Is hippocampal remapping the physiological basis for context?

Abstract

In 1980, Nadel and Wilner extended Richard Hirsh's notion that the hippocampus creates environmental representations, called “contexts,” suggesting that the fundamental structure of context was the spatial representation proposed by O'Keefe and Nadel's landmark book, The Hippocampus as a Cognitive Map (1978). This book, in turn, derives from the discovery that individual hippocampal neurons act as place cells, with the complete set of place cells tiling an enclosure, forming a type of spatial map. It was found that unique environments had unique place cell representations. That is, if one takes the hippocampal map of a specific environment, this representation scrambles, or “remaps” when the animal is placed in a different environment. Several authors have speculated that “maps” and “remapping” form the physiological substrates for context and context shifting. One difficulty with this definition is that it is exclusively spatial; it can only be inferred when an animal locomotes in an enclosure. There are five aims for this article. The first is to give an historical overview of context as a variable that controls behavior. The second aim is to give an historical overview of concepts of place cell maps and remapping. The third aim is to propose an updated definition of a place cell map, based on temporal rather than spatial overlaps, which adds flexibility. The fourth aim is to address the issue of whether the biological phenomenon of hippocampal remapping, is, in fact, the substrate for shifts in the psychological phenomenon of context. The final aim is speculation of how contextual representations may contribute to effective behavior.

1 ∣. INTRODUCTION

“Context” is a psychological term that refers to a complex set of environmental cues that influence learning and behavior. “Remapping” is a physiological term referring to an abrupt change in the otherwise stable firing properties of the set of hippocampal place cells. Over the past four decades, two related notions have cropped up. The first is that the hippocampus is the brain locus of “context,”—that is, the substrate for learned associations. This hypothesis was first proposed by Richard Hirsh, and has been substantially and repeatedly updated by Lynn Nadel, Jeffrey Wilner, and others. (Hirsh, 1974; Nadel & Willner, 1980). Support for the hypothesis comes largely from hippocampal lesion studies, in recent years focusing on the role of the hippocampus in what has been called “contextual fear learning.”

The second notion is that hippocampal place cell remapping is the neuronal process that underlies shifts in contextual frameworks. The notion that the hippocampus maintains multiple maps was first proposed by O'Keefe and Conway and the nature of remapping was described in several papers from the Brooklyn Hippocampal group in the 1980s (Kubie & Ranck, 1983; Muller & Kubie, 1987; O'Keefe & Conway, 1978). The term “remapping” was introduced and defined in 1991 (Bostock, Muller, & Kubie, 1991; Kubie & Muller, 1991). A number of authors have suggested that hippocampal remapping is the neural substrate for a shift in context. For example, in a recent review of the functions of the hippocampus, John Lisman wrote:

Now it seems that electrophysiology opens the door to a measurement-based approach with a clear definition: a new context is one that is sufficient to evoke global remapping (Lisman et al., 2017).

The goal of this article is not to defend Lisman's remark—although we are sympathetic. Rather, the goal is to examine what is meant by “context” and “remapping.” Clarifying these concepts may lay the groundwork to address two questions: what happens during remapping? Is remapping the substrate of context? We will outline two distinct notions of “context” and propose a revised definition of remapping.

2 ∣. CONTEXT

Although definitions of context vary, it typically refers to static environmental cues that influence behavior. For example, context could be a testing chamber in which one of Pavlov's dogs is conditioned to a tone. Nadel and Wilner describe two ways that context could play a role in conditioning. On the one hand, context may be an amalgam of cues, not fundamentally different from a simple conditioned cue, such as a transient tone or light. In this way of thinking, the static (contextual) cues may merge with the transient cue to form a single complex cue, both contributing to the learned association. This is the view proposed in the influential Rescorla–Wagner associative learning theory (Rescorla & Wagner, 1972). On the other hand, context may be a set of cues that acts as a foundation, or background, on which foreground cues can operate. In this second sense, context does not merge with foreground cues, rather it is a conditional; if Context A is in the background, then Stimulus B predicts event C. Nadel and Wilner argue for this hierarchical role for context in which it is both made up of and predicts the same items. They note that this view is not specific to hippocampal theories and could be found in the early 20th century in the work of Konorski (1967) and Bolles (1985).

Richard Hirsh's 1974 theory set the stage for relating contextual processing to the hippocampus. The theory suggested three things: first, that context was a background, or hierarchical cue set; second, that context supported indexed retrieval of information from memory (context was the indexing system); and third, that the hippocampus was the brain region responsible for contextual retrieval. Hirsh's theory rested on Scoville and Milner's discovery of hippocampal amnesia and a survey of animal learning theory.

3 ∣. CONTEXT AND THE HIPPOCAMPAL MAP

Six years later, Nadel and Willner (1980) extended Hirsh's notion, suggesting that the physiological basis for “context” was the “cognitive map” constructed from hippocampal place cells.

“In contrast to the traditional view that a context is merely a traditional CS, we propose that contexts are superordinate to such CSs. Within this hierarchical relation, a context both contains and predicts CSs. This approach to environmental context derives from the cognitive map theory (O'Keefe and Nadel, 1978)…”

And, in the section “Rethinking context”

“Why does a certain set of stimuli in a given situation come to have the special status of defining context? What exactly is context for an animal, and how is it learned about and internally represented? Current data suggest the answer goes something like this: when put into a new environment, animals will, after some initial caution, explore the whole of it. As a result of this exploration, the animal learns about the situation; the outcome of this learning process is the construction within the brain of a ‘map’ of the environment. This internal map represents objects and suchlike standing in appropriate relations to one another. Once an element in the situation has been noticed and explored, it will be represented in its appropriate location in the map; such maps are the embodiment of environmental context.”

