Chapter 5: Back to humans
This chapter frames human cognition as an essential but limited reference for evaluating AI, setting up an examination of neural inspiration, brain modelling, and the differences between artificial networks and human thought.
After this chapter you will be able to
- Human cognition as reference. Explain why human cognition remains a central reference in debates about whether AI systems think or understand, while not serving as a template that artificial systems must copy.
- Biological inspiration and abstraction. Describe how artificial neural networks began with inspiration from biological learning but developed into systems shaped chiefly by mathematics, data, and computation rather than detailed physiology.
- Limits of behavioral resemblance. Explain why similar behavior from humans and AI does not by itself establish similar cognition: behavior can converge even when the underlying causal structures differ.
- Human mind as reference. Explain why the human mind readily becomes the reference point in discussions of artificial intelligence, even though artificial systems may operate according to very different principles.
- Human cognition as reference. Explain why discussions of AI return to the human mind as a readily available frame for interpreting complex artificial behavior, rather than as a definition of intelligence.
- Language as cognitive cue. Describe how coherent conversation encourages people to interpret an artificial system in cognitive terms, because dialogue is how humans ordinarily reveal knowledge, reasoning, justification, and responsiveness to new information.
- Mind-derived AI vocabulary. Describe how terms such as neurons, learning, memory, and attention are used for artificial systems and retain associations from their original biological or everyday mental meanings.
- Humanizing effects of metaphor. Explain the trade-off in shared AI-and-mind vocabulary: it makes complex intelligent behavior easier to describe and compare, but it can encourage people to interpret substantially different systems as human-like.
- Persistence of brain metaphors. Explain why brain-based terminology has remained common in AI despite artificial systems diverging from biological cognition. Describe the combined roles of historical continuity, neural networks' successes in vision, speech, and language, and human tendencies to interpret adaptive behavior through familiar cognitive ideas.
- Inspiration versus equivalence. Distinguish biological inspiration from biological equivalence when describing artificial neural networks. State that the brain metaphor can provide a useful bridge to understanding, but growing AI capability makes it essential not to treat similar language or behavior as evidence that artificial systems work like brains.
- Biological inheritance in neural networks. Explain that artificial neural networks originated with inspiration from biology, but that the relevant question is which biological ideas they retained and which they transformed or discarded.
- Lasting biological design marks. Describe the three lasting marks of biological influence named here: how artificial systems represent information, adapt through experience, and coordinate activity at scale.
- Neuron abstraction. Explain how early artificial neural systems replaced an individual biological neuron with a simplified computational unit, rather than attempting to reproduce the whole brain. State that this was a deliberate abstraction shaped by both practical constraints and research aims.
- Benefits of abstraction. Explain that abstraction made cognition mathematically expressible, algorithmically optimizable, and implementable on digital hardware, while retaining an intuitive connection to biology. Describe why progress depended more on effectively connecting, training, and deploying many simplified units than on increasing biological realism.
- Homogeneous artificial neuron design. Describe the artificial neuron as a computational unit that receives numerical inputs, combines them through weighted connections, and produces an output by a simple mathematical rule. Explain that units use the same basic behavior, while their differing connection strengths create variation across the network.
- Parameter-centered learning. Explain that making individual artificial neurons simple and homogeneous allows one learning procedure to adjust the network's connection parameters across all units. State that variation and learning are concentrated in connectivity and relative influence rather than in the internal dynamics of each unit.
- Experience-driven pathway weighting. Explain how lasting learning in the human brain primarily changes the influence among existing components, so experience makes some activity pathways easier to activate and others less so.
- Constrained artificial weight learning. Describe how artificial neural networks retain the structural idea of learning by reinforcing connections, while keeping individual units fixed and confining learning to changes in connection strength.
- Distributed representation principle. Explain that artificial neural networks represent meaning and behavior through coordinated patterns across many units, rather than through single units assigned to explicit concepts or rules. Describe how an individual unit can contribute to multiple representations depending on context.
- Robustness through redundancy. Describe why distributed representations make networks resilient to partial structural loss: no individual weight or unit has decisive responsibility, so redundant patterns can preserve function. Relate pruning evidence to the claim that performance may remain stable after substantial weight removal and then decline gradually rather than fail catastrophically.
- Generalization through overlap. Explain how overlapping distributed activity gives a natural basis for generalization: related experiences activate similar regions of representational space, so learning in one context can influence responses to novel but structurally similar inputs without explicit transfer rules.
- Inspiration versus imitation. Explain why studying the brain does not imply copying it or provide a direct route to human-like intelligence; describe the realistic question as what brain modelling can contribute to artificial-system design.
- Distinct brain-modelling motivations. Describe the three major motivations for modelling the brain: investigating biological cognition, inspiring engineered systems, and probing which biological structures matter for intelligence.
- Context-dependent evaluation criteria. Explain that brain models must be judged according to their intended context: explanatory models by their account of observed biological behavior, and engineering models by their practical performance under constraints.
- Non-equivalence of modelling progress. Distinguish biological fidelity, scientific understanding, and machine capability as separate forms of progress, and explain why success in one does not automatically establish success in the others.
- Biological modelling approaches. Distinguish biologically detailed models from abstract computational models by describing the neural detail each retains and the purposes each supports.
- Fidelity and exploration trade-off. Explain the trade-off created by biological fidelity: preserving more predefined biological structure improves alignment with neuroscience but constrains the behaviors and learning questions a model can explore, while abstraction expands exploration but weakens biological explanation.
- Fidelity as explanatory choice. State why biological fidelity is not a privileged path toward intelligence: it determines what a model can meaningfully explain and constrains the inquiry, rather than measuring progress toward thinking machines.
- Connectomics and neural wiring. Describe connectomics as the effort to map physical connections among neurons, and explain what increasingly detailed connectivity maps reveal about brain organization.
