51 The Computer, the Levels, and the Limits of Both
A Back-of-Book Essay
The overview argued for an architecture-first approach to the brain: begin with the machinery evolution built, the body it regulates, and the world in which it acts, rather than with a psychological model and a search for its neural address. This essay develops that argument more carefully. It does not deny that brains compute, that abstract models can be useful, or that some functions localize. It asks a different question: which descriptions have earned biological standing, and which remain convenient summaries of behavior?
51.1 The direction of explanation
The central claim is simple: the brain is not the implementation layer of cognitive psychology.
Cognitive neuroscience often proceeds as though the major parts of the mind have already been identified. Psychology supplies constructs such as working memory, attention, executive control, autobiographical memory, or cognitive inhibition. Computational modeling specifies the operations those constructs must perform. Neuroscience is then assigned the task of finding where and how the brain implements them.
This strategy can succeed. A behavioral distinction may correspond to a real division in neural machinery. A perceptual category may recruit a reproducible set of regions. A formal model may identify a transformation that a circuit appears to perform. None of that is in dispute.
What does not follow is that the categories inherited from psychology have presumptive authority over the biology. A construct earns its place in an experiment because it helps organize behavior. It does not thereby become a component of the nervous system. The phrase working memory may summarize a requirement shared by many tasks: information must remain available after the stimulus is gone. That does not establish that the brain contains one working-memory system, or that every circuit bridging a delay is implementing a common psychological faculty. The experimenter has grouped the tasks. The brain may not have.
A brain-first approach reverses the usual presumption. It begins with the evolved brain–body–environment system: cell types, connections, recurrent dynamics, neuromodulation, development, phylogenetic history, bodily mechanics, and the ecological problems within which those mechanisms operate. It then asks which psychological descriptions survive contact with that organization. Some familiar categories may acquire clear neural foundations. Others may divide into several mechanisms, merge with functions that psychology had treated as separate, or disappear as explanatory entities.
This is not an argument for anatomy without theory. Nor is it a demand that every explanation begin at the ion channel. The relevant starting point may be a circuit motif, a recurrent network, a control loop, a population dynamic, or a coupled brain–body system. Brain first means that the decomposition of function should be constrained by the organization evolution actually produced. Psychological models may guide the search, but they do not get to supply the parts list in advance.
51.2 What the computer metaphor gets right
The claim that the brain is a computer bundles together several ideas that should be separated.
First, neural systems perform systematic transformations that can be described computationally. Neurons integrate, filter, threshold, normalize, compare, and accumulate. Networks transform sensory signals into motor commands, estimate variables that are not directly sensed, and preserve information across changes in viewpoint or time. Computational language is often the clearest way to describe what these dynamics accomplish.
Second, some organizations are multiply realizable. The same formal relation can be implemented in different materials, and similar functional outcomes can be produced by circuits with different cellular parameters. There is no reason to imagine that wet carbon has a monopoly on computation. Nor is every useful regularity tied to one exact set of synapses or conductances.
Eve Marder’s work on the crustacean stomatogastric ganglion makes the point vivid. Very similar pyloric rhythms can emerge from substantially different combinations of intrinsic conductances and synaptic strengths. This is genuine biological degeneracy, not a philosophical possibility invented to defend functionalism. The lesson, however, is not that only the output matters. The useful invariant may lie at an intermediate level: network organization, interacting time scales, dynamical regime, and robustness across perturbations. That kind of description abstracts from exact cellular parameters without floating free of the mechanism. It is precisely the sort of architecture a brain-first account should seek.
Third, and more controversially, the functional organization can be specified independently of its physical realization. On this view, psychology identifies what is computed, algorithmic theory identifies the representations and procedures, and neuroscience identifies how the procedure is implemented in tissue. The upper levels can often be developed while remaining ignorant of many lower-level details.
But ignorance is not independence. An upper-level account offered as an explanation of a biological system must be realizable by that system. It may omit conduction delays, local connectivity, energetic cost, limited precision, bodily mechanics, or developmental history when those properties do not matter to the phenomenon under study. It cannot depend on instantaneous global access, lossless copying among regions, unlimited storage, arbitrary routing, or circuitry that no plausible developmental process could construct—at least not without explaining how the biological system approximates those operations and what the approximation costs.
