The generative grammar of the brain: a critique of internally generated representations

Abstract

The past decade of progress in neurobiology has uncovered important organizational principles for network preconfiguration and neuronal selection that suggest a generative grammar exists in the brain. In this Perspective, I discuss the competence of the hippocampal neural network to generically express temporally compressed sequences of neuronal firing that represent novel experiences, which is envisioned as a form of generative neural syntax supporting a neurobiological perspective on brain function. I compare this neural competence with the hippocampal network performance that represents specific experiences with higher fidelity after new learning during replay, which is envisioned as a form of neural semantic that supports a complementary neuropsychological perspective. I also demonstrate how the syntax of network competence emerges a priori during early postnatal life and is followed by the later development of network performance that enables rapid encoding and memory consolidation. Thus, I propose that this generative grammar of the brain is essential for internally generated representations, which are crucial for the cognitive processes underlying learning and memory, prospection, and inference, which ultimately underlie our reason and representation of the world.

Introduction

Empiricism^1,2 and rationalism^3,4 have differentially predicated the foundations of human reason with regard to the role of experience in the formation of organismal representations about the external world. According to empiricist views, immediate experience with external stimuli is necessary, sufficient and fully instructive for a posteriori creation of organismal representational and adaptive processes regarding those stimuli from a blank slate^1,2. By contrast, rationalist views predicate the existence of a priori repertoires of preformed pre-representations or processes that are selected and modified following direct interaction with external stimuli to produce specific organismal representations or actions^3,4. Here, I explore the dynamic interplay between these differing doctrines and propose that a moderate rationalist–empiricist approach may best explain the intricate contributions of network preconfiguration and experience-driven plasticity to organismal representational processes.

Fundamental discoveries in the broader field of biology have revealed that dominance of either philosophical view can be problematic for advancing the understanding of biological complexities. Most notably, an earlier, empiricist-inspired dogma in the field of immuno biology had argued that the specificity and diversity of antibody responses to specific external antigens are fully instructed and created a posteriori following organismal first interaction, with the antigens viewed as direct templates^5,6 (Fig. 1a). However, a remarkable series of theoretical and experimental rationalist-inspired critical works pioneered by Burnet and Jerne^7–10 predicted and demonstrated the existence of an a priori repertoire of diverse antibodies (and lymphocyte B receptors). A first encounter with a foreign antigen induces clonal selection and amplification of the pre-existing specific lymphocytes and increased secretion of specific antibodies but does not instruct their de novo creation a posteriori^7–10. Intriguingly, experience with the antigen drives the process of clonal selection from a pre-existing lymphocyte and antibody repertoire, generated a priori via spontaneous somatic gene mutation¹¹ (Fig. 1b), which can even support a rapid immune response to a synthetic never-before-experienced antigen¹². Jerne envisioned the process of a priori internal generation of antibody diversity as a form of generative grammar of the immune system¹³, with the pre-existing antibody repertoire resembling a linguistic vocabulary.

Fig. 1 |. Instructional-to-selectionist theoretical shift in immunology and neurobiology.

a, Instructional theory of antibody diversity, which argues for an absence of pre-existing specific antibodies and for their de novo creation after the first contact with novel antigens. b, Clonal selection theory demonstrating the existence of a repertoire of pre-existing antibodies (colour coded) and B lymphocyte clones that undergo clonal selection and amplification after first contact with novel antigens. c, Instructional theory of de novo neuronal firing sequences being generated in the hippocampus on blank slate (tabula rasa) networks during a novel experience (on a linear track; middle). Earlier interpretations of exclusive externally driven hippocampal network activity have emphasized the lack of correlated temporal sequences during pre-experience sleep or rest (left) and the replay of only the recent experience-related temporal sequences during post-experience sleep (right). d, Neuronal sequence selection theory demonstrating that specific neuronal sequences expressed during novel experiences (sessions on three distinct linear tracks; colour coded) are primarily selected from a repertoire of pre-existing preplay temporal sequences expressed in the pre-experience sleep or rest as part of a preconfigured hippocampal network (left). All temporal sequences are replayed during post-experience sleep or rest (right), when they are generally stronger than during preplay. In parts c and d, letters represent individual neurons, colour coded to match participation in corresponding experiences; arrows indicate the order of neuronal firing, colour coded to match their sequential activation on the corresponding linear track. Parts c and d are adapted with permission from ref. 21, the Royal Society.

Inspired by the success of the antibody clonal selection theory of Burnet and Jerne and, more so, by the success of the broader theory of natural selection by Darwin¹⁴, Edelman¹⁵ proposed, in his 1987 book Neural Darwinism – The Theory of Neuronal Group Selection¹⁶, that processes of neuronal selection should also underlie the development and function of the nervous system. Unfortunately, the way Edelman initially envisioned somatic, synaptic and higher-order neuronal selection in his book received a cold reception from contemporary luminaries such as Barlow¹⁷ and Crick¹⁸. Decades later, the discovery of hippocampal network preconfiguration into a repertoire of temporal sequence motifs that preplay^19–21 newly expressed sequences of place cells encoding novel animal trajectories provided critical support for the neuronal selection theory coined by Edelman, for the role of preconfigured networks in rapid information encoding and for a generative grammar of the brain.

In this Perspective, I perform an in-depth investigation into the meaning of neuronal ensemble activity patterns and how they relate to the formation and expression of novel representations of the external world, ultimately envisioned as a generative grammar supporting a neural code. I first provide a historical overview of general philosophical and neurobiological views on hippocampus network activity preceding information encoding, and how the discovery of sequence preplay catalysed a transition from blank slate to network preconfiguration. I discuss how neuronal selection in preconfigured networks is hierarchically organized from individual neurons to extended neuronal sequences according to the generative rules of seemingly universal biological grammar, shared by genetics and linguistic morphology. Hippocampal network competence and performance are then envisioned as syntactic and semantic components of a generative neural grammar. Next, I discuss the postnatal development of the neural grammar and the link between neurobiological and neuropsychological perspectives on neuronal grammar and the neural code. Last, a novel role for the generative grammar of the hippocampus to support the higher-order category of internally generated representations, which include memory, prospection and inferences, is discussed. Through this process, I argue that the generative grammar of the brain could support a priori the concepts of space, time and, possibly, other types of representations about the external world.

The age of pure learning

The critical roles of the hippocampus in episodic memory formation in humans^22–24 and cognitive mapping in rodents^25–29 marked the field of neurobiology (particularly the neurobiology of contextual learning) from its beginning. Before 2011, mostly empiricist views regarding the role of experience in the creation of novel neuronal patterns for contextual encoding and consolidation predominated. Backed by early experimental findings, novel spatial experiences were postulated to create de novo neuronal activity patterns within hippocampal blank slate networks during the encoding of novel contexts^30–32.