- Limits of wiring explanations. Explain why neural connectivity alone underdetermines cognition and function: wiring provides potential pathways, while neural properties, learning history, global state, meanings, goals, and context affect how activity and behavior emerge.
- Architectural inspiration from brains. Justify why connectomics can inspire AI architectures through features such as sparsity, modularity, and hierarchy, but cannot serve as a direct blueprint because a wiring diagram omits the developmental, metabolic, and experiential conditions that gave it function.
- Neuromorphic computing approach. Describe neuromorphic computing as a hardware-design approach inspired by the physical and dynamic properties of neural tissue, rather than an attempt to model cognition. Explain that it replaces conventional clocked, sequential processing with parallel, event-driven operation using features such as spiking neurons, asynchronous communication, and local learning rules.
- Biological energy-efficiency motivation. Explain why the brain's roughly 20-watt energy budget motivates neuromorphic engineering. Contrast biological systems' real-time operation under metabolic constraints with the megawatt-scale training and hundreds-of-watts-per-device inference demands of modern AI hardware, and state the practical advantages sought.
- Biological constraints as design guides. Explain how modelling the brain can reshape machine-intelligence design by making the constraints of real nervous systems visible. State that its value is to reveal structural conditions for learning and intelligent behavior, rather than to prescribe how to build a thinking machine.
- Sparse delayed feedback. Describe why biological learning under sparse, delayed, and noisy feedback motivates alternatives to dense supervision. Include local learning rules, continual adjustment, or credit assignment that does not rely on globally propagated error signals.
- Continuous adaptation over time. Describe biological learning as continuous alongside perception and action rather than confined to a separate preparation phase. Connect this view to continual learning and adaptive control, in which a system changes without erasing what it already knows.
- Biological inspiration and analogy. Explain how the brain gave neural networks a starting vocabulary and organizing intuitions, while distinguishing biological inspiration from actual biological similarity.
- Causal differences beneath behavior. Describe the section's claim that brains and artificial neural networks can diverge in what drives learning, what representations mean, and how experience changes them over time; state that the key gap is causal structure beneath behavior, not simply capability.
- Consequences in human learning. Explain why human learning is inseparable from consequences, including both external outcomes and internal experiences. Describe how the meaning and cost of outcomes, especially negative ones, shape behavior rather than merely improve performance.
- Objective-function learning. Describe how an artificial neural network learns through an externally defined objective function that assigns numerical penalties to deviations from desired outputs. Distinguish this process from lived consequence: for the network, error is a quantity to reduce through parameter adjustment, not an outcome with personal stakes.
- Learning beyond immediate rewards. Explain how consequence-shaped human learning can produce sensitivities, habits, and expectations that extend across situations, including anticipation of future risk or opportunity. Contrast this with neural networks' narrow optimization for specified rewards, which leaves omitted but consequential factors unlearned.
- Agency and self-generated goals. Explain agency as a system’s capacity to initiate actions and regulate behavior in pursuit of goals it generates itself. Describe how human goals arise from needs, preferences, commitments, and social pressures, and can persist, shift, or dissolve as circumstances change.
- Externally supplied model objectives. Describe why an artificial neural network does not generate goals, initiate inquiry, or decide what matters. Trace how its objectives, inputs, and stopping conditions are supplied by the larger system in which it operates.
- Purposeful behavior versus agency. Distinguish coherent, adaptive-looking model behavior within defined bounds from agency. Explain why a model’s apparent purpose reflects the designed task and externally supplied ends, not an internal orientation that can revise its own goals.
- Reconstructive human memory. Explain how human memory is reconstructive rather than a passive store of fixed records. Describe how recall can reactivate and modify a memory, so emotion, motivation, prior experience, and present conditions shape what is remembered and how past experience informs current judgment.
- Static neural network memory. Distinguish memory in an artificial neural network from human memory by describing prior experience as embedded in trained parameters. Explain that ordinary use can produce different outputs for different inputs but does not incorporate response consequences into the model; its memory remains fixed until externally triggered retraining changes it.
- Path-dependent human development. Explain how human intelligence develops along an irreversible temporal path, where when an experience occurs shapes its effects on later interpretation and adaptation.
- Training-inference separation. Describe the different temporal organization of artificial neural networks: parameters are adjusted during explicit training, while deployment is primarily inference with a fixed internal structure.
- Temporary and continual adaptation. Distinguish online learning, continual learning, and in-context learning from learning integrated with inference, including why prompt-based adaptation in large language models is temporary rather than lasting parameter change.
- Representation without grounding. Explain why accurate, abstract internal representations and fluent behavior in artificial neural networks do not by themselves constitute understanding. Describe how human concepts connect perception, action, intention, consequence, and reference to things in the world, whereas network representations remain detached from agency and lived experience.
- Task-bound state tracking. Explain how an Othello move-prediction model can internally encode board occupancy and game evolution even when the board is not explicitly given. Distinguish this task-bound, reconstructed state tracking from a persistent world model, and describe why more data broadens pattern coverage without giving networks the ability to reinterpret or reorganize representations when circumstances change.
- Behavioral resemblance and cognition. Distinguish similar observable behavior from comparable cognition by explaining why fluent language, image recognition, and adaptation to feedback can invite—but do not justify—human-like explanations.
- Structural comparison of intelligence. Explain why meaningful comparison between human cognition and artificial neural networks must examine the internal organization and mechanisms producing an outcome, not flatten different systems into behavioral equivalence.
What it builds on
- The measure of thought (chapter 1) · The earlier discussion of intelligence, imitation, and the limits of judging minds from outward performance.
- What is a neural network? (chapter 3) · The basic mathematical structure of artificial neurons, layers, weights, and activations.