Thus, we have a bold triad of hypotheses: First, that context is hierarchical; second that context is critical for episodic memory storage and retrieval; and third, that the hippocampal cognitive map is the physiological basis for context. In later sections of this article, we will delve into what is a hippocampal map and what it means for the hippocampus to store multiple maps (remapping). However, first, does the Hirsh–Nadel–Wilner concept hold up to close inspection?

In the subsequent years fear conditioning, and in particular “contextual fear conditioning” has been a major battlefield to address the notion that the hippocampus is critical for contextual learning. First, consider “tone fear conditioning.” The animal is placed in an enclosure and a CS (tone) is turned on and followed by a shock. If on subsequent tests, the CS alone induces a fear-like response (freezing, defecation, etc.) there has been fear conditioning. Importantly, tone fear conditioning does not depend on the hippocampus. In contextual fear conditioning, an animal is pre-exposed to a novel enclosure. Afterward, it is placed in the enclosure and shocked. When the animal is introduced to the enclosure later, if it shows a fear-like response, it has been conditioned to the “context.” Is this a model for the Hirsh–Nadel–Wilner notion of context, and does this form of learning depend on the hippocampal mapping system? Reviews by Nadel (2008) and Rudy (2009) come to a similar conclusion that contextual fear conditioning may or may not require the hippocampus; the answer depends on the method of training and testing. According to both authors, there are two ways that a context (such as the enclosure) can be learned in the contextual fear paradigm, either by the hippocampal system or by the neocortical system. Rudy's notion is illustrated in Figure 1 (adapted from the paper). In this view, the hippocampal system supports “hierarchical associations” and the neocortical system supports “elemental associations”; Nadel refers to the hippocampal representation as “background” and the neocortical representation as “foreground.”

FIGURE 1.

Jerry Rudy's model of associative and hierarchical representations. In the elemental associative model, each individual stimulus item contributes independently to associate with an “event” (E). In the hierarchical model, a group of simple stimuli combines into a unit, much like a spatial map. This unit is the setting (or framework) in which an event can reside. Rudy speculates that each of these occurs, with the elemental relying on neocortex and hierarchical on the hippocampus (modified from Rudy, 2009)

Although this is a large literature, with varied findings, we will list three summarized by Rudy and Nadel:

1. If context fear is conditioned in an intact animal, then posttraining hippocampal lesions will eliminate the fear response; conversely, if an animal has its hippocampus removed prior to training, contextual fear can be learned. Rudy concludes that in an intact animal the hippocampal and neocortical systems are in competition to acquire contextual fear, with the hippocampal system primary, outcompeting and suppressing the neocortical system. In an animal without a hippocampus, the neocortical system is not suppressed and can form the substrate for contextual fear conditioning.

2. If an animal is pre-exposed to an environment and later exposed to the same environment paired with shock, the ensuing conditioning will be hippocampus dependent. However, if an animal is placed into a novel environment and immediately shocked, the ensuing fear conditioning will not require an intact hippocampus. Both Nadel and Rudy suggest that during the initial exposure without shock, the background context is encoded and stored by the hippocampus, forming the basis for later learning. In conditioning without pre-exposure, the environment plus shock form a single complex stimulus that is encoded and stored outside of the hippocampus.

3. “The hippocampus seems to be necessary for acquisition of context fear, and for retrieval of such fear for some days (or weeks) after initial training, but not for retrieval 28 days after training” (Nadel, 2008). Nadel proposes that in contextual fear conditioning, as in episodic memory, there are separate, parallel learning mechanisms that evolve over time. Specifically, contextual fear conditioning can be due, one the one hand, to the animal learning the relations among spatial cues to form a hippocampal map during pre-exposure prior to the US; eventually, when the shock is experienced, it is associated with the map as a CS, and in parallel, contextual fear conditioning can be simultaneously due to learning the association of the US to individual cues that comprise the context. Immediately following shock training, the “background” hippocampal map-US association predominates. After 28 days, this hippocampal representation fades leaving the multiple, independent neocortical CS–US associations to maintain control of behavior. The time course of anterograde amnesia for context fear after hippocampal lesion suggests it as an example of the “multiple-trace” theory of episodic memory described by Nadel and Moscovitch (1997).

In brief, although the contextual fear conditioning literature is complex, it provides solid support for the critical involvement of the hippocampus in a specific form of contextual learning that is learned in the background, is map-like, and is hierarchical.

4 ∣. THE EARLY DESCRIPTION OF HIPPOCAMPAL REMAPPING

The nature of place-cell maps emerged slowly after the discovery and description of hippocampal place cells (O'Keefe & Dostrovsky, 1971). Some years later, while investigating stimulus control over place-cell firing, O'Keefe and Conway noted that the hippocampal representation of space appeared to involve multiple maps (O'Keefe & Conway, 1978).

“place units were recorded in two different environments: one, a small platform where the rat had received neither training nor reward; the other, an elevated T-maze … Some units had place fields in both environments while others only had a place field in one. No relationship could be seen between the place fields of units with fields in both environments … Fifteen units had place fields on both the T-maze and the platform, 10 had a field on the T-maze alone, 7 on the platform alone, and 2 units did not have a field in either place … The finding that fifteen of 34 units had place fields on both the small platform and on the T-maze suggests that a substantial percentage of place cells can participate in the representation of more than one environment.”

The idea that hippocampal place cells provided multiple representations of space was further explored in the 1980s by the Brooklyn Hippocampal group led by Jim Ranck. Figure 2 is from an early study in which pyramidal neurons were recorded in three separate chambers: a large home cage, an operant chamber with a bar that triggered food reward at a magazine, and an eight-arm maze (Kubie & Ranck, 1983). Each was centered in the identical location of the recording room, and in each, rats could see common distal cues. Single cells were recorded across the three environments (one or two cells at a time), and virtually all cells were observed to be place cells in at least one environment. Importantly, knowing the location of a cell's place field in one environment provided no information for predicting the location in the other two environments, and cells could show location-specific discharge in one environment and be silent in another. Kubie and Ranck inferred from these findings that each environment had a “context-specific” cognitive map. Unfortunately, although behavior and cell firing were recorded on videotape, there was no objective method of creating firing-rate maps, and the conclusions were based on subjective descriptions of each cell's firing patterns.