This blank slate view was compatible with the strong bias given to memory research by the discovery of long-term potentiation of synaptic transmission, which quickly became the leading synaptic plasticity mechanism for learning and memory^33–41. Accordingly, using electrical stimulation as its model, the learning and memory field began describing synaptic plasticity induced by patterned activation of the hippocampal network⁴⁰ as in excess of the strong pre-existing activity — the evoked activity — whose contribution to learning was implicitly neglected. Consequently, a consensus grew that encoding new experiences and sequences of events in space and time crucially relied on de novo creation of corresponding neuronal sequential motifs within blank slate hippocampal networks as a form of pure learning (Fig. 1c). Consolidation of the newly created activity patterns into memory would occur via their local time-compressed reactivation and replay during subsequent slow-wave sleep^30,42–45, without adding any new or additional meaning to the encoded information, during which they are communicated to downstream neocortical areas^46–50 for long-term storage. Thus, learning and memory formation were thought to occur in two stages of neuronal ensemble activity, supporting encoding and consolidation^43,51. First, during encoding, external stimuli exert a completely instructive role over neuronal activity and create novel, unique sequential motifs on assumed blank slates^30,50,52,53. Second, during consolidation, these novel encoded patterns are recurrently replayed for several hours during subsequent sleep or resting epochs^44,54,55. By the next day, hippocampal networks would be devoid of most of the neuronal ensemble activity structure of the previous day, so would again appear as an uncorrelated, blank slate and be ready to encode new sequential information without interference^{30–32,50,53}.

This model of neuronal ensemble activity dominated the field of hippocampal learning and memory until 2011, despite earlier reports indicating that ongoing spontaneous activity in the sensory neocortex correlated with subsequent stimulus-induced neural activity^56–59. The dichotomy between the ways in which the sensory neocortex and the hippocampus would process novel information was justified in part by the need for the sensory neocortex to depict stimulus features invariably across different contexts, supported by the stronger hard-wired, less plastic network compared with the hippocampus⁶⁰.

In summary, three major factors are likely to have contributed to the predominance of this blank slate, experience-driven view of hippocampus activity. First, patterns related to future novel animal experiences were envisioned as an infinite set of possibilities; any attempt by the brain to predict them was deemed hopeless. Second, no research laboratory had empirically reported the existence of neuronal sequential motifs in the hippocampus that were predictive of or correlated with future patterns of neuronal activity expressed during encoding of novel environments. However, debates leading up to 2011 revolved around whether hippocampal activity during the sleep or rest preceding a re-exposure to a familiar context was correlated^31,61 or not³⁰ with the activity patterns that would be re-expressed during that re-exposure. Third, the theoretical, conceptual and philosophical levels predominately argued that brain activity during pure learning merely responds to instructive sensory stimuli from the external world.

Network preconfiguration and neuronal selection in the hippocampus

In 2007, results from three complementary studies extended our understanding of the role of the hippocampus in cognition beyond episodic memory and cognitive mapping. First, individuals with bilateral hippocampus damage-induced anterograde amnesia had trouble imagining new experiences⁶². Second, future thinking and mental scene construction tasks activated the hippocampus in neurotypical individuals like memory-engaging tasks⁶³. Last, the hippocampus was critical for the accelerated assimilation and consolidation of new information into pre-existing mental schemas in experimental rats⁶⁴. These results indicated that the hippocampus functions not only as a recorder of current experiences (to recall later as memories) but also as a prospective organ supporting mental constructs and frameworks for events and associations never-before experienced^21,65,66. Thus, the hippocampus supports higher-order internally generated representations of events currently not present^21,67, either from the past (recalled) or the future (imagined, inferred). Dreaming content during sleep⁶⁸, when the brain is disconnected from the external world, is impaired in individuals with bilateral hippocampal damage, which further indicates that the hippocampus has a critical role in supporting internally generated representations. Moreover, individuals with psychosis have reduced hallucinatory and delusional activities — forms of abnormal internally generated representations — following bilateral hippocampal removal^22,69,70.

Around that time, while working on understanding the organization of neuronal ensemble patterns in the hippocampal CA1 area using CA3 N-methyl-d-aspartate receptor (NMDAR) gene knockout (KO) mice⁷¹, which have abolished NMDAR-dependent synaptic plasticity of CA3–CA3 auto-associative and entorhinal cortex–CA3 pathways upstream of CA1, we made an unexpected discovery. The CA1 place cells of CA3 NMDAR KO mice exhibit reduced place field tuning during exploration of novel environments but normal tuning in familiar ones⁷¹. Surprisingly, in CA3 NMDAR KO mice these cells displayed normal place field tuning in novel environments made contiguous with a familiar one, even if the neurons had been previously inactive (‘silent’) in the familiar environment⁷². This led us to hypothesize that these silent neurons had probably been captured into ‘tuned’ neuronal ensembles during a prior activity during the familiar environment exploration (since it was the only other context experienced by mice that day outside the home cage and sleep box, and immediately preceded the novel track exploration). Thus, we inferred that co-activation of silent neurons with neurons already tuned to the familiar environment might have occurred in neuronal ensembles during the rest or sleep epochs preceding exploration of the novel environment by each animal.

Indeed, we found such co-activation within time-compressed temporal sequences during the resting epochs preceding the novel environment exposure, in both plasticity-deficient CA3 NMDAR KO mice⁷² and normal control mice¹⁹. Intriguingly, subgroups of these time-compressed temporal sequences correlated with future place-cell sequences on the novel but not the familiar segment of the contiguous environment^19,72. We called this novel phenomenon preplay in 2011 (ref. 19), to indicate its prospective nature and to distinguish it from retrospective processes such as replay. Preplay also occurred in adult experimentally naive control mice and laboratory rats during the slow-wave sleep session prior to their first (de novo) as well as subsequent (not de novo) explorations of novel extended linear environments^19,20,73. Importantly, these results observed in the absence of prior explicit experience with extended linear environments first indicated the internally generated nature of preplay and demonstrated that hippocampal networks are preconfigured into temporal sequences of firing that are used to rapidly encode future novel spatial trajectories²¹. The experimental proof of preplay and network preconfiguration in naive animals drove the conceptual transition from instructional theories of novel information encoding (Fig. 1c; experience-driven onto blank slate networks) towards a neuronal selection theory (Fig. 1d; neuronal selection from preconfigured networks).

Do network preconfiguration and preplay imply that the hippocampal network (or the animal) can predict a specific future novel experience? I caution against this interpretation. Instead, I propose that recurrent spontaneous neuronal ensemble patterns gain predictive power over the identity and order of neuronal participation in internally generated representations of future novel experiences^{19,20,73–75} that are further modified in response to specific multi-sensory inputs. In the age of pure learning, neuronal sequences were thought to be entirely created de novo during the experience within blank slate networks in response to external drives (Fig. 1c). By contrast, a neuronal selection process by which sequential neuronal activity expressed in the context of specific multi-sensory inputs during encoding of any future novel environment is primarily selected from an internal repertoire of pre-existing temporal motifs^{19–21,72,73,76–78} is consistent with network preconfiguration and preplay (Fig. 1d).