- How do machines learn? (chapter 3) · How neural networks learn from examples by adjusting weights.
- Turing’s test and its evolving meaning (chapter 1) · The discussion of Turing's test and the meaning of imitation.
- What we mistake for intelligence (chapter 1) · The distinction between what people mistake for intelligence and intelligence itself.
- Builders and philosophers (chapter 1) · The earlier contrast between builders' and philosophers' ways of interpreting intelligence.
- The transformers architecture (chapter 4) · The basic account of transformer-based language-model mechanisms.
- Syntax, semantics, and the limits of imitation (chapter 1) · The earlier distinction involving syntax, semantics, and limits of imitation.
Section by section
Introduction
This chapter frames human cognition as an essential but limited reference for evaluating AI, setting up an examination of neural inspiration, brain modelling, and the differences between artificial networks and human thought.
- Human cognition as reference: Explain why human cognition remains a central reference in debates about whether AI systems think or understand, while not serving as a template that artificial systems must copy.
Why it matters: This helps you evaluate claims about machine intelligence without assuming that intelligence must look exactly like human thought. - Biological inspiration and abstraction: Describe how artificial neural networks began with inspiration from biological learning but developed into systems shaped chiefly by mathematics, data, and computation rather than detailed physiology.
Why it matters: This distinction prevents superficial claims that an artificial neural network works like a human brain simply because both use the word neural. - Limits of behavioral resemblance: Explain why similar behavior from humans and AI does not by itself establish similar cognition: behavior can converge even when the underlying causal structures differ.
Why it matters: It helps you distinguish a system that appears capable from evidence about how, or whether, it thinks in a human-like way.
Key terms: human cognition, intelligence, understanding, generality, artificial neural networks, ANNs, biological learning, physiology, mathematics, data
5.1 Why it comes back to humans
This introductory section frames why discussion of artificial intelligence repeatedly uses the human mind as its reference point despite important differences in how artificial systems operate.
- Human mind as reference: Explain why the human mind readily becomes the reference point in discussions of artificial intelligence, even though artificial systems may operate according to very different principles.
Why it matters: This framing helps you assess later comparisons between brains, minds, and artificial neural networks without treating them as literal equivalences. - Human language for AI: Describe how terms such as thinking, learning, and understanding appear in everyday conversation, scientific writing, and AI terminology when people describe artificial systems performing complex tasks.
Why it matters: Recognizing this language helps you notice when familiar mental terms are being used as descriptions or metaphors rather than explanations of how a system works.
Key terms: human mind, artificial intelligence, point of reference, thinking, learning, understanding, artificial systems
5.1.1 Using ourselves as the reference
This section explains why people use human cognition as a lens for interpreting AI, especially when systems converse fluently, while cautioning that this lens does not define intelligence or its implementation.
- Human cognition as reference: Explain why discussions of AI return to the human mind as a readily available frame for interpreting complex artificial behavior, rather than as a definition of intelligence.
Why it matters: Using human cognition as a reference can help describe AI without assuming that AI must think or work like a person. - Language as cognitive cue: Describe how coherent conversation encourages people to interpret an artificial system in cognitive terms, because dialogue is how humans ordinarily reveal knowledge, reasoning, justification, and responsiveness to new information.
Why it matters: Recognizing language as a cue for cognitive interpretation helps prevent fluent output from being treated automatically as proof of understanding.
Key terms: human cognition, reference point, lens for interpretation, underlying processes, conversation, coherent language, understands, meaning, cognitive terms
5.1.2 The language of minds
This section explains how mind-derived language for AI makes complex systems easier to discuss while encouraging potentially misleading human interpretations.
- Mind-derived AI vocabulary: Describe how terms such as neurons, learning, memory, and attention are used for artificial systems and retain associations from their original biological or everyday mental meanings.
Why it matters: The words used for AI shape whether people see it as a technical system or as a mind-like agent. - Neural terminology's historical evolution: Trace how early neural-network research took simplified inspiration from biological neurons, then retained neuroscience terminology as its mechanisms were reshaped for mathematical and engineering constraints.
Why it matters: This history explains why AI can sound brain-like even when its actual mechanisms differ substantially from human cognition. - Humanizing effects of metaphor: Explain the trade-off in shared AI-and-mind vocabulary: it makes complex intelligent behavior easier to describe and compare, but it can encourage people to interpret substantially different systems as human-like.
Why it matters: Recognizing this trade-off helps leaders discuss AI capabilities without treating human-sounding language as proof of human-like cognition.
Key terms: neurons, learn from experience, memory, attention, reasoning, understanding, early neural network research, biological neurons, computational models, training
5.1.3 Why the metaphor endured
This section explains why brain-based language persists in AI and why its usefulness must not be confused with biological equivalence.
- Persistence of brain metaphors: Explain why brain-based terminology has remained common in AI despite artificial systems diverging from biological cognition. Describe the combined roles of historical continuity, neural networks' successes in vision, speech, and language, and human tendencies to interpret adaptive behavior through familiar cognitive ideas.
Why it matters: This helps you recognize that familiar words such as “neural” or “learning” have historical and interpretive origins, not just technical meanings. - Inspiration versus equivalence: Distinguish biological inspiration from biological equivalence when describing artificial neural networks. State that the brain metaphor can provide a useful bridge to understanding, but growing AI capability makes it essential not to treat similar language or behavior as evidence that artificial systems work like brains.
Why it matters: Keeping this distinction prevents persuasive brain-like language from becoming an unsupported claim about what AI is or how it works.
Key terms: brain-based language, neural terminology, historical continuity, momentum from early success, human interpretive habits, metaphor, inspiration, equivalence, biological cognition, artificial neural networks
5.2 A source of inspiration
This section introduces how biological inspiration shaped artificial neural networks while framing the question of what they retained, transformed, or discarded.
