How an Ancient Map Organizes Time, Tasks, Concepts, and Social Relations
46.1 The fish that expressed an untrained relation
Chapter 45 began with a mobile animal leaving a refuge and ended with population activity extending beyond the body’s present location. Movement and sensation supplied relations that remained useful when a destination was out of sight. The next question is how far that account extends when the relevant relations are not distances or routes through the environment.
The question does not begin with human abstraction. It begins with a goldfish.
A goldfish learns four overlapping discriminations: A is rewarded over B, B over C, C over D, and D over E. Each pair can be learned separately. The critical test presents B and D together for the first time. Choosing B is consistent with the order implied by the trained pairs, although that particular choice has never been reinforced.
Control fish express the preference. Fish with lesions of hippocampal pallium learn the premise discriminations but are impaired on the novel B-versus-D test [@soteloparrilla2025transitive]. The comparative significance of the lesion belongs to the preceding discussion of related systems in differently built brains. Here the result matters because it separates learning particular choices from using their overlap in a new comparison.
Transitive inference names the structure of the test. It does not establish that a fish consciously recites a syllogism, nor does the preference uniquely reveal an internal ordered line. Reward history and associative processes must also be considered when explaining performance. A relational account becomes useful when it explains transfer across trained pairs better than accounts based on their separate acquisition. The selective lesion deficit strengthens the case that the circuit contributes to this transfer; it does not settle how the relation is represented.
The fish may swim to make its choice. What makes the test nonspatial is that the critical relation concerns learned items, not a novel route through the tank. The important step is from a familiar set of encounters to a choice those encounters did not individually specify.
Navigation remains the biological starting point because it makes this problem unusually clear. A place is defined by relations among boundaries, directions, routes, recent movement, and possible outcomes. A detour preserves a destination while changing how it is reached. The goldfish test asks whether an analogous flexibility can operate over learned order. It neither reconstructs the behavior of an ancestral vertebrate nor requires every nonspatial capacity to have evolved after navigation.
The central question of this chapter is therefore precise: When does the map claim explain nonspatial performance? Time, sound, task structure, concepts, and social relations offer different answers because the experiments establish different kinds of relation.
46.2 What deserves to be called a map?
The word map is useful only if it excludes something.
A neural population that responds differently to two stimuli has distinguished them. A list has an order. A category has members. An analyst can arrange almost any collection of activity patterns in a two-dimensional plot. None of those facts alone establishes a representation that the nervous system uses to relate several states or to go beyond a familiar response.
The account of transitions in Section 45.7 provides the starting vocabulary. A state is the information needed to distinguish the current situation for the task. It may include a location, a moment within a sequence, a tone frequency, or a context defined by recent events. A transition specifies how that state changes after an action or event. Some nonspatial experiments instead test an ordered or metric relation between items without requiring a sequence of actions. A map interpretation must specify which of these structures is being proposed.
A useful map preserves relations that support prediction, traversal, recombination, or inference. It need not resemble a floor plan. An interval has temporal order; a task graph has connectivity; a deliberately constructed feature space can have distances and directions. The question is whether preserving those relations explains something that recognition of isolated items leaves unexplained.
Several forms of evidence bear on that question, but they answer different parts of it.
| Evidence | What it can establish | What remains to be shown |
|---|---|---|
| Behavioral transfer | Experience supports a new comparison or prediction. | Which representation or circuit produced the transfer. |
| Direct neural recording | Individual cells or ensembles distinguish task states and sometimes preserve their order. | Whether the animal uses the measured activity for the relevant inference. |
| Population or fMRI geometry | Neural similarity or response modulation follows a specified relational model. | The cellular implementation and the contribution of alternative models. |
| Computational model | Explicit assumptions generate a code or behavioral prediction. | Whether the biological system implements those assumptions. |
| Selective perturbation | A circuit contributes causally to performance under the tested conditions. | Which operation or represented content accounts for the deficit. |
These are complementary tests, not interchangeable rungs on a ladder. A lesion can establish a causal contribution without revealing the code. A recording can reveal structured activity without demonstrating causal necessity. A model may explain both, but only if its distinctive predictions survive comparison with alternatives. The strongest account brings the forms of evidence together while keeping their scope visible.
The goldfish example illustrates why this discipline matters. Preserved premise learning constrains a general performance explanation. Impaired transfer makes the lesion consequential for using overlapping information. Calling that contribution relational is justified; calling it proof of a particular map geometry would go further than the experiment.
