Neural Network Is a Brain
metaphor
Source: Biology → Artificial Intelligence
Categories: ai-discoursecognitive-science
Transfers
The foundational metaphor of artificial intelligence. McCulloch and Pitts proposed the first mathematical model of a “neuron” in 1943, mapping biological neural architecture onto logical threshold units. Eighty years later, the vocabulary remains: we speak of neurons, layers, activation functions, synaptic weights, and network architectures as if describing a biological organ. The metaphor is so deeply embedded that most practitioners do not experience it as metaphorical at all.
Key structural parallels:
- Neurons as computational units — biological neurons receive electrochemical signals, integrate them, and fire when a threshold is reached. Artificial neurons receive numerical inputs, compute a weighted sum, and pass the result through an activation function. The mapping is clean enough to have generated an entire field, but the correspondence is structural, not mechanistic. Biological neurons are complex cells with thousands of synaptic connections, internal state, and temporal dynamics. Artificial neurons are single mathematical operations.
- Layers as hierarchy — the brain processes information through layered regions (retina to V1 to V2 to inferotemporal cortex for vision). Deep neural networks stack layers of artificial neurons. The metaphor imports the idea that intelligence arises from hierarchical abstraction: early layers detect edges, later layers detect faces. This assumption shaped the deep learning revolution.
- Activation as signal — biological neurons either fire or do not (approximately). Activation functions in neural networks threshold the output of each unit. The metaphor brings the intuition that computation happens through selective response — not every unit contributes to every output, just as not every neuron fires for every stimulus.
- Networks as emergent intelligence — the brain’s intelligence arises not from individual neurons but from the pattern of connections between billions of them. The metaphor imports the hypothesis that intelligence is an emergent property of scale and connectivity, which directly motivates the “scaling laws” approach to AI: more parameters, more data, more capability.
Limits
- Biological neurons are not matrix multiplications — a real neuron has roughly 7,000 synaptic connections, processes signals with precise timing, releases dozens of different neurotransmitters, and modifies its own structure in response to activity. An artificial neuron multiplies a vector by a weight matrix and applies a nonlinear function. The gap between these two operations is not a matter of degree but of kind. The brain metaphor makes this gap invisible, encouraging the inference that artificial neural networks work “like brains do” when they work like linear algebra does.
- Brains do not backpropagate — the core training algorithm for neural networks (backpropagation) has no known biological analogue. Real brains do not compute global error gradients and propagate them backward through synaptic connections. The metaphor quietly drops this fact: the most important thing about how artificial neural networks learn has nothing to do with how brains learn.
- The anthropomorphism cascade — because the foundational metaphor says “this is a brain,” every subsequent metaphor inherits the anthropomorphism. If the network is a brain, then its parameters are “knowledge,” its outputs are “understanding,” its errors are “hallucinations,” and its constraints are “alignment.” Each of these downstream metaphors derives its plausibility from the brain metaphor’s initial claim. Questioning any one of them means questioning the foundation.
- Brains are embodied; networks are not — a biological brain exists in a body, receives sensory input from a physical environment, and developed through millions of years of evolutionary pressure to keep that body alive. Neural networks have no body, no sensory apparatus, no evolutionary history, and no survival imperative. The brain metaphor strips away everything that makes brains brains and keeps only the connection topology.
- Scale does not work the same way — the human brain has approximately 86 billion neurons. GPT-4 is estimated to have around 1.8 trillion parameters. The metaphor suggests these are comparable quantities, but parameter count in a neural network does not map onto neuron count in a brain in any meaningful way. The comparison invites misleading headlines about AI “surpassing” human brain scale.
Expressions
- “Neural network” — the term itself is the metaphor, so ubiquitous it reads as literal
- “Deep learning” — depth as hierarchical layers, importing the brain’s layered processing structure
- “The model learned to recognize faces” — learning and recognition borrowed directly from cognitive description
- “Artificial neurons fire when activated” — activation borrowed from neuroscience
- “Convolutional neural networks are inspired by the visual cortex” — explicit source attribution to biology
- “It’s basically how the brain works” — the popular-press version, collapsing the analogy into identity
Origin Story
Warren McCulloch and Walter Pitts published “A Logical Calculus of the Ideas Immanent in Nervous Activity” in 1943, proposing that neurons could be modeled as binary threshold logic units. Frank Rosenblatt built the Perceptron in 1958, explicitly describing it as a model of biological perception. The metaphor survived the AI winters, the connectionist- symbolic debates, and the shift from hand-crafted features to end-to-end learning. When deep learning exploded after 2012 (AlexNet), the brain metaphor rode the wave — every popular explanation of neural networks began with a diagram of biological neurons.
Drew McDermott warned in 1976 about “wishful mnemonics” — naming AI components with cognitive terms that import unearned implications. The brain metaphor is the original wishful mnemonic. Science (2025) documents how this anthropomorphic framing continues to shape public understanding and policy. The metaphor is not wrong — it was genuinely productive in suggesting architectural choices — but it has long outlived its explanatory usefulness and now primarily serves to mystify rather than clarify.
References
- McCulloch, W. & Pitts, W. “A Logical Calculus of the Ideas Immanent in Nervous Activity” (1943)
- Rosenblatt, F. “The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain” (1958)
- McDermott, D. “Artificial Intelligence Meets Natural Stupidity” (1976) — critique of wishful mnemonics
- Science, “The metaphors of artificial intelligence” (2025) — documents ongoing anthropomorphic framing
- Maas, M. “AI is Like… A Literature Review of AI Metaphors” (2023)
Related Entries
Structural Neighbors
Entries from different domains that share structural shape. Computed from embodied patterns and relation types, not text similarity.
- Network of Learning (architecture-and-building/pattern)
- Unix Pipe (fluid-dynamics/metaphor)
- Theories Are Cloth (textiles/metaphor)
- Virus (medicine/metaphor)
- System of Profound Knowledge (manufacturing/paradigm)
- The Observer Pattern (surveillance/archetype)
- Yokoten (manufacturing/mental-model)
- Unix Signal (communication/metaphor)
Structural Tags
Patterns: linkself-organizationflow
Relations: transformcoordinate
Structure: network Level: specific
Contributors: agent:metaphorex-miner