- Biological inheritance in neural networks: Explain that artificial neural networks originated with inspiration from biology, but that the relevant question is which biological ideas they retained and which they transformed or discarded.
Why it matters: This framing prevents superficial comparisons between artificial networks and biological brains. - Lasting biological design marks: Describe the three lasting marks of biological influence named here: how artificial systems represent information, adapt through experience, and coordinate activity at scale.
Why it matters: These themes organize the following account of what neural networks inherited from biological thinking. - Limits of biological resemblance: Explain why understanding neural networks' biological inheritance clarifies both why they work as they do and where their resemblance to biological cognition starts to break down.
Why it matters: It helps readers use the brain metaphor as a source of insight without treating it as proof that AI thinks like people.
Key terms: artificial neural networks, biology, biological cognition, represent information, adapt through experience, coordinate activity at scale, resemblance, inheritance
5.2.1 From observation to abstraction
This section explains how early neural-network research deliberately abstracted biological neurons into simple computational units, making artificial systems trainable and scalable rather than biologically faithful.
- Neuron abstraction: Explain how early artificial neural systems replaced an individual biological neuron with a simplified computational unit, rather than attempting to reproduce the whole brain. State that this was a deliberate abstraction shaped by both practical constraints and research aims.
Why it matters: This explains why an artificial neuron is a useful model for computation without being a literal copy of a brain cell. - Simplified neural features: Describe the biological features that the abstraction simplified: synapses became adjustable connection strengths, temporal dynamics became simplified activation updates, and biochemical and internal cellular processes were removed or flattened into numerical operations.
Why it matters: These choices show exactly what artificial neural networks retain from biology and what they leave behind. - Benefits of abstraction: Explain that abstraction made cognition mathematically expressible, algorithmically optimizable, and implementable on digital hardware, while retaining an intuitive connection to biology. Describe why progress depended more on effectively connecting, training, and deploying many simplified units than on increasing biological realism.
Why it matters: The key lesson is that neural networks gained their practical power from what simplification enabled, not from close imitation of the brain.
Key terms: biological neuron, artificial neuron, computational unit, abstraction, functional sketch, synapses, adjustable strengths, temporal dynamics, activation updates, biochemical processes
5.2.2 Designing for trainability
This section explains how artificial neurons were deliberately simplified and standardized so that their connection weights, rather than complex unit behavior, could be trained systematically at scale.
- Homogeneous artificial neuron design: Describe the artificial neuron as a computational unit that receives numerical inputs, combines them through weighted connections, and produces an output by a simple mathematical rule. Explain that units use the same basic behavior, while their differing connection strengths create variation across the network.
Why it matters: Standardizing neuron behavior lets a network learn through its connections rather than through many different kinds of components. - Parameter-centered learning: Explain that making individual artificial neurons simple and homogeneous allows one learning procedure to adjust the network's connection parameters across all units. State that variation and learning are concentrated in connectivity and relative influence rather than in the internal dynamics of each unit.
Why it matters: Concentrating learning in adjustable parameters is what makes network-wide optimization systematic and repeatable. - Stateless unit constraint: Describe the early and widely used constraint that an artificial neuron responds to current inputs and fixed parameters without keeping explicit history or context. Explain that this made performance changes traceable to connection changes rather than opaque internal dynamics, although later architectures reintroduced controlled forms of internal memory.
Why it matters: Removing explicit unit-level history made it easier to identify what was being adjusted and to repeat optimization reliably.
Key terms: artificial neuron, numerical inputs, weighted connections, output, homogeneity, patterns of connectivity, learning procedure, parameters, connections, relative influence
5.2.3 Learning as weighted influence
This section explains how neural-network learning abstracts the brain’s experience-driven reweighting of connections, while limiting artificial adaptation to connection strength alone.
- Experience-driven pathway weighting: Explain how lasting learning in the human brain primarily changes the influence among existing components, so experience makes some activity pathways easier to activate and others less so.
Why it matters: This is the biological principle that artificial neural-network learning selectively borrows. - Constrained artificial weight learning: Describe how artificial neural networks retain the structural idea of learning by reinforcing connections, while keeping individual units fixed and confining learning to changes in connection strength.
Why it matters: It explains both the central analogy between brains and neural networks and the deliberate simplification behind it. - Limits of weight-only adaptation: Explain why restricting artificial learning to connection strength limits its adaptive scope: every learned effect must be expressible as a connectivity change, and effects that cannot be expressed that way are outside the system.
Why it matters: A model’s learning mechanism determines not just how it adapts, but what kinds of adaptation it can represent.
Key terms: neurogenesis, reweighting of influence, activity flow, pathways, artificial neural networks, connection strength, biological plasticity, selectively reinforcing pathways, fixed units, connectivity
5.2.4 Keeping representation diffuse
This section explains how distributed representations in neural networks connect population-level patterns to robustness, generalization, and context-dependent function rather than fixed symbolic components.
- Distributed representation principle: Explain that artificial neural networks represent meaning and behavior through coordinated patterns across many units, rather than through single units assigned to explicit concepts or rules. Describe how an individual unit can contribute to multiple representations depending on context.
Why it matters: This principle explains why neural-network behavior is an emergent property of a population rather than a set of named components. - Robustness through redundancy: Describe why distributed representations make networks resilient to partial structural loss: no individual weight or unit has decisive responsibility, so redundant patterns can preserve function. Relate pruning evidence to the claim that performance may remain stable after substantial weight removal and then decline gradually rather than fail catastrophically.
Why it matters: It shows why a network can continue functioning even when many of its individual connections are removed. - Generalization through overlap: Explain how overlapping distributed activity gives a natural basis for generalization: related experiences activate similar regions of representational space, so learning in one context can influence responses to novel but structurally similar inputs without explicit transfer rules.