FIGURE 2.

Sketches of the firing fields of two simultaneously recorded place cells in three enclosures. The enclosures are drawn to scale, each centered in the recording room on the X (smaller apparatuses placed on the radial maze). All environments have open views of the recording room. Cell 1 depicted by light gray fields, Cell 2 by darker fields. On the radial maze, Cell 1 has fields, in arms at 6 o'clock and 1:30. Cell 2 has a field on 6 o'clock arm, near a field of Cell 1. In the large home cage, Cells 1 and 2 each have a single field. In the operant chamber, Cell 1 has two fields while Cell 2 has no fields. Note that the firing fields for each neuron do not superimpose across the chambers. Further, when each environment is rotated 90° clockwise, fields rotated with the environment for the smaller chambers, but were in fixed room location for the radial maze (from Kubie & Ranck, 1983)

Later in the decade, the Brooklyn group developed computerbased data acquisition and analytic methods and was able to convincingly demonstrate and extend the earlier findings (Muller, Kubie, & Ranck, 1987). The environments were now enclosures of various sizes with gray walls enclosing cylindrical or rectangular floor spaces. The rat's task in each was simple: to roam the floor to gather scattered food pellets. Figure 3 is of computer-generated color-coded rate maps when a single hippocampal neuron is recorded while the rat is in four chambers, 15 min in each. Again, the group found that single cells had one or two simple, clear firing fields; that a cell could have clear spatial firing in one enclosure and be silent in a second; and that knowing the location of a cell's firing field in one environment did not predict firing fields in different shaped environments. Although these findings provide the essence of the remapping concept, the term was not introduced in these papers, nor was the idea that different environments had different maps clearly presented. “Context” was not discussed.

FIGURE 3.

Color-coded firing rate maps of a single place cell recorded in four environments illustrate the foundation of the remapping idea, the observation that place cells form distinct maps of an environment. While the cell's firing fields are related in the environments of the same shape, the fields in the cylinder have no relationship to the fields in the box. This example is from Muller et al. (1987). In that study, of the 22 cells recorded in the small cylinder and the small box, 10 cells had firing fields in both and of these, only one cell had firing fields that appeared to be “related.” Of the 78 cells recorded across distinct shapes (cylinder-to-box of any size), 27 cells had firing fields in both, but only one cell appeared to have “related” firing fields

Although Kubie and Ranck in the papers from the early 1980s speculated that these apparently independent “maps” represented different contexts (Kubie & Ranck, 1982; Kubie & Ranck, 1983; Kubie & Ranck, 1984), the term “context” was not used in the later papers due to differences of opinion among the authors—specifically, Bob Muller thought “context” too vague and ill-defined. This remains a common notion among place-cell physiologists, such as John O'Keefe and Edvard Moser (personal communications).

The term “remapping” was introduced in two almost-simultaneous papers in 1991 (Bostock et al., 1991; Kubie & Muller, 1991). Remapping is described in different but complementary ways in the two papers. We will offer an updated definition of the remapping concept later in this manuscript, so let us now consider the original definitions of the concept.

Bostock et al. (1991) state:

The independent spatial firing of individual cells in sufficiently different environments strongly implies that the hippocampus can maintain independent representations in several environments (see McNaughton and Morris, 1987). The stability of the spatial patterns in each environment further implies that the representation that is active at a given time is determined by the current surroundings: a given representation is reliably reactivated by returning the rat to the appropriate environment. Each representation is therefore stationary, or in a “steady state.”

Following O'Keefe and Nadel (1978) the representation in each environment will be called a “map.” Here the term will be extended to mean that an environment-specific subset of place cells is active and that each active cell has an environment-specific firing field. It is also convenient to use the term “remapping” to describe changes of the surroundings on the place-cell population. Remapping is the transform from the hippocampal representation of a “standard” chamber to the representation of a different chamber. Remapping emphasizes changes in the active subset and in the characteristics of single cells.

The paper of Kubie and Muller (1991) provides a complementary definition of “remapping,” stating:

Remapping has two fundamental features. First, the hippocampal representation of an individual environment makes use of a surprisingly small subset of place cells, the “active subset.” The size of the active subset is about 20% of the pyramidal cell population, according to Thompson and Best (1989). Moreover, the subset used in each environment is a random sample from the population. Thus, if two environments are independent, the active subset in each is an independent sample from the total pool of hippocampal complex spike cells. As a consequence, most cells in the active subset of one representation are not in the active subset of the second. The second feature of complete remapping is that, for a cell that happens to be in both active subsets, there is no relationship between the two spatial firing patterns. For such cells, the location of the firing field in one environment is random with regard to the location in the second environment.

Although maps and remapping are inferences about the behavior of many hippocampal place cells in a single recording session, the descriptions above were based on what was then state of the art, but nonetheless limited technology—only one or two neurons were recorded at a time. In 1991, the remapping of sets of neurons had never been directly observed, it was inferred. The inferences were based on two observed properties of single neurons. The first was stability: firing patterns were stable from session to session. The second was location specificity: the location of firing fields was reliably specified by the particular features of an environment. These observations from single cell recordings were sufficient to make explicit inferences about the population of place cells.

5 ∣. TEMPORAL AND SPATIAL PROPERTIES OF HIPPOCAMPAL MAPS

Our current view is that a map is better defined by the temporal discharge relationships among its component neurons rather than their spatial relationships. We contend that time is primary because the timing of synaptic timing is what neurons directly encode—other relationships are secondary. This view is summarized by Gyuri Buzsaki in The Brain From Inside Out, in which he argues that the “outside in” framework as the mistake of current neuroscience (Buzsaki, 2019).

