[1] 7w No Pattern Separation in the Human Hippocampus

[1] 30w Trends in Cognitive SciencesTICS 2095 No. of Pages 14 How are Concept Cells created and consolidated? The understanding of this process should help to clarify the Trends in Cognitive Sciences

[1] 162w In humans, recordings of brain signals are typically limited to noninvasive techniques such as fMRI, which gives only an indirect and vague measure of the aggregate activity of large numbers of neurons [11]. However, it is sometimes possible to record from individual neurons in patients with epilepsy refractory to medication, who are implanted with intracranial electrodes in the MTL to identify the seizure-originating area and evaluate the possibility of its surgical resection [12][13][14]. With these recordings, we have found Concept Cells responding to specific concepts [15,16]. Figure 1A shows one of these neurons, which fired selectively to three different pictures of actor Jackie Chan and to his written and spoken name. In line with the fact that these neurons represent concepts rather than specific features of the stimuli, the responses to the different pictures of Jackie Chan were undistinguishable from each other, in terms of both their strength and their latency, as was the case for most (~80%) MTL responsive neurons [17].

[2] 213w Human MTL neurons can also respond to associated concepts, as shown with the neuron in Figure 1B, which responded to actor Leslie Nielsen and to two pictures of an airplane. In terms of visual features, these images are very dissimilar, but at the conceptual level they are related due to the movie Airplane!, featuring Nielsen. To quantify this observation, we used two different association metricsasking the subjects to rank how much pairs of pictures were related to each other and using the normalized number of hits obtained when doing an internet search of two concepts togetherand found that [18]: (i) if MTL neurons fire to more than one concept, these concepts tend to be related; (ii) MTL neurons respond to a small fraction of associated concepts (4%); (iii) the neurons' responses match the association scores provided by the subjects more closely than those obtained with the internet-based metric, thus showing that MTL neurons represent subjective relationships, based on personal experiences; (iv) MTL neuron responses represent specific associations that cannot be explained by the familiarity of the stimuli, visual similarities between the pictures, or broad semantic categorizations (e.g., actors, musicians); and (v) MTL neurons do not show a topographic organization, since the concepts to which nearby neurons fired were not related to each other.

[3] 89w As for the responses to Jackie Chan in Figure 1A, the responses to Leslie Nielsen and the two airplane pictures were indistinguishable from each other in terms of both their strength and their latency. This suggests a nearly binary code, with MTL neurons mostly firing in the same way to pictures of the same person [17] since they convey the same meaning for memory functionsor to a few different but associated persons [18,19], keeping distinct representations that are, however, linked by the neurons firing to both concepts (Figure 1C).

[4] 538w The coding of associations described in the previous section was observed during passive viewing. Given these results, we then asked whether we could modulate the responses of MTL neurons by having the subjects learn new associations [20]. Figure 2A shows a neuron initially responding to actor Josh Brolin [the preferred (P) stimulus] and not to the Eiffel Tower [the nonpreferred (NP) stimulus]. After seeing the image of Brolin by the Eiffel Tower, the subject learnt the association between the two stimuli and the neuron started firing to the Eiffel Tower (without Brolin) as well. Figure 2B shows the grand average population responses, where we observe that, after learning, the neurons started firing to the associated (NP) pictures, in contrast to a decay in the responses to the preferred stimuli due to repetition suppression [13,21]. Note also that, in contrast to the long-term associations described in the previous section, responses to the associated (NP) pictures were smaller than those to the original (P) pictures. Differences in the responses to the P and NP pictures were found in about 80% of the neurons, whereas such Glossary Concept Cells: neurons in the human MTL that respond to specific concepts, such as a familiar person. They are characterized by two distinctive features -(i) high selectivity, firing to relatively very few persons (places or objects) and (ii) high invariance, firing to completely different views of the same person and even to the person's name. Conjunctive coding: a coding mechanism by which neurons respond to conjunctions of features/inputs and show context modulations in their responses. Pattern separation and conjunctive coding are two different, although related, notions. Conjunctive coding is defined at the single-neuron level and refers to a modulation of the neuron's responses to differentiate information in different contexts, whereas pattern separation can be seen as an extreme version of conjunctive coding and refers to orthogonalization (of similar inputs) at the network level to avoid memory interference. Declarative memory: also called explicit memory; the memory of things that can be named (facts and events). Declarative memory depends on the MTL and is subdivided into episodic and semantic memory. Engram: a term introduced by Richard Semon and later popularized by Karl Lashley that refers to neuronal assemblies encoding memories. Concept Cells are engrams representing specific concepts (e.g., Jennifer Aniston, Jackie Chan) in the human MTL. Episodic memory: the memory of personal experiences (e.g., remembering what we did on our last birthday). Episodic memory relies on the rapid formation of arbitrary associations, to be able to encode and later recall single experiences (e.g., remembering having seen a movie with a particular actor on our birthday). Neuronal assembly: a group of interconnected neurons that tend to be activated together, referred to as an 'engram', when representing a specific memory. Partially overlapping assemblies: a simple model to encode associations in the human MTL, with which a relatively small percentage (~4%) of the assembly of neurons responding to a particular concept also responds, in most cases with the same strength and latency, to an associated one. This overlap in the assemblies representing associated Trends in Cognitive Sciences differences were present in only 20% of the consolidated, long-term associations, suggesting that consolidation involves a unitization of the neuronal responses [19].

