Hippocampal replays under the scrutiny of reinforcement learning models

CoBeL-RL: A neuroscience-oriented simulation framework for complex behavior and learning

Reinforcement learning (RL) has become a popular paradigm for modeling animal behavior, analyzing neuronal representations, and studying their emergence during learning. This development has been fueled by advances in understanding the role of RL in both the brain and artificial intelligence. However, while in machine learning a set of tools and standardized benchmarks facilitate the development of new methods and their comparison to existing ones, in neuroscience, the software infrastructure is much more fragmented. Even if sharing theoretical principles, computational studies rarely share software frameworks, thereby impeding the integration or comparison of different results. Machine learning tools are also difficult to port to computational neuroscience since the experimental requirements are usually not well aligned. To address these challenges we introduce CoBeL-RL, a closed-loop simulator of complex behavior and learning based on RL and deep neural networks. It provides a neuroscience-oriented framework for efficiently setting up and running simulations. CoBeL-RL offers a set of virtual environments, e.g., T-maze and Morris water maze, which can be simulated at different levels of abstraction, e.g., a simple gridworld or a 3D environment with complex visual stimuli, and set up using intuitive GUI tools. A range of RL algorithms, e.g., Dyna-Q and deep Q-network algorithms, is provided and can be easily extended. CoBeL-RL provides tools for monitoring and analyzing behavior and unit activity, and allows for fine-grained control of the simulation via interfaces to relevant points in its closed-loop. In summary, CoBeL-RL fills an important gap in the software toolbox of computational neuroscience.

How our understanding of memory replay evolves

Memory reactivations and replay, widely reported in the hippocampus and cortex across species, have been implicated in memory consolidation, planning, and spatial and skill learning. Technological advances in electrophysiology, calcium imaging, and human neuroimaging techniques have enabled neuroscientists to measure large-scale neural activity with increasing spatiotemporal resolution and have provided opportunities for developing robust analytic methods to identify memory replay. In this article, we first review a large body of historically important and representative memory replay studies from the animal and human literature. We then discuss our current understanding of memory replay functions in learning, planning, and memory consolidation and further discuss the progress in computational modeling that has contributed to these improvements. Next, we review past and present analytic methods for replay analyses and discuss their limitations and challenges. Finally, looking ahead, we discuss some promising analytic methods for detecting nonstereotypical, behaviorally nondecodable structures from large-scale neural recordings. We argue that seamless integration of multisite recordings, real-time replay decoding, and closed-loop manipulation experiments will be essential for delineating the role of memory replay in a wide range of cognitive and motor functions.

Internally Triggered Experiences of Hedonic Valence in Nonhuman Animals: Cognitive and Welfare Considerations

Do any nonhuman animals have hedonically valenced experiences not directly caused by stimuli in their current environment? Do they, like us humans, experience anticipated or previously experienced pains and pleasures as respectively painful and pleasurable? We review evidence from comparative neuroscience about hippocampus-dependent simulation in relation to this question. Hippocampal sharp-wave ripples and theta oscillations have been found to instantiate previous and anticipated experiences. These hippocampal activations coordinate with neural reward and fear centers as well as sensory and cortical areas in ways that are associated with conscious episodic mental imagery in humans. Moreover, such hippocampal “re- and preplay” has been found to contribute to instrumental decision making, the learning of value representations, and the delay of rewards in rats. The functional and structural features of hippocampal simulation are highly conserved across mammals. This evidence makes it reasonable to assume that internally triggered experiences of hedonic valence (IHVs) are pervasive across (at least) all mammals. This conclusion has important welfare implications. Most prominently, IHVs act as a kind of “welfare multiplier” through which the welfare impacts of any given experience of pain or pleasure are increased through each future retrieval. However, IHVs also have practical implications for welfare assessment and cause prioritization.