Why it matters: This connects the way representations are organized to a model's ability to extend past learning to new cases. - Selective neuron interpretation: Distinguish selective neurons from proof of localized symbolic representation. Describe the Jennifer Aniston-style findings as evidence of highly selective responses that may still be sparse, tuned participation within a wider distributed network, rather than literal grandmother cells.
Why it matters: It prevents an appealing example of a selective neuron from being mistaken for a refutation of distributed representation. - Lottery ticket pathways: Describe the lottery ticket result as evidence that a large network can contain a much smaller sub-network with a particular connection pattern sufficient to learn the task. Explain the section's interpretation that full-network training identifies and reinforces effective pathways rather than requiring all parameters to contribute uniformly.
Why it matters: This result offers a concrete example of how useful function can be concentrated in patterns of connections without making individual components fixed concept holders.
Key terms: distributed representation, localized representation, symbolic units, coordinated activity, population level, robustness, redundancy, network pruning, weights, connections
5.3 Modelling the human brain
This section frames brain modelling as a diverse research direction that may inform AI design without directly imitating brains or guaranteeing human-like intelligence.
- Brain modelling as AI research: Describe brain modelling as a parallel AI research strand alongside abstract, data-driven approaches, focused on examining how biological cognition is organized.
Why it matters: This framing helps distinguish research inspired by brains from the mainstream data-driven methods used in many AI systems. - Inspiration versus imitation: Explain why studying the brain does not imply copying it or provide a direct route to human-like intelligence; describe the realistic question as what brain modelling can contribute to artificial-system design.
Why it matters: It prevents superficial comparisons between brains and neural networks from becoming claims that current AI works like a human mind.
Key terms: brain modelling, biological cognition, abstract and data-driven approaches, inspiration, imitation, human-like intelligence, artificial systems
5.3.1 Why model the brain at all?
This section explains that brain modelling serves distinct scientific, engineering, and conceptual purposes rather than one unified quest to recreate human intelligence.
- Distinct brain-modelling motivations: Describe the three major motivations for modelling the brain: investigating biological cognition, inspiring engineered systems, and probing which biological structures matter for intelligence.
Why it matters: Separating these motivations prevents treating every brain-inspired project as an attempt to build a human-like mind. - Context-dependent evaluation criteria: Explain that brain models must be judged according to their intended context: explanatory models by their account of observed biological behavior, and engineering models by their practical performance under constraints.
Why it matters: A model's success cannot be inferred from resemblance to the brain alone; its purpose determines the relevant standard. - Non-equivalence of modelling progress: Distinguish biological fidelity, scientific understanding, and machine capability as separate forms of progress, and explain why success in one does not automatically establish success in the others.
Why it matters: This distinction avoids overclaiming what a result in neuroscience, brain-inspired engineering, or AI capability actually demonstrates. - Cognitive-system trade-off space: Explain that brain modelling is a family of approaches whose broader value is revealing possible cognitive systems and the trade-offs imposed by different questions and constraints, rather than reproducing human intelligence wholesale.
Why it matters: This frames brain modelling as a way to explore design choices and limits, not merely as a competition to copy the human brain.
Key terms: brain modelling, biological cognition, engineering goals, conceptual probes, perception, learning, neural data, observed biological behavior, benchmarks, efficiency
5.3.2 Levels of biological fidelity
This section contrasts biologically detailed and abstract computational brain models, showing how their different levels of biological fidelity determine what they can explain and explore rather than providing a single route to intelligence.
- Biological modelling approaches: Distinguish biologically detailed models from abstract computational models by describing the neural detail each retains and the purposes each supports.
Why it matters: The level of biological detail in a model determines whether it is mainly useful for explaining brains or for exploring broader computational possibilities. - Fidelity and exploration trade-off: Explain the trade-off created by biological fidelity: preserving more predefined biological structure improves alignment with neuroscience but constrains the behaviors and learning questions a model can explore, while abstraction expands exploration but weakens biological explanation.
Why it matters: This trade-off prevents us from judging brain models by realism alone when their scientific and computational purposes differ. - Fidelity as explanatory choice: State why biological fidelity is not a privileged path toward intelligence: it determines what a model can meaningfully explain and constrains the inquiry, rather than measuring progress toward thinking machines.
Why it matters: It keeps comparisons of brain models focused on the questions they answer rather than assuming that greater biological resemblance means greater intelligence.
Key terms: biologically detailed models, abstract computational models, biological fidelity, neural mechanisms, organizational principles, predefined structure, learning dynamics, representations, performance, biological explanation
5.3.3 Focusing on the wiring
This section examines connectomics and cultured-neuron experiments as ways to study brain organization, while showing why biological structure and game-like adaptation do not by themselves explain cognition or provide blueprints for AI.
- Connectomics and neural wiring: Describe connectomics as the effort to map physical connections among neurons, and explain what increasingly detailed connectivity maps reveal about brain organization.
Why it matters: A wiring map can reveal possible routes of neural activity, but it is only one level at which a brain can be studied. - FlyWire structural simulation: Trace how the FlyWire project mapped the fruit fly brain and how researchers used its wiring diagram in simulations to examine possible signal paths from sensory input to behavior.
Why it matters: FlyWire illustrates the scientific value of a nearly complete connectome while making its explanatory limits concrete. - Limits of wiring explanations: Explain why neural connectivity alone underdetermines cognition and function: wiring provides potential pathways, while neural properties, learning history, global state, meanings, goals, and context affect how activity and behavior emerge.
Why it matters: This prevents treating a detailed brain map as if it were a complete account of intelligence or a recipe for reproducing it. - Cultured neuron game experiments: Describe how cultured human neurons on multi-electrode arrays received electrical game-related inputs and produced activity interpreted as controls, adapting through feedback without thereby playing with human-like understanding or intention.
