Neural ensembles in navigation: from single cells to population codes

Abstract

The brain can represent behaviorally relevant information through the firing of individual neurons as well as the coordinated firing of ensembles of neurons. Neurons in the hippocampus and associated cortical regions participate in a variety of types of ensembles to support navigation. These ensemble types include single cell codes, population codes, time-compressed sequences, behavioral sequences, and engrams. We present the physiological basis and behavioral relevance of ensemble firing. We discuss how these traditional definitions of ensembles can constrain or expand potential analyses due to the underlying assumptions and abstractions made. We highlight how coding can change at the ensemble level while underlying single cell codes remain intact. Finally, we present how ensemble definitions could be broadened to better understand the full complexity of the brain.

Introduction

Ensembles are collections of one or more neurons whose activity co-varies or explains variance in a decoded variable, such as spatial position. Ensembles are proposed to be the default state of the brain, formed by co-activity which exists before content is mapped onto them[1–3]. Through changing membership and temporal order, ensembles can represent more information than individual neurons alone, and exhibit a potentially infinite number of combinations of neural activity. These ensembles are involved collectively in neural computations, as most neurons cannot cause the discharge of a downstream neuron alone[4]. Furthermore, ensembles can encode variables beyond what can be encoded by single neurons.

Ensemble coding is observed in many brain systems[5], but is perhaps most thoroughly studied and at the widest variety of scales in the navigation system. Neural ensembles in brain regions involved in navigation can be generally viewed from six perspectives. These perspectives are largely differentiated by the analysis approach used, as any given neuron can participate in more than one ensemble type. First, single neurons can be studied by training a decoding model (Figure 1a), which learns from training data how to receive spike trains as input and predict a spatial variable as output[6]. Each neuron is said to encode or represent the spatial variable decoded by the model. Many CA1 neurons, for instance, are place cells, as their activity increases when the animal is inside the location decoded by the model, the neuron’s place field[7]. Second, population codes can be deduced by collectively analyzing all recorded neurons (Figure 1b), commonly followed by dimensionality reduction[5]. Within the dimensionally reduced subspace, the geometry of activity reveals the structure of the population code. The axes of this subspace reflect the greatest co-variation and can be mapped onto latent variables, which can reflect position but also abstract variables like choice. The population activity occupies one point in the subspace at each moment in time; that point represents the value of latent variables encoded by the population. Third, cell assemblies are commonly defined by identifying patterns of co-activity within a brief time window. Fourth, time-compressed sequences of multiple neurons can fire sequentially on the scale of 100 ms bounded by the power or phase of local field potentials (LFP), extracellular currents that reflect the summation of activity from neurons (Figure 1c). These sequences can represent a path through space faster than the speed of the animal. They are most commonly observed during running, bounded by a single theta cycle (theta sequences)[8], and during rest, bounded by a sharp-wave ripple (replays)[9]. A third LFP oscillation, slow gamma, chunks theta sequences and replays into shorter trajectories[10,11].

Figure 1.

Six views of neural ensembles in navigation. Each neuron can participate in multiple ensembles and in multiple ensemble types, even simultaneously. Triangles represent neurons and boxes represent environments. (a) A single neuron (top) can represent a variable decoded from its firing field, for instance a place field within an environment (bottom). (b) Activity from all simultaneously recorded neurons can be combined into a population, which can for instance be dimensionally reduced from the N dimensions of N neurons to a 3-dimensional subspace. (c) Neurons which are repeatedly co-active can be grouped into assemblies. (d) Sequential activity can be organized by LFPs, for instance a replay (top) during a sharp-wave ripple (bottom). (e) Neurons can also be ordered by their peak activity over a single trial. (f) Neurons which express immediate early genes (top) during exposure to a context (bottom) are classified as engrams.