Not every relation is a map, and not every map is Euclidean. The map claim becomes explanatory when a specified relational structure accounts for what the organism can predict or infer beyond an isolated present cue.
46.3 Time: when the route is a sequence
Physical travel supplies a succession of states because the body moves. A temporal task can require a succession even when location offers little information about how far the task has progressed.
In delay conditioning, a predictive stimulus remains present when the event it predicts begins. In trace conditioning, the stimulus ends before that event, leaving an interval to bridge. Goldfish hippocampal-pallium lesions impair trace conditioning while sparing delay conditioning [@gomez2022trace]. The spared condition argues against a wholesale inability to detect the events, produce the response, or learn their association. The contrast identifies a contribution under temporal separation; it does not, by itself, reveal a sequence of time cells in fish.
Mammalian recordings address the representational question more directly. Christopher MacDonald and colleagues trained rats to use an object to determine their response to an odor encountered after a delay. Different CA1 neurons were active at successive moments in the interval. Analyses considered location and ongoing behavior as well as elapsed time; the rats were not motionless. Changing delay duration left some timing relations stable while other cells changed their temporal firing patterns, a response termed retiming [@macdonald2011time].
The resulting time-cell sequence makes successive portions of the interval distinguishable. It is not a dedicated stopwatch with one permanently assigned neuron for each second. Activity can also depend on the preceding object, location, and behavior. Those dependencies are relevant to the task: the animal must retain which sequence is underway, not merely register that time has passed.
The map interpretation is useful to the extent that the population represents a position within an organized episode. The earlier event is absent, yet the current state can remain related to it and to what is expected next. That is a possible contribution to temporal bridging, rather than proof that the recorded sequence is its only mechanism. The fish lesion and rat recording converge on a functional problem through different experiments; they do not demonstrate an identical cellular implementation across species.
This interval-spanning sequence must also be distinguished from the compressed theta sequences described in Section 45.8 and ripple-associated activity in Section 45.10. All express temporal order, but their timescales, physiological states, and measurements differ. The shared idea is that population activity can preserve relations among events that are not simultaneously present.
46.4 Moving through a tone
Time changes whether the animal acts or not. Dmitriy Aronov, Rhino Nevers, and David Tank designed a task in which an animal controlled progress through a nonspatial variable.
A rat deflected and held a joystick as a tone rose along a logarithmic frequency axis, then released it within a target frequency range to obtain reward. Randomly varying the sweep rate across trials separated frequency from elapsed time. Slower sweeps elicited later releases, consistent with frequency-guided performance. Recordings in dorsal CA1 and medial entorhinal cortex revealed fields tied to frequency as well as activity tied to task events. Some participating neurons also had place or grid fields during foraging [@aronov2017nonspatial].
The playback controls sharpen the interpretation. Unrewarded passive sweeps recruited little CA1 activity; adding reward recruited more, although modulation was less prevalent and temporally precise than during active performance. Action–sound coupling therefore enhanced the representation but was not strictly necessary to engage it. Varying sweep rate also allowed frequency alignment to be compared with timing relative to joystick press or release.
The result is striking because tone frequency is not physical location. A high tone is not north of a low tone. Yet frequency became a state variable within a controlled task. Holding the joystick allowed progress; auditory feedback indicated the changing state; release converted a suitably timed encounter with the target into an outcome. The action did not provide unrestricted movement toward and away from arbitrary auditory destinations. The task imposed a particular progression and a stopping decision.
This distinction gives the map interpretation both substance and limits. Fields along the relevant variable make it possible to distinguish progress through the task. Their occurrence in populations also used during navigation supports shared representational resources. It does not establish a hexagonal auditory lattice, an untrained shortcut through sound, or causal necessity for the measured cells.
The control perspective remains useful without those stronger claims. A controller needs information that predicts the consequence of continuing or stopping an action. Here that information concerned sound. The experiment shows how a nonspatial variable can acquire the operational importance that location has during navigation.
46.5 Learning a graph that was never shown
A continuous axis is only one kind of map. Many problems are better described as graphs. A graph contains nodes, representing states, and links, representing possible transitions. Two subway stations can be physically close but poorly connected; distant stations can be joined by an express line. Reachability depends on the network.