[5] 72w Subjects needed different numbers of trials to remember each of the P-NP association pairs presented in each experiment. When aligning the trials to the time of learning (Figure 2C), we observe a marked increase of the neuronal responses to the associated images at the precise time of learning (trial 0). Moreover, Figure 2D shows that after learning there were no differences between the P and NP responses in the different tasks performed.

[6] 126w The distinction between episodic and semantic memory proposed by Tulving [74] seems in principle very natural and can be traced down to the thoughts of William James and even Aquinas and Aristotle, among others (see Table 1 in [85]). It is supported by double dissociation studies showing that MTL lesions produce deficits in episodic but not semantic memory, whereas lesions in the neocortex, and particularly in the temporal pole, give rise to semantic dementia but with a relatively preserved episodic memory [86,87]. However, the distinction between episodic and semantic memory is not clear cut (see main text) and these are two, not independent, but rather interacting systems with which the episodic recall of a particular experience may be supported by semantic information and vice versa [28].

[7] 264w At the neuronal level, the main difference between MTL and neocortical representations is that MTL neurons encode arbitrary associations, in contrast to the more ordered associations in the neocortex [34,69]. This is in line with computational models proposing the existence of complementary systems in the hippocampus and neocortex, with sparse representations supporting episodic memory and the fast formation of arbitrary associations in the former and distributed representations storing organized hierarchical information supporting semantic memory in the latter [33,34]. Some of the associations initially stored in the MTL may consolidate in the neocortex but at a relatively slow rate, so that the new links will not destroy the previously stored relationships and hierarchies representing semantic information. Associations between more disparate conceptswhich are natural in episodic memory and can be rapidly stored in the MTL because in this area there are no hierarchical organizations to be brokenmay not fit and, therefore, consolidate, within the organized structure in the neocortex, which explains why arbitrary associations, and therefore episodic memory, always depend on the hippocampus, as proposed by multiple trace theory. The long-term coding of arbitrary associations by MTL neurons allows jumps in a memory narrative, like the transitions between different scenes in a movie, which is in line with evidence showing a role of this area in coding associations between discontiguous, incongruent events [23,88,89] and the fact that patients with MTL lesions have very limited recall and imagination, being able to provide only fractional accounts that are supported by neocortical structures [80], as when remembering a few isolated scenes from a movie but not the movie plot.

[8] 130w Neurons like Concept Cells have so far not been found in other species; to clarify, animals may have notions of concepts (e.g., cat, cheese), but they lack such an explicit and invariant representation of concepts at the single-neuron level. In particular, a study in monkeys analyzed the responses of hippocampal neurons to face and vocal identities familiar to the animals (pictures and voices of other monkeys in the colony, the experimenters interacting with them, etc.) but did not find neurons with such a degree of selectivity and abstraction, and, unlike with Concept Cells, there was no correlation between facial and vocal identities [52]. Another study analyzed hippocampal responses in rats while the animals interacted with conspecifics and reported that none of the recorded neurons responded selectively to individual rats [53].