Fifth, behavioral sequences are usually defined by sequential firing by all recorded neurons with a discrete firing field over the span of a single trial of a behavioral task[12] (Figure 1d). Finally, engrams can be defined as neurons which express immediate early genes (such as Fos) during the temporal window of several minutes of exposure to an environmental context[13,14] (Figure 1e). Collectively, these six views of ensembles are most commonly described in the CA1 region of the hippocampus, but have also been observed throughout related cortical regions.

Here, we compare neural ensemble codes across these six views. We first discuss how the brain’s physiology and experimenters’ analysis methods define the boundaries of which neurons are included in ensembles. We then detail how the ensemble lens through which we interpret our findings constrains or expands potential hypotheses due to the assumptions and abstractions made at each ensemble scale. Next, we explain how different views of ensembles are behaviorally relevant and yet can lead to differing conclusions. Overall, we propose that our definitions of ensembles are fundamentally circumscribed by analysis choices and therefore can only provide multiple perspectives on neural activity rather than identifying a true underlying biological grouping. Finally, we discuss how future research can broaden the definition of an ensemble to better understand how the brain supports navigation.

What determines ensemble boundaries?

Physiology

Neural ensembles can arise from physiological properties that spatially and temporally join or segregate neurons into groups. For instance, excitability, which is in part regulated externally by inhibitory inputs and internally by biophysical membrane properties, plays a key role in the regulation of ensemble membership. Inhibitory inputs limit how many CA1 neurons in the hippocampus can be induced to represent a place field[15] and constrain the spatial size of each place field[16]. A rise in CA1 membrane potential predicts the formation of a place field[17]. Inhibitory neurons also pace the oscillations that organize time-compressed sequences[18]. Expression of genes which increase[19] or decrease[20] excitability of cells temporally joins or segregates cells from engrams, respectively. In parallel, inhibitory inputs to these excitatory ensemble neurons modulate the maintenance[21], size[22,23], and memory recall of the represented context[24] of an engram. Thus, excitability can constrain ensemble size across multiple ensembles views and could be a rapidly adaptable substrate for changing ensemble membership.

The boundaries of ensemble membership may also be defined by the temporal window during which its outputs could be integrated by a downstream reader, as defined by membrane time constants and plasticity windows. The membrane time constant of many excitatory cells is 1–100ms[25], approximately the length of a sharp-wave ripple or single theta cycle which encapsulates a time-compressed sequence[26]. The time window over which inputs can induce long-term potentiation in many excitatory cells is 1–30ms[27], matching the length of a single slow gamma oscillation, which chunks theta sequences and replays into shorter trajectories[10,11]. Cell assemblies defined on this timescale have recently been shown to drive activity in downstream reader neurons within the prefrontal cortex-amygdala reciprocal circuit[28]. Finally, CA1 pyramidal cells can potentiate inputs on the scale of seconds[29], matching the length of behavioral sequences. How ensembles over longer time scales could be integrated is unclear, though downstream readers may themselves combine as an ensemble to integrate inputs from several upstream sub-ensembles.

Which neurons are active in any given ensemble could be regulated by excitability, and the temporal window over which neural activity could be collectively received could be regulated by integration by downstream readers. While classically defined ensemble types are circumscribed by these physiological properties, how we determine which neurons participate in a given ensemble is often restricted by the analysis methods selected.

Analysis methods

Most navigational studies analyze neural activity through one of six views, which are restricted by the experimental conditions and analysis methods used. First, single neuron activity can be observed using an adjacent electrode or a calcium indicator, thus neurons are only included in analysis if they meet spike sorting or voltage thresholds or region of interest (ROI) selection criteria. Whether a cell is defined to represent a particular variable, such as a spatial field, can depend on experimenter-defined tuning thresholds or comparison to a shuffled distribution. Simultaneously recorded single neurons which pass at least the activity detection thresholds are then analyzed as part of the larger ensemble sizes.