In the language of the neural code, the outside-in framework has shown only that decoding of stimulus properties from neural activity is possible in principle and only if an observer has the code book (i.e., the set of stimulus–response correlations). However, the brain only has its own neural “responses,” and the outside-in framework does not mention how such a code would be generated or read by neural responses alone. All any neuron in the brain ever “sees” is that some change occurred in the firing patterns of its upstream peers. It cannot sense whether such change is caused by a (particular) external perturbation … Thus, neurons embedded in networks of other neurons do not “know” what brain sensors are sensing; they simply respond to their upstream inputs” (p. 16)

Therefore, the map is ideally defined by the temporal discharge relationships among its component neurons, without regard to external features like the environmental location of a firing field or its relationship to stimuli. The temporal discharge, in turn, has a natural and direct correspondence to the spatial relationships of firing fields (Hampson, Byrd, Konstantopoulos, Bunn, & Deadwyler, 1996). This view can be explained with a simple prototype. Imagine a small rectangular environment that is represented by the discharge of hippocampus place cells. In small environments, perhaps 30% of potential hippocampal place cells have simple firing fields—for each neuron, a single discrete place where it fires strongly. Each place cell fires when the animal (rat) crosses its field in any direction, and will be virtually silent when the animal is not in the field.

The rat walks continuously for perhaps 30 min, with no fixed spatial goal—similar to pellet chasing as in Muller et al. (1987). A temporal cross correlation for each cell pair is constructed on the seconds time scale of locomotor behavior by taking the firing of one member of the pair as the reference (Cell 1), and looking at the temporal lead or lag of action potentials of the other member of the pair (Cell 2). When this is done for all spikes for Cell 1 and the lead and lag spikes for Cell 2, it produces a histogram that indicates whether the firing of Cell 2 is independent, positively, or negatively correlated with the firing from Cell 1. Figure 4 illustrates examples for a cell pair recorded in two chambers. During the session in the cylinder (top row), the fields overlap and the temporal cross correlation peaks at offset zero. In the next session, in a triangular enclosure (bottom row) the fields do not overlap and the cross correlations are very low at the zero time lag. Now imagine computing cross correlations for cell pairs with overlapping firing fields and disjoint firing fields in a single session.

FIGURE 4.

Two simultaneously recorded hippocampal place cells that illustrate two things. The first is that when cell pairs have overlapping firing fields their temporal cross correlations overlap at lag zero (top row), and when the cell pairs have nonoverlapping fields their temporal cross correlations are low at the zero time offset (bottom row). The second feature is the scrambling of the location of cell pairs characteristic of remapping. In this case, Cells 1 and 2 have overlapping firing fields when recorded in a 15 min session in the cylindrical enclosure. When the rat was removed and minutes later recorded in a triangular enclosure each cell had a firing field, but the fields did not overlap. Thus, even with this limited sample we can infer that the map fundamentally changed or remapped. These data were recorded in the early 1990s by Kubie and Muller. The color codes of the firing rate maps indicate regions of space (pixels) with different firing rates; the highest rates, coded purple, about 15 spikes/s. Yellow indicates visited pixels with zero spikes

The first point is that cell pairs with overlapping firing fields will cofire because they have positive spike train cross correlations, and that cell pairs with disjoint fields will not cofire, having negative spiketrain cross correlations. Second, and most important, cell pairs with positive cross correlations come from the set of cell pairs with firing fields that overlap, while cell pairs with negative or zero cross correlations will come from cell pairs with disjoint or neighboring firing fields, respectively. The second inference is critical; namely, we can infer the spatial relations across the firing fields of place-cell pairs solely from their behavioral-time scale temporal properties. In other words, on the time scale of locomotor behavior, there is mutual information between the spatial and temporal features of spike train cofiring relationships.

Now consider all of the cell pairs in the map. These can be divided into two categories based solely on temporal relations: those with positive cross correlations and those with negative or zero cross correlations. The pairs with positive correlations are spatially continuous, while those with negative correlations are discontinuous. Our contention is that this list is the essential description of the hippocampal map.

The temporal criterion does not mean that space is unimportant; indeed the temporal relations predict the spatial relations and vice versa. In our view, it is important to consider both time and space, because to an experimental observer they are interrelated, but the temporal relationships are primary. The spatial aspect of the representation can be considered as a continuous topological surface. A topological surface is continuous if a reference point can move from any one point on the surface to any other point without leaving the surface (Zeeman, 1966). This type of spatial representation has no inherent shape or metric. (Our guess, consistent with others, is that entorhinal grid cells provide the metric, or distance measure). Now consider a smaller topographic surface represented by a pair of place cells. This smaller surface is continuous if the firing fields of the two cells overlap or discontinuous if they do not overlap or touch each other.

There are two reasons for considering time primary, rather than space. The first is that neuron-to-neuron functional interactions are temporal, not spatial. In our view, all neuronal codes are temporally organized codes. The second is that, as we will describe below, the hippocampus can exhibit map-like temporal patterns without an immediate spatial reference.

To digress momentarily with an important detail, we note that things are more complicated on subsecond time scales and the time scale of single passes through firing fields because place cells with overlapping firing fields discharge in cell assemblies that organize cofiring cells into subgroups that signal the same type of information, such as the animal's position on an inbound or outbound journey along a track, or the animal's position in the spatial frame established by local cues as opposed to the spatial frame established by distal cues (Gothard, Skaggs, Moore, & McNaughton, 1996; Harris, Csicsvari, Hirase, Dragoi, & Buzsaki, 2003; Huxter, Burgess, & O'Keefe, 2003; Kelemen & Fenton, 2010; Kelemen & Fenton, 2013b). For the present discussion, we focus on the temporal organization of place cell discharge on the behavioral time scale of several seconds, during which discharge can nonetheless be dynamically organized into functionally defined subgroups, often according to hidden cognitive variables like attention and choice of goal (Fenton et al., 2010; Gothard, Skaggs, & McNaughton, 1996; Johnson, Fenton, Kentros, & Redish, 2009; Johnson & Redish, 2007). As we will discuss below, these apparent departures from expectations based on location, are easily accommodated by the concept of a fundamentally temporally organized hippocampal map.