[9] 109w Arguably, the closest to human Concept Cells are neurons in the anterior 'face patches' of the macaque visual system, which respond selectively to relatively few faces [54,55]. However, a more recent study showed that, rather than responding to specific individuals, these neurons are activated by complex visual features shared among the faces eliciting the neurons' responses [56,57]. It could still be argued that perhaps more experiments are needed to find whether, and to what extent, Concept Cells may be uniquely human. However, it is not just absence of evidence, because strong hippocampal responses have been reported in other animals but with a different type of coding compared to humans.

[10] 190w No Pattern Separation in the Human Hippocampus Consider two overlapping memories, both involving the actor Jackie Chan: the first related to the movie Rush Hour with Chris Tucker and the second to Enter the Dragon with Bruce Lee. According to pattern separation, the associations between Chan and Tucker and between Chan and Lee are represented by two distinct, non-overlapping assemblies (Figure 3, left panel). However, the human single-neuron data described in the previous sections suggest an alternative model, with individual concepts represented by specific and invariant engrams and the association between related concepts encoded via partially overlapping assemblies (Figure 3, right panel), which explains the finding of neurons firing to associated concepts [16]. This model is different to pattern separation because: first, each engram encodes a specific concept (e.g., Jackie Chan) and not conjunctions of concepts (e.g., Jackie Chan and Bruce Lee); and second, the recall of the associations involves coactivation of the involved engrams (facilitated by the overlap in their representations), and different overlapping memories, like the two movies with Jackie Chan, do not produce orthogonal representations because they both share the activation of the same 'Jackie Chan engram'.

[11] 66w Data in support of partially overlapping assemblies (instead of pattern separation) come from the fact that, first, very different pictures of the same person, such as those of Jackie Chan in Figure 1, may evoke different episodic memories (e.g., Jackie Chan in tuxedo or with a menacing face may elicit the memories of different movies in which he played), but contrary to the findings with fMRI

[12] 325w Trends in Cognitive Sciences activations using lures [37], tend to give the same neuronal responses [17]. The relatively lower chance of getting responses to the names of the persons (Figure 1C) can be attributed to the fact that processing the names involves different pathways that may fail to trigger the representation of the concept (also, subjects may be familiar with a person's face but not their name), but not to pattern separation (i.e., having orthogonal representations for the pictures and the name of the person) because, unlike in the monkey hippocampus, where no correlations between the facial and vocal identities were found [52], in no case did a neuron fire to a person's name and not to their pictures in the human MTL [58]. Second, in the face-place association paradigm (Figure 2), the associations between faces and places were not encoded by initially nonresponsive neurons and, instead, neurons that at first fired to a person started firing to the . At the single-neuron level, pattern separation is implemented via conjunctive coding, with the neurons firing to Chan AND Tucker together in the first case and to Chan AND Lee together in the second. The right panel illustrates the idea of partially overlapping assemblies, which is in line with findings from single-neuron recordings in the human medial temporal lobe (MTL) (see Figure 1). In this case, there is an invariant, context-independent representation of each of the actors and the association corresponding to one or the other memory is established by having relatively few neurons firing to both concepts: the association between Chan and Tucker is given by the neurons in blue/yellow and that between Chan and Lee by the neurons in blue/orange. In this case, there is no pattern separation since both memories trigger the same activation of the 'Jackie Chan engram'. There is also no conjunctive coding, since the neurons encoding the associations fire equally to one OR the other concept presented alone or together.

[13] 199w associated place, and also fired to the composite image of the person in the place (Figure 2A) [20]. In other words, the associations were encoded by expanding the tuning of the neurons initially responding to a concept and not by recruiting new neurons (at least from those we recorded from), as it would have been expected with conjunctive coding. Third, the different tasks and conditions in this experiment (after learning) did not change the neuronal responses (Figure 2D), as is typically the case in the rat and monkey hippocampus [36,[59][60][61]. In general, Concept Cells fire to a particular person irrespective of whether the subjects are passively looking at pictures of the person during 'screening sessions' (used to determine which pictures to use in follow-up tasks [12]), seeing morphed versions of the pictures in a perceptual discrimination task [62], during working memory [63], in a pair association task [20], and even when recalling [20,64] or thinking [65,66] about the person eliciting the neuron's responses. In line with this evidence, a recent study showed that, in contrast to task modulations observed in frontal lobe neurons, the representation by MTL neurons was invariant across a recognition memory and a categorization task [67].

