---
id: FOUND-013
title: "The Brain and Biological Neurons"
alternative_names:
  - "Neuroscience Foundations of AI"
  - "Biological Neural Computation"
category: "Foundations → Neuroscience"
volume: I
chapter: 4
entry_type: "Standard Article"
level: 3
status: "Draft"
last_reviewed: 2026-10-15
version: "0.1.0"
prerequisites:
  - "FOUND-002"
related_entries:
  - "FOUND-006"
  - "FOUND-014"
  - "TECH-NN-001"
  - "TECH-DL-001"
  - "HIST-003"
---

# The Brain and Biological Neurons

## 1. Concise Definition
The **brain** is the organ of the nervous system that processes sensory information, controls behaviour and supports cognition. Its basic signalling cells are **neurons**, which receive inputs through dendrites, integrate them in the cell body, and transmit electrical impulses (action potentials) along axons to other cells at junctions called synapses. Artificial neural networks were originally inspired by this organization, though they differ from biological circuits in many respects.

## 2. Overview & Significance
The brain is the only system generally agreed to exhibit general intelligence in the full human sense, and it has served as both inspiration and benchmark for AI. Neuroscience has contributed ideas such as distributed representation, hierarchical feature processing and learning by synaptic modification, while AI has in turn provided models used to interpret neural data (Hassabis et al., 2017; Richards et al., 2019).

The relationship should be stated carefully. Contemporary artificial neural networks ([[TECH-NN-001] Artificial Neural Networks](../vol03_technical/TECH-NN-001_artificial_neural_networks.md)) are not models of the brain in a detailed sense. Their units are simple weighted-sum functions, their training algorithm differs from known biological learning, and they are typically organized and trained quite differently from brain circuits. The relationship is better described as partial, selective inspiration, together with ongoing research on how much further the correspondence extends.

For AI forecasting and the study of general intelligence, the brain also provides reference quantities, such as cell counts and energy budgets, though mapping these to computational requirements is uncertain (Section 7).

## 3. Historical Development
McCulloch and Pitts (1943) proposed a logical calculus of nervous activity in which idealized neurons with threshold behaviour could implement logical operations, an early formal model that fed into the history of neural networks ([[HIST-003] Symbolic AI and the Perceptron](../vol02_history/HIST-003_symbolic_ai_perceptron.md)). Hodgkin and Huxley (1952) gave a quantitative description of membrane current and its application to conduction and excitation in nerve, a mathematical model of how action potentials arise from ion-channel dynamics. Hebb's proposal that connections strengthen when pre- and post-synaptic cells are active together became the conceptual basis for "Hebbian" learning rules.

Hubel and Wiesel (1962) described receptive fields and functional architecture in cat visual cortex, including neurons selective for oriented edges, findings that influenced hierarchical models of vision and convolutional architectures. Bi and Poo (1998) reported that the direction and magnitude of synaptic modification in cultured hippocampal neurons depended on the relative timing of pre- and post-synaptic spikes (spike-timing-dependent plasticity).

Quantification of brain composition was refined by Azevedo et al. (2009), who counted cells in adult human brains by isotropic fractionation and reported on average 86.1 ± 8.1 billion neurons and 84.6 ± 9.8 billion non-neuronal cells, revising the commonly quoted 100 billion neurons and ten-to-one ratio of glia to neurons. Herculano-Houzel (2009) reviewed these results and argued that the human brain is a linearly scaled-up primate brain in cellular composition. The standard neuroscience reference is Kandel et al. (2021).

More recent work connects deep learning and neuroscience in both directions. Yamins and DiCarlo (2016) reviewed how goal-driven deep networks provide models of sensory cortex. Lillicrap et al. (2020) reviewed whether backpropagation-like credit assignment could occur in the brain. Beniaguev, Segev and London (2021) examined the computational complexity of single neurons by fitting deep networks to simulated neuron models. Dorkenwald et al. (2024) published a wiring diagram of an adult fruit fly brain, a result in connectomics.

## 4. Technical / Conceptual Explanation

### 4.1 Neuron structure and signalling
A typical neuron has dendrites (receiving inputs), a soma (cell body), an axon (output), and synaptic terminals. The membrane potential changes as ion channels open and close; when depolarization crosses a threshold, a regenerative all-or-none **action potential** (spike) propagates along the axon. Hodgkin and Huxley (1952) described this process with coupled differential equations for voltage-dependent conductances:
$$C_m \frac{dV}{dt} = -\sum_k g_k(V,t)\,(V - E_k) + I_{\text{ext}},$$
where $V$ is membrane potential, $C_m$ capacitance, $g_k$ the conductance of ion species $k$ and $E_k$ its reversal potential. This model, developed from measurements on the squid giant axon, remains the template for biophysical neuron models.