Mona Garvert, Raymond Dolan, and Timothy Behrens exposed participants to object sequences generated by random walks through an undisclosed graph. Cover tasks directed attention to the objects without instructing participants to learn the network. On the next day, objects appeared in random order during fMRI. Adaptation of the BOLD response varied with relations learned from the earlier sequence, including distances beyond immediate neighbors. Participants did not report explicit knowledge of the graph [@garvert2017map].
The blood-oxygen-level-dependent, or BOLD, signal is an indirect measure of neural activity. Here, adaptation concerns how the response to an object depends on the preceding object; greater suppression is used as an index of related representations. This was not a recording of neurons traversing a graph. A separate behavioral sample supplied an implicit response-time signature. Among the tested distance measures, expected future occupancy provided a particularly good account of the neural and behavioral effects. The experiment therefore supports learned structure extending beyond isolated adjacent pairs, while leaving open how that structure is read out for a new goal-directed choice.
The successor representation in Section 45.7 makes the model comparison concrete. It summarizes expected, discounted visits to future states under a particular transition process or policy [@stachenfeld2017predictive]. States can resemble one another because their expected futures overlap. That is different from arranging them by physical distance and different from storing only the next observed item.
In the hidden-graph experiment, the sequence generator supplied the transition statistics. In active behavior, a policy helps determine which transitions are experienced. This difference matters when transferring the model to control. A learned prediction under one policy does not automatically describe every available route. A new goal can change value without changing connectivity; a new action policy can change expected occupancy even when the connections remain the same. A blocked connection requires another kind of revision. The formal account is useful because it distinguishes these changes rather than treating them all as a map becoming stronger.
Anna Schapiro and colleagues addressed a related question: how can local encounters produce larger event groupings? Participants experienced sequences from a graph containing communities. Each available immediate transition had the same probability, so a community boundary was not simply an unusually improbable next step. Nevertheless, behavioral segmentation, fMRI adaptation, and cortical pattern similarity reflected the community organization [@schapiro2013neural]. Items could become grouped through their wider temporal relationships.
The two studies should not be collapsed into one experiment. Garvert tested relational distances and compared predictive metrics; Schapiro tested community structure and event organization. Neither requires participants to reconstruct an explicit drawing. Nor does either, on its own, prove that a particular hippocampal cellular mechanism is necessary for all graph learning.
For a controller, the attraction of graph structure is practical. A bottleneck controls access to many futures. A community can identify a context in which transitions tend to remain. A loop permits return to a previous state. These properties are not captured by straight-line distance alone. The map claim explains nonspatial performance when such learned structure improves prediction or transfer, and when its measured consequences distinguish it from simpler accounts of familiarity and pairwise association.
46.6 From feature space to concept space
Graphs establish that a map can be discrete. Human experiments also ask whether a continuous nonspatial domain recruits a signature associated with spatial grid coding.
Alexandra Constantinescu, Jill O’Reilly, and Timothy Behrens constructed a domain of cartoon birds varying in neck length and leg length. Participants learned relations within this domain and, during scanning, watched a bird change before imagining the continuation of that change. The two rates of change defined a direction through the feature space. Entorhinal and connected cortical BOLD signals showed sixfold directional modulation [@constantinescu2016conceptual].
The geometry matters because the prediction was more specific than increased activity during a difficult judgment. Directions separated by approximately 60 degrees were treated as related by the grid-like model. Inferences about a shared orientation could be tested against the measured responses. This provides a bridge between the spatial physiology in Chapter 45 and a learned nonspatial domain.
The bridge does not erase the difference between methods. A grid cell is identified from its firing fields. A grid-like fMRI signal is a population-level signature compatible with a proposed organization. BOLD imaging does not directly resolve the underlying fields, and sixfold modulation alone does not uniquely establish their cellular generator or its causal necessity. Nor does a deliberately constructed two-dimensional task demonstrate that all concepts have hexagonal geometry.
There is another issue. Neck and leg length are features whether or not they organize a useful concept. Stephanie Theves, Guillén Fernández, and Christian Doeller separated the full feature space from the dimensions relevant to categorization. Participants learned abstract stimuli varying along three dimensions and associated them with everyday objects. Two dimensions defined the concept; the third remained relevant to remembering the stimulus. During scanning, hippocampal adaptation to the associated objects was better explained by distance in the concept-relevant plane than in the full three-dimensional space [@theves2020concept].
Comparisons with alternative two-dimensional planes argued against a simple preference for two dimensions. Recall testing also gave no evidence that participants had merely forgotten the irrelevant feature. The result concerns task-dependent distance coding inferred from fMRI adaptation, not direct observation of a neural coordinate sheet.