Second, all recorded neurons can be analyzed as a population. The types of variables that can be extracted depend on what analysis method was used, such as linear combinations or nonlinear dimensionality reductions. However, it is unclear which types of methods are physiologically relevant, or how a downstream region could perform such a transformation on an entire upstream population to decode a variable. Moreover, even if only a subset of a population is decoded downstream, we cannot determine whether all active cells belong to a primary, initiating ensemble or whether some cells are firing in feedback. For instance, decoding spatial variables from a subset of the ensemble can be as informative as decoding from the entire population[30]. It is unclear whether the remainder of the population is informative or redundant behaviorally.

Third, cell assemblies are bounded by the time window selected by the researcher, most commonly 10–50ms to align with the timescale of long-term potentiation and a single gamma cycle[31], but can also be bound by longer windows to align with a single theta cycle[32] or sharp-wave ripple[33]. Thus, the membership of a cell assembly relies heavily on the experimental hypothesis. To identify cell assemblies, the pairwise correlations of binned, normalized spike trains undergo principal component analysis, then each component above a set threshold is labeled an assembly. These then undergo independent component analysis to extract vectors of weights, which reflect how much each cell’s firing contributes to an assembly. Cells with weights above a certain threshold are labeled members of that assembly[34]. Thresholds for assembly detection and membership can be set by an experimenter-defined value or null hypothesis distribution. Neither of these thresholds has yet been linked to physiology. Thus, it remains incompletely understood what timescale, assembly resolution, or assembly size is most relevant to downstream readers or behavior.

Fourth, time-compressed sequences are usually bounded by the length of a single theta cycle or sharp-wave ripple[8,9] (but see also [35]). Such sequences are often only examined if they coincide with LFP or multi-unit activity (MUA) of a sufficient amplitude and if the decoded spatial locations of the neurons progress in a linear trajectory over time[36]. However, LFP or MUA amplitude and the correlation coefficient of decoded space over time exist along a continuum, thus similar sequences just below these two thresholds may be used by the brain, but discarded in analyses[37]. As discussed later, discarded spiking can be physiologically relevant[38], and such restrictive definitions can change the results of analyses[39]. In addition, sequential firing is identified by ordering neurons by their activity peak, only considering a single firing field from each neuron. This approach can introduce complexities when dealing with the multiple firing fields found in navigational brain regions such as CA1[40] and entorhinal cortex[41].

Fifth, behavioral sequences are captured over the span of a single trial of a behavioral task, which often spans a few seconds[12], though they have also recently been observed to span up to 100 seconds and occur independent of trials[42]. Their membership is less restrictive than time-compressed sequences, as they can include all cells active during a trial. However, like time-compressed sequences, they are restricted to neurons that, when sorted by peak firing field, progress linearly over time, discarding additional firing fields.

Finally, engrams are bounded by the length of exposure to an environmental context, thus they are defined by their association with that context[13,14]. A neuron must express immediate early genes to be included in an engram, and so functional engram analysis requires tagging of these neurons. Engram experiments thus discard neurons that don’t express these genes at a high enough level to elicit sufficient opsin expression, which can vary based on viral expression magnitude. It is also unclear how a downstream neuron would integrate engram codes, which are captured on the scale of minutes. Finally, it is unclear which subset of cells are required for the engram, as memory strength is unrelated to engram size[43].

In sum, systems neuroscience research is both expanded and constrained by examining neural spiking through the lens of classical definitions of ensembles. This syntax arises from underlying physiology but is defined by how we design experiments and analyses.

Comparing ensemble views

The ensemble definition adopted in a single experiment can restrict the questions that can be asked by abstracting away relevant variables. At the same time, each ensemble view can also expand the types of questions that can be asked by analysis methods that reveal new dimensions of representations. In many cases, analyzing data across multiple views reveals novel insights not found at a single view, as many ensemble views are complementary.