The existence of a neuronal selection process is critically supported by our key discoveries. First, preplay^19–21 provided the crucial evidence that hippocampal networks preceding novel experiences are preconfigured into sequential motifs and not blank slate networks. Newly expressed place-cell sequences encoding novel animal trajectories correlate with pre-existing temporal sequence motifs expressed during the preceding sleep or awake rest, indicating that experience-driven neuronal selection occurs during encoding. Second, we revealed in animals exposed to multiple distinct linear environments that a pre-existing repertoire of temporal sequences existed (as opposed to a single dominant generic sequence)^20,72,74,79. This exposure led to the expression of unique place-cell sequences across each novel environment. Furthermore, individual place cells in each unique spatial sequence had participated in distinct sets of corresponding temporal sequence preplay (with reduced overlap) from a larger pre-existing repertoire expressed during the preceding sleep^20,21,72,74. Importantly, most of the pre-existing repertoire remained unallocated, indicating a larger preconfigured network capacity available for selection during the rapid and distinct encoding of numerous experiences to be encountered (Fig. 1d, grey arrows). Last, we uncovered spontaneous neuronal organization (3 ± 1 neurons or distinct groups of co-firing neurons called cell assemblies) into recurring short, time-compressed temporal sequence motifs — neuronal tuplets — expressed during sleep^74,79. Distinct neuronal tuplets cannot only be selected but can also be multiplexed in a generative manner to generate novel distinct extended temporal sequences, which can then be decompressed into unique spatial-temporal sequences during the encoding of future novel experiences⁷⁴.

The generative neuronal selection process of multiplexing short temporal sequence motifs into a repertoire of distinct extended place-cell sequences during novel experiences is analogous to the linguistic generative process of combining words into numerous longer distinct sentences (Fig. 2a). Similarly, neuronal tuplets (3 ± 1 neurons) are envisioned as neuro-codons (analogues to the codons with three nucleotides found in the genetic code)^74,80. In the generative grammar of the neural system, the role of neuronal tuplets could be analogous to that of genetic codons or words in the generative formation of a repertoire of extended peptide chains during translation or extended sentences, respectively, which suggests that a universal generative grammar is used for information processing in biology (Fig. 2a). Compared to the better-understood linguistic and genetic generative grammars, and to the generative grammar from the repertoire of pre-existing antibodies, which resembles a linguistic vocabulary in the immune system (Fig. 1b), the generative grammar of the brain has had a protracted appreciation.

Fig. 2 |. Unified hierarchical generative grammar and coding across genetics, neurobiology and linguistics.

a, Comparative generative grammar across genetics, neurobiology and linguistics comprised of individual constituents (nucleotide, neurons, letter or phoneme), intermediate coding chunks (codons or amino acids, neuronal tuplets or neuro-codons, word or morpheme), and extended coding sequences (chains of amino acids or peptides, neuronal sequences of multiplexed tuplets, sentences). The size of the repertoire of distinct possible ‘codes’ increases from left to right (individual constituents to extended sequences). Extended neuronal sequences are formed from concatenation, multiplexing and editing of pre-existing chunks of short neuronal sequence motifs. b, Three representative examples of the repertoire of preconfigured neuronal temporal sequences (left) of an animal during sleep or rest that match distinct future novel extended place-cell sequences when the naive animal experiences each novel linear environment (right). The general temporal-to-spatial transformation from time-compressed neuronal firing sequences to matching spatial place-cell sequences is envisioned as a network competence (which has the same neural syntax for all three linear environments). The specificity of temporal sequences that match distinct spatial sequences (trajectories), which can be interpreted as specific representations of individual spatial sequences (a form of ground truth or neural code), is envisioned as network performance (different neural semantic for the three linear environments). c, A sample of the repertoire of preconfigured temporal sequences from the rat hippocampus CA1 area that shows three clusters of temporal sequences, each specifically matching the spatial context of one track but not the other two. Parts b and c are adapted with permission from ref. 20, PNAS.

Competence and performance of neuronal networks

Pre-existing cell assemblies and their sequential motifs in adult naive animals are predictive of and correlated with the future place-cell sequences expressed during the exploration of novel environments, which signifies that the naive hippocampal network has the competence to represent future animal trajectories through novel space prior to its exploration^19,20,73,75. Competence in this context is the ability of hippocampal neuronal ensembles to be spontaneously organized into time-compressed temporal sequences of firing before^19,20, during^{19,45,81–84} and after novel experiences^30,53,73, which can be decoded as possible trajectories through physical (and mental) space (Fig. 3).

Fig. 3 |. Neural competence and neural performance in neuronal ensemble representations of the external world.

Network preconfiguration expressed during sleep or rest produces a repertoire of temporal firing sequences and virtual trajectories (left), which preplay future novel place-cell sequences and depict future novel spatial trajectories (middle). Network preconfiguration or preplay is envisioned as a form of neural competence to express time-compressed sequences that can undergo a temporal-to-spatial transformation to express place-cell sequences during novel experiences. The ability of a network to replay experiences during post-experience sleep or rest (plasticity in replay) with higher fidelity than preplay (right) is envisioned as a form of neural performance (but not as a new competence). The neural competence to operate in (correlated) time-compressed sequences before, during and after a novel experience is akin to a neural syntax (see Fig. 2). The neural performance to represent a novel experience with higher fidelity during and after its occurrence compared to before is akin to a neural semantic, which adds specific experience-related meaning to the pre-existing patterns of activity. The bottom three panels demonstrate a Bayesian decoding approach to trajectory depiction before (left), during (middle) and after (right) the occurrence of the experience. Several key features of network preconfiguration or preplay and plasticity in replay are also listed. Adapted with permission from ref. 73.

After the encoding of novel experiences, cell assemblies that participated (and their sequential motifs) undergo subtle, selective plasticity in replay over preplay during post-exploration sleep or rest (Fig. 1d). This plasticity in replay is expressed as a higher fidelity and incidence in depiction of animal trajectory compared to the pre-exploration sequential cell assemblies, preferentially during replay events with high neuronal participation^{20,73,74,76,85}. Plasticity in this context is merely a gain in hippocampal network performance to internally represent the trajectory of an animal from the pre-existing (preconfigured) network competence (Figs. 2b and 3). It should not be confounded with the acquisition of a new competence to represent spatial trajectories following a new experience. Training-induced increases in human running performance provide one analogy for this network competence–performance relationship, whereby a pre-existing competence to run slowly on a flat 100 m track can, through training, lead to increased performance in running 100 m uphill or fast enough to win the 100 m race in a sports competition. By contrast, no training amount can induce increased performance for unassisted flying, for which the human body has no pre-existing competence. Consistently, the relative gain in hippocampal network performance is smaller and subtler, conceptually and in the magnitude of plasticity in replay^73–75, than the overall pre-experience network competency for depicting novel sequential spatial trajectories (for instance, the difference between proportions of replay and preplay events is much smaller than the proportion of preplay events).