This selectivity is what makes the map account useful for control. A world contains more variables than a controller can represent with equal precision. A useful state description preserves differences that predict consequences and compresses differences that do not alter the current judgment. The auditory task made one sound dimension consequential. Concept learning made a relation between selected features consequential. The representational question is which differences survive that selection, and why.
A conceptual map is a task-shaped relational structure. Its value lies in preserving the dimensions needed for a prediction or inference, rather than reproducing every property of an object.
46.6.1 Development of relational performance
Yukun Qu and colleagues extended this approach in 2026 to 203 participants aged 8–25. Participants learned neighboring rank relations among 25 objects on attack and defense dimensions, without seeing their latent two-dimensional arrangement. Later judgments required untrained comparisons; a second task introduced new objects through local relations and tested inference to more distant ones [@qu2026development].
The neural measures were fMRI measures. The investigators tested sixfold directional modulation, cross-validated consistency with an estimated orientation, and representational similarity analysis, which compares multivoxel patterns with a specified relational model. An RSA-derived entorhinal grid-like measure covaried with age and inference performance. The reported demonstration in high performers involved 75 participants; age-related analyses also examined the broader sample. Medial prefrontal distance-related BOLD responses supplied a distinct measure of relational organization. These were not recordings of individual developing grid cells.
Controls addressed several simpler explanations. Participants had to reach a memory criterion; analyses considered memory performance and training exposure. Motion was screened and modeled. Other angular periodicities and Euclidean distance were examined as alternatives to the sixfold effect. The youngest participants could still make some successful inferences, with performance depending on the comparison’s demands. The results therefore do not mark an age below which relational inference is absent.
The developmental conclusion is an association among age, performance, and neural organization in this task. Because different people were compared across ages, it does not establish an individual’s maturational sequence. Statistical mediation does not turn that association into a causal prerequisite. The reported relationships with standardized reasoning measures likewise do not establish that strengthening a grid-like signal would increase intelligence. No evolutionary chronology follows from these age differences.
Within those limits, the interpretation remains substantial: increasingly effective inference can accompany a more coherent organization of learned relations, rather than merely better recall of the trained pairs. Together with the feature-selection result, this suggests that development may change how information is organized for use. It leaves open which mechanisms produce the change and whether other tasks depend on the same geometry.
46.8 Evolutionary continuity without a simple chronology
The extension from location to time, task, concept, and social relation raises a historical question. Which operations were inherited, which were elaborated, and which acquired new uses?
Navigation has explanatory priority because a mobile animal must estimate its relation to a world extending beyond current sensation. Food, water, shelter, and escape give that ability immediate biological consequence. The preceding chapters established the comparative and physiological basis of the argument. Their evidence does not need to be repeated to recognize the continuity: distinguishing states, preserving relations among them, and updating those relations can be useful in several domains.
The continuity is an interpretation of converging evidence, not a recovered inventory of the first vertebrate’s mental capacities. Living goldfish have their own evolutionary history. Their performance constrains claims that flexible nonspatial relations require uniquely human machinery, but it does not date the origin of each capacity. Similarly, developmental order within a rat or age-related fMRI differences in humans cannot be substituted for an evolutionary sequence. Section 45.6 explains why temporal precedence and causal dependence must also remain separate within development itself.
Expanded association systems and culturally transmitted knowledge greatly change the information available to human hippocampal–entorhinal circuitry. Language can specify an unseen relation; a diagram can expose a structure that would otherwise require many encounters; an institution can define roles and obligations. These inputs alter what can count as a relevant state. The experiments considered here support an inherited circuit’s participation in such tasks without showing that every form of structured thought is implemented by that circuit.
Stephen Jay Gould and Elisabeth Vrba introduced exaptation to distinguish a feature’s present use from the historical circumstances in which it arose [@gouldvrba1982exaptation]. Some uses considered here are plausible examples. An ancestral vertebrate circuit was not selected for learning fictional birds or navigating an organizational chart.
It does not follow that temporal bridging and every nonspatial relation appeared only after a wholly spatial organ was borrowed for other purposes. Modern performance can combine conserved operations, lineage-specific elaboration, and recruitment for novel tasks. Navigation supplies the biological anchor of this account. Establishing the chronology of those different contributions requires comparative evidence beyond the human laboratory task.