Reinforcement learning with associative or discriminative generalization across states and actions: fMRI at 3 T and 7 T

The model-free algorithms of "reinforcement learning" (RL) have gained clout across disciplines, but so too have model-based alternatives. The present study emphasizes other dimensions of this model space in consideration of associative or discriminative generalization across states and actions. This "generalized reinforcement learning" (GRL) model, a frugal extension of RL, parsimoniously retains the single reward-prediction error (RPE), but the scope of learning goes beyond the experienced state and action. Instead, the generalized RPE is efficiently relayed for bidirectional counterfactual updating of value estimates for other representations. Aided by structural information but as an implicit rather than explicit cognitive map, GRL provided the most precise account of human behavior and individual differences in a reversal-learning task with hierarchical structure that encouraged inverse generalization across both states and actions. Reflecting inference that could be true, false (i.e., overgeneralization), or absent (i.e., undergeneralization), state generalization distinguished those who learned well more so than action generalization. With high-resolution high-field fMRI targeting the dopaminergic midbrain, the GRL model's RPE signals (alongside value and decision signals) were localized within not only the striatum but also the substantia nigra and the ventral tegmental area, including specific effects of generalization that also extend to the hippocampus. Factoring in generalization as a multidimensional process in value-based learning, these findings shed light on complexities that, while challenging classic RL, can still be resolved within the bounds of its core computations.

Freezing revisited: coordinated autonomic and central optimization of threat coping

Animals have sophisticated mechanisms for coping with danger. Freezing is a unique state that, upon threat detection, allows evidence to be gathered, response possibilities to be previsioned and preparations to be made for worst-case fight or flight. We propose that - rather than reflecting a passive fear state - the particular somatic and cognitive characteristics of freezing help to conceal overt responses, while optimizing sensory processing and action preparation. Critical for these functions are the neurotransmitters noradrenaline and acetylcholine, which modulate neural information processing and also control the sympathetic and parasympathetic branches of the autonomic nervous system. However, the interactions between autonomic systems and the brain during freezing, and the way in which they jointly coordinate responses, remain incompletely explored. We review the joint actions of these systems and offer a novel computational framework to describe their temporally harmonized integration. This reconceptualization of freezing has implications for its role in decision-making under threat and for psychopathology.

Model-Based and Model-Free Replay Mechanisms for Reinforcement Learning in Neurorobotics

Experience replay is widely used in AI to bootstrap reinforcement learning (RL) by enabling an agent to remember and reuse past experiences. Classical techniques include shuffled-, reversed-ordered- and prioritized-memory buffers, which have different properties and advantages depending on the nature of the data and problem. Interestingly, recent computational neuroscience work has shown that these techniques are relevant to model hippocampal reactivations recorded during rodent navigation. Nevertheless, the brain mechanisms for orchestrating hippocampal replay are still unclear. In this paper, we present recent neurorobotics research aiming to endow a navigating robot with a neuro-inspired RL architecture (including different learning strategies, such as model-based (MB) and model-free (MF), and different replay techniques). We illustrate through a series of numerical simulations how the specificities of robotic experimentation (e.g., autonomous state decomposition by the robot, noisy perception, state transition uncertainty, non-stationarity) can shed new lights on which replay techniques turn out to be more efficient in different situations. Finally, we close the loop by raising new hypotheses for neuroscience from such robotic models of hippocampal replay.

Planning in the brain

Recent breakthroughs in artificial intelligence (AI) have enabled machines to plan in tasks previously thought to be uniquely human. Meanwhile, the planning algorithms implemented by the brain itself remain largely unknown. Here, we review neural and behavioral data in sequential decision-making tasks that elucidate the ways in which the brain does-and does not-plan. To systematically review available biological data, we create a taxonomy of planning algorithms by summarizing the relevant design choices for such algorithms in AI. Across species, recording techniques, and task paradigms, we find converging evidence that the brain represents future states consistent with a class of planning algorithms within our taxonomy-focused, depth-limited, and serial. However, we argue that current data are insufficient for addressing more detailed algorithmic questions. We propose a new approach leveraging AI advances to drive experiments that can adjudicate between competing candidate algorithms.

Replay in minds and machines

Experience-related brain activity patterns reactivate during sleep, wakeful rest, and brief pauses from active behavior. In parallel, machine learning research has found that experience replay can lead to substantial performance improvements in artificial agents. Together, these lines of research suggest replay has a variety of computational benefits for decision-making and learning. Here, we provide an overview of putative computational functions of replay as suggested by machine learning and neuroscientific research. We show that replay can lead to faster learning, less forgetting, reorganization or augmentation of experiences, and support planning and generalization. In addition, we highlight the benefits of reactivating abstracted internal representations rather than veridical memories, and discuss how replay could provide a mechanism to build internal representations that improve learning and decision-making.

