The Brain as a Control System

Overview

How to read this book

This textbook takes an opinionated approach to the human brain. I should say at the outset what the opinion is, because it shapes every chapter that follows.

Neuroscience is usually taught from the top down. We are told about the cognitive achievements that fascinate us — language, abstract reasoning, conscious perception — and then we go looking in the tissue for the machinery that produces them. This is an understandable way to organize a course. It begins with what students already care about. But I have come to think it gets the biology backward, and that starting there quietly installs a set of assumptions that are difficult to dislodge later.

The brain did not evolve to think. It evolved to govern a body and to solve the physical problems of staying alive long enough to reproduce. Every circuit we will study exists because, at some point in a very long history, it helped an organism do something concrete: find food, avoid being eaten, regulate its internal state, locate and choose a mate. Cognition, in the sense we usually mean it, is a late and heavily elaborated outgrowth of these control functions — not a separate faculty bolted on top of them. If we begin with control and movement and work upward, the more abstract capacities look less like miracles and more like what they are: extreme quantitative extensions of machinery that is, at bottom, about regulating a body in a world.

So throughout this book we will view the brain as a control system, and we will take that phrase seriously rather than as a loose metaphor. The chapters in this first unit lay the groundwork for that view in three steps. Chapter 1 asks why an organism should have a brain at all, and why human brains in particular became so large and so energetically expensive. Chapter 2 turns to evolution and development together — evo-devo — and to the deeply conserved vertebrate brain plan that we share with fish and frogs. Chapter 3 gives an anatomical tour of the human brain, but one organized around function rather than around names, so that the parts mean something before we dissect them in detail.

Before any of that, though, it is worth being explicit about what kind of thing we are claiming the brain is — and, just as importantly, what it is not. That requires a short detour through the history of how scientists have tried to picture the brain, because the picture we inherit is doing more work than we usually notice.

A warning about metaphors

Throughout history, scientists have reached for the most sophisticated technology of their era to explain the brain. The Greeks and Romans, who were superb hydraulic engineers, imagined vital spirits flowing through hollow nerves the way water moved through their aqueducts. Descartes, surrounded by the elaborate mechanical automata of his age, pictured the brain as a clockwork mechanism with animal spirits inflating the muscles like balloons. The industrial revolution gave us the brain as a telegraph network, and then as a telephone switchboard with operators connecting one region to another.

By the middle of the twentieth century, as electronic computers emerged, the metaphor changed again — and this time it stuck. For more than seventy years the dominant image of the brain has been an information-processing device that computes with neural signals. The vocabulary is now so familiar that we forget it is metaphorical at all. We “store” and “retrieve” memories, “process” stimuli, “encode” experiences. We describe working memory as a “buffer” and attention as “bandwidth.” The endurance of this language is striking, but endurance is not the same as accuracy. It may tell us more about the flexibility of computational vocabulary than about how brains actually work.

It is worth being precise about what we are even comparing the brain to. When people say the brain is “like a computer,” they almost always mean a particular kind of computer — one built on the von Neumann architecture, formalized by John von Neumann in 1945, which underlies nearly every digital machine from a phone to a supercomputer. That architecture has a defining feature: it separates memory from processing. Data and instructions sit passively in a memory unit; a central processing unit fetches them, performs an operation, and writes the result back. This fetch–decode–execute cycle, repeated billions of times per second, is enormously powerful precisely because of the separation. You can change what the machine does just by loading different instructions. But that same separation is exactly where the analogy to the brain breaks down.

The brain violates nearly every assumption of the von Neumann design.

There is no separation of memory and processing. In a computer, memory chips hold information while a processor manipulates it. In the brain, the very same synapses that hold information — in the strengths of their connections — are also what transform the signals passing through them. A synapse simultaneously remembers, through its weight, and computes, by scaling what passes across it. There is no neural act of fetching data from a passive store.

