Replay, the default mode network and the cascaded memory systems model.

The spontaneous replay of patterns of activity related to past experiences and memories is a striking feature of brain activity, as is the coherent activation of sets of brain areas - particularly those comprising the default mode network (DMN) - during rest. We propose that these two phenomena are strongly intertwined and that their potential functions overlap. In the 'cascaded memory systems model' that we outline here, we hypothesize that the DMN forms the backbone for the propagation of replay, mediating interactions between the hippocampus and the neocortex that enable the consolidation of new memories. The DMN may also independently ignite replay cascades, which support reactivation of older memories or high-level semantic representations. We suggest that transient cortical activations, inducing long-range correlations across the neocortex, are a key mechanism supporting a hierarchy of representations that progresses from simple percepts to semantic representations of causes and, finally, to whole episodes.

Hierarchical gradients of encoded spatial and sensory information in the neocortex are attenuated by dorsal hippocampal lesions.

During navigation, the neocortex must actively integrate learned spatial context with current sensory experience to guide behaviours. However, the relative encoding of spatial and sensorimotor information among cortical cells, and whether hippocampal feedback continues to modify these properties in familiar environments, remains poorly understood. Thus, two-photon microscopy of male and female Thy1-GCaMP6s mice was used to longitudinally image neurons spanning superficial retrosplenial cortex and layers II-Va of primary and secondary motor cortices before and after bilateral dorsal hippocampal lesions. During behaviour on a familiar cued treadmill, the locations of two added obstacles were interchanged to decouple place-tuning from cue-tuning among the position correlated cells with fields at those locations. The subpopulations of place- and cue-tuned cells each formed interareal gradients such that higher-level cortical regions exhibited higher fractions of place cells, whereas lower-level regions exhibited higher fractions of cue cells. Position correlated cells in motor cortex also formed translaminar gradients; cells closer to the cortical surface were more likely to exhibit fields and were more sparsely and precisely tuned than deeper cells. After dorsal hippocampal lesions, a neural representation of the learned environment persisted but retrosplenial cortex exhibited significantly increased cue-tuning and, in motor cortices, both position correlated cell recruitment and population activity at the unstable obstacle locations became more homogeneously elevated across laminae. Altogether, these results support that the hippocampus continues to modulate cortical responses in familiar environments, and the relative impact of top-down feedback obeys hierarchical interareal and interlaminar gradients opposite to the flow of bottom-up sensory inputs.Significance statement During learning, the hippocampus imparts spatial context to memory representations throughout the superficial neocortex. However, the post-learning role of the hippocampus has not been well defined. The results of this study suggest that, during navigation of a familiar environment, the hippocampus continues to link unreliable sensory attributes to a stable contextual framework, effectively updating the learned model of the environment. The results are also consistent with top-down suppression of sensory-evoked activity during behaviour, which varied in strength according to hierarchical proximity to the hippocampus. This effect was abolished by bilateral lesions of the dorsal hippocampus, supporting that the hippocampus plays an ongoing role in propagating context-dependent predictions throughout the cortical hierarchy, a core hypothesis of the predictive coding theoretical framework.

Learning in deep neural networks and brains with similarity-weighted interleaved learning.

Understanding how the brain learns throughout a lifetime remains a long-standing challenge. In artificial neural networks (ANNs), incorporating novel information too rapidly results in catastrophic interference, i.e., abrupt loss of previously acquired knowledge. Complementary Learning Systems Theory (CLST) suggests that new memories can be gradually integrated into the neocortex by interleaving new memories with existing knowledge. This approach, however, has been assumed to require interleaving all existing knowledge every time something new is learned, which is implausible because it is time-consuming and requires a large amount of data. We show that deep, nonlinear ANNs can learn new information by interleaving only a subset of old items that share substantial representational similarity with the new information. By using such similarity-weighted interleaved learning (SWIL), ANNs can learn new information rapidly with a similar accuracy level and minimal interference, while using a much smaller number of old items presented per epoch (fast and data-efficient). SWIL is shown to work with various standard classification datasets (Fashion-MNIST, CIFAR10, and CIFAR100), deep neural network architectures, and in sequential learning frameworks. We show that data efficiency and speedup in learning new items are increased roughly proportionally to the number of nonoverlapping classes stored in the network, which implies an enormous possible speedup in human brains, which encode a high number of separate categories. Finally, we propose a theoretical model of how SWIL might be implemented in the brain.