By analogy with linguistic grammar, hippocampal network competence can be envisioned as a form of basic neural syntax with similar general organization of extended time-compressed neuronal sequences before and after any experience. Similarly, network performance as a form of neural semantic enables increased specificity and meaning to internally generated representations of novel experiences for subsequent use (Figs. 2b and 3). Hippocampal network competence and performance have not always been understood in this way. For instance, the network performance gain via plasticity in replay^73–76 has been interpreted as the acquisition of a new competence to represent the novel experience offline, but only after that experience has already occurred. Historically, prior to encoding of novel information, the blank slate (tabula rasa) model deemed the hippocampal network incompetent to represent an experience until during or after its occurrence^30,50,52,53 (Fig. 1c). This model of gaining competence over a blank slate (Fig. 1c) may superficially be understood as similar to the model of gaining in performance over preconfigured competence (Fig. 1d) because, in post-experience compared to pre-experience sleep or rest, a higher-fidelity network representation is induced in both models⁸⁶ (Fig. 1c,d). Whereas the relative gain is true in both models, the critical difference relies on the strength of the pre-existing patterns in comparison to chance (how significant is the competence, the preplay) and on the increased magnitude of the pre-experience (and the post-experience) patterns compared to the actual gain in performance (competence versus gain in performance, preplay versus plasticity in replay). Contrary to the classical blank slate model, my view is that network competence is expressed in both pre-experience and post-experience hippocampal networks and is generally stronger than the gain in performance, which contradicts the intuition of the classical model where the gain in performance was misinterpreted as a new competence^{19–21,72–75,79}.

Thus, this view of non-binary gain in performance over pre-existing competence on experience encoding and representation explicitly integrates the empiricist and rationalist approaches to network learning (Fig. 1d). Neuronal ensemble activity patterns expressed during new learning are partly selected from pre-existing large repertoires (not created de novo; rationalist view) and gain new representational meaning during the novel context experience^20,21, partly instructed by multi-sensory inputs from the external world (experience-driven; empiricist view), which add to the new representational meaning and experience-related specificity of the overall neuronal activity. The specificity of the neuronal selection process — the identity of neurons and neural motifs selected out of the large repertoire available — is determined partly by the intrinsic network activity^{19,74,76,79,87} (prior excitability, firing rate, neuronal coordination; rationalist view) and partly by the specificity of instructional signals from the external world^27,79,88,89 (multi-sensory and contextual stimuli; empiricist view).

Specific ‘selected’ neurons have been variably envisioned as part of a ‘backbone’, ‘rigid’, ‘inside-out’ and ‘top-down’ network^{20,21,74–76,90–92} that underlies network competence, which manifests as preconfiguration, predictive coding and preplay within hippocampal networks. Meanwhile, specific ‘experience-instructed’ neurons have been considered part of a complementary ‘flexible’, ‘plastic’, ‘outside-in’ and ‘bottom-up’ subnetwork^{74–76,79,90–92} that primarily drives hippocampal network performance gain, which manifests as experience-related plasticity in replay⁷³ and prediction-error signals⁷⁴. It has been proposed that this neuronal functional diversity has a preferential distribution across hippocampus CA1 sublayers, suggestive of the existence of parallel, partly anatomically segregated, rigid and plastic hippocampal subnetworks^92–94. While sublayer anatomical distribution of CA1 neurons might favour their differential experience-driven plasticity under certain contextual conditions, it is more conceivable that the assumed roles of rigid and plastic neuronal responses are more dynamic and interchangeable across these neuronal groups in a context-dependent manner^74,75.

As briefly introduced above, neuronal network competence and performance are not unlike syntactic and semantic rules used to describe linguistic grammar⁹⁵. What are the major components of the generative neural grammar that govern hippocampal network organization? There are compelling similarities between a linguistic sentence and extended neuronal sequence organization: the repertoire of preconfigured neuronal firing sequences is analogous to a linguistic vocabulary repertoire; preplay or preconfiguration of a generic trajectory representation is a form of neural competence similar to the linguistic competence⁹⁵ to form a sentence; and plasticity in replay over preplay generates the neural performance gain to discriminate across representations of distinct contexts, similar to distinct new sentences of an analogous syntax that generate the linguistic performance⁹⁵ gain to effectively communicate. Next, I discuss the key components of this generative neural grammar.

The neural grammar syntax

Similar to linguistics, where the language syntax can be better understood through speech — as noted by von Humboldt, “language in actuality only exists in spoken discourse”⁹⁶ — the neural syntax underlying neural competence can be better understood through neural performance. Operationally, network preconfiguration and preplay can be better understood by investigating the use of pre-existing patterns during encoding and their subsequent gain in performance via plasticity in replay during consolidation. Thus, neural syntax is a neuronal organization that is similar in essence before, during and after a novel neural representation and that governs both neural competence and performance.

I envision a tight analogy between linguistic syntax (or morphology) and neural syntax (Fig. 2a). Accordingly, individual neuron activity parallels a letter or a phoneme, with neuron identity matching letter identity (Fig. 2a, left) and its variable firing rate matching the variable duration or emphasis of a phoneme. Cell assemblies — neuronal ensembles co-firing within brief time epochs^73,85,97,98 — are analogous to a linguistic morpheme and syllables composed of tightly assembled phonemes, whereas short sequential neuronal tuplet motifs^74,79 are comparable to words (single or multiple functionally connected morphemes). Individual neurons can participate in multiple distinct cell assemblies^20,73 and neuronal tuplets⁷⁴ in the same way that individual phonemes participate in multiple distinct morphemes and words (Fig. 2a, middle). Thus, for neural syntax, the cell assembly and, more so, the tuplet (as the single and multiple-morpheme words, respectively) generate intrinsic meaning in preconfigured networks, and not the individual neuron (as the more promiscuous phoneme).

The possible tuplet repertoire expressed (or preplayed) during sleep and rest (offline states in a neural network)^20,21,51,74 is analogous to a linguistic repertoire of words, the vocabulary of a language used to represent and communicate experience. The organization of neurons into cell assemblies and neuronal tuplets is generally similar before and after a novel experience that engages those neurons^73,74. These short neuron activity sequences become combinatorial chunks of functionally connected individual neurons (Fig. 2a, middle). This organization enables a generative process that exponentially increases the representational capacity of the network from the size of the repertoire of active neurons (phoneme repertoire) to that of the repertoire of possible tuplets (vocabulary repertoire)⁷⁴. In addition, such organization provides the critical argument for favouring neuronal ensemble coding over the single neuron doctrine^99,100, which is consistent with a generative neural grammar syntax.

The higher-order multiplexing of tuplets into numerous extended neuronal sequences (Fig. 2a, right) that are expressed along distinct novel linear tracks^20,74,79 (Fig. 2b) mirrors the generative process of combining words into numerous distinct sentences⁹⁵. That is, variation in tuplet identity and order rank within multiplexed tuplets generates numerous distinct sequences. During an impromptu speech, new sentences are constantly formed by recombing specific words from an otherwise constant vocabulary while preserving the phoneme (letter)–morpheme (syllable)–word–sentence linguistic syntax. Likewise, during novel experiences, new, distinct, extended, sentence-like neuronal sequences are formed by recombining individual cell assemblies and neuronal tuplets (Fig. 2a, right). This higher-order organizational level of the neural grammar is generative, which combinatorically increases the neuronal network representational capacity from being restricted to the tuplet repertoire size to an almost infinite repertoire of possible neuronal sequences (sentences repertoire size).