There is no bus, and almost no data movement. Computers depend on physical wires that shuttle data between memory and processor, and this shuttling creates the famous “von Neumann bottleneck” that limits their speed. The brain does not move data around in this sense. When you recall your grandmother’s face, you do not copy a file from long-term storage into a working-memory buffer. The pattern of activity that is your recognition of her arises in the very populations of neurons that learned to recognize her in the first place.

There is massive parallelism with no central processor. A von Neumann machine has one CPU, or a handful of cores, stepping through operations in sequence. The brain has on the order of 86 billion neurons computing at once, with no central executive handing out instructions to passive subsystems. Every neuron is at once a processor and a memory element, responding to thousands of inputs while shaping thousands of outputs.

And physical structure embodies the computation. In a computer the spatial layout of transistors is essentially arbitrary; what matters is the logical wiring diagram. In the brain, physical architecture is destiny. The neurons that track an insect’s heading are literally wired in a ring, and it is the ring shape that lets a bump of activity follow the animal’s rotation. The architecture does not implement an algorithm for heading; it is the mechanism. This is a theme we will return to constantly: in nervous systems, the wiring and the computation are not separable layers.

Here is the deepest version of the worry. The computer metaphor encourages a particular and, I think, misguided research program in which we take a psychological construct — working memory, attention, an “executive” — and go hunting for the place in the brain that implements it. Consider Baddeley’s influential model of working memory, with its phonological loop, its visuospatial sketchpad, its central executive. As an account of behavior it is excellent: it predicts why you can hold roughly as many words as you can pronounce in about two seconds, why verbal and visual tasks interfere differently, why particular lesions produce particular deficits. But Baddeley did not arrive at the phonological loop by looking at neurons. He inferred it from reaction times and error patterns. The behavioral phenomena are real; the box-and-arrow diagram proposed to explain them need have no more neural reality than the vital spirits of the Greeks. When neuroscientists then ask “where is the phonological loop?”, they are smuggling in the von Neumann assumption that a model derived from behavior must map onto distinct hardware.

The philosopher David Marr tried to discipline this with his famous three levels of analysis — the computational level (what problem is being solved?), the algorithmic level (by what steps?), and the implementation level (in what physical stuff?) — and he argued we could study each more or less independently. This is a genuinely useful framework, and I do not want to wave it away. But notice the assumption underneath it: that brains, like computers, cleanly separate what is computed from how it is physically realized. If neural tissue does not work that way — if the physical dynamics are the computation — then the levels do not stay neatly stacked. An algorithmic story worked out without any neural constraints may simply fail to correspond to anything the tissue is doing.

There is an even more radical possibility, associated with the ecological psychologist James J. Gibson, that the brain does not “compute” in the symbol-manipulating sense at all. Gibson was often frustratingly vague about mechanism, but his core intuition is worth holding onto. Ask: what is the best system for working out how water flows around an obstacle? It is not a supercomputer solving the Navier–Stokes equations. It is an actual stream. The water does not calculate its path; its physical properties simply are the dynamics we describe with those equations. By analogy, a neural circuit may not implement an algorithm for perception or movement so much as possess physical dynamics, shaped by evolution and development, that produce adaptive behavior directly. The mathematics we use afterward — Bayes’ theorem, control theory, information theory — captures real regularities in those dynamics. But the brain no more “runs” Bayes’ theorem than the river “runs” fluid dynamics.

I do not insist you accept the strongest form of this argument, and we will meet plenty of places in this book where computational language is the most useful tool we have. The point of raising it here, at the very start, is prophylactic. The computer metaphor has given us a rich and genuinely productive vocabulary, and I will use parts of it without apology. But its core assumption — that brains, like computers, separate data from processing — is almost certainly false to the biology, and it can quietly steer a whole field toward the wrong questions. So when we reach for computational language later, I want us to do it knowingly, as a convenience and not as a discovered truth.

The alternative: architecture first

If cognitive constructs do not map cleanly onto neural hardware, and if brains do not compute like von Neumann machines, then we need a different starting point. The one this book adopts is to reverse the usual direction of inference: begin with the architecture that evolution actually built, and work forward to function, rather than beginning with a psychological theory and searching for its neural correlate.