Coupling between slow waves and sharp-wave ripples engages distributed neural activity during sleep in humans.

Hippocampal-dependent memory consolidation during sleep is hypothesized to depend on the synchronization of distributed neuronal ensembles, organized by the hippocampal sharp-wave ripples (SWRs, 80 to 150 Hz), subcortical/cortical slow-wave activity (SWA, 0.5 to 4 Hz), and sleep spindles (SP, 7 to 15 Hz). However, the precise role of these interactions in synchronizing subcortical/cortical neuronal activity is unclear. Here, we leverage intracranial electrophysiological recordings from the human hippocampus, amygdala, and temporal and frontal cortices to examine activity modulation and cross-regional coordination during SWRs. Hippocampal SWRs are associated with widespread modulation of high-frequency activity (HFA, 70 to 200 Hz), a measure of local neuronal activation. This peri-SWR HFA modulation is predicted by the coupling between hippocampal SWRs and local subcortical/cortical SWA or SP. Finally, local cortical SWA phase offsets and SWR amplitudes predicted functional connectivity between the frontal and temporal cortex during individual SWRs. These findings suggest a selection mechanism wherein hippocampal SWR and cortical slow-wave synchronization governs the transient engagement of distributed neuronal populations supporting hippocampal-dependent memory consolidation.

Spatial Information Encoding across Multiple Neocortical Regions Depends on an Intact Hippocampus.

There has been considerable research showing populations of neurons encoding for different aspects of space in the brain. Recently, several studies using two-photon calcium imaging and virtual navigation have identified "spatially" modulated neurons in the posterior cortex. We enquire here whether the presence of such spatial representations may be a cortex-wide phenomenon and, if so, whether these representations can be organized in the absence of the hippocampus. To this end, we imaged the dorsal cortex of mice running on a treadmill populated with tactile cues. A high percentage (40-80%) of the detected neurons exhibited sparse, spatially localized activity, with activity fields uniformly localized over the track. The development of this location specificity was impaired by hippocampal damage. Thus, there is a substantial population of neurons distributed widely over the cortex that collectively form a continuous representation of the explored environment, and hippocampal outflow is necessary to organize this phenomenon.SIGNIFICANCE STATEMENT Increasing evidence points to the role of the neocortex in encoding spatial information. Whether this feature is linked to hippocampal functions is largely unknown. Here, we systematically surveyed multiple regions in the dorsal cortex of the same animal for the presence of signals encoding for spatial position. We described populations of cortical neurons expressing sequential patterns of activity localized in space in primary, secondary, and associational areas. Furthermore, we showed that the formation of these spatial representations was impacted by hippocampal lesion. Our results indicate that hippocampal inputs are necessary to maintain a precise cortical representation of space.

Hippocampus-dependent emergence of spatial sequence coding in retrosplenial cortex.

Retrosplenial cortex (RSC) is involved in visuospatial integration and spatial learning, and RSC neurons exhibit discrete, place cell-like sequential activity that resembles the population code of space in hippocampus. To investigate the origins and population dynamics of this activity, we combined longitudinal cellular calcium imaging of dysgranular RSC neurons in mice with excitotoxic hippocampal lesions. We tracked the emergence and stability of RSC spatial activity over consecutive imaging sessions. Overall, spatial activity in RSC was experience-dependent, emerging gradually over time, but, as seen in the hippocampus, the spatial code changed dynamically across days. Bilateral but not unilateral hippocampal lesions impeded the development of spatial activity in RSC. Thus, the emergence of spatial activity in RSC, a major recipient of hippocampal information, depends critically on an intact hippocampus; the indirect connections between the dysgranular RSC and the hippocampus further indicate that hippocampus may exert such influences polysynaptically within neocortex.

Packet-based communication in the cortex.