6 ∣. MAP REGISTRATION

To be clear, in the proposed view, the hippocampal map is defined by a set of temporal discharge relationships, not by firing fields. However, what about space? As noted above, there are spatial properties of some hippocampal maps, but they are primarily organized in time and in the head of the rat, rather than the external world. If there is a temporal or spatial map, it is an internal representation. In our view, this is consistent with the title of O'Keefe and Nadel's book, The Hippocampus as a Cognitive Map—that is, the map is in the hippocampus. The nature of an internal representation is that it can exist unto itself and is self-sufficient.¹ In the example of Figure 4, each cell is a place cell that fires at a location in the box. The relationship between the set of place cells and physical space is a property we term “spatial map registration.” A hippocampal map can be registered, such that cells fire in specific spatial locations, or the map can be unregistered, such that cell firing has no direct relationship to current location. In some situations, such as during sleep or quiet wakefulness, the map is unregistered, but virtually identical to a registered map that is expressed during awake exploration of an enclosure (Dragoi & Tonegawa, 2011; Foster, 2017; Louie & Wilson, 2001; Skaggs & McNaughton, 1996). In other cases, such as when rats walk on a treadmill, hippocampal maps are unregistered to space although they could be considered registered to temporal events and even sequences of time (Kraus, Robinson II, White, Eichenbaum, & Hasselmo, 2013; MacDonald, Lepage, Eden, & Eichenbaum, 2011; Pastalkova, Itskov, Amarasingham, & Buzsaki, 2008) as well as other nonspatial events like jumping (Lenck-Santini, Fenton, & Muller, 2008).

This is a simple description of a two-dimensional (2D) spatial map. In principle, one could construct a 2D layout of firing fields from the list of cell-pair temporal discharge relationships. In this view, the hippocampus does not directly know about space, but, by analyzing hippocampal activity one can (or the brain can) infer a 2D topographic map. This view suggests that the set of temporal relations across hippocampal neurons is also consistent with higher dimensional spatial maps and nonspatial maps (Aronov, Nevers, & Tank, 2017; Jeffery, Jovalekic, Verriotis, & Hayman, 2013; Lenck-Santini et al., 2008). This notion of a map only requires that the set of spike train discharge relationships be cell-pair specific and consistent. Although the spatial constraints of locomotion create temporal patterns of discharge relationships, in our view, spatial constraints are not critical. Any situation, such as functional neural connections, synaptic effectiveness, and network dynamics that reliably induces consistent cell-pair specific patterns of spike train cofiring is sufficient to produce a distinct map. The essence is that the set of cofiring relationships, measured as cross correlations should be consistent. It is unimportant if they are internally generated, induced by the spatial constraints of locomotion, or the influence of sensation or any other combination of factors.

7 ∣. REMAPPING

According to the temporally defined hippocampal map, described above, remapping occurs when the set of cofiring relationships reorganizes amongst the cell pairs; in the extreme these relationships can scramble but they may merely reorganize into one of multiple stable patterns of cofiring due to the relative constant and slow changing influences of neural connectivity and synaptic effectiveness. That is, we say there is a remap between Conditions 1 and 2 when the set of neuron pairs with temporal overlap in Condition 1 has no relation to the set of neuron pairs with temporal overlap in Condition 2. Conversely, we say that two conditions share the same map when the set of neuron pairs with temporal overlap is very similar. Figure 4 illustrates a cell pair that appears to be a component of remapping. When the cell pair is recorded in a cylindrical enclosure, the fields overlap (top row); when the cell pair is recorded in a second environment, a triangular enclosure, the fields are disjoint (bottom row). Notice that the scrambling of relations between the cells can be seen both in the spatial maps and in the temporal cross correlations.

When a map is registered, there is a one-to-one relationship between temporal and spatial overlap between neuron pairs. This is simply because it takes time to walk from one location to another. If the locations overlap, on some occasions, an animal will take a path across the overlap and cells are likely to cofire, as captured by a positive cross correlation for that pair of cells. For a pair of ideal place cells, when the spatial locations of firing fields have space between them, the minimal time it takes the animal to move from one field to the other will make cofiring impossible. The temporally defined hippocampal map functions as a topological surface, where the likelihood of cofiring describes the relatedness, and for a spatially registered map that describes the likelihood the subject will move from the location of one firing field to the other. Cells pairs with overlapping and adjoining place fields cofire; in topological terms, they share an edge, and cell pairs with nonadjacent firing fields never cofire; in topological terms, they share no edge. In principal, and to a point, the hippocampal map can be stretched or rotated and remain the same map. For example, consider the cue card rotations in the cylindrical enclosure used in Muller and Kubie (1987). In Session 1, the white cue, as viewed from the overhead camera was centered at 0° and in Session 2, it was rotated 90° counterclockwise along the wall. All cells recorded with this manipulation had their firing field rotated 90° counterclockwise. This was interpreted as no change in the map; what had changed, using current terminology, is the registration of the map to the veridical world, but note the registration to the cue card also did not change. The inference was that the topographic relationships among the set of place cell firing fields was unaltered (Muller & Kubie, 1987). The related inference—not noted in the paper—was that the fundamental cofiring relations among cell pairs were unaltered, and thus from this inside out perspective, nothing had changed, whereas from the outside in perspective there was a substantial change.