Interestingly, the probabilistic, within-sequence relative order rank allocated to individual neurons within extended multiplexed preplay temporal sequences in sleep or rest can match their order rank taken within the extended sequence of place cells on subsequent novel exploratory sessions^19,20. Moreover, some neurons have a probabilistic activation preference towards the start of multi-neuronal sequences during both pre-experience and post-experience sleep, whereas others have a higher incidence towards the middle or the end⁷⁵. These offline organizational features largely withstand neuronal activation during novel experiences^75,79 and appear probabilistically (pre)configured into a generic multi-neuronal primitive sequence. This generic primitive sequence is a flexible syntactic framework onto which new neuronal patterns are embedded and which is the primary driver of preplay and replay sequences⁷⁵. This is not unlike the syntactic probabilities for specific morphemes and words to occur preferentially at the beginning, middle or end of sentences (such as ‘the’, ‘and’ or ‘too’, respectively). The generic organization of individual neurons into cell assemblies, tuplets and extended sequences in an equivalent manner before (preplay), during (theta sequences) and after representing a novel experience (replay) is akin to a neuronal syntax that underlies the neural competence to rapidly represent a sequential experience.

The specifics of a novel experience are recorded as increased neuronal firing rates, co-firing, tuning to cell assembly, tuplet editing and participation within an extended sequence^19,20,73,74 (Fig. 3). These plasticity changes reflect aspects of the experience-dependent gain in network performance operating within the grammatical rules of the pre-existing network competence. Thus, our understanding of network performance depends on our understanding of network competence and its generative grammar.

My assessment of network competence and performance reflects a neurobiological, network-centric perspective on brain function, which aims to elucidate the general neural grammar rules that govern the hierarchical organization and function of neural networks irrespective of their immediate roles in specific cognitive functions. However, a parallel cognition-centric, neuropsychological approach interprets neuronal patterns as a neural code that critically serves a specific cognitive function or behaviour such as memory or cognitive mapping (and their behavioural readouts; Fig. 2b). In the next section, I discuss the idea that a better understanding of the neuronal grammar as a form of functional organization of a network is critical for understanding the neural codes employed.

The phonological and semantic components of the neural grammar

A generative grammar is composed of syntactic, phonological and semantic components⁹⁵. As discussed above, the hierarchical functional organization of individual neurons into cell assemblies, tuplets and extended sequences provides the neural syntax underlying network competence to represent sequential information. Below, I briefly introduce an approach to decipher the phonological component during communication between brain areas before focusing on how different understandings of the neural semantic inform perspectives on the nature of neural codes employed in internally generated network representations. Through this process, the neural generative grammar used by the hippocampal network would be fully elucidated.

The phonological component, analogous to communication during speech, could be envisioned as the functional interaction between connected brain areas, with an upstream area transmitting information and a downstream area decoding it^{48,50,101–106}. A neuroscience approach to deciphering this phonological-type interaction between brain areas would be successful if (1) both brain areas use a neuronal grammar; (2) the neuronal grammars of each area, although different from each other, are both understood well; and (3) the information transmitted from the upstream to the downstream areas is transformed within the grammatical rules governing their network performance. The generative grammar described based on time-compressed cell assemblies, tuplets and extended neuronal sequences seems to best characterize the hippocampus network (and probably some brain areas (multi)synaptically connected to it)^104,107,108. The existence of analogous word or syllable-like and sequential functional organization has been reported in the rodent dorsolateral striatum motor network^109,110, avian song motor network^111,112 and rodent visual system^113,114. When the neural grammar of areas upstream and downstream of the hippocampus, such as the entorhinal cortex, prefrontal cortex and the septum, become better understood, the phonological component of neural grammar would be an interesting way to study the communication between these brain areas.

Inspired by linguistic studies, it has been argued that the neural semantic component could be understood as the pure representation (ground truth) of a certain unique experience, which is distinctive from similar yet different experiences^19,20,115. With the advent of Bayesian decoding techniques, in addition to current animal trajectories through space, past (replay) and future (preplay) novel animal trajectories depicted during sleep can now be visualized. Some proposals speculated that a pure representation might be accessible only by simultaneously observing neuronal activity in the downstream areas related to that same unique experience as a correlated readout of hippocampal network activity^86,90,115. I think this does not always need to be the case, for several reasons. First, the plasticity in replay over preplay that is specific to one experience but not to others may already qualify as a semantic component of the neural grammar. This neural semantic would be supported by the specific network performance gain to represent a specific and distinct experience during post-experience compared to pre-experience. Second, replay of a specific experience is sparse, and many such replay events are highly correlated with different aspects of the experience but not always correlated among themselves (for example, neuronal functional dependency length is short, one to two preceding cells⁷⁴). Thus, many replay events important for the same experience appear as ‘synonyms’ with similar syntax (made of cell assemblies and multiplexed tuplets) yet partly different content, which would be decoded differently by downstream areas despite being correlated with the same experience. Last, downstream areas simultaneously receive a convergence of inputs from multiple subregions and areas, some not involved in representing the exact same aspect of the given experience, which can still influence activity in the downstream areas. I propose that the communication between the ‘internal language’ of the brain as the decoder and the external world as the transmitter is more approachable. In this framework, the intrinsic network competence can be envisioned as a context-independent source of generic predictive codes^19,74 and the context-specific bottom-up information from the external world would act as a prediction-error signal that can improve network performance in representing the world^74,79.

Generative neural grammar underpins the ability of basic neuronal patterns (individual cells, cell assemblies, tuplets) to combine in numerous ways to generate a large repertoire of distinct preconfigured neuronal sequences that support representations of numerous experiences in neural networks^20,74,79. Despite the differences across the resulting sequences, the neural syntax (organizational rules) and the neural vocabulary (overall repertoire of motifs) are preserved across all novel sequences being generated. Generative grammar serves two distinct representational processes: de novo preplay with subtle editing during future novel experiences^{19,20,73,75,76} and generative preplay, a novel combination of pre-existing (forward and reverse) sequential motifs previously created in relation to past documented experiences^77,78. Both generative processes are organized in compressed temporal sequences, occur prior to (and in the absence of) a directly matching sequential experience and are instantiations of neural competence (and flexibility). Thus, de novo and generative preplay are internally created, generative processes that ‘set the brain free’ from dependency on prior exactly matching experience.

Network preconfiguration critically enables other processes in addition to de novo and generative preplay. First, a higher speed of encoding that supports one-trial learning (compared to the slower process of de novo creation of neural patterns within a blank slate network). Second, the explicit use of prior information and knowledge for new encoding or learning (not explicit in the replay function, which has been mostly linked to post-experience memory consolidation). Last, a capacity for making specific predictions enabled from the preservation of the past or current identity and specificity of the neural network when responding to a generic stimulus. Moreover, these processes critically distinguish network preconfiguration (and network competence) from the more specific increase in network performance to represent a recent novel experience, which merely expands the pre-existing generative process and is expressed as plasticity in replay^51,73,76.