Cortical circuits work through the generation of coordinated, large-scale activity patterns. In sensory systems, the onset of a discrete stimulus usually evokes a temporally organized packet of population activity lasting ∼50-200 ms. The structure of these packets is partially stereotypical, and variation in the exact timing and number of spikes within a packet conveys information about the identity of the stimulus. Similar packets also occur during ongoing stimuli and spontaneously. We suggest that such packets constitute the basic building blocks of cortical coding.

Immediate-early gene Homer1a intranuclear transcription focus intensity as a measure of relative neural activation.

Immediate-early genes (IEGs) exhibit a rapid, transient transcription response to neuronal activation. Fluorescently labeled mRNA transcripts appear as bright intranuclear transcription foci (INF), which have been used as an all-or-nothing indicator of recent neuronal activity; however, it would be useful to know whether INF fluorescence can be used effectively to assess relative activations within a neural population. We quantified the Homer1a (H1a) response of hippocampal neurons to systematically varied numbers of exposures to the same places by inducing male Long-Evans rats to run laps around a track. Previous studies reveal relatively stable firing rates across laps on a familiar track. A strong linear trend (r2 > 0.9) in INF intensity was observed between 1 and 25 laps, after which INF intensity declined as a consequence of dispersion related to the greater elapsed time. When the integrated fluorescence of the entire nucleus was considered instead, the linear relationship extended to 50 laps. However, there was only an approximate doubling of H1a detected for this 50-fold variation in total spiking. Thus, the intranuclear H1a RNA fluorescent signal does provide a relative measure of how many times a set of neurons was activated over a ~10 min period, but the dynamic range and hence signal-to-noise ratios are poor. This low dynamic range may reflect previously reported reductions in the IEG response during repeated episodes of behavior over longer time scales. It remains to be determined how well the H1a signal reflects relative firing rates within a population of neurons in response to a single, discrete behavioral event.

Place navigation in the Morris water task results in greater nuclear Arc mRNA expression in dorsal compared to ventral CA1.

Previous work has shown that the dorsal hippocampus has greater activity than ventral regions during place navigation. Exposure to a novel context has also been found to increase hippocampal activation, possibly due to increased spatial demands. However, activation patterns in dorsal and ventral regions have not been investigated in the Morris water task (MWT), which remains the most popular assay of place memory in rodents. We measured activity in a large population of neurons across the CA1 dorsal-ventral axis by estimating nuclear Arc mRNA with stereologic systematic-random sampling procedures following changes to goal location or spatial context in the MWT in rats. Following changes to goal location or spatial context in the MWT, we did not find an effect on Arc mRNA expression in CA1. However, Arc expression was greater in the dorsal compared to the ventral aspect of CA1 during task performance. Several views might account for these observed differences in dorsal-ventral Arc mRNA expression, including task parameters or the granularity of representation that differs along the dorsal-ventral hippocampal axis. Future work should determine the effects of task differences and required memory precision in relation to dorsal-ventral hippocampal neuronal activity.

Hippocampal coupling with cortical and subcortical structures in the context of memory consolidation.

Memory consolidation is a gradual process through which episodic memories become incorporated into long-term 'semantic' representations. It likely involves reactivation of neural activity encoding the recent experience during non-REM sleep. A critical prerequisite for memory consolidation is precise coordination of reactivation events between the hippocampus and cortical/subcortical structures, facilitated by the coupling of local field potential (LFP) oscillations (slow oscillations, sleep spindles and sharp wave/ripples) between these structures. We review the rapidly expanding literature on the qualitative and quantitative aspects of hippocampal oscillatory and neuronal coupling with cortical/subcortical structures in the context of memory reactivation. Reactivation in the hippocampus and cortical/subcortical structures is tightly coupled with sharp wave/ripples. Hippocampal-cortical/subcortical coupling is rich in dimensionality and this dimensionality is likely underestimated due to the limitations of the current methodology.

Evidence for an Evolutionarily Conserved Memory Coding Scheme in the Mammalian Hippocampus.