### 4.2 Synapses and plasticity
Chemical synapses transmit signals by neurotransmitter release; each synapse has a variable efficacy ("weight"). **Plasticity** is the activity-dependent change of synaptic strength, regarded as a principal basis of learning and memory. Spike-timing-dependent plasticity (Bi & Poo, 1998) is one experimentally characterized form: whether a synapse strengthens or weakens depends on the order and interval between spikes. A simple Hebbian rule can be written $\Delta w_{ij} \propto x_i x_j$, a schematic of correlation-based learning, which by itself is unstable without normalization.

### 4.3 From neurons to artificial units
The standard artificial neuron computes $y = \phi\left(\sum_i w_i x_i + b\right)$, abstracting the neuron as a weighted sum followed by a nonlinearity. This echoes the threshold units of McCulloch and Pitts (1943). It omits dendritic computation, spike timing, neuromodulation, diverse cell types and dynamics on multiple timescales.

### 4.4 Large-scale organization
Cortex is organized into areas with hierarchical, recurrent connectivity; subcortical structures contribute to memory, action selection and regulation. According to Azevedo et al. (2009), only 19% of neurons are in the cerebral cortex despite its representing the large majority of brain mass. Sensory cortex shows hierarchical feature extraction (Hubel & Wiesel, 1962), a pattern reflected in layered artificial networks.

### 4.5 Energy and computation
Lennie (2003) analysed the energetic cost of cortical computation and argued that the metabolic cost of neural signalling limits the fraction of neurons that can be strongly active at any time, implying sparse activity. [VERIFICATION NEEDED: Lennie (2003) argument summarized from the paper's thesis without retrieving its text] Biological computation is energetically constrained in ways that differ from digital hardware ([[TECH-HW-001] AI Hardware](../vol03_technical/TECH-HW-001_ai_hardware.md)).

## 5. Types, Components & Classifications
- **Cell types**: excitatory and inhibitory neurons with many morphological and molecular subtypes; non-neuronal glial cells (Azevedo et al., 2009).
- **Synapses**: chemical and electrical; excitatory and inhibitory; with short- and long-term plasticity.
- **Brain regions**: cortex, hippocampus, cerebellum, basal ganglia, thalamus, brainstem, among others (Kandel et al., 2021).
- **Levels of description**: molecular, cellular, circuit, systems and behavioural; models differ in which level they capture.
- **Model classes**: biophysical (Hodgkin-Huxley type), simplified spiking, rate-based, and abstract artificial units.
- **Brain-inspired computing**: spiking neural networks and neuromorphic hardware (not covered by a source here).

## 6. Applications & Concrete Examples
- **Architectural inspiration.** Hierarchical visual processing (Hubel & Wiesel, 1962) is a historical motivation of convolutional networks ([[TECH-DL-001] Deep Learning](../vol03_technical/TECH-DL-001_deep_learning.md)).
- **Models of sensory cortex.** Yamins and DiCarlo (2016) review work in which goal-optimized deep networks predicted neural responses in visual cortex better than earlier models; this is a research result about specific visual areas, not a general claim that deep networks reproduce the brain.
- **Single-neuron computation.** Beniaguev et al. (2021) trained temporally convolutional networks to replicate the input/output behaviour of a detailed biophysical model of a cortical layer 5 pyramidal neuron, and reported that five to eight layers were needed, and that a much simpler network sufficed when NMDA receptors were removed from the model. The finding concerns a simulated neuron model, not recordings, so its implication for real neurons depends on the model's fidelity.
- **Connectomics.** Dorkenwald et al. (2024) report a neuronal wiring diagram of an adult fruit fly brain, the kind of data used to test hypotheses about circuit computation.
- **Neuroscience-inspired AI.** Hassabis et al. (2017) argue that neuroscience can inspire AI algorithm and architecture design and validate AI approaches; Richards et al. (2019) propose a framework in which neuroscience studies objective functions, learning rules and architectures.

## 7. Advantages, Limitations & Unresolved Questions
**Advantages of the brain as a reference.** An existence proof of general intelligence running on low power; a source of hypotheses about learning, memory and perception.

**Limitations of the analogy.**
- *Learning rules.* Backpropagation requires information flows (such as symmetric weights and separate forward and backward phases) not obviously present in the brain. Lillicrap et al. (2020) review proposals by which brain-like circuits might approximate backpropagation using neural activity differences; whether the brain implements such credit assignment is unresolved.
- *Unit complexity.* If single neurons perform computation as complex as the fitted networks in Beniaguev et al. (2021) suggest, counting neurons or synapses as simple units may understate biological complexity.
- *Quantifying computation.* There is no consensus on the number of operations the brain performs, so comparisons of brain "compute" with AI hardware are uncertain by a wide margin. Cell counts (Azevedo et al., 2009) do not determine computational capacity.
- *Substrate.* Whether intelligence or consciousness depends on biological properties beyond computation is a philosophical dispute ([[FOUND-005] Philosophy of Mind and the Turing Test](FOUND-005_philosophy_of_mind.md); [[FOUND-014] Consciousness and Machine Understanding](FOUND-014_consciousness_and_machine_understanding.md)).