8 ∣. ASSESSING REMAPPING, PAST AND PRESENT

While the evidence for remapping has come primarily from outside-in observations of changes in the spatial firing patterns of individual neurons, the concept of remapping has always been about inside-out changes of an internally organized brain map. The reason for this is largely technical: the ability to record the relationship between single neurons and position emerged in the 1980s. With this approach, changes in an internal map could not be directly measured and so only inferred. Only since 1993 have neuroscientists been able to record the simultaneous discharge of sufficiently large numbers of hippocampal neurons during behavior in multiple environments (Wilson & McNaughton, 1993). Even so, in the recording with the greatest proportion of place cells reported by Wilson and McNaughton, only 34 of 82 cells were active place cells, meaning that only a few cells were ever coactive as the rat explored the space. To simultaneously record a sufficient number of place cells to directly assess the hippocampal map has required novel recording technologies including microfabricated, dense electrode arrays, and optical recording technologies such as Ca²⁺ imaging with multiphoton microscopes or with miniscopes (Csicsvari et al., 2003; Dombeck, Harvey, Tian, Looger, & Tank, 2010; Meshulam, Gauthier, Brody, Tank, & Bialek, 2017; Pnevmatikakis et al., 2016; Rossant et al., 2016; Ziv et al., 2013).

Not only has there been technological advance, but our understanding of what constitutes location-specific discharge has also evolved. It is now clear that many principal cells that do not discharge in what looks like classic, unimodal firing fields, in fact signal location-specific information, and in some cases, cells with broadly distributed spatial firing signal more location-specific information that cells with unimodal firing fields (Meshulam et al., 2017). This is especially easy to see in more natural conditions, such as larger than standard recording environments, with linear dimensions over 1 m, in which single place cells discharge in multimodal firing fields (Fenton et al., 2008; Jeffery et al., 2013; Park, Dvorak, & Fenton, 2011) such that a pair of cells can cofire in one location because they have overlapping firing fields in that location, and not cofire in another location because they have disjoint firing fields in that other location.

Although the cross-correlation plots are a useful tool in identifying remapping, they are inherently flawed. The cross correlation requires averaging over extended periods, many minutes—this is not what the nervous system does. In addition, in large environments, it is common to find single neurons with multiple firing fields, one of which may overlap a second cell's field and the other being disjoint (Fenton et al., 2008; Park et al., 2011). With the ability to record many neurons at a time, two more direct tools are possible for assessing cofiring: activity vectors and firing sequences. Activity vectors describe the cofiring of multiple cells in a small time window (Wilson & McNaughton, 1993). If one imagines a map represented by 2,000 place cells in a specific environment, perhaps a cylinder, each location in the environment will be covered by many place fields, and, when the animal is at a particular location, specific activity vectors will occur. In our cylinder example, we can imagine perhaps 100 distinct 2,000-dimensional activity vectors, each in register with a location in the environment. The vectors at neighboring locations will tend to be similar and the population activity will typically evolve smoothly in time as the subject moves, which is why dimensionality reduction algorithms like principal component analysis will be able to project these 100 vectors onto a manifold-like subspace with a dimensionality much lower than 2,000. In other words, these 100 vectors represent a tiny subset of the possible permutations of the place-cell population activity, and each set of vectors is identified with a unique map, that has the form of a manifold defined in dimensions of cofiring. When the animal is transferred to a second chamber, perhaps a rectangular enclosure, there will be a second map defined by a distinct set of observed activity vectors, also having the form of a manifold defined in a distinct set of cofiring dimensions. In other words, if this second map is a complete remap, there will again be 100 activity vectors covering each location in the environment, with this second set of vectors independent of the activity vectors in the cylinder. Again, each of the vectors in the rectangular enclosure can uniquely identify the second map. Such an “ensemble code” is physiological, instantaneous and can be sufficient for map identification, even on the subsecond time scales of mental operations (Kelemen & Fenton, 2010; Kelemen & Fenton, 2013a; Neymotin, Talbot, Jung, Fenton, & Lytton, 2017; Park et al., 2011).

Somewhat analogous to the set of activity vectors, is the set of firing sequences. Again, imagine the recording of thousands of neurons while a rat is in a cylindrical enclosure. A firing sequence is a series of cell firing over a period of time, perhaps 5 s, with each sequence representing a path through the cylinder. While there is a seemingly unlimited set of permutations of sequential firing of neurons, the observed series, representing possible paths through space is much smaller. As part of a series, Cell A may be followed in firing by Cell B, whose field overlaps, but will never be followed in firing by Cell C, whose field is distant from the Cell A field. Critically, the vast majority of firing sequence permutations represent impossible paths and will not occur in a given environment, and those that could occur represent possible but not necessarily experienced sequences (Gupta, van der Meer, Touretzky, & Redish, 2010). Thus, again, a specific map supports a strictly defined, restricted subset of firing sequences, which, in a spatially registered environment, represent possible paths. Figure 5 is from preliminary data recording using mini-scope Ca²⁺ imaging that illustrates the differences in hippocampal firing vectors when a mouse is recorded on a cylindrical platform or in a box enclosure. The activity reflects Ca²⁺ transients and not the Na⁺-based action potentials, although of course the two signals are related (Chen et al., 2013). The basic finding is that firing vectors in one environment (the box) have no similarities to the vectors when the mouse is on the circular platform.

FIGURE 5.