Multiple realization does not mean that any lower-level organization can realize any upper-level function. It means that more than one organization may do so. The lower level may underdetermine the upper-level account while still delimiting the space of admissible accounts. Abstraction permits omission; it does not permit contradiction.
The first two claims are fully compatible with a brain-first approach. The third is sometimes true, but its degree of truth must be established rather than assumed. The autonomy of a computational description is not granted merely because an input–output relation can be written as an equation.
Among computational neuroscientists, few now imagine a literal von Neumann machine inside the skull, with a central processor, passive memory stores, an address bus, and symbolic files shuttled into a prefrontal workspace. Contemporary models are often recurrent, distributed, dynamical, stochastic, and sometimes explicitly embodied. But the older architecture has not disappeared from cognitive psychology or from ordinary cognitive-neuroscience explanation. It survives less as a hardware diagram than as a grammar.
Information is said to be stored in one region, copied or transferred to another, loaded into working memory, retrieved from a buffer, read out by a downstream system, or broadcast to a central workspace. These expressions can be harmless and useful shorthand. They can also carry substantive assumptions: that content exists independently of the activity realizing it, that it can be duplicated without transformation, that regions act as addressable stores, and that routing can be specified independently of anatomical pathways, delays, and the state of the receiving system. A theory can call itself distributed while retaining a surprisingly classical picture of storage, transmission, and control.
A claim that information is copied from one brain area to another should therefore be treated as a hypothesis, not as a transparent description. What is preserved? What is transformed? Through which pathway? With what delay, loss, recoding, and dependence on the current state of the receiving circuit? Rejecting a literal desktop computer does not settle these questions.
The important issue is not whether brains can be described computationally. They can. It is whether an abstract task description has explanatory priority over the biological mechanism: whether we can specify what the system is doing first and then treat anatomy, dynamics, timing, development, and embodiment as details of implementation. In some domains that separation is extremely productive. In others it obscures the phenomenon we are trying to explain.
51.3 Marr’s levels are questions, not a staircase
No framework has had more influence on how cognitive scientists describe the relation between theory and brain than David Marr’s levels of analysis. The computational level asks what problem the system solves and why that problem matters. The algorithmic level asks which representations and procedures accomplish it. The implementation level asks how those procedures are physically realized.
These are valuable questions. The difficulty begins when they are converted into an order of explanatory authority: first specify the computation, then identify the algorithm, and only afterward descend to the brain. That order fits an engineered device whose purpose is already known. If I open a cash register, I do not need its circuitry to discover why it adds prices. The task was assigned by a designer, and the correctness conditions exist independently of the machine.
Marr’s work on early vision showed why analysis from the top can be powerful. Marr and Hildreth began with the structure of the visual problem and proposed mathematical operations that could reveal intensity changes across scale. Their theory did not simply predict the final organization of retinal and cortical circuitry, but it generated precise, biologically informed hypotheses before that circuitry was fully understood. In problems strongly constrained by optics, geometry, or mechanics, analysis of the task can narrow the space of possible solutions dramatically.
That success should be retained. What should not be retained is the assumption that every function of an evolved nervous system arrives with an equally clear task specification. A scientist may design a working-memory experiment, define the relevant delay, and decide what counts as correct performance. Evolution did not design a working-memory task. It modified circuits that already controlled orienting, action, navigation, vocal behavior, and bodily regulation. Those circuits may keep information available when current behavior requires it, but the shared experimental demand does not prove that evolution built them as implementations of one abstract faculty.
Even where the upper levels can be developed first, they remain answerable to the lower ones. A proposed representation or algorithm does not have to mention membrane channels or individual synapses. It does have to be compatible with the connectivity, time constants, noise, precision, and causal pathways of a system capable of realizing it. The lower levels need not uniquely dictate the higher-level theory, but they can rule theories out.
For an evolved system, knowledge also runs upward. Connectivity may reveal that two behaviors recruit the same ancient control circuit. Comparative anatomy may show that a structure long associated with a human cognitive category is an elaboration of machinery that served a different problem in earlier vertebrates. Population dynamics may disclose a variable that psychology had never named. The body may show why a neural signal has the time constant or coordinate frame it does. In such cases, implementation does not merely answer how. It helps reveal what the system is doing and where the functional boundaries should be drawn.