Forest or the trees: population codes vs single neurons

Single neuron spike trains can be examined independently to decode the variables each neuron represents or collectively to decode the variables the population represents (e.g. the spatial position of the animal). Collections of neural spike trains can be analyzed through linear combinations such as population decoding, where all neurons are used to train a model to decode a variable, and similarity analyses, where population vectors are compared to reveal variables distinguished by the whole population. Both analyses can decode variables not detectably coded by individual neurons, such as environmental context or the order, valence, or identity of cues[30,44–48]. Linear combinations can reveal phenomena which are represented by individual neurons, but below detection thresholds, as their combination increases statistical power. Alternatively, collections of spike trains can be analyzed through linear or nonlinear dimensionality reduction, where the N-dimensional spike trains of N neurons can be constrained to a lower-dimensional subspace, and the dynamics of the population through that subspace can represent variables[49]. Nonlinear methods can reveal phenomena which aren’t represented by individual neurons, but only by the nonlinear dynamics of the population. For instance, Shahbaba et al used a neural network approach to reduce CA1 spike trains to two-dimensions and found the geometry in this subspace could distinguish the order of stimuli[44]. Nieh et al used a manifold inference approach on CA1 activity and found that the resulting manifold was a conjoined cognitive map of physical and abstract variable spaces[50,51] (Figure 2a). The manifold was so robust that it could be used to decode physical and abstract variables as well as the raw spike trains, even when a manifold from a different mouse was used. Population ensembles can thus be analyzed through linear and nonlinear methods to decode below thresholds or abstract variables.

Figure 2.

Different ensemble views reveal different aspects of the neural code. Unless specified, heatmaps range from red (high) to blue (low) activity or firing rate. (a) CA1 pyramidal cells jointly encoded place and accumulated evidence as mice ran down a linear track accumulating evidence to decide whether to turn left or right (top). Example place fields on the two trials (bottom). Reproduced with permission from [50] (b) Grid cells in the entorhinal cortex changed their firing rate, orientation, and spatial scale when a rewarded location was introduced (left vs right). Reproduced with permission from [52]. (c) (Left) A single neuron recorded as a mouse navigated a virtual linear track remaps between two representations on a trial-by-trial basis. (Right) Dimensionally reducing all neurons to the first three principal components revealed two circular manifolds, each representing one progression along the circular linear track. The population activity progressed along the manifolds and jumped between them during remap events. Adapted with permission from [53]. (d) Mice were trained on a collection of virtual linear tracks whose appearance morphed along a continuum from Track A to Track B. Mice were divided into two groups: those who mostly saw Tracks A and B (rare morph) and those who saw an even sample along the continuum (frequent morph). When exposed to all morphed tracks, place cells did not appear to remap differently between the two groups (left). Representational similarity analyses revealed a more abrupt remap in the rare morph group (center). The remapping dynamics of the two groups (discrete vs continuous) was made clear by examining the two principal eigenvectors of the similarity matrix (left). Adapted with permission from [55]. (e) A neural network model received velocity inputs (v) of a simulated animal exploring a 2D arena and outputted the animal’s location through place fields (right). The recurrently connected hidden units had firing fields that were similar to actual grid cells (bottom). Adapted with permission from [57]. (f) The activity of 149 grid cells reduced into three dimensions revealed a twisted torus geometry, which had been predicted by continuous attractor models[60]. A trajectory through space (left) matched a trajectory along the torus (middle), and activity from a single grid cell (top right) mapped to a single region on the torus (bottom right). Reproduced with permission from [58]. (g) Engram cells are place cells. Engram cells were active exclusively in one context, but were spatially unstable. Place cells were active in multiple contexts, but were spatially stable within a single context. Reproduced with permission from [61].