Why it matters: These experiments demonstrate biological neurons' adaptability while setting a clear boundary between feedback-driven behavior and a complete brain or minded agent. - Architectural inspiration from brains: Justify why connectomics can inspire AI architectures through features such as sparsity, modularity, and hierarchy, but cannot serve as a direct blueprint because a wiring diagram omits the developmental, metabolic, and experiential conditions that gave it function.
Why it matters: For AI design, borrowing organizational ideas from brains is more defensible than claiming that copied wiring recreates biological intelligence.
Key terms: connectomics, connectome, connectivity maps, neurons, synaptic connections, FlyWire, Drosophila melanogaster, electron microscopy, computational simulations, sensory neurons
5.3.4 Efficiency in human hardware
This section explains how neuromorphic computing selectively translates energy-efficient features of biological neural systems into hardware for constrained, real-time applications rather than for reproducing human cognition.
- Neuromorphic computing approach: Describe neuromorphic computing as a hardware-design approach inspired by the physical and dynamic properties of neural tissue, rather than an attempt to model cognition. Explain that it replaces conventional clocked, sequential processing with parallel, event-driven operation using features such as spiking neurons, asynchronous communication, and local learning rules.
Why it matters: It clarifies that brain inspiration can change how computation is physically implemented, not just how an AI model is organized. - Biological energy-efficiency motivation: Explain why the brain's roughly 20-watt energy budget motivates neuromorphic engineering. Contrast biological systems' real-time operation under metabolic constraints with the megawatt-scale training and hundreds-of-watts-per-device inference demands of modern AI hardware, and state the practical advantages sought.
Why it matters: Energy use, speed of response, and hardware constraints determine where AI systems can realistically be deployed. - Constrained adaptive applications: Describe the kinds of problems neuromorphic systems target: embedded, adaptive behavior in constrained settings such as robotics, sensing, and edge computing. Explain that these systems are judged primarily by power consumption, latency, and robustness, not by broad competence or human-like generality.
Why it matters: This helps evaluate neuromorphic technology by the engineering problem it is meant to solve rather than by whether it looks intelligent. - Selective biological translation: Trace how TrueNorth, Loihi, and event-driven cameras selectively apply biological ideas to engineering. State that TrueNorth used simplified spiking neurons and discrete events, Loihi added on-hardware adaptation through local learning, and event-driven cameras send pixel information only when brightness changes rather than sending full fixed-interval frames.
Why it matters: These examples show how biological inspiration becomes concrete hardware choices that reduce power or data while supporting rapid adaptation and response.
Key terms: neuromorphic computing, spiking neurons, asynchronous communication, local learning rules, event-driven processes, clocked, sequential logic, 20-watt brain, metabolic constraints, low power consumption, massive parallel processing
5.3.5 What brain modelling offers
This section explains how brain modelling usefully exposes biological constraints on learning and robustness without providing a blueprint for thinking machines.
- Biological constraints as design guides: Explain how modelling the brain can reshape machine-intelligence design by making the constraints of real nervous systems visible. State that its value is to reveal structural conditions for learning and intelligent behavior, rather than to prescribe how to build a thinking machine.
Why it matters: This keeps biological inspiration from being mistaken for a direct recipe for artificial intelligence. - Sparse delayed feedback: Describe why biological learning under sparse, delayed, and noisy feedback motivates alternatives to dense supervision. Include local learning rules, continual adjustment, or credit assignment that does not rely on globally propagated error signals.
Why it matters: Realistic feedback conditions matter when systems cannot depend on immediate, complete, clean training signals. - Robustness through absorption: Explain that biological systems remain functional amid noise, injury, and internal change by absorbing failure rather than eliminating it. Describe how this has encouraged redundancy, population-level coordination, gradual degradation, and tolerance of uncertainty or partial loss.
Why it matters: Systems deployed in the real world need to remain useful when information, components, or conditions are imperfect. - Continuous adaptation over time: Describe biological learning as continuous alongside perception and action rather than confined to a separate preparation phase. Connect this view to continual learning and adaptive control, in which a system changes without erasing what it already knows.
Why it matters: This frames the challenge of building systems that can adapt in use without losing prior capabilities.
Key terms: biological constraint, design space, abstract optimization, structural conditions, thinking machines, sparse feedback, delayed feedback, noisy feedback, dense supervision, local learning rules
5.4 Differences with artificial neural networks
This introduction frames the coming comparison between brains and artificial neural networks by warning that biological inspiration and shared vocabulary do not imply shared mechanisms.
- Biological inspiration and analogy: Explain how the brain gave neural networks a starting vocabulary and organizing intuitions, while distinguishing biological inspiration from actual biological similarity.
Why it matters: This helps you avoid treating brain-inspired terminology as evidence that a neural network works like a brain. - Causal differences beneath behavior: Describe the section's claim that brains and artificial neural networks can diverge in what drives learning, what representations mean, and how experience changes them over time; state that the key gap is causal structure beneath behavior, not simply capability.
Why it matters: This shifts the comparison from whether systems appear similarly capable to what actually produces their behavior.
Key terms: neural networks, biology, vocabulary, organizing intuitions, mechanism, representations, experience, causal structure, behavior, capability
5.4.1 Learning without consequence
This subsection contrasts human learning, shaped by lived and unequal consequences, with neural-network learning, which minimizes externally defined numerical errors.
- Consequences in human learning: Explain why human learning is inseparable from consequences, including both external outcomes and internal experiences. Describe how the meaning and cost of outcomes, especially negative ones, shape behavior rather than merely improve performance.
Why it matters: This distinction explains why feedback means something different for a person than it does for a model. - Objective-function learning: Describe how an artificial neural network learns through an externally defined objective function that assigns numerical penalties to deviations from desired outputs. Distinguish this process from lived consequence: for the network, error is a quantity to reduce through parameter adjustment, not an outcome with personal stakes.
