Drawing the Map
Place Cells, Grid Cells, and the Machinery of Knowing Where You Are
We have an old structure and a clear function. The hippocampus, we have argued, is the ancient organ of navigation — a derivative of the medial pallium, conserved from fish to human — and its job, in the deep lineages and arguably still in us, is to build a map of the world in the service of predicting where the body’s needs can be met. But so far the “map” has been a metaphor, or at most an inference from behavior: the goldfish acts as if it holds a map; Tolman’s rat takes shortcuts as if consulting one. In this chapter the metaphor becomes literal. We are going to open the skull, lower an electrode into the hippocampus of a moving animal, and watch the map being drawn, one neuron at a time.
The question that organizes everything here is deceptively simple. How can a sheet of neurons — cells that do nothing but receive inputs and fire — represent something as abstract as a place? A place is not a stimulus. It is not a color or a tone or a touch. You can be in the same place facing different directions, in light or dark, smelling different things. What stays constant is the location itself, and somehow the brain holds it. The answer, worked out over fifty years, turns out to require not one kind of cell but a small society of them, each answering a question the previous one leaves open. We will meet them in that order — as answers, not as a list — and at the end we will watch the finished map do the one thing that makes it matter for this unit: run forward, into a future the animal has not yet reached.
The cell that knows where you are
In 1971, John O’Keefe — a psychologist, as we keep noting, because it matters to this unit’s argument that the cognitive map was recovered by Tolman’s intellectual heirs and not handed down from anatomy — lowered a microelectrode into the hippocampus of a freely moving rat and recorded from single neurons while the animal wandered an enclosure. Most of what a brain does while an animal explores is a blur of activity with no obvious pattern. But some hippocampal cells did something startling and specific. A given neuron would fall silent almost everywhere — and then fire in a hard burst whenever, and only whenever, the rat occupied one particular patch of the enclosure. Move the rat to that spot and the cell spoke; move it away and the cell went quiet. A different neuron had a different spot. A third, another.
O’Keefe called them place cells, and the patch of space that drives a given cell is its place field. Here was Tolman’s cognitive map, made of neurons. The hippocampus was not storing a picture of the room; it was tiling the room with cells, each claiming a location, so that wherever the animal stood, some specific subset of place cells was active and that pattern of activity was the animal’s position. The brain represents a place by dedicating cells to it. Knowing where you are is knowing which place cells are firing.
This was the discovery that, more than any other, turned the cognitive map from a psychological idea into a physiological fact, and it is the reason O’Keefe shares a Nobel Prize. But it immediately raises a question it cannot answer on its own. A place cell tells you that you are in a particular spot — it is a label for a location. It does not tell you how the brain measures space in the first place: how it knows that this location is a certain distance and direction from that one, how the map has metric structure rather than being a scattered set of unrelated labels. A map is not just a list of named places. It is named places in measured relation to one another. Where does the measurement come from? For that we have to leave the hippocampus proper and step one synapse upstream, into the entorhinal cortex — and into one of the most beautiful findings in all of neuroscience.
The cell that measures space
If you record from the medial entorhinal cortex — the spatial gateway we met in the last chapter, the one built from the same medial pallium as the hippocampus itself — while a rat crosses a large open floor, you find cells that also fire according to the animal’s location. But where a place cell fires in one spot, these cells fire in many — and the many spots are not scattered. They are laid out in a stunningly regular pattern: a triangular, hexagonal lattice, tiling the entire floor like the cells of a honeycomb. As the animal crosses the enclosure, the cell falls silent and fires, silent and fires, at perfectly spaced intervals, and if you plot every spot where it fired, the dots form a grid.
These are grid cells, discovered by Edvard and May-Britt Moser, and they are the brain’s answer to the measurement problem. A hexagonal lattice is, in effect, a coordinate system — a ruler the brain lays over space, an internally generated metric that does not depend on any particular landmark because it is regular everywhere. Different grid cells have lattices of different sizes — some fine-grained, their firing spots close together; some coarse, spots far apart — and lattices offset from one another in position. Layer these many grids of different scales and offsets on top of one another and you have something remarkable: a representation in which every location in the environment produces a unique combination of grid-cell activity, a kind of coordinate readout precise enough to specify where the animal is to fine resolution. The grid cells supply the measured space; the place cells, receiving grid input among other signals, mark the salient locations within it. The Mosers shared O’Keefe’s Nobel Prize for this, and rightly: between them, place cells and grid cells are the map and the ruler that makes the map metric.