[14] 113w With current technology, it is not possible to pinpoint in which subhippocampal structure the neurons are located [12] and it could therefore be argued that pattern separation may be found if a specific structure is focused on (Box 2). This is, however, unlikely since the lack of pattern separation is a ubiquitous finding in the human recordings, covering different subhippocampal areas. Moreover, human MTL neurons tend to show binary responses to ambiguous pictures according to the subjects' perceptual decisions [62,68], which could in principle be taken as evidence for pattern separation. However, this is a perceptual discrimination function performed in the neocortex, and MTL neurons just get the outputs of these computations [69].

[15] 10w How Can We Reconcile the Human fMRI and Single-Neuron Findings?

[16] 145w The engrams representing each concept have different numbers of neurons with different spatial distributions along the hippocampus. Therefore, it is not surprising that the coactivation of two or more engrams related to a particular memory (e.g., Jackie Chan and Chris Tucker) gives a specific pattern of voxel activations that is different to the one obtained with another overlapping memory (Jackie Chan and Bruce Lee), although the activation of the common element (Jackie Chan's engram) is the same in both cases. The coactivation of context-invariant engrams coding overlapping memories can explain the differential activations obtained not only with pairwise associations ([47-51] among many others), but also with movies with shared content [43,44] (given that each movie has a specific context that triggers different coactivations of associated concepts), in different spatial layouts [45,46], where each layout has distinct item-location associations, and with overlapping sequences of items [40][41][42],

[17] 27w given the natural tendency to associate contiguous items within sequences [70] and that MTL neurons were shown to encode associations between contiguous items within sequence presentations [71].

[18] 111w The fMRI differential responses to similar 'lure' views of the same items [37] are in principle more surprising, since they are in direct contraposition to the invariant responses found with single-cell recordings in the same area (Figure 1A) [17]. However, as mentioned previously, it has been argued that these differences could be attributed to processes other than pattern separation [38]. In particular, neurons in the human MTL were shown to respond nonselectively to novel pictures [14,72,73] and, therefore, the differential fMRI responses could be due to the activation of the same invariant engram encoding a particular item, together with nonselective novelty responses triggered by the presentation of the novel lure pictures.

[19] 95w The highly selective and invariant representation by Concept Cells is ideal for the rapid encoding of meaningful associations between arbitrary items, which is the basis not only of episodic memory but also of our creativity and imaginationto conceive non-existing things, like a unicorn, or relationships between, in principle, unrelated concepts. This facilitates focusing on meaningful abstractions and getting rid of irrelevant details that would just clutter these relationships. In line with this view, it is not surprising that patients with MTL damage have not only episodic memory but also imagination [80] and creative-thinking deficits [81].

[20] 138w The context-independent representation of Concept Cells is also ideal for generalizations and transferring knowledge across contextsto use information in novel situations without the need for extensive training, since the same neuronal representation of concepts applies to different circumstances. Converging evidence shows that the hippocampus encodes inferences and generalizations [82][83][84], something that is computationally challenging due to the conflicting requirements for generalization, which in principle tends to merge representations, and pattern separation, which tends to differentiate them [10,34]. In this respect, the model of partially overlapping assemblies described above (Figure 3) may offer an optimal trade-off, by keeping separate and partially overlapping representations for associated concepts and by having concepts that were not experienced together (Chris Tucker and Bruce Lee) develop an overlapping representation, if they tend to be regularly coactivated due to their shared association (with Jackie Chan).

[21] 49w Finally, the context-invariant representation by Concept Cells is ideal for building nested and elaborated thoughts that are not necessarily attached to any specific context. This allows us to go beyond the particular things that are present in our immediate surroundings and think about thoughts, about concepts and their relationships.

