitious vision for whole‑brain emulation, a concept then largely theoretical. The intervening years have seen substantial scientific and technological developments that now justify a careful reexamination of progress toward this objective.
The motivations for undertaking this complex, multi-stage process are diverse. For many scientists, the payoff is basic insight: a faithful model would provide unprecedented understanding of how perception, memory, and decision-making emerge from neural circuits under controlled conditions. Clinicians see a rapid-testing platform for neurological and psychiatric interventions, cutting the time and cost of in vivo trials. Others anticipate a path to AI systems whose native architecture mirrors our own, potentially easing the problem of aligning machine behaviour with human values.
Experts hold even more disparate views on what will ultimately be required for faithful brain emulation. Will it demand neuron-by-neuron reconstruction, or must we descend to individual molecules and their conformational states? Should the scope encompass the entire body and nervous system, or can we isolate the brain, or even specific regions like the cortex or cerebellum? Numerous factors beyond neurons and synapses could prove necessary: glial cells, neuromodulatory peptides, hormones, ion concentrations, gap junctions, synaptic plasticity rules, and perhaps even the diffusion dynamics of signalling molecules. The relative importance of each remains uncertain, and this report does not attempt to cover every conceivable emulation possibility. Ultimately, determining which variables are essential is a strictly empirical question that must be resolved through the interplay between modeling and experimentation.
Here, with the input of countless experts in the field, we provide a detailed assessment organized around three fundamental capabilities: recording brain function, mapping brain structure, and emulation of both on computers. We focus on the ingredients most commonly identified as necessary: electrical activity at single-neuron resolution, synaptic connectivity, molecular annotations that distinguish cell types and synaptic properties, and the computational frameworks needed to integrate them. Our analysis concentrates on five commonly studied organisms, representing the primary systems that have driven most research relevant to whole-brain emulation.
In the following we present organism-level investigations across the end-to-end brain emulation pipeline (see figure AbroadStark2Gladiator) first, followed by detailed methodological chapters on neural dynamics, connectomics, and computational neuroscience. The upcoming paragraphs and definitions aim to build a shared terminology we can leverage across the report. Readers new to the field may prefer to read the Glossary and the methodological chapters before turning to the organism sections.
To emulate a brain, researchers must first understand its neural dynamics: the patterns of activity that emerge as organisms perceive, decide, and act. This functional mapping occurs in vivo, using recording methods that span a wide range of scales and invasiveness. Non-invasive techniques like fMRI and EEG can monitor whole-brain indicators of activity but lack the spatial and temporal precision needed to resolve individual neurons or single spikes. While valuable for human neuroscience and clinical applications, these approaches do not provide the detail necessary for neuron-scale emulation and thus receive limited coverage in this report. For sufficient resolution, researchers turn to invasive techniques. Optical methods track activity in thousands of neurons simultaneously using fluorescent indicators and microscopy, while electrophysiological approaches insert electrode arrays to capture precise electrical signals from hundreds to thousands of cells. Beyond passive recording, perturbation methods like optogenetics allow researchers to selectively activate or silence specific neurons while observing network responses, enabling causal rather than merely correlational mapping of circuit function. These measurements fulfill two indispensable functions for emulation. First, they supply the constraints needed for fitting parameters of computational brain models: providing the firing rates, temporal patterns, and synaptic properties that connectivity alone cannot reveal. Second, they establish validation criteria: once model parameters have been fitted, the same recordings provide quantitative benchmarks that brain emulations must match to demonstrate biological fidelity.
Once functional data has been collected in vivo, the next step is mapping the brain's physical architecture ex vivo. This is the domain of connectomics, which aims to reconstruct the organism's complete neural wiring diagram, and potentially its molecular composition, at synaptic resolution. Today this is an inherently destructive process conducted post-mortem. The brain is first chemically fixed to preserve its structure, then extracted and sliced into ultrathin sections, often just tens of nanometers thick. These sections are systematically imaged using techniques like electron microscopy (EM), which provides the resolution necessary to visualize individual synapses and subcellular structures. The resulting massive image datasets are computationally processed through AI-driven algorithms that align sections and trace neuron boundaries, followed by extensive human proofreading to correct tracing errors. However, EM-based pipelines face two major limitations: proofreading is labor-intensive and expensive, and EM captures minimal molecular information about neurotransmitter identities, synaptic receptor types, ion channel distributions, or the protein markers that define cell classes. These molecular features determine functional properties like synaptic strength and neural excitability, which are likely necessary for accurate model parameterization. These limitations have motivated the development of alternative approaches. X-ray microscopy can image much thicker tissue sections, which reduces the number of slices and simplifies reconstruction, though it remains an emerging technology. Expansion microscopy (ExM) has reached greater maturity and addresses both problems: it physically expands tissue to enable light-based imaging at synaptic resolution while still allowing molecular labeling techniques that can identify properties of interest. Furthermore, molecular labeling combined with genetic barcoding techniques assign and read out unique identifiers to neurons, allowing computational algorithms to match disconnected fragments and massively reduce manual proofreading.