Whole-body procedural learning benefits from targeted memory reactivation in REM sleep and task-related dreaming

Sleep facilitates memory consolidation through offline reactivations of memory traces. Dreaming may play a role in memory improvement and may reflect these memory reactivations. To experimentally address this question, we used targeted memory reactivation (TMR), i.e., application, during sleep, of a stimulus that was previously associated with learning, to assess whether it influences task-related dream imagery (or task-dream reactivations). Specifically, we asked if TMR or task-dream reactivations in either slow-wave (SWS) or rapid eye movement (REM) sleep benefit whole-body procedural learning. Healthy participants completed a virtual reality (VR) flying task prior to and following a morning nap or rest period during which task-associated tones were readministered in either SWS, REM sleep, wake or not at all. Findings indicate that learning benefits most from TMR when applied in REM sleep compared to a Control-sleep group. REM dreams that reactivated kinesthetic elements of the VR task (e.g., flying, accelerating) were also associated with higher improvement on the task than were dreams that reactivated visual elements (e.g., landscapes) or that had no reactivations. TMR did not itself influence dream content but its effects on performance were greater when coexisting with task-dream reactivations in REM sleep. Findings may help explain the mechanistic relationships between dream and memory reactivations and may contribute to the development of sleep-based methods to optimize complex skill learning.

Deep reinforcement learning to study spatial navigation, learning and memory in artificial and biological agents

Despite the recent advancements and popularity of deep learning that has resulted from the advent of numerous industrial applications, artificial neural networks (ANNs) still lack crucial features from their biological counterparts that could improve their performance and their potential to advance our understanding of how the brain works. One avenue that has been proposed to change this is to strengthen the interaction between artificial intelligence (AI) research and neuroscience. Since their historical beginnings, ANNs and AI, in general, have developed in close alignment with both neuroscience and psychology. In addition to deep learning, reinforcement learning (RL) is another approach that is strongly linked to AI and neuroscience to understand how learning is implemented in the brain. In a recently published article, Botvinick et al. (Neuron, 107:603-616, 2020) explain why deep reinforcement learning (DRL) is important for neuroscience as a framework to study learning, representations and decision making. Here, I summarise Botvinick et al.'s main arguments and frame them in the context of the study of learning, memory and spatial navigation. I believe that applying this approach to study spatial navigation can provide useful insights for the understanding of how the brain builds, processes and stores representations of the outside world to extract knowledge.

Neural signatures underlying deliberation in human foraging decisions

Humans have a remarkable capacity to mentally project themselves far ahead in time. This ability, which entails the mental simulation of events, is thought to be fundamental to deliberative decision making, as it allows us to search through and evaluate possible choices. Many decisions that humans make are foraging decisions, in which one must decide whether an available offer is worth taking, when compared to unknown future possibilities (i.e., the background). Using a translational decision-making paradigm designed to reveal decision preferences in rats, we found that humans engaged in deliberation when making foraging decisions. A key feature of this task is that preferences (and thus, value) are revealed as a function of serial choices. Like rats, humans also took longer to respond when faced with difficult decisions near their preference boundary, which was associated with prefrontal and hippocampal activation, exemplifying cross-species parallels in deliberation. Furthermore, we found that voxels within the visual cortices encoded neural representations of the available possibilities specifically following regret-inducing experiences, in which the subject had previously rejected a good offer only to encounter a low-valued offer on the subsequent trial.