Marr’s levels are therefore better treated as a coordinate system than as a staircase. We should ask all three questions, but we should not assume that psychology answers the first one before biology begins. The degree to which a level can stand on its own is an empirical achievement.
51.4 Development is not another implementation detail
Marr’s framework is largely synchronic. It asks how an organized system performs a task. It does not directly ask how that organization comes into existence.
The evo-devo perspective developed by Peter Robin Hiesinger and others makes this omission difficult to ignore. The genome is not a blueprint containing a miniature picture of the mature brain or a complete point-to-point wiring diagram. It contributes molecular resources, regulatory relations, and local developmental rules whose consequences unfold through time. Cells divide, migrate, differentiate, recognize local signals, extend axons, compete for targets, generate spontaneous activity, and change their behavior in response to the structures already produced. Each developmental step alters the conditions under which the next step occurs.
Hiesinger describes this as a problem of algorithmic growth. The complexity of the mature architecture need not be explicitly listed in inherited instructions. It can be generated by the repeated operation of local rules within a growing, active organism. Development is therefore not the execution of a blueprint. It is a history-dependent process of construction and self-organization.
Development can itself be described at Marr-like levels. One might ask what development accomplishes, which cellular and molecular processes generate the result, and how those processes are physically realized. But this would be a second analysis whose output is the architecture analyzed by the first. The developmental “algorithm” is not simply running inside a fixed adult machine. It builds, prunes, and reorganizes the machine. Development therefore cuts across Marr’s three levels rather than fitting neatly beneath them.
This distinction has consequences for explanation. A circuit may be mathematically coherent and physically possible yet biologically unreachable. It may require globally addressed connections, exact wiring specified independently at every synapse, or a sequence of construction that no local developmental process can achieve. Conversely, relatively simple developmental interactions can produce ordered maps, repeated motifs, and robust connectivity without specifying every mature detail in advance.
Biological realizability is therefore more demanding than physical possibility. A proposed architecture must be dynamically workable, developmentally reachable, and modifiable through evolutionary changes that preserve viability along the way. The adult brain is not neutral hardware onto which a cognitive program has been loaded. Its possible computations are shaped by the generative history through which its architecture came to exist.
Evolution adds a second history. Selection modifies genes, regulatory systems, developmental interactions, bodies, and behavior; it does not redesign the mature circuit directly from a blank page. Function, algorithm, implementation, development, and evolution are not rungs on one ladder. They are intersecting constraints on one another.
51.5 A construct is not a part
Before asking where a psychological construct is located, it is worth asking what kind of thing the construct names. Cognitive psychology places very different entities in grammatically similar boxes.
A face is a recurring class of objects in the world. Faces have a characteristic visual structure, move in lawful ways, and carry unusually important social information. Human ventral visual cortex contains regions that respond selectively to faces, including reproducible territory in lateral fusiform and inferior occipital cortex, while superior temporal regions are especially sensitive to changing facial information. The exact boundaries vary across individuals, and face perception depends on a wider network, but the specialization is real enough to be found repeatedly, disrupted by focal damage, and altered by direct stimulation.
That is meaningful localization. A brain-first approach has no reason to reject it. Localization is not the enemy; unearned localization is.
Now compare face perception with working memory. Working memory does not name a class of objects. It names a requirement imposed by many tasks: some information must remain usable after it is no longer present. Visual cortex can preserve task-relevant visual information. Auditory and speech-motor circuits can preserve phonological structure. Spatial networks can maintain a location or intended movement. Prefrontal populations can stabilize a goal, bias other systems, or organize a sequence. These systems need not bridge a delay by the same mechanism; persistent firing, short-term synaptic change, recurrent population state, and later reactivation can all contribute. Their common appearance in working-memory tasks does not establish a common working-memory component.
Attention is more heterogeneous still. The word can refer to orienting the eyes, enhancing sensory gain, selecting one object among competitors, prioritizing a memory, maintaining a task rule, or committing shared effectors to one action. These phenomena interact, but a shared label does not prove a shared mechanism. The broad construct may be useful for describing behavior while concealing several biologically distinct operations.
Autobiographical memory is different again. It names a complex behavioral accomplishment: reconstructing events, situating them in time and place, linking them to semantic knowledge about the self, recovering affective significance, and often organizing the result into a verbal narrative. It would be remarkable if such an accomplishment corresponded to a single neural part in the way that a retinotopic visual field corresponds to organized sensory cortex.