De novo and generative preplay are critically different from prospective replay (when replay of a familiar context is re-expressed prior to re-exploration of that same context^82,83,116). Replay has traditionally been envisioned as an external-input unifactorial discriminator and classifier that models hard boundaries between different inputs (non-transformative and non-predictive in nature), which support a neuropsychological perspective on the neural code. By contrast, de novo and generative preplay have been envisioned as an external-input multifactorial integrator into pre-existing internal network dynamics that models input distributions with partly overlapping features, which continuously evolve and are somewhat predictive of future patterns of activity expressed during novel experiences. These preplay features support a neurobiological perspective whereby the neural syntax described in the previous section is a prerequisite for expression of the neural semantic as the neural code (Figs. 2b and 3). Thus, processes called ‘generative or transfer replay’⁷⁸ and, based on one non-peer-reviewed preprint, ‘re-tuning’¹¹⁷ — whereby hippocampal networks generate unique novel sequential patterns by recombining parts of previous replays to match future novel sequential experiences — are merely instantiations of generative preplay (the perspective is shifted towards future-oriented representations in generative preplay from past-oriented representations in replay). A critical requirement of a pure generative replay process is to preserve the prior representational meaning of the previous replay parts being recombined or re-tuned, which does not necessarily apply to generative preplay. During novel future experiences, the pre-existing internally generated temporal motifs undergo an across-modality temporal-to-spatial transformation when the preceding temporal sequences are partly matched by the future spatial sequences^19,20,74. This temporal-to-spatial transformation is likely to be enabled by the similarity in the intrinsic organization of space and time, two modalities characterized by their quantal, sequential and contiguous nature. The conversion of time to space could likely be generalized to the conversion of time to any cognitive task and/or modality that is quantal, sequential and contiguous in nature^118,119.

During a novel future experience, the syntactic and semantic components of the neural generative grammar constrain the network patterns expressed, which enables context-independent, generic top-down predictive coding^{74,79,120–123}. The extent of this predictive coding is constrained by the identity of participating neurons, the size of recombining tuplets and their combinatorial rules. This is analogous to linguistics where, during a speech, the next sentence spoken is generally constrained and predicted at the level of the pre-existing phonemes, vocabulary and grammatical syntax of the language. To generalize from similar past to future novel experiences, the previous network activity patterns can augment, in a context-dependent bottom-up manner, the overall predictive power of the preconfigured network for related versus unrelated future novel experiences. Indeed, bottom-up neuronal patterns expressed during similarly oriented contexts augmented top-down predictive coding of hippocampal neuronal activity during future novel contexts (calculated from pre-experience sleep) more than patterns expressed during unrelated orthogonal contexts^79,123. This is analogous to sentences used in previous speeches being more predictive of the sentences used in related versus unrelated future speeches, in addition to the more generic, context-independent predictive power of the phonemes, vocabulary and syntax of the language.

The postnatal development of the neural grammar

Internally generated representations in the hippocampus have a role in the rapid relational binding of spatial locations and events into spatial and mental trajectories and memory episodes^{25,81,124–126}. In particular, the rapid encoding of sequential spatial experiences into memory episodes is achieved during navigation by internally generated representations of spatial trajectories within time-compressed hippocampal place-cell sequences (for example, theta sequences^81,127–129). Theta sequences are partly selected from a pre-existing sequence repertoire^19,20,73,74, detectable during preceding sleep and rest periods, and partly contributed by the multi-sensory-motor information during novel experiences¹²⁸. A memory consolidation phase follows the rapid encoding of place-cell sequences, initially supported by the plasticity in replay during rest and sleep following novel experience^{51,54,55,73,130}. Encoding and consolidation are the two neuropsychological stages of memory formation, which map onto the proposed neurobiological syntax (neural competence, preplay, theta sequences and replay) and semantic (neural performance, experience-specific theta sequences and their plasticity in replay) components of neural grammar described above.

These parallel neuropsychological and neurobiological attributions of neuronal patterns to memory stages and grammatical components, respectively, as well as their dependency on experience, are understood primarily on the basis of studies on adult animals. An important question became when and how these neuronal patterns emerge during animal development and to what extent are they innate or induced by structured early-life experience. We revealed that network competence and performance emerged in three distinct stages using eye-opening in rodents — which initiates visually guided movement through space and time-compressed representation of spatial trajectories — as our first neurodevelopmental time point¹³¹. In stage 1, prior to becoming competent in expressing time-compressed neuronal sequences, persistent, stationary and location-depicting neuronal ensembles dominated network activity during sleep occurring on postnatal day (P) 15–16; additionally, preplay, theta sequences and spatial trajectory replay were all below chance levels^131,132. In stage 2 (P17–22), the gradual age-dependent and recent experience-independent expression of preconfigured trajectory-like preplay and of replay sequences during sleep and rest mark the emergence of network competence¹³¹. During stage 2, the network clearly has the competence to represent sequential trajectories during sleep and rest but network performance does not exhibit increased experience-dependent trajectory representation, either by theta sequences during experience or by plasticity in replay post-experience.

Stage 3 (P23–24 onward) is dominated by increased network performance for experience-dependent trajectory representation and is marked by the emergence of theta sequences during navigation and plasticity in replay during rest and sleep^131,132. Interestingly, this neurobiological developmental stage in rodents also marks the emergence of the adult-like neuropsychological stage of rapid memory encoding and consolidation, which ends the prevailing infantile amnesia period^51,133–135. In adult animals, increased neuronal performance (Figs. 2b,c and 3) manifests in numerous ways. First, discrimination across different contexts increases gradually from early to late laps of sessions⁷⁹. Second, increased representation of an experienced location, event or trajectory during replay versus preplay^73,74,76 manifests as increases in the number, tuning, and temporal coordination of neurons participating in cell assemblies and increased reactivation or sequence replay specifically in spiking events with the largest proportions of neuronal participation. Third, incidence, not size, of neuronal tuplets selected in the novel experience increases during replay versus preplay⁷⁴. Last, neuronal sequence length and proportion of participating neurons increase during the replay of longer experiences versus preplay^55,73. The exact neurodevelopmental timeline for all these aspects of increased network performance remains to be studied.