That reversal rests on a few principles that recur throughout the chapters ahead.

Evolution shapes architecture for behavior, not for computation. Nervous systems were selected to control movement, maintain the body’s internal state, and anticipate environmental change. Whatever “cognitive” capacities we possess are elaborations of these control functions. The architectures we keep finding — the layered cortex, the cerebellar microcircuit, the loops through the basal ganglia — are solutions to control problems, not general-purpose computing modules.

Development embodies assumptions about the world. The program that builds a circuit encodes expectations about the environment it will meet. Visual cortex develops neurons tuned to oriented edges not because it learns to run an edge-detection algorithm, but because the statistics of natural images, fed through a few simple learning rules, make those features nearly inevitable. The architecture, in effect, assumes that edges exist, and grows to detect them. This is the bridge to Chapter 2, where development and evolution become a single story.

Dynamics replace algorithms. Rather than asking “what algorithm does this circuit run?”, we will more often ask “what dynamics does this architecture produce?” The grid cells of the entorhinal cortex do not execute a coordinate-geometry routine; their recurrent wiring creates attractor dynamics whose natural consequence is a hexagonal map of space.

Control, not representation, is the organizing idea. Instead of a brain that builds representations to be processed, we will picture a control system that holds the body stable while pursuing goals. On this view, much of what we call “working memory” may be sustained activity that bridges a temporal gap in control, and much of what we call “attention” may be the dynamics that arise when a control system has to choose among competing possible actions.

None of this means ignoring cognition. The phenomena Baddeley catalogued are real and a theory of the brain must eventually explain them. The claim is only that the explanation is more likely to be found by starting from the machinery evolution built than by assuming that machinery must contain a tidy implementation of our psychological vocabulary.

Control, homeostasis, and allostasis

With that framing in hand, we can state more precisely what we mean by calling the brain a control system — and here we meet the one piece of control theory that will follow us through the entire book.

The familiar image of a control system is a household thermostat. It operates by feedback: it waits until the room temperature falls below a set point, detects the error, and switches on the furnace to correct it. In biology, this reactive defense of an internal variable is called homeostasis, and it is real and indispensable. Your body regulates blood pH, core temperature, and a dozen other variables in just this error-correcting way.

But pure homeostasis — waiting for an error before acting — is often a poor survival strategy, because biological correction is slow. An animal that waits until it is severely dehydrated before it begins to look for water, or until its core temperature is already dangerously low before it seeks shelter, has waited too long. The error itself is the threat. What the brain does better than a thermostat is to act before the error occurs. Using sensory cues and the residue of past experience, it predicts an upcoming need and initiates behavioral or physiological change in advance. This anticipatory, predictive mode of control is called allostasis — stability through change, or more loosely, stability through prediction. The thermostat reacts; the brain forecasts.

I want to plant a flag on this distinction now because it is the through-line of the book. Almost everything that follows can be read as a story about a control system shifting, over evolutionary time, from reactive homeostasis toward predictive allostasis — and about the machinery, and the metabolic cost, that this shift required. Chapter 1, in asking why brains grew large and expensive, is at bottom asking what it costs to buy prediction.

One more commitment: the conserved blueprint

Finally, a promise about method. The human brain has unique functional elaborations, and we will not pretend otherwise. But it is not built from scratch. It is an elaborated version of the same vertebrate brain plan that has existed for more than 500 million years, and evolution is a deeply conservative engineer — it modifies what exists far more often than it invents something new. The same basic neural architectures found in a fish or a frog are running inside you right now. For that reason this book will move freely between species, using a bird, a monkey, or a sea slug whenever that animal makes a piece of conserved control circuitry easier to see. When we finally turn to what, if anything, makes the human brain special (a question Chapter 1 takes up directly), it will be against this backdrop of deep conservation — and the answer, I will argue, turns out to be quantitative rather than magical.

With the frame established, we can ask the first and most basic question of all: why have a brain at all?