Decades of research identify the hippocampal formation as central to memory storage and recall. Events are stored via distributed population codes, the parameters of which (e.g., sparsity and overlap) determine both storage capacity and fidelity. However, it remains unclear whether the parameters governing information storage are similar between species. Because episodic memories are rooted in the space in which they are experienced, the hippocampal response to navigation is often used as a proxy to study memory. Critically, recent studies in rodents that mimic the conditions typical of navigation studies in humans and nonhuman primates (i.e., virtual reality) show that reduced sensory input alters hippocampal representations of space. The goal of this study was to quantify this effect and determine whether there are commonalities in information storage across species. Using functional molecular imaging, we observe that navigation in virtual environments elicits activity in fewer CA1 neurons relative to real-world conditions. Conversely, comparable neuronal activity is observed in hippocampus region CA3 and the dentate gyrus under both conditions. Surprisingly, we also find evidence that the absolute number of neurons used to represent an experience is relatively stable between nonhuman primates and rodents. We propose that this convergence reflects an optimal ensemble size for episodic memories.SIGNIFICANCE STATEMENT One primary factor constraining memory capacity is the sparsity of the engram, the proportion of neurons that encode a single experience. Investigating sparsity in humans is hampered by the lack of single-cell resolution and differences in behavioral protocols. Sparsity can be quantified in freely moving rodents, but extrapolating these data to humans assumes that information storage is comparable across species and is robust to restraint-induced reduction in sensory input. Here, we test these assumptions and show that species differences in brain size build memory capacity without altering the structure of the data being stored. Furthermore, sparsity in most of the hippocampus is resilient to reduced sensory information. This information is vital to integrating animal data with human imaging navigation studies.

Involvement of fast-spiking cells in ictal sequences during spontaneous seizures in rats with chronic temporal lobe epilepsy.

See Lenck-Santini (doi:10.1093/awx205) for a scientific commentary on this article. Epileptic seizures represent altered neuronal network dynamics, but the temporal evolution and cellular substrates of the neuronal activity patterns associated with spontaneous seizures are not fully understood. We used simultaneous recordings from multiple neurons in the hippocampus and neocortex of rats with chronic temporal lobe epilepsy to demonstrate that subsets of cells discharge in a highly stereotypical sequential pattern during ictal events, and that these stereotypical patterns were reproducible across consecutive seizures. In contrast to the canonical view that principal cell discharges dominate ictal events, the ictal sequences were predominantly composed of fast-spiking, putative inhibitory neurons, which displayed unusually strong coupling to local field potential even before seizures. The temporal evolution of activity was characterized by unique dynamics where the most correlated neuronal pairs before seizure onset displayed the largest increases in correlation strength during the seizures. These results demonstrate the selective involvement of fast spiking interneurons in structured temporal sequences during spontaneous ictal events in hippocampal and neocortical circuits in experimental models of chronic temporal lobe epilepsy.

Place field expansion after focal MEC inactivations is consistent with loss of Fourier components and path integrator gain reduction.

Both hippocampal place fields and medial entorhinal cortex (MEC) grid fields increase in scale along the dorsoventral axis. Because the connections from MEC to hippocampus are topographically organized and divergent, it has been hypothesized that place fields are generated by a Fourier-like summation of inputs over a range of spatial scales. This hypothesis predicts that inactivation of dorsal MEC should cause place field expansion, whereas inactivation of ventral MEC should cause field contraction. Inactivation of dorsal MEC caused substantial expansion of place fields; however, as inactivations were made more ventrally, the effect diminished but never switched to contraction. Expansion was accompanied by proportional decreases in theta power, intrinsic oscillation frequencies, phase precession slopes, and firing rates. Our results are most consistent with the predicted loss of specific Fourier components coupled with a path integration gain reduction, which raises the overall place field scale and masks the contraction expected from ventral inactivations.

Reactivation of Rate Remapping in CA3.