[1] 231w Theoretical and modelling studies, as well as evidence from single-neuron recordings in rats and monkeys, have demonstrated that pattern separation constitutes a fundamental principle of memory coding in the hippocampus. Differential fMRI activations in humans were also interpreted to reflect pattern separation processes, but contrary to this view, I have argued that evidence from human single-neuron recordings suggests that pattern separation may not be present in the human hippocampus, and that the different way in which we store memories could, at least in part, explain human cognitive abilities. A note of caution, however, is in order, as it is possible that, instead of all-or-none distinctions, humans may also show some degree of pattern separation (or conjunctive coding), although not as much as in other animals, and other animals may have some degree of invariance and context-independent representations, although not as much as in humans. More experiments are needed to establish whether and to what extent neuronal representations in humans differ from those described in other species (see Outstanding Questions). Moreover, it is possible that the human hippocampus may also generate other representations that could be useful to temporarily deal with a particular task at hand (e.g., to remember a list of words) but that, if consolidated into long-term hippocampal memories, should end up forming invariant and partially overlapping assemblies, as we have found for familiar concepts with the human-single neuron data.

[2] 123w The human single-neuron findings also prompt a reconsideration of the notion of episodic memory, which, more than a personal time travel or the replay of sequences of events (as described in the rat hippocampus), could be seen as a reconstruction process based on few associations. This rough and high-level coding of associations in the hippocampus should certainly be complemented by more detailed information stored in the neocortex and, in this respect, I have argued that the main difference between the hippocampus and the neocortex is that the hippocampus can rapidly form and store arbitrary associations, which are characteristic of episodic memory, whereas the neocortex provides a more ordered and hierarchically organized repository of information, which is the basis of semantic memory (Box 1).

[3] 40w Animal models have given invaluable insights into the functioning of the brain, with a level of mechanistic understanding that is typically not accessible in humans. Compared to the richness and complexity of human behavior, recordings in animalsperforming well-controlled tasks under

[4] 199w Outstanding Questions Are Concept Cells exclusively human? More experiments are needed to quantify whether, and to what extent, the selective and invariant representation of Concept Cells may be present in other species, comparing how neurons modulate (or not) their responses with changes in the task and environment. What are the behavioral consequences of not having pattern separation? Comparisons across species may show that a lack of pattern separation offers advantages for generalization, creativity, and abstract thinking, but also a tendency to forget details and to create false memories, mixing information from different contexts. What are the computational consequences of not having pattern separation in terms of memory flexibility and capacity? Do partially overlapping assemblies offer an optimal trade-off to avoid interference on the one hand and facilitate generalization on the other? Can the explicit and abstract neuronal representations by Concept Cells have an impact in AI developments and help tackling the major challenge of 'general intelligence'; that is, transferring knowledge to novel tasks and situations with minimal training? The representation of concepts and associations by Concept Cells in the MTL should be complemented by neocortical activations. But what are the specific roles of the MTL and neocortex in memory?

[5] 141w Should we revisit the notion of episodic memory and its difference with semantic memory? In the human hippocampus, episodic memory seems to be supported by coactivations of engrams, being more a reconstruction than a replay of sequences of events. Moreover, at the neuronal level the main difference between hippocampal (episodic) and neocortical (semantic) representations seems to be given by the storage of arbitrary and ordered associations, respectively. laboratory conditionsprovide a tractable and simplified version of how different functions are implemented in the brain, and there is a natural tendency to extrapolate these findings to humans. However, I argue that, complementary to efforts trying to understand how principles described in other species may also apply to humans, we should also seek to find differences and specific neuronal mechanismssuch as a lack of pattern separationthat may explain our unique cognitive abilities and intelligence.

[1] 122w Pattern separation is a basic principle of neuronal coding that precludes memory interference in the hippocampus. Its existence is supported by numerous theoretical, computational, and experimental findings in different species. However, I argue that recent evidence from single-neuron recordings suggests that pattern separation may not be present in the human hippocampus and that memories are instead coded by the coactivation of invariant and contextindependent engrams. This alternative model prompts a reassessment of the definition of episodic memory and its distinction from semantic memory. Furthermore, I propose that a lack of pattern separation in memory coding may have profound implications that could explain cognitive abilities that are uniquely developed in humans, such as our power of generalization and of creative and abstract thinking.