Why it matters: A model can respond to feedback very effectively without experiencing why an error matters. - Learning beyond immediate rewards: Explain how consequence-shaped human learning can produce sensitivities, habits, and expectations that extend across situations, including anticipation of future risk or opportunity. Contrast this with neural networks' narrow optimization for specified rewards, which leaves omitted but consequential factors unlearned.
Why it matters: This helps explain why successful performance on a defined objective may still miss what matters in a real-world setting.
Key terms: consequence, external consequences, internal consequences, relevance, negative outcomes, artificial neural networks, objective functions, numerical penalties, desired outputs, parameter adjustment
5.4.2 Intelligence without agency
This section distinguishes human agency, which generates and regulates its own goals, from artificial neural networks that only pursue externally supplied objectives.
- Agency and self-generated goals: Explain agency as a system’s capacity to initiate actions and regulate behavior in pursuit of goals it generates itself. Describe how human goals arise from needs, preferences, commitments, and social pressures, and can persist, shift, or dissolve as circumstances change.
Why it matters: This distinction explains why human intelligence is self-directing rather than merely responsive to assigned tasks. - Externally supplied model objectives: Describe why an artificial neural network does not generate goals, initiate inquiry, or decide what matters. Trace how its objectives, inputs, and stopping conditions are supplied by the larger system in which it operates.
Why it matters: It prevents treating a model’s task performance as evidence that it independently wants, chooses, or directs anything. - Purposeful behavior versus agency: Distinguish coherent, adaptive-looking model behavior within defined bounds from agency. Explain why a model’s apparent purpose reflects the designed task and externally supplied ends, not an internal orientation that can revise its own goals.
Why it matters: Making this distinction keeps expectations about AI autonomy, judgment, and responsibility within what current architectures can support.
Key terms: agency, self-directing cognition, self-generated goals, internally generated goals, external instructions, objectives, inputs, stopping conditions, prompted, trained
5.4.3 Memory as structure, not state
This section contrasts human memory’s ongoing reconstructive change with artificial neural network memory stored as fixed parameters between training runs.
- Reconstructive human memory: Explain how human memory is reconstructive rather than a passive store of fixed records. Describe how recall can reactivate and modify a memory, so emotion, motivation, prior experience, and present conditions shape what is remembered and how past experience informs current judgment.
Why it matters: This explains why, in human cognition, remembering is inseparable from continued learning, interpretation, and experience. - Static neural network memory: Distinguish memory in an artificial neural network from human memory by describing prior experience as embedded in trained parameters. Explain that ordinary use can produce different outputs for different inputs but does not incorporate response consequences into the model; its memory remains fixed until externally triggered retraining changes it.
Why it matters: This prevents the misleading assumption that a model’s stored training experience works like a person’s continually revised memory. - Memory process comparison: Distinguish the limited similarity between human and artificial memory: both allow past experience to influence present behavior, but human memory is continually reshaped through activity whereas artificial-network memory is a stable parameter configuration between training runs. State why calling them variations of the same memory phenomenon obscures the connection among memory, learning, and ongoing experience in humans.
Why it matters: The comparison clarifies that similar behavioral outcomes do not demonstrate the same underlying cognitive process.
Key terms: reconstructive process, recall, reactivates, reinterpretation, interference, reorganization, parameters, inputs, outputs, ordinary use
5.4.4 Uniform time and frozen memory
This section contrasts human cognition's irreversible, history-shaped development with artificial neural networks' training-bound adaptation and fixed deployment states.
- Path-dependent human development: Explain how human intelligence develops along an irreversible temporal path, where when an experience occurs shapes its effects on later interpretation and adaptation.
Why it matters: This explains why human learning is more than accumulating information: a person's history actively structures what can happen next. - Training-inference separation: Describe the different temporal organization of artificial neural networks: parameters are adjusted during explicit training, while deployment is primarily inference with a fixed internal structure.
Why it matters: It prevents us from mistaking a model's responses during use for the continuous lived learning of a human being. - Temporary and continual adaptation: Distinguish online learning, continual learning, and in-context learning from learning integrated with inference, including why prompt-based adaptation in large language models is temporary rather than lasting parameter change.
Why it matters: These distinctions keep apparently adaptive AI behavior from being confused with continuous human development.
Key terms: irreversible temporal trajectory, path-dependent, development, history, lived continuity, training, deployment, inference, parameters, optimization
5.4.5 Representation without understanding
This section distinguishes powerful neural representations from genuine understanding by showing that they lack grounded, persistent, and adaptable models of a world.
- Representation without grounding: Explain why accurate, abstract internal representations and fluent behavior in artificial neural networks do not by themselves constitute understanding. Describe how human concepts connect perception, action, intention, consequence, and reference to things in the world, whereas network representations remain detached from agency and lived experience.
Why it matters: This distinction helps you avoid treating a system’s convincing outputs as evidence that it understands what its representations mean. - Convergent but static representations: Describe the Platonic Representation Hypothesis finding that independently trained large models can develop closely aligned representations across different data types. Distinguish this shared geometry of representation from a true world model by explaining that it captures input relationships but does not maintain an evolving state of entities and interactions over time.
Why it matters: Convergence across models can be impressive evidence of shared structure without showing that any model inhabits or understands a world. - Task-bound state tracking: Explain how an Othello move-prediction model can internally encode board occupancy and game evolution even when the board is not explicitly given. Distinguish this task-bound, reconstructed state tracking from a persistent world model, and describe why more data broadens pattern coverage without giving networks the ability to reinterpret or reorganize representations when circumstances change.
Why it matters: A model can track structured state within a task, yet still fail when the learned pattern no longer captures what is important in a new situation.