**Open questions.** How the brain assigns credit across many layers; how much detail of dendritic and molecular processing is necessary for a faithful model; how memory is stored and consolidated; and what organizational principles produce general-purpose learning.

## 8. Current State of Knowledge
As of 2026-10-15:
- The structure of neurons, action-potential generation and chemical synaptic transmission are **Established** (Hodgkin & Huxley, 1952; Kandel et al., 2021).
- Revised cell counts for the human brain (Azevedo et al., 2009) are **Well Supported** within the method's limits; the figures are for adult male brains in that study.
- Spike-timing-dependent plasticity is **Well Supported** as a phenomenon in studied preparations (Bi & Poo, 1998); its role in learning in vivo is less certain.
- Goal-driven deep networks as models of sensory cortex are **Well Supported** for visual areas studied (Yamins & DiCarlo, 2016), while their status as mechanistic explanations is debated.
- Whether the brain implements backpropagation-like learning is **Contested** (Lillicrap et al., 2020).
- Complete connectomes exist for small animals such as the fruit fly (Dorkenwald et al., 2024); no comparable map exists for a human brain as of this review. [VERIFICATION NEEDED: status of mouse and human large-scale connectome projects as of 2026-10]

## 9. Future Trajectory & Strategic Outlook
No registered forecast is made. Observed directions: larger connectomes and the use of AI to analyse them; models linking dendritic computation to network function; and continued use of AI systems as models of neural responses. Speculative proposals, such as whole-brain emulation, are hypothetical scenarios, not engineering programs with demonstrated feasibility, and are addressed in [[PHIL-CONSC-001] Machine Consciousness](../vol11_philosophy/PHIL-CONSC-001_machine_consciousness.md) and the forecasting volume.

## 10. Related Entries & Knowledge Graph
- **Foundational Concepts**:
  - [[FOUND-002] Definitions of Intelligence](FOUND-002_definitions_of_intelligence.md): what the brain is taken to exemplify.
  - [[FOUND-006] Collective and Animal Intelligence](FOUND-006_collective_and_animal_intelligence.md): cognition in other species.
- **Related Methods & Technologies**:
  - [[TECH-NN-001] Artificial Neural Networks](../vol03_technical/TECH-NN-001_artificial_neural_networks.md): the engineering abstraction.
  - [[TECH-DL-001] Deep Learning](../vol03_technical/TECH-DL-001_deep_learning.md): deep networks and their brain-inspired roots.
  - [[HIST-003] Symbolic AI and the Perceptron](../vol02_history/HIST-003_symbolic_ai_perceptron.md): early neural models.
  - [[FOUND-010] Optimization and Gradient Descent](FOUND-010_optimization_and_gradient_descent.md): the learning algorithm compared with biological plasticity.
- **Advanced Implications & Future Trajectories**:
  - [[FOUND-014] Consciousness and Machine Understanding](FOUND-014_consciousness_and_machine_understanding.md): neural correlates and machine minds.
  - [[PHIL-CONSC-001] Machine Consciousness](../vol11_philosophy/PHIL-CONSC-001_machine_consciousness.md): substrate questions.