The problem, then, is not that psychological constructs are always false. It is that they do not all have the same logical status. Some name stimulus domains. Some name task demands. Some collect a family of operations. Some summarize complex achievements. Treating them all as comparable boxes in an information-processing diagram creates a false expectation that each should have a corresponding neural address.
Double dissociation does not fully solve this problem. Suppose damage to one system selectively impairs verbal maintenance while damage to another impairs spatial maintenance. That is important evidence that the two tasks depend differently on the damaged machinery. It does not uniquely show that the brain contains two subdivisions of a general working-memory system. The first deficit might arise from damage to an auditory–motor circuit used for covert rehearsal; the second from damage to machinery for spatial orientation and action. The dissociation is real. The category under which the experimenter grouped the tasks remains open to revision.
A top-down construct can be a useful searchlight. It can organize experiments, reveal regularities, and generate hypotheses. It should not be mistaken for a parts list. The appropriate rule is not never localize. It is:
Localization is a result, not a premise. Describe freely; reify only when the biology earns it.
51.6 A controller is a coupled dynamical system
The limits of explanatory autonomy become clearest in feedback control. A feedback circuit is not best pictured as a device that receives an input, performs an arithmetic operation, and hands an answer to the body. It is better pictured as one part of a coupled dynamical system.
A formal operation such as addition has correctness conditions that can be specified independently of the device performing it. The relation expressed by \(27 + 35 = 62\) remains the same whether it is realized with pencil marks, an abacus, transistors, or trained neural activity. A feedback controller must also transform signals, but its success is not exhausted by the correctness of one transformation. It depends on what happens as controller, body, and environment continually change one another through time.
In a simple continuous-time description, the body or controlled plant might be written as
\[ \dot{x}=f(x,u,d), \qquad y=h(x), \]
while the controller is written as
\[ \dot{z}=g(z,y,r), \qquad u=k(z,y,r). \]
Here, \(x\) is the state of the body or plant, \(z\) is the internal state of the controller, \(y\) is the sensed state, \(u\) is the control action, \(d\) represents disturbances, and \(r\) is a reference value or viable target range. The object of interest is the evolution of the joint system:
\[ \frac{d}{dt} \begin{bmatrix} x\\ z \end{bmatrix} =F(x,z,d,r). \]
The equations may instead be discrete, stochastic, delayed, or hybrid. The central point is unchanged. The controller’s action alters the body; the altered body changes the controller’s next input; disturbances affect the trajectory; and delays determine whether correction stabilizes the system or drives it into oscillation. The relevant properties include stability, attractors, transients, overshoot, settling time, robustness, and the ability to remain inside a viable region of state space.
A thermostat offers a simple illustration. Strictly speaking, the thermometer is only the sensor. The thermostat compares sensed temperature with a reference and changes the state of a heater or cooling system. We can write an abstract control law, but that law alone does not tell us whether temperature will be regulated successfully. The outcome depends on the size and insulation of the room, the power and latency of the heater, the placement of the sensor, outside disturbances, and the delay between action and measured effect.
Move the same controller from a small room with a fast furnace to a large building with slow radiant heating and it may overshoot or oscillate. Add enough delay and a stable loop becomes unstable. The control law may be unchanged, but its role in the coupled system has changed. Timing is not merely a measure of performance added after the computation has been identified. An action that arrives too late may be the wrong action.
Neural control is embedded more deeply than a household thermostat. The mechanics of the body help determine which neural commands are effective. Tendons store and return energy. Limb geometry constrains possible movements. Viscoelastic tissues absorb perturbations before a reflex could respond. Passive dynamic walkers can produce surprisingly lifelike steps with little or no active control because the mechanics of the legs and gravity do part of the work. In running animals, rapid mechanical stabilization can precede any plausible neural correction. Behavior is produced by the coupled system, not by a program in the brain followed by a body that merely executes it.
This does not mean that the body is secretly performing symbolic calculations. It means that some of the regularity an observer might assign to a neural controller is generated by the physical organization of the plant. Drawing the computational boundary at the skull can therefore misdescribe the mechanism.