This figure illustrates the activity of an ensemble of 48 simultaneously recorded CA1 place cells out of an ensemble of 119 cells recorded while a mouse chased scattered pellets in two environments, a box and a circular platform. The neurons had been virally infected to express the fluorescent calcium indicator GCaMP6f. The fluorescent image was recorded using a head-mounted miniature microscope, through a GRIN lens implanted in CA1. The images were analyzed using the CNMF algorithm to simultaneously separate the cells fluorescence and deconvolve the calcium transients to infer spiking activity. (a) Representative rate maps. (b) Activity correlation matrix with each small square representing the correlation between two different 10-s epochs across a 5-min span. The mouse spent first 5 min in the box (upper left) and next 5 min on the circular platform. Correlations approaching +0.3 are yellow, and −0.3 are blue. Correlations between vectors for the same period, with each 10-s epoch correlated with the entire set of 10-s epochs in the 5 min on the circular platform (red line) or box (gray line). (c) The vector pCorr (Neymotin et al., 2017) of cell-pair correlations sorted from strongest (yellow) to weakest in the box, with the same rank order for the circular platform period. The vectors are self-similar (strong correlations) within the same environment, but distinct (weak correlations) between environments. This directly demonstrates remapping, without regard to any position information and is by-definition independent of how the activity is registered to an environment

9 ∣. PARTIAL REMAPPING AND RATE REMAPPING

The remapping described above, where a neuron's transition from Environment A to B appears unpredictable and unrelated to other neurons, is termed complete (or global) remapping. Complete remapping of Environment A to B is not always the case. Maps between two environments, like the environments themselves, can be different but retain similarities. Partial remapping, based on traditional analysis of place cell firing fields, exists between a pair of environments when, in pairwise analysis, a substantial proportion of cell pairs retain their spatial relations, while other cell pairs do not; similar also for temporally defined hippocampal maps. A powerful method of reliably eliciting partial remapping is a cue-dissociation paradigm developed by Shapiro, Tanila, and Eichenbaum (1997) (Tanila, Shapiro, & Eichenbaum, 1997) and exploited by Knierim (2002) (Lee, Yoganarasimha, Rao, & Knierim, 2004). In these studies, an animal learns a familiar environment with many cues: roughly half attached to the distal walls and the other half attached to the floor. On probe trials, the set of wall cues is rotated one direction and the floor of the apparatus is rotated in the opposite direction. One can imagine three possible results: first, complete remapping, the original map will be scrambled; second, no remapping, the set of place cells will remain “coherent” and maintain relations to each other; third, partial remapping, where the result bears resemblance to the original map, but the set does not maintain coherence. The common observation is partial remapping. On a given probe trial, a large subset of place cells remains in register with the floor cues while another large subset remains in register with the wall cues. In the standard cell-pair firing field analysis, many cell pairs will maintain their spatial firing field relations, and presumably also their temporal cofiring relations, while many cell pairs have their relations change.

10 ∣. RATE REMAPPING

Rate remapping described in Hayman, Chakraborty, Anderson, and Jeffery (2003) and Leutgeb et al. (2005) is a second type of partial spatial remap. In rate remapping, when comparing the firing patterns of the set of place cells across two environments, the relative spatial locations of the firing fields remain fixed, while each neuron's firing rate within its fields can vary greatly. For example, a given place cell will show comparable locations of its firing fields in the two environments, but different firing rates in the corresponding fields. In the Leutgeb et al.'s study, they found complete (global) remapping when the set of place cells was compared across similar shaped enclosures in two separate recording rooms, but the same set exhibited rate remapping when compared across two enclosures with different cues on the wall but in the same recording room.

Although partial remapping and rate remapping are clear phenomena, at present, it is hard to determine their significance without corresponding behavioral studies. For example, when the hippocampus exhibits a partial remap across two environments, does the rat treat these two as completely distinct, partially distinct or identical (Jeffery, Gilbert, Burton, & Strudwick, 2003)? Intuitively, the ability to treat as “partially distinct” would seem to make responses more flexible, but this remains to be determined. Another possibility is that rate remapping is specialized for distinguishing between two occasions of experience in the same environment, useful for episodic memory encoding, as opposed to navigational opportunities (Leutgeb et al., 2005). While these notions need further investigation, we point out that in the present context, both partial and rate remapping distinguishes hippocampal maps that are defined on manifolds of temporal cofiring relationships, but because cofiring is partially preserved across the two conditions, similarities in cofiring opportunities between the two environments are also preserved, although not homogeneously across the population of activity vectors. Such conjectures also merit exploring, but the focus of the current paper is on the better studied phenomenon of complete (or global) remapping: the tendency for the hippocampus to orthogonalize pairs of maps. This is a remarkable phenomenon, which we believe permits the hippocampus to store independent representations of a vast number of environments and, perhaps, contexts.

11 ∣. DO HEAD-DIRECTION CELLS AND GRID CELLS REMAP?

As we have expressed, remapping occurs when the cofiring relationships of neurons change—this is independent of the particular registration of the hippocampal map to the environment that is assessed by the physical location of firing fields, and spatially defined tuning curves. A critical point, described above, is that in the hippocampus, complete remapping is a common phenomenon, permitting a very large number of orthogonal maps. While one can argue that subtle, partial changes may be observed in other areas, at present, outside of the hippocampus we know of nothing equivalent to complete remapping. A dramatic dissociation was demonstrated by the response to the psychotomimetic phencyclidine (PCP; Kao et al., 2017). PCP did not change firing fields but it caused remapping of the cofiring relationships of CA1 place cells as well as impairments of a familiar place avoidance task. In contrast to the PCP effect, there is no remapping if the cofiring relationships of neurons remain constant within a brain structure. Accordingly, head-direction cells do not remap. It was noted early on that pairs of head-direction cells maintain their preferred angular discharge relationships, even when registration with the environment shifts or drifts (Ranck, 1985; Taube, Muller, & Ranck, 1990). Ensemble recordings of ~10 head-direction cells in rodents confirm that they retain their temporal discharge relationships during all known conditions including sleep (Peyrache, Lacroix, Petersen, & Buzsaki, 2015). In fact, temporal relations were observed to maintain in the head-direction cell population recorded from medial entorhinal cortex (MEC) across navigation tasks designed to test the relationship between cofiring and spatial tuning. Despite, excellent navigation on a rotating arena that continuously dissociated the environment into a stationary space and a rotating space, the temporal cofiring relationships maintained within the head-direction cell ensemble (and within the ensemble of all cells recorded in MEC), but head-direction spatial tuning was nonetheless degraded (Park, Keeley, Savin, Ranck, & Fenton, 2019). This degradation occurred because the registration to the environment was intermittent and variable, which dramatically demonstrates that the temporally defined map can dissociate from registration with the environment.