## 11. References & Foundational Literature
1. McCulloch, W. S., & Pitts, W. (1943). A logical calculus of the ideas immanent in nervous activity. *The Bulletin of Mathematical Biophysics*, 5(4), 115-133. https://doi.org/10.1007/BF02478259. [Early formal model of neurons as logical threshold units. Checked against Crossref.]
2. Hodgkin, A. L., & Huxley, A. F. (1952). A quantitative description of membrane current and its application to conduction and excitation in nerve. *The Journal of Physiology*, 117(4), 500-544. https://doi.org/10.1113/jphysiol.1952.sp004764. [Biophysical model of the action potential. Checked against Crossref.]
3. Hubel, D. H., & Wiesel, T. N. (1962). Receptive fields, binocular interaction and functional architecture in the cat's visual cortex. *The Journal of Physiology*, 160(1), 106-154. https://doi.org/10.1113/jphysiol.1962.sp006837. [Receptive-field organization in visual cortex. Checked against Crossref.]
4. Bi, G.-q., & Poo, M.-m. (1998). Synaptic modifications in cultured hippocampal neurons: Dependence on spike timing, synaptic strength, and postsynaptic cell type. *The Journal of Neuroscience*, 18(24), 10464-10472. https://doi.org/10.1523/JNEUROSCI.18-24-10464.1998. [Spike-timing-dependent plasticity. Checked against Crossref.]
5. Lennie, P. (2003). The cost of cortical computation. *Current Biology*, 13(6), 493-497. https://doi.org/10.1016/S0960-9822(03)00135-0. [Energetic constraints on cortical activity. Bibliographic record checked against Crossref; abstract not retrieved, so the stated argument is from author knowledge.]
6. Azevedo, F. A. C., Carvalho, L. R. B., Grinberg, L. T., et al. (2009). Equal numbers of neuronal and nonneuronal cells make the human brain an isometrically scaled-up primate brain. *Journal of Comparative Neurology*, 513(5), 532-541. https://doi.org/10.1002/cne.21974. [86.1 ± 8.1 billion neurons; 19% in cerebral cortex. Checked against Crossref abstract; Crossref lists nine authors.]
7. Herculano-Houzel, S. (2009). The human brain in numbers: A linearly scaled-up primate brain. *Frontiers in Human Neuroscience*, 3, Article 31. https://doi.org/10.3389/neuro.09.031.2009. [Review of brain cellular composition across primates. Checked against Crossref and Europe PMC abstract; article number from DOI.]
8. Yamins, D. L. K., & DiCarlo, J. J. (2016). Using goal-driven deep learning models to understand sensory cortex. *Nature Neuroscience*, 19(3), 356-365. https://doi.org/10.1038/nn.4244. [Review of deep networks as models of sensory cortex. Checked against Crossref.]
9. Hassabis, D., Kumaran, D., Summerfield, C., & Botvinick, M. (2017). Neuroscience-inspired artificial intelligence. *Neuron*, 95(2), 245-258. https://doi.org/10.1016/j.neuron.2017.06.011. [Perspective on the interplay of neuroscience and AI. Checked against Crossref.]
10. Richards, B. A., et al. (2019). A deep learning framework for neuroscience. *Nature Neuroscience*, 22(11), 1761-1770. https://doi.org/10.1038/s41593-019-0520-2. [Proposes studying objective functions, learning rules and architectures. Checked against Crossref (32 listed authors).]
11. Lillicrap, T. P., Santoro, A., Marris, L., Akerman, C. J., & Hinton, G. (2020). Backpropagation and the brain. *Nature Reviews Neuroscience*, 21(6), 335-346. https://doi.org/10.1038/s41583-020-0277-3. [Review of whether and how the brain could approximate backpropagation. Five authors confirmed on Crossref.]
12. Beniaguev, D., Segev, I., & London, M. (2021). Single cortical neurons as deep artificial neural networks. *Neuron*, 109(17), 2727-2739.e3. https://doi.org/10.1016/j.neuron.2021.07.002. [Deep networks fitted to a biophysical neuron model; five to eight layers needed. Checked against Europe PMC abstract.]
13. Dorkenwald, S., et al. (2024). Neuronal wiring diagram of an adult brain. *Nature*, 634(8032), 124-138. https://doi.org/10.1038/s41586-024-07558-y. [Connectome of an adult fruit fly brain. Checked against Crossref (294 listed authors, including consortium members).]
14. Kandel, E. R., Koester, J. D., Mack, S. H., & Siegelbaum, S. A. (Eds.). (2021). *Principles of Neural Science* (6th ed.). McGraw-Hill. ISBN 978-1259642234. [Standard neuroscience textbook. ISBN checked against Open Library (editors from author knowledge, Open Library lists the first three).]

## 12. Evidential Classification Matrix
| Section / Claim | Evidence Category | Supporting Source | Confidence / Status |
| :--- | :--- | :--- | :--- |
| Action potential mechanism and synaptic transmission | Established | Hodgkin & Huxley (1952); Kandel et al. (2021) | High |
| Human brain: about 86 billion neurons, 19% in cortex (adult male sample) | Well Supported | Azevedo et al. (2009) | Moderate to High |
| Hierarchical receptive-field organization in visual cortex | Established | Hubel & Wiesel (1962) | High |
| Spike-timing-dependent plasticity in cultured neurons | Well Supported | Bi & Poo (1998) | Moderate to High |
| A simulated L5 neuron's I/O needs a five-to-eight-layer network | Preliminary | Beniaguev et al. (2021) | Moderate; simulation-based |
| Goal-driven deep nets predict visual cortex responses | Well Supported | Yamins & DiCarlo (2016) | Moderate |
| The brain implements backpropagation-like credit assignment | Contested | Lillicrap et al. (2020) | Low to Moderate |
| Whole-brain emulation is feasible | Hypothetical | None cited | Low |

## 13. Revision History
- **v0.1.0** (2026-10-15): Initial draft. Citations checked against arXiv/Crossref/publisher pages where marked; not independently fact-checked.