[2] 199w In the African savannah, three little figures departed from their gang dazed by the eruptions of a nearby volcano, leaving footprints that remained engraved in the fallen volcanic ashes and that, nearly 4 million years later, we recognize as belonging to our hominin ancestor, the Australopithecus. The Australopithecines that produced the famous Laetoli footprints, however, were likely no more clever than chimpanzees, with about the same brain size [1]. Over the course of evolution, the hominin brain went through successive changes, and in anatomically modern humans is about three times larger than that of a chimp [1]. However, this increase of brain size cannot solely explain the development of unique cognitive abilities such as language, imagination, nested thoughts, an unmatched creativity, etc. [2]. In fact, there is a more than 50% brain size variability in people with comparable intelligence, and other animals have brains comparable to and even larger than that of humans [1]. So, rather than number of neurons, neurons in a particular area, or specific anatomical differences [3], it seems more likely that the cognitive gap between humans and other species is due to differences in the neuronal coding principles underlying the functioning of the human brain.

[3] 130w Our thoughts rely on our memories, and the way we store them should certainly have an impact on our cognitive abilities. Many studies have demonstrated the involvement of the medial temporal lobe (MTL) (the hippocampus and surrounding cortex) in declarative memory (see Glossary) [4][5][6][7][8]. In this area, a general coding principle is pattern separation, which orthogonalizes memory representations [9,10]. Several works have provided compelling evidence of pattern separation in the rat and monkey hippocampus, and human fMRI studies reported divergent hippocampal activations also consistent with pattern separation [9,10]. However, after presenting evidence from single neuron-recordings, I will argue that there is no pattern separationthat means, an orthogonalization of memory representationsin the human MTL and that this lack of pattern separation may be a key component underlying our unique cognitive abilities.

[4] 37w In contrast to claims based on theoretical and modelling studies, findings with single-neuron recordings in rats and monkeys, and indirect evidence from fMRI activations, I argue that pattern separation may not be present in the human hippocampus.

[5] 26w Relatively recent evidence from human single-neuron recordings shows that, contrary to pattern separation, episodic memories are coded by contextindependent and invariant engrams in the human hippocampus.

[6] 36w Associations constitute the skeleton of episodic memories and are coded with partially overlapping assemblies, which prompts us to reconsider the view of episodic memory as mental time travel, and the distinction between episodic and semantic memory.

[7] 23w A lack of pattern separation may explain human cognitive abilities, such as our unique powers of generalization and of creative and abstract thinking.

[8] 121w Associations constitute the skeleton of episodic memory (e.g., to remember having seen Jackie Chan and Bruce Lee in a Kung Fu movie). Several works on humans and other animals have shown the involvement of the MTL in coding associations [16,18,20,[22][23][24][25][26][27], but the specific role of this area in episodic memory continues to be debated. In particular, while there is a general consensus that semantic memory initially involves the MTL and then consolidates in the neocortex, the standard consolidation model [6,7] argues that the MTL encodes episodic memories only during learning and not after their consolidation in the neocortex, whereas multiple trace theory [5,8] argues that the MTL provides a long-term representation that continues to be critical for episodic memory after learning.

[9] 93w Evidence supporting one or the other model has been mainly provided by behavioral studies in patients with MTL lesions, which, due of the variability of the precise location and the extent of the lesions, have given mixed results [4]. Moreover, the distinction between episodic and semantic memory is not always clear cut [28][29][30] (Box 1). For example, does an association between Jackie Chan and Bruce Lee represent a semantic memory, knowing that they are both Kung Fu fighters, or the episodic memory of having watched a movie in which they fought each other?