Analyses at the single neuron and population scale can answer different questions. For example, Butler, Hardcastle, and Giocomo observed changes in single entorhinal neuron tuning of multiple spatial properties only of neurons whose firing fields were near a rewarded location[52] (Figure 2b). Population analyses would have abstracted away the multiple spatial properties and firing field locations, and thus missed this finding. In contrast, Low et al and Sheintuch et al used single neuron and population analyses to find that medial entorhinal cortex and hippocampus, respectively, contained multiple maps of a single linear track, which switched between trials[53,54]. This was initially revealed by raster plots of individual neuron spike trains, but was best captured and understood when population analyses revealed that the whole population of neurons was alternating between two stable attractors (Figure 2c). Finally, Plitt et al showed that three dimensionality reduction methods found a CA1 remapping dynamic that could not be fully captured from individual neuron fields[55] (Figure 2d). Thus, single neuron analyses capture a wider variety of input dimensions, while population analyses produce a wider variety of output dimensions.

Population analyses are also used in artificial neural networks, a modeling architecture based on the brain, which lead to insights into how the brain might perform computations. For instance, optimizing a recurrent neural network to estimate head direction from angular velocity inputs creates a network layer whose weights match the firing patterns of Compass and Shifter neurons, two neural types in the fly head direction circuit[56]. Training a biologically plausible recurrent neural network to estimate position based on velocity inputs creates a network layer whose weights match firing patterns of grid cells[57] (Figure 2e). Thus, neural networks can recapitulate the structure and function of real neurons. These models have also led to testable hypotheses, such as grid-structured firing being generated by neurons organized into a continuous attractor network, which was recently demonstrated to be biologically accurate in entorhinal grid cells[58] (Figure 2f).

Population analyses can reveal structure in how neurons interact with each other, over time[59], and with underlying, abstract variables coded by the whole population. Single neuron coding models, by contrast, consider heterogeneous coding and properties beyond spike trains, such as tuning specificity. Both analyses are complementary, and experiments which study both scales can reveal the true complexity underlying spatial variable representation.

Place or context: sequences vs engrams

The intermediate ensemble sizes - sequences and engrams - are almost never examined together due to the different experimental conditions that define them. The CA1 place cell code and the CA1 engram code have long been studied in parallel as memory storage mechanisms in the hippocampus, with place cells encoding a cognitive map of space and engrams encoding a contextual index. Two recent studies compared these codes by examining spatial representations in cells expressing Fos at high (engram cells) or low levels. The first, conducted in novel environments where Fos expression is higher, found that high Fos-expressing cells are in fact place cells, but ones with sparse firing, low spatial stability, and activity specifically in one context[61] (Figure 2g), thus encode context more faithfully than space. The second, conducted in familiar environments where Fos expression is lower, found that high Fos-expressing cells had more stable and accurate spatial coding, and Fos was required to stabilize this code[62]. Thus, contextual indexing and spatial coding are different representation-oriented views of the same ensembles.

Sequences contain several additional dimensions of variables due to their temporal organization. Sequences encode temporal relationships between variables[63], and integrating over this temporal order can represent additional variables, such as movement direction[64]. In behavioral sequences, the order of peak firing of neurons can change between trials to distinguish variables not encoded by individual neurons, such as the trial type[44,65], animal’s choice[12,66,67], and distance traveled[68]. The oscillations that organize time-compressed sequences can both chunk sequences into sub-ensembles readable by a downstream neuron[69] and coordinate this reading by downstream brain regions[70]. These oscillations also provide a second layer of representation, as adding the theta phase at which a neuron fires to a decoding model improves accuracy[71].

Engrams, on the other hand, have been extensively causally linked to memory. Myriad studies have manipulated them to suppress or enhance recall, or even to induce an artificial association[19]. Similar manipulations have been made of place cells[15,72,73] and recall has been suppressed by disrupting replays[26]. However, just one study thus far has successfully enhanced recall by manipulating replays[74], and no studies yet have altered recall through manipulating behavioral sequences.

Engrams and sequences contain the same neurons yet encode different but complementary variables. Sequences reveal how the neural code adapts and is structured to represent temporal order, while engrams reveal how the neurons are recruited by gene expression changes and represent context. Future studies should examine the temporal activity patterns of engram cells, which could provide insight into how this minutes-long timescale code could be processed by the brain on the milliseconds-long timescale of dendritic integration and synaptic plasticity.