The Neuroscience of Spatial Navigation and the Relationship to Artificial Intelligence

Recent advances in artificial intelligence (AI) and neuroscience are impressive. In AI, this includes the development of computer programs that can beat a grandmaster at GO or outperform human radiologists at cancer detection. A great deal of these technological developments are directly related to progress in artificial neural networks-initially inspired by our knowledge about how the brain carries out computation. In parallel, neuroscience has also experienced significant advances in understanding the brain. For example, in the field of spatial navigation, knowledge about the mechanisms and brain regions involved in neural computations of cognitive maps-an internal representation of space-recently received the Nobel Prize in medicine. Much of the recent progress in neuroscience has partly been due to the development of technology used to record from very large populations of neurons in multiple regions of the brain with exquisite temporal and spatial resolution in behaving animals. With the advent of the vast quantities of data that these techniques allow us to collect there has been an increased interest in the intersection between AI and neuroscience, many of these intersections involve using AI as a novel tool to explore and analyze these large data sets. However, given the common initial motivation point-to understand the brain-these disciplines could be more strongly linked. Currently much of this potential synergy is not being realized. We propose that spatial navigation is an excellent area in which these two disciplines can converge to help advance what we know about the brain. In this review, we first summarize progress in the neuroscience of spatial navigation and reinforcement learning. We then turn our attention to discuss how spatial navigation has been modeled using descriptive, mechanistic, and normative approaches and the use of AI in such models. Next, we discuss how AI can advance neuroscience, how neuroscience can advance AI, and the limitations of these approaches. We finally conclude by highlighting promising lines of research in which spatial navigation can be the point of intersection between neuroscience and AI and how this can contribute to the advancement of the understanding of intelligent behavior.

Modeling awake hippocampal reactivations with model-based bidirectional search

Hippocampal offline reactivations during reward-based learning, usually categorized as replay events, have been found to be important for performance improvement over time and for memory consolidation. Recent computational work has linked these phenomena to the need to transform reward information into state-action values for decision making and to propagate it to all relevant states of the environment. Nevertheless, it is still unclear whether an integrated reinforcement learning mechanism could account for the variety of awake hippocampal reactivations, including variety in order (forward and reverse reactivated trajectories) and variety in the location where they occur (reward site or decision-point). Here, we present a model-based bidirectional search model which accounts for a variety of hippocampal reactivations. The model combines forward trajectory sampling from current position and backward sampling through prioritized sweeping from states associated with large reward prediction errors until the two trajectories connect. This is repeated until stabilization of state-action values (convergence), which could explain why hippocampal reactivations drastically diminish when the animal's performance stabilizes. Simulations in a multiple T-maze task show that forward reactivations are prominently found at decision-points while backward reactivations are exclusively generated at reward sites. Finally, the model can generate imaginary trajectories that are not allowed to the agent during task performance. We raise some experimental predictions and implications for future studies of the role of the hippocampo-prefronto-striatal network in learning.

Real-time sensory-motor integration of hippocampal place cell replay and prefrontal sequence learning in simulated and physical rat robots for novel path optimization

An open problem in the cognitive dimensions of navigation concerns how previous exploratory experience is reorganized in order to allow the creation of novel efficient navigation trajectories. This behavior is revealed in the "traveling salesrat problem" (TSP) when rats discover the shortest path linking baited food wells after a few exploratory traversals. We have recently published a model of navigation sequence learning, where sharp wave ripple replay of hippocampal place cells transmit "snippets" of the recent trajectories that the animal has explored to the prefrontal cortex (PFC) (Cazin et al. in PLoS Comput Biol 15:e1006624, 2019). PFC is modeled as a recurrent reservoir network that is able to assemble these snippets into the efficient sequence (trajectory of spatial locations coded by place cell activation). The model of hippocampal replay generates a distribution of snippets as a function of their proximity to a reward, thus implementing a form of spatial credit assignment that solves the TSP task. The integrative PFC reservoir reconstructs the efficient TSP sequence based on exposure to this distribution of snippets that favors paths that are most proximal to rewards. While this demonstrates the theoretical feasibility of the PFC-HIPP interaction, the integration of such a dynamic system into a real-time sensory-motor system remains a challenge. In the current research, we test the hypothesis that the PFC reservoir model can operate in a real-time sensory-motor loop. Thus, the main goal of the paper is to validate the model in simulated and real robot scenarios. Place cell activation encoding the current position of the simulated and physical rat robot feeds the PFC reservoir which generates the successor place cell activation that represents the next step in the reproduced sequence in the readout. This is input to the robot, which advances to the coded location and then generates de novo the current place cell activation. This allows demonstration of the crucial role of embodiment. If the spatial code readout from PFC is played back directly into PFC, error can accumulate, and the system can diverge from desired trajectories. This required a spatial filter to decode the PFC code to a location and then recode a new place cell code for that location. In the robot, the place cell vector output of PFC is used to physically displace the robot and then generate a new place cell coded input to the PFC, replacing part of the software recoding procedure that was required otherwise. We demonstrate how this integrated sensory-motor system can learn simple navigation sequences and then, importantly, how it can synthesize novel efficient sequences based on prior experience, as previously demonstrated (Cazin et al. 2019). This contributes to the understanding of hippocampal replay in novel navigation sequence formation and the important role of embodiment.