46.9 Where the map analogy stops
The success of the map idea creates its own danger. Once any structured representation is called a map, the word explains nothing.
A neural pattern does not become a cognitive map because an analyst can plot it. Spatial language is insufficient, and a category hierarchy is not automatically traversed as a state-transition structure. Familiar words and established knowledge do not become hippocampal simply because they are related to other words and facts.
The useful tests concern consequences. Does the proposed geometry explain neural responses beyond sensory similarity? Does a changed transition structure alter prediction? Can a learned relation support a new comparison? Does circuit disruption selectively affect that use? An experiment need not answer every question to contribute. It must identify which question it has answered.
Different domains need not share one geometry. A time-cell sequence is ordered but not a two-dimensional plane. The auditory task supplies a continuous axis. A graph supplies connectivity. The bird task deliberately constructs a metric feature space, while the social task assigns coordinates to evolving relations. Even within one domain, the organism’s representation need not preserve all properties of the experimenter’s diagram.
Nor should the analogy be limited by false mathematical restrictions. Graphs can contain cycles, and a circular variable returns to its starting value after a full rotation. Nearness need not be transitive: if A is near B and B is near C, A need not be near C. The relevant constraint is whether the proposed structure preserves the relations needed for the behavior, not whether it fits a familiar sheet of graph paper.
High dimensionality poses a selection problem. The controller must retain enough information for the present demand without representing every feature equally. Theves’s experiment supplies a concrete example of relevant dimensions shaping a measured distance signal. But such compression can become costly when the task changes. A previously irrelevant feature may become decisive, requiring the organism to recover or learn distinctions that its earlier state description ignored.
Formal proof, grammar, and deeply nested rules cannot be assigned to the hippocampus merely because their elements stand in relations. Reading this sentence depends on visual processing, language, established knowledge, attention, and task control. Hippocampal relational organization may contribute when new information must be integrated or a particular context recovered. That contribution does not make the hippocampus the organ that reads, reasons, or understands.
Finally, structure alone does not determine what should be done. A map can preserve routes to both water and shelter. Thirst and threat alter their current consequences. A policy determines which action is selected, and that action changes subsequent experience. Their interaction makes the map useful to a body; it does not dissolve the distinctions among relation, value, and choice.
Shared operations do not require identical representations. The map claim explains behavior only to the extent that a specified structure helps account for a defined prediction, inference, or action.
46.10 Coda: from structured domains to particular events
This chapter began with a fish expressing an untrained relation and moved through intervals, tones, hidden graphs, concepts, and social encounters. Across these examples, the map interpretation had to earn its place by identifying what was related, how that relation was measured, and what it could explain.
A structured domain is not yet a memory for one particular occurrence. The same room can contain breakfast today and an argument tomorrow. The same person can cooperate in one context and defect in another. A familiar route can acquire a barrier. A useful controller must sometimes preserve the occurrence that qualifies the general relation, rather than immediately averaging it into what usually happens.
The population distinctions and ordered activity developed in Chapter 45 supply possible ingredients for keeping overlapping encounters distinguishable. The next question is how particular encounters remain available when bodily need or circumstances change.
Chapter 47 begins with exploration and the acquisition of information before a need makes it urgent. It then turns to the loss of event continuity in Clive Wearing and, subsequently, to H.M. and the interpretation of his amnesia. Its problem is how particular events keep a general map useful in a changing world. The history of memory theory and the competing accounts of consolidation belong to that discussion.
Established findings. The cited goldfish experiments separate trained or temporally contiguous performance from more demanding relational use. Rat recordings reveal interval-related sequences and fields within an auditory task. Human fMRI studies identify responses consistent with learned graph relations, selected conceptual dimensions, and experimentally defined social geometry. Developmental comparisons link some of these neural measures with age and inference performance.
The chapter’s organizing interpretation. Navigation is the biological starting point for understanding how a regulated body can use relations beyond the present cue. Related operations can organize nonspatial states, but the explanatory claim depends on the actual structure and task. Map structure, current value, and policy interact within a distributed controller.
Open questions. Behavioral inference does not uniquely identify a code; fMRI geometry does not resolve its cellular implementation; and age-related association does not establish a causal prerequisite. Which representations are necessary for particular nonspatial inferences, how goals select relevant dimensions, and when predictive models outperform simpler alternatives remain questions for experiments. Comparative evidence does not yet establish a universal chronology from spatial to nonspatial cognition.