The same point applies within the nervous system. Membrane and synaptic time constants influence which temporal patterns can be integrated. Neuromodulators can reconfigure the effective circuit. Energetic constraints influence firing rates, wiring length, and coding strategies. Noise can degrade performance in one setting and contribute to exploration or probabilistic behavior in another. These properties are not always part of the most useful model, but neither can they be dismissed in advance as implementation details. A low-dimensional dynamical model may omit most of the underlying biophysics while remaining an excellent explanation. Its attractors, gains, delays, and transitions must nevertheless be realizable by the circuit it describes.
This is why the decisive contrast is not analog versus digital. A digital controller embedded in a real-time loop is still a controller; it cannot be paused without consequence. An analog system can be used merely as an offline model. The important distinction is between a formal description of a trajectory and a mechanism causally responsible for producing and regulating that trajectory in the world.
A simulation may numerically integrate the same equations and reproduce a neural circuit’s behavior across many inputs and perturbations. That is scientifically valuable. But formal equivalence is not automatically explanatory equivalence. The simulation may preserve the geometry of the dynamics while omitting the body that makes the controlled state useful, the delays that determine stability, the energetic constraints that shaped the architecture, and the causal coupling through which the system regulates anything at all. An embedded digital implementation could restore some of those relations; an offline simulation does not possess them merely by matching the equations.
The conclusion is not that simulations fail to compute or that biological tissue possesses some special essence. It is that computation in a feedback circuit is naturally described as the evolution of a coupled dynamical system. For an embodied controller, properties assigned to “implementation” may help determine the dynamics, define success, and even specify the problem. The upper-level description can abstract from the lower level, but it cannot violate the principles that make stable control possible.
51.7 Evolution does not begin with our categories
Developmental systems are among the principal things evolution modifies. Selection cannot redesign an adult circuit directly or replace an organism with a wholly new architecture in one step. Heritable changes alter the generative processes through which nervous systems and bodies are built, while the organism must remain viable at every intermediate stage. New circuits are assembled from old ones, duplicated structures diverge, and existing pathways acquire additional roles.
This history does not imply that the brain is uniformly distributed or that specialization is an illusion. Recurrent and consequential problems can favor specialized machinery. Sensory surfaces become topographically organized. Motor systems retain orderly relations to the body. Particular classes of biologically important stimuli can recruit reproducible cortical territory. Evolution can produce partial modules, interfaces, and divisions of labor.
But evolution does not owe us a decomposition that matches the vocabulary of cognitive psychology. The same circuit may contribute to several behaviors because each behavior depends on a shared control operation. A behavior may depend on several ancient systems that arose at different times. A region recruited by a modern laboratory task may have been shaped long before that task, or anything like its psychological description, existed.
The basal ganglia illustrate the difficulty. They contribute to movement, action selection, reinforcement, habit, and aspects of affect and cognition. One response is to declare them a general-purpose implementation of some abstract faculty. Another is to ask how a conserved architecture for gating and organizing behavior was progressively elaborated and coupled to new cortical and subcortical systems. The second question begins with phylogeny, development, and circuit organization rather than with a modern psychological box.
The same discipline should apply to the objectives attributed to the brain. An engineered algorithm may optimize a single explicit function. Organisms face multiple, shifting viability constraints: obtain food without becoming food, conserve energy without missing opportunity, regulate temperature and fluid balance, compete, cooperate, mate, care for offspring, and prepare for conditions that have not yet arrived. These demands cannot usually be reduced to one stationary target known in advance. The variables that matter, the tradeoffs among them, and the range of acceptable solutions are properties of a particular body in a particular niche.
A brain-first account therefore uses evolution as a source of constraints, not as a license for adaptive storytelling. Claims about ancestral function should be grounded in comparative anatomy, developmental continuity, conserved connectivity, and behavior across species. The goal is to reconstruct how present mechanisms could have emerged through modifications of earlier developmental and control systems. That reconstruction may reveal genuine specialization. It may also reveal that a familiar cognitive category cuts across several systems with different histories.
The point is not that evolution always produces a palimpsest and never a module. It is that the outcome must be discovered. Evolution and development constrain the decomposition; our vocabulary does not.
51.8 When physiology supplies the categories
Spatial navigation is often presented as a triumph of construct-first neuroscience: psychology proposed a cognitive map, and neuroscience later found the cells that implemented it. The actual history is less tidy and more revealing.