Neural grammar comes of age

Over the last twelve years since 2011, the idea of a hippocampal neural grammar organized into a repertoire of preconfigured temporal sequence motifs that preplay neuronal recruitment during future novel experiences has received strong experimental support. At least nine different methodological and conceptual approaches across several laboratories and animal species (mouse, rat and human) have successfully demonstrated network competence, preconfiguration and/or preplay in the hippocampus prior to a novel experience. These nine approaches cover the hierarchical functional organization of the neural grammar, from individual neurons to cell assemblies and neuronal tuplets, to extended neuronal sequences. At the sentence-like level of extended neuronal sequences, at least three distinct statistical approaches to electrophysiological neuronal ensemble activity of the hippocampus (CA1 area) at the millisecond timescale demonstrated network preconfiguration: (1) rank order correlation between neuronal firing sequences during slow-wave sleep or awake rest and future novel place-cell sequences on a previously unexplored linear context^{19,20,72,73,76–78,91}; (2) Bayesian decoding of future animal trajectory on a novel linear context from neuronal ensemble activity during the preceding sleep^19,20,73,76; and (3) multi-neuronal primitive sequence based on the probabilistic organization of individual neurons into one general sequence, akin to backbone preconfiguration, which correlated with future novel place-cell sequences and was mostly preserved as a post-experience backbone replay sequence^75,136. At the intermediate, morpheme or word-like level, neuronal co-firing^72,73,137, cell assembly temporal coordination^73,85,138 (particularly between neurons with the same birthdate^139,140) and short triplet motif (tuplet) organization^74,79,141 occurring during sleep constitute the building blocks of a predictive code^74,79 for future novel place-cell ensemble activity. At the individual neuron phoneme-like level, enhanced neuronal excitability expressed as increases in firing rate, burst propensity and activation of immediate early genes^19,73,87 in specific neurons generally increased the probability of their recruitment to encode a future novel experience. Last, computational models of preplay^142,143 simulated the role of network preconfiguration in place-cell sequence formation and animal navigation in new mazes.

Often perceived as rationalist thoughts challenging empiricist-dominated neurobiological dogmas of learning and memory, network preconfiguration and preplay in the hippocampus understandably experienced some resistance in the beginning. The geneticist Haldane famously outlined a generic four-stage acceptance process of a novel scientific view¹⁴⁴, as follows: “1) This is worthless nonsense. 2) This is an interesting, but perverse, point of view. 3) This is true, but quite unimportant. 4) I always said so.” The critiques of network preconfiguration or preplay have by now been enlisted, with variable progress, under each of the four Haldane stages of acceptance^{52,53,77,90,145}. We debunked several pitfalls that prevented researchers from reporting network preconfiguration in the past^19,73 (versus the first and second Haldane stages), which included two classes of statistical fallacies, namely circularity and inadequacy (Box 1).

Box 1. Preplay, replay and statistics.

The critique of preconfiguration rested on existing empiricist thoughts that novel experiences create de novo neuronal firing sequences in a pre-existing blank slate hippocampal network, followed by their post-experience replay during subsequent sleep or rest. Replay is generally stronger (has higher fidelity) in representing a recent experience than preplay (plasticity in replay)^73,76, given that preplay is more generic and experience-expectant whereas replay is more experience-specific (the strength of replay is also state-dependent, and greater during awake-rest states than during slow-wave sleep^73,169). These replay and preplay differences and thoughts, like “the existence of preplay undermines the existence of replay”⁵³, inadvertently tuned the statistical parameters for computing chance levels towards reaching the boundaries separating preplay and replay, which placed preplay and replay on the opposite sides of inferred statistical chance. Two main classes of statistical fallacies resulted: (1) the statistical method itself operates under the assumption that networks lack preconfiguration (for example, it ignores the preconfigured sequence motif repertoire and the functional neuronal activity diversity relative to future experiences — backbone versus plastic) and thus faults owing to circularity; and (2) the statistical method strips preconfiguration of its key constituents, which disproportionately reduces the strength of preplay compared to replay, and thus faults owing to inadequacy.

Network preconfiguration and preplay are biologically supported by the existence of (1) a pre-existing repertoire of temporal sequence motifs that probabilistically reflect the functional network connectivity²¹; (2) functional diversity of neuronal groups⁷⁶, some operating in a top-down context-independent manner (backbone or non-plastic) and others in a bottom-up context-specific manner (plastic); and (3) functional neuronal organization into millisecond-timescale temporally coordinated or co-firing neuronal ensembles (cell assemblies)^73,85. The circularity statistical fallacy is first exemplified when using neuronal cell ID shuffles (NCISs) to test preplay or replay incidence significance. In many iterations, NCISs that test for the chance level for a single ‘A’ experience will artificially resample from the larger repertoire of valid neuronal sequences, many of which are not at chance because they are pre-representations of (or parts of) additional possible future experiences20 with partially overlapping elements (for example, A₁, A₂ … A_n experiences). Given that preplay is more generic and multifarious, whereas replay is more experience-specific, the threshold for exceeding the set 95th percentile of NCIS matches is raised artificially higher for preplay than for replay. Furthermore, NCISs imbalance the proportions of neurons across both top-down, non-plastic and bottom-up, plastic neuronal sequences within specific preplay or replay events towards more plastic neurons (or events), which decreases the significance of preplay. When NCISs were performed within the corresponding plastic or non-plastic neuronal sequence groups, the significance of preplay re-emerged⁷⁶. Since replay has a higher incidence and strength than preplay, the circularity fallacy will also disproportionately affect preplay more than replay. Finally, an arbitrary tuning and allocation of significance threshold parameters, as in the case of multi-parameter thresholds⁵², also favoured replay over preplay incidence significance.

The inadequacy fallacy is first exemplified by statistical methods that artificially decorrelate the temporal coordination or co-firing of neuronal ensembles such as the ‘leave-one-cell-out’ procedures (decorrelation inadequacy). As context-dependent neuronal co-firing is increased in replay compared to preplay, these types of procedures disproportionately affect preplay. Additional statistical fallacies that prevented past studies from recognizing network preconfiguration and preplay have been detailed in our previous publications^19,73. These include inadequate use of rank order correlation methods with densities of place cells per spatial bins outside the calculated accuracy range^30,52; inadequate use of data shuffle methods, their combinations (such as Poisson-like combined with circular space-bin shuffle) in Bayesian decoding methods and their interpretation⁵²; and inadequate preplay or replay event comparisons across different states and contexts (for example, preplay during sleep outside the context compared to replay during awake states within the context⁵²).

In sum, when studying the phenomena of hippocampal network preconfiguration and preplay — the core of the network competence to generate time-compressed sequences — the listed conceptual, technical and interpretational fallacies^73,76 need to be critically considered.

Parallel to the increasing direct experimental evidence for network preconfiguration and neural grammar in the hippocampus and other brain areas, a growing number of studies have implicitly provided evidence for possible network preconfiguration mechanisms and roles (versus the third Haldane stage). A first set of studies concerned the potentially underlying cellular mechanisms of network preconfiguration and documented that long-term potentiation induction in intrahippocampal synapses can subsequently remap the representation of an externally unchanged familiar context³⁸ and affect unstimulated synapses on (the same) postsynaptic neurons via metaplasticity^146,147 or synaptic tagging and capture of key plasticity molecules^37,148, which would preconfigure and bias future neuronal responses. A second group of studies on behavioural tasks of rapid learning demonstrated improved animal performance given that a related yet different prior animal experience likely preconfigured the underlying network in a somewhat predictable manner. Some of these cases are memory linking of temporally overlapping experiences¹⁴⁹, selection and allocation of specific (preconfigured) neurons to newly formed memory engrams^150–153 and rapid learning via assimilation of new information into pre-existing, related mental schemas^{38,64,72,150–152,154}. Plasticity in replay, a gain in network performance, has been critically tied to increased temporal coordination of participating neuron firing during or after the experience^73,85,155 and has been proposed to be critical primarily for memory consolidation^{54,130,156,157}; this complements and underscores the role of a neural grammar in memory formation.