With structural and functional data in hand, computational neuroscience seeks to instantiate accurate brain models in silico. This modelling work happens entirely in the computer, using previously acquired experimental data as constraints. Neurons can be modelled digitally with varying complexity, from simple integrate-and-fire units to Hodgkin-Huxley formulations that explicitly track ion-channel kinetics. Synapses range from static weights to activity-dependent plasticity rules that strengthen or weaken according to spike history. All these characteristics are abstracted into mathematical expression, consisting of countless variables, or parameters, per neuron and synapse. When functional recordings are available for the target organism, parameters are tuned to reproduce observed firing rates, spike timing, and perturbation responses. When such recordings are absent, parameters are inferred only indirectly, using proxy constraints such as structural connectivity, cell-type specific statistics, and functional measurements from related organisms, which still leaves many degrees of freedom unconstrained. The resulting network can then be evaluated in two ways: first, against held-out neural activity to test predictive accuracy; second, through embodiment in a simulated or robotic environment where sensory inputs drive the model and motor outputs are compared to the original behaviour. Discrepancies at either level reveal gaps in the model, whether missing structural details, insufficient functional constraints, or inappropriate modeling choices. This feedback is what makes emulation the integrative test of the entire pipeline: it identifies where data collection must be refined or expanded, closing the loop by directing future experimental work toward the measurements that are most likely to improve the next generation of emulations.
------------------------------------------------------------------------
Figure 6 - Overview of steps required for brain emulation. Using the complete toolkit of neural dynamics to measure variables influencing brain activity while the organism is still alive (1), then destructive reconstruction by connectomic methods to map the brain’s structure (2), and finally computational neuroscience approaches to use all of the data to create an emulation (3). The computational model predicts electrical neuronal activity constrained by reconstructed connections between neurons and potentially molecular substructure (e.g., individual transmitters and proteins). During embodiment, neural activity is decoded in behavior such as language or movement, and information from the environment, such as light and sound, as well as local or body-wide signal molecules like hormones or neuropeptides, encoded back into neural activity.
### Machine-readable figure record: `main-fig-06`
```yaml
id: main-fig-06
document_id: sobe-2025-main-ai
document_path: report.md
anchor: main-fig-06
canonical_url: https://brainemulation.mxschons.com/ai/report.md#main-fig-06
figure: "Brain emulation pipeline"
assets:
- "../assets/current-repository/hand-drawn/brain-emulation-pipeline-overview.svg"
data_status: unavailable_in_repository
publication_export_members:
- "images/image18.jpg"
canonical_pdf_pages:
- 26
note: "Semantic SVG is available; no numeric plot input exists."
```
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.
------------------------------------------------------------------------
## Definitions
We provide a full glossary at the end of the report. But for a few terms and concepts we want to add clarifications for all readers.
Specifically, the terms ‘simulation’ and ‘emulation’ are often used interchangeably. In this report, however, we draw a sharp distinction, adopting the formal framework established in 2008 by Sandberg and Bostrom ([Sandberg and Bostrom, 2008](https://ora.ox.ac.uk/objects/uuid:a6880196-34c7-47a0-80f1-74d32ab98788)). In particular, we consider a model a simulation if, given the same inputs and initial conditions, it matches a target system’s outputs without necessarily reproducing the internal causal dynamics that produced them at a chosen level of biophysical detail. We consider a model an emulation if it matches the target system’s outputs by implementing the same internal causal dynamics at a chosen level of biophysical detail, ensuring that it behaves like the original for the same underlying reasons.
The practical importance of this distinction is well-illustrated by considering a large language model trained on an individual's diary. Such a model functions as a simulation: its objective is to learn and reproduce the statistical patterns of the source text, enabling it to convincingly replicate the author's writing style and expressed views. It achieves high performance on this reference dataset (in-distribution behavior) but is likely to fail when generalizing to novel scenarios because it has not captured the causal structure of the brain that produced the text. This is due to a fundamental information bottleneck: the diary, as a low-bandwidth output, contains far less information than is required to uniquely specify the complex internal state of the brain that wrote it ([Sandberg, 2012](https://www.aleph.se/andart/archives/2012/04/how_many_persons_can_there_be_brain_reconstruction_and_big_numbers.html)). As a consequence, when prompted about a childhood memory not mentioned in the diary, the simulation would most likely generate a plausible fabrication, whereas a true emulation would access the causal memory trace itself, or report its absence.