Are ensemble codes relevant to behavior?

To examine whether classically defined ensembles might form a behaviorally relevant syntax, we must test whether these ensembles affect navigation behavior. Multiple definitions of ensembles can be related to behavior to predict spatial paths and learning. Theta sequences[75] and replays[76–79] (but see also [80]) can predict subsequent trajectories to goals. During correct trials as compared to error trials, behavioral sequences have steeper slopes[81] and replays have more co-active cell pairs[82]. As learning progresses, the strength of cell assemblies in the hippocampus[83] and prefrontal cortex[33] increases, either through more cells participating or greater participation of existing member cells. Humans with more flexible cell assembly membership have greater recall[84]. Mice with more distinct representations of different conditions, through context-specific place cells[85] or remodeled assemblies[32], learn faster. Rats with more place cells representing the goal location learn to find this goal faster[86]. Finally, Alzheimer’s model mice with fewer sharp-wave ripples learn spatial goal approach and goal avoidance slower[87]. Therefore, the content, structure, and incidence of some ensembles correlates with goal-directed navigation performance.

Beyond correlations, optogenetic manipulations provide compelling, causal evidence that ensemble codes support spatial navigation behaviors (Table 1). Driving just a few dozen CA1 place cells linked to reward at an encoded location can induce behavior that indicates the animal has associated that location with reward. Targeting place cells whose fields were inside the rewarded zone on a linear track, Robinson et al activated these neurons in a different location, inducing reward consumption at the new location[73]. Inducing an ensemble of neurons to form a place field at a future reward location improves learning of that reward location[15]. Closed-loop stimulation of the medial forebrain bundle upon spiking of a target place cell during wake or sleep induces place preference for the location that cell encodes[72]. In sum, place cell representations of rewarded locations are causally related to acquiring a goal.

Table 1.

Manipulating CA1 ensembles alters navigation behavior.

Manipulation	Target	Condition	Effect	Reference
Drive	Place cells with fields in the reward zone at end of track	At an earlier location	Reward consumed at earlier location	[73]
Drive	Place cells with fields at start of track	At a later location	Overshoot goal without consumption	[73]
Induce place coding	Pyramidal cells	At future goal location	Faster acquisition when new goal introduced	[15]
Stimulate medial forebrain bundle	Place cells with fields at target location	When these place cells fire, during sleep or wake	Conditioned place preference	[72]
Drive	Engram	In context not associated with footshock	Ectopic recall	[19]
Inhibit	Engram	In context associated with footshock or during sleep	Impaired recall	[19,89]
Interrupt	Sharp-wave ripples	At goal location or during sleep	Slower learning	[26,88]
Interrupt	Sharp-wave ripples	When replay encodes goal location during sleep	Impaired recall	[90]
Lengthen	Sharp-wave ripples	During inter-trial interval	Faster learning	[74]

Manipulating ensembles on the scale of CA1 replays and engrams also alters spatial memory. Driving activity in engram cells induces freezing outside of the associated fear context, while inhibiting these cells suppresses freezing inside the associated context[19]. Similarly, suppressing sharp-wave ripples specifically at a goal location impairs learning[88]. Closed-loop manipulations have shown that disrupting replays during sharp-wave ripples[26] or reactivation of engrams during sleep[89] impairs learning. However, most studies have disrupted all sharp-wave ripples, including those replaying content from other environments. To address this, Gridchyn et al used an online decoder to disrupt only replays of a specific environment to show that it only impairs recall of that environment[90]. Finally, artificially extending sharp-wave ripples also extends their replay trajectories and improves spatial working memory[74]. Thus, disrupting engrams and replays, during recall or consolidation, disrupts spatial memory, while enhancing engrams and replays enhances spatial memory.