UNLABELLED: The hippocampus is thought to contribute to episodic memory by creating, storing, and reactivating patterns that are unique to each experience, including different experiences that happen at the same location. Hippocampus can combine spatial and contextual/episodic information using a dual coding scheme known as "global" and "rate" remapping. Global remapping selects which set of neurons can activate at a given location. Rate remapping readjusts the firing rates of this set depending on current experience, thus expressing experience-unique patterns at each location. But can the experience-unique component be retrieved spontaneously? Whereas reactivation of recent, spatially selective patterns in hippocampus is well established, it is never perfect, raising the issue of whether the experiential component might be absent. This question is key to the hypothesis that hippocampus can assist memory consolidation by reactivating and broadcasting experience-specific "index codes" to neocortex. In CA3, global remapping exhibits attractor-like dynamics, whereas rate remapping apparently does not, leading to the hypothesis that only the former can be retrieved associatively and casting doubt on the general consolidation hypothesis. Therefore, we studied whether the rate component is reactivated spontaneously during sleep. We conducted neural ensemble recordings from CA3 while rats ran on a circular track in different directions (in different sessions) and while they slept. It was shown previously that the two directions of running result in strong rate remapping. During sleep, the most recent rate distribution was reactivated preferentially. Therefore, CA3 can retrieve patterns spontaneously that are unique to both the location and the content of recent experience. SIGNIFICANCE STATEMENT: The hippocampus is required for memory of events and their spatial contexts. The primary correlate of hippocampal activity is location in space, but multiple memories can occur in the same location. To be useful for distinguishing these memories, the hippocampus must be able, not only to express, but also to retrieve both spatial and nonspatial information about events. Whether it can retrieve nonspatial information has been challenged recently. We exposed rats to two different experiences (running in different directions) in the same locations and showed that even the nonspatial components of hippocampal cell firing are reactivated spontaneously during sleep, supporting the conclusion that both types of information about a recent experience can be retrieved.

Long-term recordings improve the detection of weak excitatory-excitatory connections in rat prefrontal cortex.

Characterization of synaptic connectivity is essential to understanding neural circuit dynamics. For extracellularly recorded spike trains, indirect evidence for connectivity can be inferred from short-latency peaks in the correlogram between two neurons. Despite their predominance in cortex, however, significant interactions between excitatory neurons (E) have been hard to detect because of their intrinsic weakness. By taking advantage of long duration recordings, up to 25 h, from rat prefrontal cortex, we found that 7.6% of the recorded pyramidal neurons are connected. This corresponds to ∼70% of the local E-E connection probability that has been reported by paired intracellular recordings (11.6%). This value is significantly higher than previous reports from extracellular recordings, but still a substantial underestimate. Our analysis showed that long recording times and strict significance thresholds are necessary to detect weak connections while avoiding false-positive results, but will likely still leave many excitatory connections undetected. In addition, we found that hyper-reciprocity of connections in prefrontal cortex that was shown previously by paired intracellular recordings was only present in short-distance, but not in long distance (∼300 micrometers or more) interactions. As hyper-reciprocity is restricted to local clusters, it might be a minicolumnar effect. Given the current surge of interest in very high-density neural spike recording (e.g., NIH BRAIN Project) it is of paramount importance that we have statistically reliable methods for estimating connectivity from cross-correlation analysis available. We provide an important step in this direction.

Interaction of egocentric and world-centered reference frames in the rat posterior parietal cortex.

Navigation requires coordination of egocentric and allocentric spatial reference frames and may involve vectorial computations relative to landmarks. Creation of a representation of target heading relative to landmarks could be accomplished from neurons that encode the conjunction of egocentric landmark bearings with allocentric head direction. Landmark vector representations could then be created by combining these cells with distance encoding cells. Landmark vector cells have been identified in rodent hippocampus. Given remembered vectors at goal locations, it would be possible to use such cells to compute trajectories to hidden goals. To look for the first stage in this process, we assessed parietal cortical neural activity as a function of egocentric cue light location and allocentric head direction in rats running a random sequence to light locations around a circular platform. We identified cells that exhibit the predicted egocentric-by-allocentric conjunctive characteristics and anticipate orienting toward the goal.

Offline reactivation of experience-dependent neuronal firing patterns in the rat ventral tegmental area.

In a rest period immediately after a task, neurons in the hippocampus, neocortex, and striatum exhibit spatiotemporal correlation patterns resembling those observed during the task. This reactivation has been proposed as a neurophysiological substrate for memory consolidation. We provide new evidence that rodent ventral tegmental area (VTA) neurons are selective for different types of food stimuli and that stimulus-sensitive neurons strongly reactivate during the rest period following a task that involved those stimuli. Reactivation occurred primarily during slow wave sleep and during quiet awakeness. In these experiments, VTA reactivation patterns were uncompressed and occurred at the firing rate level, rather than on a spike-to-spike basis. Mildly aversive stimuli were reactivated more often than positive ones. The VTA is a pivotal structure involved in the perception and prediction of reward and stimulus salience and is a key neuromodulatory system involved in synaptic plasticity. These results suggest new ways in which dopaminergic signals could contribute to the biophysical mechanisms of selective, system-wide, memory consolidation, and reconsolidation during sleep.