Why a triangular/hexagonal lattice specifically? Because the hexagonal packing is the most efficient way to tile a plane with regularly spaced points — it places each firing field equidistant from six neighbors, the densest symmetric arrangement possible. A nervous system trying to lay down a uniform spatial metric with the fewest cells lands on this geometry for the same reason bees do.
The grid is organized into modules. Grid cells are not a smooth continuum of scales but cluster into a handful of discrete spacings, each module sharing a common lattice scale and orientation, with scale increasing in roughly geometric steps along the dorsal-to-ventral extent of the medial entorhinal cortex (dorsal modules fine, ventral modules coarse). This modular, multi-scale organization is what gives the population its enormous representational capacity: because the modules have incommensurate scales, the combination of phases across modules does not repeat over any reasonable distance, so a relatively small number of cells can specify position uniquely over a large space — the same trick by which a few gears of different sizes in an odometer count a very large number of turns without repeating.
Alongside the grid cells, the same region and its neighbors contain the rest of the surveying toolkit: head-direction cells, which fire according to the direction the animal’s head is pointing, like a neural compass; and border or boundary-vector cells, which fire when the animal is at a particular distance and direction from an environmental edge — a wall, a drop, the limit of the enclosure. Direction, distance-to-boundary, and metric grid together constitute a genuine surveying instrument: a compass, a set of range-finders to the walls, and a coordinate lattice.
The map that exists before the world
We now have place cells marking locations and grid cells, head-direction cells, and boundary cells supplying the metric, the compass, and the edges. The natural assumption — the one a blank-slate view of the brain would make — is that the animal learns all of this from experience: that it explores an environment, gradually discovers its layout, and builds the map from the sensory evidence the way you might survey unfamiliar terrain. That assumption is wrong, and the way it is wrong is one of the strongest pieces of evidence in this entire unit.
When you record from the spatial system of a rat pup as it first begins to venture out of the nest, the cell types do not all appear together, and they do not appear in the order a learning account would predict. The head-direction cells come first — and they are present, mature, and adult-like before the pup has opened its eyes, before it has had any patterned visual experience of the world at all. The neural compass is online before the animal has seen the environment it will navigate. Boundary cells follow close behind, essentially as soon as the pup is exploring. And the grid cells mature later and more slowly, with the precision of the place-cell map improving in step with grid maturation over the following days.
Read that sequence again, because it inverts the blank-slate story completely. The brain does not build its sense of direction from experience of the world; it brings a sense of direction to the world, ready-made, driven by the vestibular system’s report of the body’s own movement and rotation. Heading and self-motion come first because they are built from the body, not from the scene. The animal is equipped, before its first journey, with the machinery to track which way it is facing and how far it has moved. Only then, layered onto that bodily scaffold, does the richer map of external space mature.
This is the developmental face of a claim this unit has made from the beginning: that the map is, at its root, a sensorimotor structure — built from the body in motion, not assembled from sensory snapshots of the world. We will see in a moment why that matters so much. But first we need the piece that connects self-motion to the map, the computation that is doing the real work underneath head-direction and grid cells alike.
Path integration: the dead reckoning beneath the map
There is an ancient navigational trick, older than maps, that a sailor without instruments or a desert ant returning to its nest both rely on: dead reckoning, or path integration. If you know which way you started, and you keep track of every turn you make and every step you take, then at any moment you can compute, from that running tally of your own movements alone, exactly where you are relative to where you began — without any landmark, in total darkness, by bookkeeping on your own motion.
This is the computation the self-motion machinery performs, and it is why head-direction cells and the vestibular sense are the foundation the rest of the map is built on. The desert ant, foraging a meandering path across featureless sand, can turn at any moment and run in a straight line directly home — because it has been integrating its own movements the entire way. The grid cell’s metric is, in large part, this path-integration signal made into a coordinate lattice: a way of keeping a continuously updated estimate of position from the stream of self-motion, anchored and corrected by landmarks when they are available but not dependent on them. The map is not, at bottom, a record of what the world looks like. It is a record of where the body has gone. And a system that tracks where the body has gone is, by its nature, poised to compute where the body will go if it keeps moving — which is the bridge to everything that follows.