Collections of neurons can be manipulated to alter navigation behavior. Although we restrict our targeting and interpretation of manipulations to specific ensemble definitions, these likely affect neurons outside the target ensemble and could be viewed from different ensemble perspectives, perhaps to different conclusions. Future work should try to manipulate population structure or demonstrate a correlation between changes in population structure and navigation.

Different conclusions from different views

Conclusions are often drawn from just viewing neural activity through the lens of one or two ensemble definitions. However, coding can change - due to altered environments or disease - through one ensemble lens but not another. For instance, when barriers are introduced or goal locations are changed, different ensembles emerge without individual neurons remapping - changing their tuning to reflect the environment change. Replays adapt to represent paths around changing barriers without place cells remapping[78], and novel goal locations are overrepresented in replays despite not being overrepresented by the place cell population’s mean firing rate[91]. In a recent study, Mau et al identified that cell assemblies reduced their co-activity over a single reversal learning session in the same environment, while place fields of the cells in the assemblies did not remap[32]. Animals can acquire new goals without remapping of single neurons, while remapping of assemblies and replays emerges early in learning[32,91], suggesting these ensembles may be the substrate for spatial learning. In contrast, when a larger change like a different context or different task is introduced, representations can be stable at the population level while dynamic at the single neuron level, allowing a coherent map to be preserved while individual neurons remap to convey context[30,92,93]. Similarly, population representations of a context can be stable over weeks while individual neuron representations drift[94]. This suggests that when single neurons do remap, they may do so in a coordinated manner, such that the underlying dimensionally reduced manifold retains the same shape but is rotated or translated in the subspace, as has been previously seen[53]. In sum, small environmental changes can alter replays and assemblies without changes to individual neurons, while larger changes can remap individual neurons while keeping population codes intact. Studies that examine only one of these views may not capture the full complexity of remapping.

Coding changes can be also observed through one ensemble view but not another in disease models. Many disease models with impaired spatial memory have normal place coding at the single neuron level, but impaired ensemble coding. First, Suh et al found that schizophrenia model mice have normal place fields, but no sequential relationship between spike timing and decoded spatial position in replays[95]. Disorganized replays have since also been found in schizophrenia patients[96]. Second, mice hemizygous for Scn2a, a gene involved in several neurodevelopmental disorders, have normal place fields but truncated replays with fewer place cells recruited[97]. Third, Poll et al identified sub-ensembles within the contextual engram classified by the conditions in which they were active. While control and Alzheimer’s model mice had similarly sized engrams, Alzheimer’s model mice had a subensemble which encoded novelty. When this sub-ensemble was suppressed, it prevented the novelty signal from interfering with the engram, rescuing recall[98]. Finally, Viana da Silva et al found that mice expressing amyloid precursor protein in CA3 had normal place fields yet disordered theta sequences[99]. Thus, disorganized time-compressed sequences and aberrant novelty ensembles are present in disease models lacking spatial tuning impairments. Examining place coding across a broad range of ensemble definitions has revealed that coordination of neurons may be a target for therapeutics and a biomarker for spatial memory rescue.

Beyond classic ensemble definitions

Looking beyond experimenter-defined ensemble classifications, new methods of detecting relationships between spike trains have uncovered novel ensemble structures. Repeated subsetting approaches have identified that neurons in CA1 sequentially fire in small, stable ensembles, which later combine into behavioral and replay sequences[100,101]. This organization could explain why theta sequences and replays are chunked into smaller trajectories, each within a single slow gamma oscillation[10,11]. Two models which approximate neural activity using spatiotemporal motifs have been used to order neurons into spiking sequences, which extracted behavioral sequences spanning a linear track run without the model having access to the decoded location[102,103]. This could identify new forms of sequential firing which don’t map cleanly onto a sequential variable such as time or location. Finally, our understanding of replays has been extended by new classification algorithms that identified that most replays don’t meet traditional definitions of replay, yet represent clear spatial trajectories[39,104]. This not only revealed that these previously discarded replay events might contain useful, decodable information, but also that an earlier finding about replay content depended on which subset of replays was considered[39]. Thus, restricting analyses to only ensembles that meet the stricter, classic definitions can change findings and ignore information the brain is integrating. However, classically defined ensembles are easier to target and manipulate, thus whether these new ensembles alter behavior is unknown. In sum, unsupervised methods reveal many activity structures previously unknown, without making assumptions about how the brain works, but their biological relevance is still unstudied.