Bridging Neuroscience and AI: Environmental Enrichment as a Model for Forward Knowledge Transfer.

Continual learning (CL) refers to an agent's capability to learn from a continuous stream of data and transfer knowledge without forgetting old information. One crucial aspect of CL is forward transfer, i.e., improved and faster learning on a new task by leveraging information from prior knowledge. While this ability comes naturally to biological brains, it poses a significant challenge for artificial intelligence (AI). Here, we suggest that environmental enrichment (EE) can be used as a biological model for studying forward transfer, inspiring human-like AI development. EE refers to animal studies that enhance cognitive, social, motor, and sensory stimulation and is a model for what, in humans, is referred to as 'cognitive reserve'. Enriched animals show significant improvement in learning speed and performance on new tasks, typically exhibiting forward transfer. We explore anatomical, molecular, and neuronal changes post-EE and discuss how artificial neural networks (ANNs) can be used to predict neural computation changes after enriched experiences. Finally, we provide a synergistic way of combining neuroscience and AI research that paves the path toward developing AI capable of rapid and efficient new task learning.

The Nonclassic Psychedelic Ibogaine Disrupts Cognitive Maps.

Background: The ability of psychedelic compounds to profoundly alter mental function has been long known, but the underlying changes in cellular-level information encoding remain poorly understood. Methods: We used two-photon microscopy to record from the retrosplenial cortex in head-fixed mice running on a treadmill before and after injection of the nonclassic psychedelic ibogaine (40 mg/kg intraperitoneally). Results: We found that the cognitive map, formed by the representation of position encoded by ensembles of individual neurons in the retrosplenial cortex, was destabilized by ibogaine when mice had to infer position between tactile landmarks. This corresponded with increased neural activity rates, loss of correlation structure, and increased responses to cues. Ibogaine had surprisingly little effect on the size-frequency distribution of network activity events, suggesting that signal propagation within the retrosplenial cortex was largely unaffected. Conclusions: Taken together, these data support proposals that compounds with psychedelic properties disrupt representations that are important for constraining neocortical activity, thereby increasing the entropy of neural signaling. Furthermore, the loss of expected position encoding between landmarks recapitulated effects of hippocampal impairment, suggesting that disruption of cognitive maps or other hippocampal processing may be a contributing mechanism of discoordinated neocortical activity in psychedelic states.

Postnatal development of Homer1a in the rat hippocampus.

Homer1a (H1a) is an immediate early gene involved in multiple forms of synaptic plasticity. It exhibits a postnatal increase in the rat forebrain (Brakeman et al. (1997) Nature 386:284-288) and reduces the density and size of dendritic spines in hippocampal neurons (Sala et al. (2003) J Neurosci 23:6327-6337). We evaluated hippocampal H1a expression at different postnatal ages (P3, P5, P7, P9, P15, P19, P23, P35, and adult) using Fluorescence In Situ Hybridization (FISH) and qRT-PCR. Maximal electroconvulsive shock (MECS) was used to induce maximal expression relative to home cage (HC) controls. Large scale images and confocal z-stacks from dorsal subiculum (DS), CA1, CA3, and dentate gyrus (DG) were analyzed by both manual and automated methods. In DS, CA1, and CA3 a significant proportion of cells (40%) expressed small but detectable levels of H1a from P3; however, MECS did not up-regulate H1a during the first postnatal week. MECS induced H1a positive cells during the second postnatal week and induction reached adult levels at P9. H1a-Intra Nuclear Foci (INF) size and intensity varied with age, increasing at P19-23 in CA1 and CA3 and from P9 to P23 in DS. In DG, H1a expression exhibited a lamination pattern and an H1a-INF size and intensity gradient across the granule cell layer, consistent with the outside-in maturation of DG granule cells. The developmental progression of H1a corresponds to the synaptic refinement period supporting the conclusion that H1a could play an important role in this process.