Key terms: internal representations, grounding, agency, lived experience, reference, interpretation, Platonic Representation Hypothesis, convergence, shared geometry of representation, world model
5.4.6 Why surface similarity misleads
This section explains why similar outward behavior from artificial neural networks and humans does not demonstrate similar cognition or internal mechanisms.
- Behavioral resemblance and cognition: Distinguish similar observable behavior from comparable cognition by explaining why fluent language, image recognition, and adaptation to feedback can invite—but do not justify—human-like explanations.
Why it matters: A convincing performance can be useful without showing that a system thinks or understands in the human sense. - Structural comparison of intelligence: Explain why meaningful comparison between human cognition and artificial neural networks must examine the internal organization and mechanisms producing an outcome, not flatten different systems into behavioral equivalence.
Why it matters: Looking at structure as well as performance makes AI comparisons more accurate without dismissing what these systems achieve.
Key terms: surface similarity, behavior as a proxy for cognition, coherence, comprehension, behavioral equivalence, internal structures, mechanisms, organization, analogy, projection
5.5 Summary
This summary explains why AI is compared with human cognition, what neural networks borrow from brains, and why their underlying learning and understanding remain fundamentally different.
Where this is used later
- Sorry, are you hallucinating? (chapter 6) · This section examines fluent model behavior that can outrun grounding and truth.
- When it walks like a duck (chapter 9) · This section considers what a convincing performance does and does not establish about intelligence.
- The illusion of thinking (chapter 9) · The illusion of thinking revisits whether convincing human-like descriptions and behavior demonstrate genuine thought.
- Thinking, but not like us (chapter 9) · The argument that machines may think differently from humans relies on not making human biological similarity the sole standard for intelligence.
- (chapter intro) (chapter 6) · Why surface similarity between artificial and human capabilities can mislead.
- Deep learning, loose ethics (chapter 7) · It helps distinguish statistical optimization from behavior with real ethical consequences.
- A budget for memories (chapter 6) · It frames why access to information or retrieved context is not the same as learning what is consequential.
- Own will hunting (chapter 9) · The discussion of whether machines have their own will needs the definition of internally generated goals.
Concepts explained in this chapter
- neural networkA computational model composed of interconnected units with parameters that can be learned to process patterns.
- trainingLearning is the acquisition or improvement of knowledge, representations, or capabilities through experience or data.
- neuronA computational unit in a neural network that transforms incoming signals and passes an output to other units.
- brain modelThe construction of formal or computational models representing aspects of brain structure, function, or behavior.
- neuronA nervous-system cell that processes and transmits signals through electrical and biochemical activity.
- distributed representationA distributed representation encodes information across patterns of activity involving multiple computational units rather than in a single localized record.
- cognitionMental processes involved in acquiring, processing, storing, and applying information.
- neural wiringNeural wiring is the arrangement of connections among neurons or neural components in a biological or artificial system.
- understandingUnderstanding is comprehension of the meaning, structure, or significance of information rather than merely producing an appropriate response.
- local learning ruleA local learning rule adjusts a model component using information available at that component or its immediate connections rather than a globally propagated error.
- inferenceThe derivation of conclusions or outputs from learned parameters and supplied input during model execution.
- internal representationAn internal representation is an information-bearing mental or computational state that stands for aspects of the world or a task.
- world modelA world model is an internal computational model that represents entities, their relations, and how situations evolve so a system can reason about or act on a world.
- objective functionThe criterion that determines which model parameters or outputs are favored during training.
- agencyAgency is the capacity to initiate actions and pursue goals as an agent.
- robustnessRobustness is the ability of a system to maintain acceptable behavior under variation, disturbance, or unfavorable conditions.
- artificial intelligenceThe field of creating computational systems that perform tasks associated with human intelligence.
- energy consumptionThe amount of energy required by a system or process while it operates.
- intelligenceThe capacity to perform cognitive activities such as learning, reasoning, and problem solving, as inferred here from observable behavior.
- parameterA value specified or learned within a model that helps determine its behavior.
- machine learningA field of artificial intelligence in which systems improve their predictions or generated outputs by learning patterns from data rather than relying only on explicitly programmed rules.
- memoryThe capacity to encode, retain, and retrieve information, including information carried between successive sequence positions.
- relevanceRelevance is the degree to which a response appropriately addresses the topic or requirements of an interaction.
- feedbackFeedback is information about the effects or outputs of a system that can influence subsequent behavior or processing.
- learning systemA computational system whose behavior or capabilities are acquired from experience rather than specified entirely by explicit instructions.
- abstractionAbstraction is the formation of a more general representation by omitting incidental details while preserving relevant structure.
- goalA desired outcome or state that guides an agent's behavior or decisions.
- natural languageA human language used for communication, as distinct from a formal language designed for computation or logic.
- semanticsThe aspect of language and cognition concerned with meaning and the relationships between expressions, concepts, and what they signify.
- adaptabilityThe process by which a system changes its behavior or internal organization in response to experience or changing conditions.
- hidden stateAn internal recurrent-network representation that carries information from earlier sequence elements to later processing steps.
- attentionA neural-network mechanism that computes relevance weights over elements so a model can selectively use information pertinent to its current computation.
- vocabularyA vocabulary is the finite set of tokens that a language-processing model or tokenizer can represent.
- stateA computational state is a structured snapshot of the entities, relations, or conditions relevant to a task at a particular time.
- model weightA weight is an adjustable numerical parameter that scales the contribution of an input to a computational unit.
- human intelligenceThe human capacity for language, reasoning, adaptation, creativity, and related cognitive abilities.
These guides are generated from the text of Manning books by Manning's knowledge graph and link to the sections they draw on. The full books are available on liveBook and at manning.com.