Dynamics of Awake Hippocampal-Prefrontal Replay for Spatial Learning and Memory-Guided Decision Making

Spatial learning requires remembering and choosing paths to goals. Hippocampal place cells replay spatial paths during immobility in reverse and forward order, offering a potential mechanism. However, how replay supports both goal-directed learning and memory-guided decision making is unclear. We therefore continuously tracked awake replay in the same hippocampal-prefrontal ensembles throughout learning of a spatial alternation task. We found that, during pauses between behavioral trajectories, reverse and forward hippocampal replay supports an internal cognitive search of available past and future possibilities and exhibits opposing learning gradients for prediction of past and future behavioral paths, respectively. Coordinated hippocampal-prefrontal replay distinguished correct past and future paths from alternative choices, suggesting a role in recall of past paths to guide planning of future decisions for spatial working memory. Our findings reveal a learning shift from hippocampal reverse-replay-based retrospective evaluation to forward-replay-based prospective planning, with prefrontal readout of memory-guided paths for learning and decision making.

New perspectives on dimensionality and variability from large-scale cortical dynamics

The neocortex is a multi-scale network, with intricate local circuitry interwoven into a global mesh of long-range connections. Neural activity propagates within this network on a wide range of temporal and spatial scales. At the micro scale, neurophysiological recordings reveal coordinated dynamics in local neural populations, which support behaviorally relevant computations. At the macro scale, neuroimaging modalities measure global activity fluctuations organized into spatiotemporal patterns across the entire brain. Here we review recent advances linking the local and global scales of cortical dynamics and their relationship to behavior. We argue that diverse experimental observations on the dimensionality and variability of neural activity can be reconciled by considering how activity propagates in space and time on multiple spatial scales.

Hippocampal Sequences During Exploration: Mechanisms and Functions

Although the hippocampus plays a critical role in spatial and episodic memories, the mechanisms underlying memory formation, stabilization, and recall for adaptive behavior remain relatively unknown. During exploration, within single cycles of the ongoing theta rhythm that dominates hippocampal local field potentials, place cells form precisely ordered sequences of activity. These neural sequences result from the integration of both external inputs conveying sensory-motor information, and intrinsic network dynamics possibly related to memory processes. Their endogenous replay during subsequent sleep is critical for memory consolidation. The present review discusses possible mechanisms and functions of hippocampal theta sequences during exploration. We present several lines of evidence suggesting that these neural sequences play a key role in information processing and support the formation of initial memory traces, and discuss potential functional distinctions between neural sequences emerging during theta vs. awake sharp-wave ripples.

Interactions of spatial strategies producing generalization gradient and blocking: A computational approach

We present a computational model of spatial navigation comprising different learning mechanisms in mammals, i.e., associative, cognitive mapping and parallel systems. This model is able to reproduce a large number of experimental results in different variants of the Morris water maze task, including standard associative phenomena (spatial generalization gradient and blocking), as well as navigation based on cognitive mapping. Furthermore, we show that competitive and cooperative patterns between different navigation strategies in the model allow to explain previous apparently contradictory results supporting either associative or cognitive mechanisms for spatial learning. The key computational mechanism to reconcile experimental results showing different influences of distal and proximal cues on the behavior, different learning times, and different abilities of individuals to alternatively perform spatial and response strategies, relies in the dynamic coordination of navigation strategies, whose performance is evaluated online with a common currency through a modular approach. We provide a set of concrete experimental predictions to further test the computational model. Overall, this computational work sheds new light on inter-individual differences in navigation learning, and provides a formal and mechanistic approach to test various theories of spatial cognition in mammals.