Conclusion

Neural ensembles arise from intrinsic neural dynamics and can reveal structure, variables, and remapping independent of individual neuron representations. Studies thus far have restricted themselves to viewing ensembles through just one or two definitions, which constrains which hypotheses can be tested and can miss changes that are visible through one view but not others. Notable recent exceptions have uncovered incredibly rich coding by examining single neurons, behavioral sequences, population decoding, and dimensionally reduced subspaces in the same dataset[44,50]. No one view of ensembles can explain computation. Each ensemble definition provides a unique perspective of the collective underlying neural activity. Broadening our definition of ensemble membership, looking across scales of temporal, neuronal, and representational restrictions on definitions, while remaining couched in physiological plausibility, and subsequent testing for behavioral relevance will help the field better understand the brain from a less experimenter-biased perspective.

Future research directions

Future research could further expand our definition of ensembles. Identifying ensembles in a variety of animal models performing naturalistic navigation strategies, such as seed caching in birds and hunting in bats, will reveal what aspects of ensemble codes are unique to goal-oriented navigation in rodents and which are ecologically common (Figure 3a). Technological advances will continue to increase the number of simultaneous neurons which can be recorded, which will clarify the size of ensembles and reveal new, sparser codes. Recordings across multiple brain regions will enable identification of how neural codes in sensory cortex translate to associational areas and then to motor cortex, or how ensembles may span multiple regions. Ensemble membership should also be expanded to include oft excluded cell types which can nonetheless encode spatial variables - such as inhibitory neurons[105,106] and astrocytes[107,108], as well as electrical activity besides action potentials, such as dendritic activity[109] and subthreshold activity (Figure 3b).

Figure 3.

Future research directions. Future research should: (a) examine ensembles during more behaviors, in more animal models (such as seed caching in chickadees), across more brain regions, and with more simultaneously recorded neurons, (b) expand ensemble membership beyond excitatory cell spiking, such as to astrocytic activity, inhibitory neuron spikes, or non-spiking activity, (c) identify sequences and clusters with the fewest assumptions, while considering pre-existing connectivity, multiple variables, and multiple dimensions of variables, and (d) selectively silence or drive ensembles to examine the effects in downstream brain regions and on behavior. Chickadee reprinted with permission from [113].

Future research should also test the breadth of ensemble representations and their biological relevance. Using unsupervised approaches can identify ensembles without experimenter-imposed structure (Figure 3c). Ensemble analyses should consider how pre-existing dynamics constrain potential single neuron representations, rather than assuming that all neurons are equally likely to represent all variables[1,2]. Models that decode variables from ensembles should consider that a single neuron can encode multiple variables[110,111] and that multiple dimensions of a variable, such as peak firing rate and specificity, may be behaviorally relevant. Finally, the biological relevance and plausibility of ensembles should be tested by artificially altering their activity (Figure 3d). Future tools may be able to artificially induce sequential ensembles[74] or alter the formation of ensembles independent of individual neuron tuning. The biological plausibility of the resolution of decoded signal should be considered. For instance, population decoding can discriminate visual contrasts at 100x the resolution of what the mouse behaviorally indicates it can distinguish[112]. To test biological relevance, future work should continue to assess how behavior is casually altered by ensembles and should begin to study how codes are integrated by downstream readers.

Highlights.

• Neural ensembles support navigation

• Multi-neuron ensembles can represent variables independent of the single neuron code

• Different views of ensembles can represent different variables

• Broader definitions of ensembles reveals the full richness of coding