Tolman argued that rats could learn relational properties of an environment that were difficult to reduce to a chain of stimulus–response associations. His cognitive map was an important theoretical proposal about behavior. He did not predict a hippocampal neuron that would fire when an animal occupied one restricted location. Place cells were discovered when John O’Keefe and Jonathan Dostrovsky recorded neurons in freely moving rats and found firing related to the animal’s location and orientation. The physiological pattern supplied a new kind of entity for theory to explain.
Grid cells were not a predicted construct either. Recordings in medial entorhinal cortex revealed neurons with multiple firing fields arranged in a striking triangular lattice. The hexagonal pattern was a biological discovery, not a box carried down from cognitive psychology. Once discovered, it transformed computational accounts of path integration, metric representation, and the relation between entorhinal and hippocampal systems.
The same is true of head-direction cells, border-related responses, remapping, theta phase relations, and the organization of spatial scale. Behavioral theory helped determine which experiments seemed interesting and how the findings were interpreted. The discoveries were not theory-free. But the brain did not merely fill in components that psychology had specified beforehand. Neural recordings disclosed an architecture and a vocabulary that psychology did not possess.
Place cells and grid cells are therefore not simply locations where spatial memory lives. They are components of interacting systems whose dynamics distinguish location, direction, boundary, movement, context, and scale more finely than the umbrella term spatial memory ever did. Their discovery changed the behavioral questions, generated new models, and suggested new experiments. Explanation moved in both directions.
This is what a brain-first success looks like. Begin with a real problem faced by an animal, but do not assume that the experimenter’s description of the problem identifies the mechanism. Record the system, perturb it, map its inputs and outputs, compare its organization across species, and allow the biological regularities to revise the psychological account.
51.9 Cognition as increasingly decoupled control
If the brain is an embodied controller, where does cognition enter? One useful answer is that cognition extends control beyond the immediately sensed present.
At one end are short loops in which a current stimulus rapidly changes current action. A perturbation stretches a muscle and evokes correction. A rise in body temperature recruits responses that dissipate heat. The relevant state is present, and the loop closes quickly through the body and world.
Other behaviors depend on conditions that are no longer directly sensed. An animal must preserve a location after the landmark disappears, continue an action sequence after the initiating cue is gone, or choose now on the basis of a delayed consequence. Some state of the system must bridge the gap. The phenomena grouped under working memory appear here, but no single buffer is required. Different control systems can maintain or recover task-relevant information by different mechanisms.
Wider separation in space produces navigation to places that are not currently visible. Wider separation in time produces anticipation and allostasis: present behavior organized around a future bodily state. Planning extends control to actions not yet performed. Counterfactual reasoning compares possible outcomes in situations that do not presently exist. Episodic construction brings together elements of past experience to guide choices beyond the original episode.
Language and other symbolic practices extend the distance further. Words, diagrams, numerals, and written procedures allow one person to manipulate information about absent objects and share it with others. Cultural systems can create formal tasks whose correctness conditions are unusually independent of the body performing them. Long division is a genuine algorithm whether executed on paper, in silicon, or through trained neural activity.
The computer metaphor is most at home in this far territory because some human activities have been deliberately organized to support portable symbols and explicit procedures. But it is a mistake to take those culturally refined activities as the template for the entire nervous system. A person performing long division is not evidence that a spinal circuit, a hypothalamic regulator, or a heading network is best understood as a symbol-processing program.
The movement away from immediate coupling is also not one simple ladder. Control can be decoupled in time, space, sensory contact, counterfactual possibility, or social transmission. New organizational regimes may emerge along the way. Language is not merely a longer reflex. The continuity claim is more modest: capacities for planning, recollection, and symbolic thought were built by elaborating, combining, and culturally scaffolding older systems for perception, valuation, action, and regulation. Their evolutionary history matters to their present organization.
A brain-first account therefore approaches the far end by climbing from the near end. It asks how machinery that originally controlled bodies in immediate environments was extended to absent states, hypothetical actions, and shared symbols. It does not begin with the abstract procedures visible at the end of that history and project them backward onto every circuit from which they arose.