The hippocampus during active movement is dominated by a roughly 8 Hz oscillation in its local field potential called the theta rhythm. Place-cell firing is organized with respect to this rhythm in a way that is among the most remarkable temporal codes known in the brain, and it reveals that the map is forward-looking at the finest timescale of its operation.
The phenomenon, discovered by O’Keefe and Recce in 1993, is called theta phase precession. As an animal runs through a place cell’s field, the cell does not simply fire whenever the animal is in the field. When in the theta cycle it fires shifts systematically: on first entering the field, the cell fires late in the theta cycle; with each successive cycle as the animal advances through the field, the firing moves to progressively earlier phases; and as the animal exits, the cell is firing at the very start of the cycle. Firing phase, in other words, encodes position within the field — a temporal code riding on top of the rate code, so that the cell signals not just that the animal is in its field but how far through it has progressed.
The population consequence is the part that matters for this unit. When phase precession is aligned across many place cells with neighboring fields, each theta cycle contains a compressed sequence: within that ~125-millisecond window, the active cells fire in the order of their fields along the animal’s path, sweeping from positions just behind the animal, through its current location, to positions just ahead of it. These are theta sequences, and they mean that on every single cycle of the ongoing rhythm — many times a second, continuously, as a matter of the map’s baseline operation — the hippocampus is representing not a static point but a short trajectory that leans into the future, reaching slightly ahead of where the body actually is. The forward-looking character of the map is not a special mode reserved for difficult decisions. It is built into the clock.
(The mechanism generating phase precession remains debated, and recent work suggests phase precession and theta sequences may depend on partly distinct network states rather than one strictly producing the other. The forward-sweeping population code is robust; its detailed origin is still being worked out — an honest seam, like the teleost homologies of the last chapter.)
From reading the map to running it
Everything so far describes a map the animal reads as it moves: place cells reporting where it is, grid cells measuring the space, the whole system continuously updated by self-motion. But a map you can only read is inert — a position display, no more. The thing that makes the hippocampal map the engine of prediction this unit has promised is that the animal can run the map offline, decoupled from its actual body, to represent places it is not and journeys it has not taken. There are two great instances of this, and together they convert the map from a display into a simulator.
The first we met in the overview, and we can now place it precisely in the machinery. When a rat pauses at a junction in a maze — Tolman’s vicarious trial and error, the pause-and-look that so embarrassed the behaviorists — the place-cell representation does not sit still at the animal’s body. It sweeps forward, down first one arm of the maze and then the other, the cells representing positions ahead of the animal activating in sequence as though the rat were mentally running each possible route before choosing. This is the theta-sequence machinery of the deeper-dive box, recruited and extended into deliberation: the same forward-leaning population code that leans slightly ahead on every theta cycle, now thrown far down the unchosen paths to evaluate them. The map is being used to simulate — to generate the experience of going somewhere the body has not gone, in order to decide whether to go there. And tellingly, these sweeps are most prominent early in learning and at hard choices, when the future is genuinely uncertain, and fade as the route becomes habitual and no simulation is needed. This is what a deliberative, predictive system should look like: it runs the map forward exactly when prediction is worth the effort.
The second instance happens when the animal stops moving altogether — resting, or asleep. The theta rhythm gives way to brief, explosive bursts of synchronized activity called sharp-wave ripples, and during these bursts the place cells fire again — but compressed in time, sped up many-fold, replaying the sequences they expressed when the animal actually ran a path earlier. This is replay: the map running a journey at high speed while the body lies still. Replay runs forward and backward; it revisits paths recently travelled and, sometimes, paths never travelled at all, stitched together from pieces. And it is not idle. When replay is disrupted, animals fail to consolidate what they learned about a maze; when it is permitted, they learn. The offline map is training the rest of the brain, broadcasting its compressed journeys to the cortex — which connects this chapter directly to the consolidation machinery we will examine when we finally turn to human memory and the patient who could not form it.
Put the two together — the awake forward sweep and the resting replay — and the static map has become a generative one. The hippocampus does not merely tell the animal where it is. It runs journeys the animal is not taking: ahead, to evaluate a choice; offline, to learn from the past and rehearse the future. This is the mechanism beneath the thesis of the whole unit. The map predicts because the map can be run, decoupled from the body, forward into places and times the animal has not yet reached.