Whether grid cells remap is a bit more complicated. Grid cells in the MEC are organized in a series of roughly horizontal anatomical “modules” (Barry, Hayman, Burgess, & Jeffery, 2007; Brun et al., 2008; Walling, Bromley, & Harley, 2006). For each module, the spacing of all of the firing field bumps of grid cells is the same, but as one proceeds ventrally, to other modules, the grid spacing increases. Consequently, within a module the so-called “phase relations” between pairs of cells is fixed. That is, for cell pair AB, a firing field bump of cell B will be a certain distance and direction from the closest firing bump of cell A. A reasonable surmise is that the temporal cofiring of all cell pairs within a module will be fixed across environments, thus no remapping (Stensola & Moser, 2016; Wernle et al., 2018). We also note that the cofiring of pairs of grid cells from different modules can vary across environments and for different regions within an environment, which would also generate changes in coactivity that define remapping. It seems the jury is out as recent work also suggests that the entire grid cell system operates as one, very large map, even after rat subjects are disoriented, with grid cell firing vectors from two locations indicating distance and direction connecting these two locations (Bush, Barry, Manson, & Burgess, 2015; Weiss et al., 2017).

Is remapping a useful concept for other regions of cortex? One might surmise that primary sensory and motor cortical areas, such as V1 and M1 do not remap. We feel it a useful question to ask whether other cortical regions, such as S1 and premotor cortex contain discrete replaceable maps and, therefore, remap, to exchange one for another, which is difficult to assess using the conventional measures of remapping that rely on spatial tuning curves, but is naturally assessed using the cofiring-based assessments we are arguing for. Remapping has proved to be a useful concept for the hippocampus because spatial map registration is so easy to see. With better tools and more universal concepts, remapping can also be assessed and perhaps revealed in other “cognitive” brain regions.

12 ∣. ARE MAPS AND REMAPPING THE PHYSIOLOGICAL BASIS FOR CONTEXT?

Inherent in the concept of “context” is the notion that a mind must be able to have multiple possible contexts and to switch among them, with one context presiding at a time, as pointed out in Nadel and Willner (1980). This suggests that for a brain region to process context it must be able to represent multiple contextual states and exhibit a process of switching from one to another.

Was John Lisman correct? Are there “multiple maps” in the hippocampus and is “remapping” evidence of the process of switching between contextual states? This is undoubtedly an attractive hypothesis, but at this stage, there is no certain answer. With the clear, formally expressed notions of remapping as we have proposed, and equally clear notions of context, the issue can be addressed, and perhaps resolved. We expect that remapping is no longer a matter of speculation and inference, as it can now be measured directly from the coactivity patterns within ensembles of neurons within and between brain areas. Indeed, with the advent of techniques that can record hundreds of neurons, answers will emerge (Jun et al., 2017). Moreover, we expect that these studies will go beyond a simple hypothesis test and provide a roadmap to unveiling the subtleties and mechanisms that are likely to involve diverse cell types and population features of the network activity, including oscillatory activity (Buzsaki, 2010).

We close with speculations about context itself and the relation of hippocampal remapping to context. We support the idea that context is hierarchical, as suggested by Nadel, Willner, and Rudy, and is a feature of a cognitive, spatial map. Context is the framework on which episodes are experienced, stored, and expressed.

Although we are not yet convinced that hippocampal maps are the neurobiological realization of contexts, we conclude by assuming maps are the substrate of context, and we speculate on the function. First, perhaps least controversial, a hippocampal map is a discrete representation, providing spatial information of a specific environment. Second, we accept the notion proposed by Hirsh and Nadel and Willner that hippocampal maps serve as an indexing system useful for episodic memories.

A third notion, related to the previous two, is that each map (context) provides information about the limited but expansive range of behavioral possibilities. We have found three, virtually overlapping terms describing this third function. One term is “affordance.” According to Gibson (1986)

“The affordances of the environment are what it offers the animal, what it provides or furnishes, either for good or ill.”

According to this notion, hippocampal context presents to the animal the set of available paths, or opportunities, in a particular environment. The affordances largely refer to the sensory features of an environment that contribute to possible actions. A second term is “kinematic.” Muller and Kubie suggest a kinematic hypothesis for the hippocampus in their 1987 paper (Muller & Kubie, 1987)

“ … a kinematic analysis can produce a list of possible states of the system of allowed trajectories … We would like to suggest that place cell firing represents a solution to the kinematic problem.”

A final term is “model.” Daw et al. have suggested that the hippocampus is used for “model-based planning” (Vikbladh et al., 2019). A model in this sense is an internal representation of the environment that provides the opportunity to deliberate, to perform vicarious trial- and-error behaviors prior to action, correlates of which are observed in hippocampus place cell activity (Redish, 2016).

In brief, our notion is that remapping is an expression of the ability of the hippocampus to produce one map at a time from a library of many maps. Further, that each map, rather than being a determinant of specific behavioral outcome, is a prediction machine, indicating the range of behavioral possibilities.

13 ∣. FINAL COMMENT

The hippocampal revolution was triggered by three explosions: Scoville and Milner's discovery that hippocampal damage produced profound amnesia, O'Keefe and Dostrovsky's discovery of hippocampal place cells and O'Keefe and Nadel's remarkable synthesis and proposal found in “The Hippocampus as a Cognitive Map.” Starting with the book and continuing to the present, Lynn Nadel has played a critical role in linking the physiological findings of place cells to psychology and behavior. Lynn is motivated not only to understand how place cells contribute to an animal's self-localization and navigation, but a second issue: how the cognitive map forms the basis for understanding of declarative memory. This has been a monumental, revolutionary project. We are not there yet, but progress has been made, with great debt to Lynn.

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are available from the corresponding author upon reasonable request.