Conclusion: a neuro-critique of pure reason

In his Critique of Pure Reason⁴, Kant argued that our knowledge of the world is built, in part, on intuitions and concepts such as those of space and time that are expressed a priori and pre-structure our representation of experiences about the world. The research findings highlighted from the separate fields of linguistics, immunology, molecular biology and neurobiology are consistent with the Kantian thought that biological systems exhibit a priori states that pre-structure experiences and, in part, the a posteriori effects^{8,10,13,95,131}. In neurobiology, these a priori states take the form of functional network preconfiguration into temporal sequence motifs and predictive coding, which have been reported in the hippocampus^19,20,74,76 and the neocortex^56,58,158 across species. An important question is whether this organization is innate or dependent on similar early-life experiences. In the rodent hippocampus, network preconfiguration and preplay capable of predictive coding emerge during the third week of postnatal development in experimentally naive animals and are initially not modifiable by direct experience, which suggests that they could, at least in part, be innate¹³¹. Age-dependent maturation of the network enables experience-dependent changes in time-compressed sequential patterns of activity only later, during the fourth postnatal week¹³¹. Depriving animals of critical geometric experience by rearing them from birth inside non-Euclidian spherical environments¹⁵⁹ did not prevent the emergence of neural competence and performance (place cells, cell assemblies, preconfigured preplay, theta sequences and plasticity in replay) in the fourth postnatal week (the beginning of neurodevelopmental stage 3; see above). Furthermore, hippocampal neurons born on the same prenatal day were consistently more likely to be recruited together by a specific experience during adulthood¹³⁹. Altogether, these findings indicate that network preconfiguration starts before birth, is partly innate and may last a lifetime.

The generative grammar described in this Perspective relates to the hippocampal network preconfiguration expressed as the neural competence to operate in time-compressed sequential motifs that can combine in both an experience-independent and experience-dependent manner. This generative grammar supports a rapid experience-dependent selection of pre-existing patterns during encoding and incorporation of new information from external stimuli expressed as a gain in network performance that underlies early memory consolidation stages, which enables traditional hippocampus function as a memory recorder of the direct experience^{22,30,160,161}. However, in this Perspective, I propose that the generative grammar of the hippocampus has a novel role in supporting the higher-order category of internally generated representations, to which memory is a subordinated class (Fig. 4). The specific experience-independent, internally generated patterns of activity acquired via spontaneous generative combination of pre-existing sequential motifs could support the proposed roles of the hippocampus in a vast group of internally generated representations: past oriented (memory, mental travel)^19,21,30; future oriented (imagining, planning, predictive coding, schema-based accelerated learning)^{62,64,65,72,74,79,122}; sleep and resting related (dreaming, insight)^68,162; abnormal (hallucinations)²²; and inferential (cognitive mapping, transitive inference, spontaneous thought, insight and creativity)^25,163–166. It could also inspire the successful evolution of generative forms of artificial intelligence (such as large language models). Future research should explore, in detail, how the generative grammar of the brain supports this rich range of related higher-order cognitive functions and what the underlying^48,167 brain networks and states are. This generative grammar of the brain is proposed to be at the foundation of many of our cognitive processes, which ultimately underlie our reason, will and representation of the world^4,168.

Fig. 4 |. Internally generated representation forms.

Episodic memory²², imagining⁶², dreaming⁶⁸, transitive inference¹⁶³ and cognitive mapping²⁵ are forms of internally generated representation for which hippocampus integrity is critical. Hippocampus activity has also been associated with insight¹⁶⁵, whereas bilateral hippocampus removal in individuals with psychosis diminished their abnormal positive symptoms such as hallucinations²². Evolving artificial intelligence algorithms that use an analogous generative grammar to the hippocampus, such as large language models including ChatGPT, have the potential to test the capacity of generative grammars in assisting internally generated representations.

Glossary

Backbone

Robust pre-existing organization of neuronal (sequential) activity, on the framework of which new information is encoded.

Blank slate

An original state of neural networks (or the mind) believed to be devoid of any functional organization, which will be configured only because of direct experience. Also known as tabula rasa.

Cell assemblies

Groups of interconnected neurons repeatedly co-activated within short timeframes (of a few tens of milliseconds) thought to underlie an associative representational code.

Generative grammar

A set of organizational rules that can explain current patterns and predict the expression of future new patterns from (re)combination of existing elements.

Morpheme

The smallest unit of semantic meaning within a word.

Multi-neuronal primitive sequence

Overall probabilistic organization of neurons into one generic or backbone neuronal sequence characteristic to a network that is mostly preserved across experiences.

Network preconfiguration

A state of neural network functional organization often related to, constraining or predictive of future patterns of network activation; in contrast to a blank slate.

Neural code

Patterns of neuronal and neuronal ensemble activity believed to encode and represent specific stimuli or states, which could be decoded by a downstream entity.

Neural grammar

A set of hierarchical organization rules for the generation of current and future novel neural patterns from (re) combination of pre-existing cellular and circuit neural motifs.

Neural semantic

A representational feature gained by specific neural patterns for the depiction of specific objects or events from the external world, often interpreted as a neural code.

Neural syntax

A set of rules for the combinatorial hierarchical organization of the neuronal activity of individual neurons into cell assemblies and short circuit motifs towards extended neural sequences.

Neurobiological

A level of explanation of neuronal ensemble activity with a focus on the rules for the combinatorial organization of neural activity (neural syntax).

Neuro-codons

Recurring short sequence motifs of 3 ± 1 neurons, also called neuronal tuplets, that are further multiplexed into extended neuronal sequences; part of neural syntax.

Neuronal selection

A theory in neuroscience postulating that patterns of neural activity are selected from a larger pre-existing repertoire, rather than exclusively created de novo, during a novel experience.

Neuropsychological

A level of explanation of neuronal ensemble activity with a focus on the representational value of neuronal activity as a form of neural semantic or code.

Phoneme

The smallest perceptually distinct unit of sound that composes a word (and can separate it from another word), often represented by one letter or a small cluster of letters.

Place cells

A neuron type functionally tuned primarily to the position of an animal in external space; most abundant in the hippocampus.

Plasticity

The property of neural activity of being modified within normal ranges owing to a certain experience, lasting for a variable amount of time after the experience has ended.

Plasticity in replay

Recent experience-related changes in sequence replay that increase its fidelity, incidence or extent in depicting the experience compared with preplay (plasticity in replay over preplay). Considered a form of gain in network performance, not a new competence.

Preplay

Significant correlations between sequential patterns of neuronal ensemble activity during a novel (spatial) experience and those expressed during the preceding sleep or rest, prospective in nature. Demonstrates that preconfigured network patterns contribute to encoding of future novel experiences.

Pre-representations

Decoded neuronal ensemble activities during sleep and rest that resemble virtual trajectories through subsequently explored novel spaces.

Universal biological grammar

A set of rules for the hierarchical organization of elements into increasingly complex assemblies with added meaning, common across different aspects of biology.