A complication we will not duck: maps of things that are not places
Before we leave the machinery, honesty requires opening a door that leads into the final unit of this book, because the cells we have just described turn out not to be quite as purely spatial as the story so far implies — and this is the single most important challenge to the way this unit has framed things.
The cells of the hippocampal-entorhinal system, it turns out, do not only map physical space. When humans learn the structure of an abstract space — for instance, a two-dimensional space of stimuli that vary along two continuous features — functional imaging reveals a grid-like, six-fold-symmetric code in the entorhinal cortex and prefrontal cortex, the same hexagonal signature the grid cells use for physical rooms. The finding has been extended to other non-spatial continua, and in rodents, hippocampal and entorhinal cells have been shown to map a non-spatial variable — the frequency of a tone the animal controls — with the same kind of structured firing they use for location. The map machinery, it appears, can be applied to conceptual and relational structure, not only to space.
This is the strongest evidence for a rival framing of everything in this unit, and an honest text must say so. On that rival view — associated with the relational-memory and “cognitive spaces” traditions — physical space is not the fundamental thing the hippocampus represents at all; it is merely the most studied special case of a general capacity to map relational structure of any kind, spatial or not. If that view is right, then this unit has the priority backwards: the system would not be a spatial map later borrowed for abstraction, but a general relational engine of which spatial mapping is one application.
Why, then, has this book insisted on spatial primacy? Not on the grounds that conceptual maps are mere metaphor — the grid-in-concept-space findings show the actual metric transfers, which is more than metaphor. The case for spatial primacy rests instead on what we saw in this chapter and the last: the system is, in its developmental and evolutionary origins, sensorimotor. The head-direction compass is built from the vestibular sense of the body’s own motion, online before the eyes open; the metric is grounded in path integration over real physical movement; the structure is identifiable, and spatial in function, in fish that diverged from us half a billion years ago. The rival view has no comparable account of why, if the code is fundamentally abstract-relational, that particular instance — physical space, built from self-motion — should be the one that is innate, vestibular, and evolutionarily ancient. The most defensible reading is therefore that the map is spatial at its root and exapted for abstraction: a metric built for a body moving through the world, later turned on conceptual spaces by a brain that had the machinery lying ready. The grid in concept-space is the spatial system doing new work, not evidence that it was never spatial.
This is a genuine and unsettled debate at the frontier of the field, and you should hold it as one. We flag it here, at the end of the mechanism chapter, because it is precisely the hinge on which the next unit turns — the unit in which the human brain takes this ancient spatial machinery and runs it forward into language, into other minds, into past and future selves. The map that began by finding food in the water is about to be asked to navigate far stranger spaces.
The short version, for the main thread: the same machinery that maps rooms can map abstract structure too, and that fact is the bridge into the human material that closes the book. Whether you read it as a spatial system exapted for abstraction (this book’s view) or as a general relational system of which space is one case (the principal rival), the cells we have met in this chapter are the substrate. The disagreement is about priority — which came first, and what was built on what — and that is exactly the kind of question the deep-history chapter equipped us to weigh.
Where we have arrived
We set out to answer how a sheet of neurons can represent a place, and we have the answer in full. Place cells mark locations; grid cells lay down the metric that makes those locations a measured space; head-direction and boundary cells supply the compass and the edges; and beneath all of it, path integration over self-motion keeps a running estimate of where the body has gone — the sensorimotor foundation on which the whole map is built, online before the animal has even seen the world. And the map is not inert. Through theta sequences, forward sweeps, and offline replay, it runs — decoupled from the body, ahead into choices and back through experience — which is the mechanism that makes it the predictive engine this unit has claimed it to be.
We now hold all three legs of the argument. The function: prediction, the use of the past to prepare for the future. The deep history: the medial-pallial map, ancient and conserved from fish to human. And now the mechanism: the cells and dynamics that draw the map and run it forward. Only one thing remains, and it is the thing every other textbook puts first and this one has deliberately saved for last. We have built up the ancient, universal, spatial machinery of memory from the bottom — its purpose, its origin, its cells. Now, at last, we are ready to ask what happens when it is damaged in a human being, and what the most famous patient in the history of neuroscience can tell us when we meet him not as the beginning of the story of memory, but as its culmination.