[10] 149w The results with newly formed [20] and long-term [18,19] associations described in the previous section points towards a role of MTL neurons in encoding episodic memories rather than semantic categorizations. This is because the associations reflected specific and subjective relationships [18] that could be established rapidly between arbitrary items [20], as is the case with unique experiences in episodic memory. Moreover, the fact that responses were obtained during passive viewing, the first time the subjects saw the images in the experiments [13,21], indicates that the neurons were already encoding these concepts and associations before the experiments took place, thus supporting the idea of long-term coding in line with multiple trace theory; without such long-term coding, it would indeed be difficult to conceive how the MTL could form de novo representations (with complex links to specific sets of neurons in neocortical areas) every time we create or recall a memory.

[11] 100w Theoretical and modelling work has proposed the existence of two complementary processespattern separation and pattern completionunderlying the encoding and recall of memories in the hippocampus [31][32][33][34][35]. Pattern separation is a network property that at the single-neuron level is implemented via conjunctive coding [36], with neurons firing to conjunctions of features to create distinct representations. A large number of studies with single-neuron recordings have consistently shown the presence of pattern separation or conjunctive coding in the rodent and monkey hippocampus [i.e., the neurons' responses tend to change when the environment or the task performed by the animals is changed (Box 2)].

[12] 129w In humans, indirect evidence for pattern separation (or conjunctive coding) has been given by the finding of differential activations obtained with fMRI data [9,10]. (However, in the next section we concepts is relatively small to avoid confusing the concepts with each other, but sufficiently large to effectively link the two concepts. Pattern completion: a process that allows the retrieval of a memory, activating the corresponding neuronal representation from partial cues. Pattern separation: a process by which the neuronal representation of similar memories is orthogonalized, creating distinct, non-overlapping assemblies to avoid interference. Semantic memory: the memory of facts and general knowledge. Semantic memory relies on the storage of ordered and hierarchically organized associations (e.g., knowing that Paris is a city that is the capital of France, which is in Europe).

[13] 227w see that a model other than pattern separation or conjunctive coding can generate such differential fMRI responses.) One of these studies used an incidental encoding task in which pictures of everyday objects were either presented once (novel pictures) or repeated at a later time (repeated pictures), also including pictures that were similar to previously presented ones (lure pictures) [37]. The analysis of high-resolution fMRI showed adaptation effects for the picture repetitions and peristimulus time histograms on the bottom. Despite their differences (shown in different postures, with different expressions, etc.), the responses to the pictures of Jackie Chan were indistinguishable from each other in terms of both their strength and their latency. (B) Responses of another neuron in the hippocampus that fired to actor Leslie Nielsen and to two airplane pictures. The neuron did not respond to any of the other stimuli presented and encodes the association between Nielsen and the airplane pictures, given the movie Airplane! featuring the actor. As before, the responses to Nielsen and the airplane pictures were indistinguishable from each other, in terms of both their strength and their latency. (C) Empirical probabilities that MTL neurons fire to different pictures of the same person (Jackie Chan), to the picture and the name of a person, and to different but associated (Jackie Chan and Bruce Lee) and not associated (Jackie Chan and Lionel Messi) persons.

[14] 25w similar activations to the novel and lure pictures in the CA3/dentate gyrus, indicating that the lure pictures were encoded differently than their similar-looking counterparts [37].

[15] 29w The different activations to the lure and the repeated presentations were taken as evidence of pattern separation [10,37]. However, it was later argued that such differences could also be

[16] 61w It has long been established that the rodent hippocampal formation is involved in spatial navigation, particularly after the discovery of place [90] and grid [91] cells representing specific locations [59]. Several works have provided clear evidence of pattern separation in the rodent hippocampus, describing changes in the representation of spatially tuned neurons that are observed when the environment is altered [92][93][94][95][96][97].

[17] 116w Within the hippocampal formation, the dentate gyrus has been proposed to be the main storehouse for pattern separation, having a very large number of neurons that facilitates the formation of non-overlapping neuronal representations [32,34]. Moreover, neurogenesis is prevalent in the dentate gyrus, thus providing a continuous repository of available neurons to perform pattern separation [98][99][100], and deficits in this process have been linked to memory problems due to aging or pathologies such as Alzheimer's disease [9,10]. In line with this view, it was shown that pattern separation is prevalent in the dentate gyrus [101,102] (but the role of distinct types of neurons in the dentate gyrus for pattern separation is a subject of active research [103]).