While the conceptual distinction is clear, operationalizing it requires specific criteria. To this end, we introduce the term minimal brain emulation to define the minimum properties or “biophysical components” we consider necessary for a model to be classified as an emulation rather than a simulation. It is crucial to note that meeting this baseline does not, by itself, say anything about whether an emulation is accurate. Accuracy is a separate, quantitative measure of how well a model’s predictions match a biological reference. Instead, this baseline establishes the necessary, but not sufficient, foundation for a model’s architecture to even be considered a candidate for achieving high-fidelity emulation. The specific criteria that constitute a minimal brain emulation based on our minimal brain emulation model components are detailed in the table below. The core properties are:
- Approximately covering all neurons in the organism’s brain
- Based on accurate synaptic-level connectome
- Modelling at least some cell type diversity, e.g., at the neuronal subclass level (such as Pvalb, Sst, Vip inhibitory subclasses, or differences between cortical layers)
- Modelling at least at the level of point neurons
- Consistent with the scale of neuronal spiking activity.
Finally, we wish to note that the term ‘whole brain’ lacks a standardized definition and is applied inconsistently across the literature. To ensure precision within this report, we adopt a quantitative threshold: a model is considered 'whole brain' if it incorporates at least 95% of neurons and at least 95% of brain volume for the respective organism.
### Table 3 — Scales for Defining Brain Emulations
Table 3 - Scales for Defining Brain Emulations: A) Dimensions of brain representations spanning from pure simulation to full emulation. The requirements in bold constitute “minimal brain emulation components” as defined by the authors. B) ) Observable Outputs of Simulations / Emulations.
| | | | | | | | |
|----|----|----|----|----|----|----|----|
| A\) | Connectivity accuracy | Percentage of neurons | Cell types | Plasticity | Neuronal Accuracy | Neuromodulation | Temporal resolution |
| Pure simulation | Artificial (no real connectivity data, using random or artificial connections) | Small circuit – substantially less than a brain area. | No cell types (only one neuron type, e.g., like in typical ANNs). | Completely static network structure and synaptic parameters. | \- | None (no neuromodulation signaling systems modeled) | No concept of time (ANNs) or very coarse representation of time (like “before” and “after”). |
| | Simple statistical rules (inspired by a few biological observations) | A whole brain area or multiple interacting brain areas. | Most basic cell types, e.g., excitatory vs. inhibitory neurons. | Some limited plasticity, such as short-term plasticity. | Neuronal networks (ANNs with units like ReLU, firing rate units, no spiking) | Single modulator (single neuromodulation signaling system, like dopamine only) | Slow relative to neuronal spiking activity (hundreds of milliseconds to seconds and more). |
| | Complex statistical rules (from systematic biological studies and/or partial connectomic information) | ~Whole brain | Some cell type diversity, e.g., at the neuronal subclass level (such as Pvalb, Sst, Vip inhibitory subclasses, or differences between cortical layers). | More extensive plasticity (e.g., STDP). | Point Neurons (LIF) | Multiple modulators (interacting neuromodulation signaling systems) | Consistent with the scale of neuronal spiking activity. |
| Minimal brain emulation | Accurate connectome (complete, including verified synaptic connections) | ~Whole body | Diverse cell types, possibly including non-neuronal cells. | Full dynamic plasticity, including growth and pruning of connections. | Neuronal Compartments | Complex system (interacting neuromodulation systems with feedback) | Fast (microseconds or faster, capturing details of fast subcellular processes). |
| | | | | |
|----|----|----|----|----|
| B\) | Behavior Complexity | Personality- defining characteristics | In / Out-of-distribution | Learning horizon |
| No representation | None (no emergent behaviors modeled) | None (no personality traits represented) | No behavior | No Learning |
| | Simple outputs (basic stimulus-response patterns) | Basic traits (emerging from simple rules) | Limited in distribution behavior | Short term (seconds) |
| | Feedback loops (simple feedback between different neural systems) | Scattered traits (emerging from network interactions) | Full in-distribution behavior | Medium Term (days) |
| Complete representation | Complex behaviors (emerging from system interactions) | Complete profile (personality profile emerging from neural dynamics) | Out-of-distribution behavior | Long-term (years) |