51.10 What a brain-first psychology would do
A brain-first psychology is not a ban on top-down models. It changes their status. A psychological construct becomes a proposal to be tested against biology, not a component to be located.
The program begins by identifying the organization of the system. What are its inputs and outputs? Which cell types and connections recur? What are its characteristic time scales and population dynamics? Which other systems modulate it? What body does it control, and through which sensors and effectors? How is the circuit assembled during development? Which features are conserved across related species, which are later elaborations, and which arise only through individual experience? What happens when the circuit is perturbed during naturalistic behavior rather than only during the task that supplied its conventional label?
From that basis, one can ask what variables the system tracks, what disturbances it corrects, which future states it anticipates, and which tradeoffs it resolves. Computational models remain indispensable here. They can reveal whether a proposed mechanism is sufficient, identify hidden variables, and generate predictions. But the model is answerable to the architecture, the developmental route that produced it, and the organism’s closed-loop behavior. A good fit to an experiment does not by itself establish that the model’s internal boxes are the brain’s parts. A model may omit lower-level detail; as an account of the brain, it may not rely on operations the brain cannot support.
At the timescale of behavior, brain, body, and environment form a coupled dynamical system. At the timescale of development, that system is assembled through a generative process rather than read from a blueprint. At the timescale of evolution, developmental processes, bodies, and mature architectures are modified together under the constraint of continued viability. These are not decorative details to be added after a cognitive model has been specified. They delimit the space of biologically admissible models.
This approach will not always make the brain look less organized. Face perception may retain a recognizable network of specialized regions. Retinotopy, somatotopy, and other orderly mappings will remain. In other cases, familiar categories may fragment. Attention may resolve into orienting, sensory gain, action competition, memory prioritization, and task maintenance, each with partially distinct mechanisms. Working memory may become a description of several ways in which different systems keep information available. Autobiographical memory may remain a useful name for an achievement assembled from event construction, semantic knowledge, affect, spatial context, and narrative machinery.
The result cannot be decided in advance. That is the point. Brain-first inquiry does not replace universal localization with universal distribution, or modules with undifferentiated networks. It makes the architecture earn whichever description turns out to be appropriate.
Nor does brain first mean brain alone. For an embodied controller, the explanatory unit may include the nervous system, endocrine signals, bodily mechanics, environmental structure, and artifacts used in action. The boundary should follow the mechanism, not the skull.
Psychology still has a central role. Behavior is where the organism’s problems and solutions become visible. Carefully designed tasks can isolate dependencies that anatomy alone would never reveal. The change is in the direction of authority. Behavioral categories organize observations; biology determines whether they correspond to natural divisions in the machinery.
51.11 The brain before the model
The dispute is not whether brains compute. They do. It is not whether abstract models can explain. They can. It is not whether functions ever localize. Some do, with impressive reliability.
The dispute concerns what we are entitled to assume before examining the system. A computational model does not acquire biological reality merely because it predicts behavior. A double dissociation does not guarantee that the brain is divided along the lines of the task labels. A reliable activation does not turn every psychological noun into a module. The organism may solve the experimenter’s task by recruiting machinery organized for a different set of problems.
Abstraction is indispensable because no explanation can include every physical detail. But the freedom to ignore detail is not the freedom to contradict the system being explained. A biologically credible account must be compatible with the architecture and dynamics of the nervous system, with the body and environment that close its control loops, and with a developmental and evolutionary route by which that organization could arise.
Marr’s questions remain valuable, but no level has permanent priority. In engineered formal systems, the separation of what from how may be strong. In embodied biological controllers, the body, timing, architecture, and niche may help specify what the computation is. Development explains how the architecture becomes possible, and evolution constrains which developmental changes can be retained. These forms of explanation meet in the same organism even though they do not occupy one simple hierarchy.
A brain-first psychology begins from the opposite direction from the familiar implementation search. It asks what evolution modified, how development assembled it, what the body requires, what variables circuits regulate, what dynamics their architecture permits, and what behavior emerges when those circuits are coupled to a world. It then builds a psychological vocabulary adequate to that organization.
Some old categories will survive. Some will be divided, combined, or discarded. The purpose is not to decide which outcome we prefer. It is to stop deciding the outcome before the brain has had its say.
The brain is not the implementation layer of a psychology specified somewhere else. A biologically adequate psychology is one of the things neuroscience must discover.