[18] 103w Hippocampal representations in rodents vary not only when the environment is modified but also after other manipulations: when the task performed by the animals is changed [104][105][106], when associations with odors or items at specific locations are introduced [107-109], or when the animal's overall trajectory [110,111], trajectory planning [75], reward location [112], motivational state [113], etc. is changed (for reviews see [36,59,60,114]). Grid cells, with regular geometrical patterns of activations, have been considered to provide a more stable representation [59], but they also change their firing patterns following physical changes in the environment [115][116][117] and other cognitive factors, such as reward location [118][119][120].

[19] 107w Similar results were found in the monkey hippocampus, with neurons firing to specific conjunctions of objects, viewpoints, positions, reward locations, contexts, tasks, etc. [61,[121][122][123][124][125][126]. More generally, the changes in the activation of hippocampal neurons, given by physical or cognitive factors, provide contextual information related to the experience of the animal in the environment. Therefore, it has been argued that the hippocampus has a relational memory role, linking together the different elements constituting an experience [36,60]. Within this framework, the representation of spatial locations, which is particularly prevalent in rodents (given the importance of knowing their surroundings for their behavior), is one of many components constituting a memory.

[20] 90w attributed to match-mismatch processes [38]. More generally, differentiating objects from lures is a perceptual discrimination task, which is not much related to episodic memory and the differentiation of overlapping memories attributed to pattern separation. For this reason, besides variations of the adaptation paradigm with lures [10,37,39], other studies have reported differential fMRI activations disambiguating overlapping sequences of items [40][41][42], movies with shared content [43,44], spatial layouts [45,46], associations between items and contexts [47,48], faces and scenes [49], and pictures with different reward outcomes [50], among various types of associations [51].

[21] 36w Having proposed that a lack of pattern separation is a key, unique feature of how memories are coded in the human MTL, let us now discuss how this may have an impact on human cognitive abilities.

[22] 159w Pattern separation leads to combinatorial explosion (i.e., the fact that we run out of neurons to encode separately all possible combinations constituting our memories). This may not be an issue for animals performing just a handful of tasks in the laboratory, but it is problematic when dealing with human memory, considering the richness and complexity of our experiences. In fact, we tend to forget details unless they are very salient and, rather than encoding separately different views of a person or object, as suggested by the fMRI studies with 'lures' [37], it seems more plausible to have invariant representations with MTL neurons not even coding these aspects (e.g., what a person looked like when meeting them) [17]. Likewise, it makes sense that there is no pattern separation in the coding of associations in the human MTL, because otherwise, having every experience involving a person we know represented by different, non-overlapping sets of neurons would also lead to combinatorial explosion.

[23] 208w Episodic Memory as Reconstruction rather than Replay Episodic memory has been traditionally seen as a 'mental time travel', in which sequences of events constituting episodic experiences are revisited [74]. In line with this view, sequences of place cells in the rat hippocampus fired not only when the animals physically travelled the environment, but also when they planned or recalled these trajectories [75][76][77]. However, while it is plausible to conceive how largely non-overlapping assemblies may encode and replay sequences in the rat hippocampus [76], having memories encoded by coactivations of invariant engrams makes a replay of sequences more unlikely, given the difficulty in establishing chains of complex engram coactivations that do not interfere with each other and that are robust to the coding of new contexts and associations. For this reason, instead of a mental time travel or a replay of event sequences, it seems more plausible to view episodic memory as an active reconstruction process based on a schema [78] given by engrams coactivations, which is supported by more detailed activations in the neocortex. Making an analogy with a music score, the human hippocampus would then tend to store chords (i.e., notes played together) or rough chord progressions rather than the detailed sequence of notes constituting a melody.

[24] 90w Direct evidence in favor of or against sequential coding with single-neuron recordings in the human hippocampus is scarce. However, two studies, in which subjects had to remember pictures presented in sequences, showed that the position of the pictures in the sequences had no effect on the firing of MTL neurons during reactivations in a maintenance period (after the encoding of the picture presentations) [63,79], in line with a sustained activation of the neurons and in contrast to the time-locked sequential activations observed during delay periods in the rat hippocampus [75].