Why neurons mix: high dimensionality for higher cognition

Mapping Low-Dimensional Dynamics to High-Dimensional Neural Activity: A Derivation of the Ring Model From the Neural Engineering Framework

Empirical estimates of the dimensionality of neural population activity are often much lower than the population size. Similar phenomena are also observed in trained and designed neural network models. These experimental and computational results suggest that mapping low-dimensional dynamics to high-dimensional neural space is a common feature of cortical computation. Despite the ubiquity of this observation, the constraints arising from such mapping are poorly understood. Here we consider a specific example of mapping low-dimensional dynamics to high-dimensional neural activity-the neural engineering framework. We analytically solve the framework for the classic ring model-a neural network encoding a static or dynamic angular variable. Our results provide a complete characterization of the success and failure modes for this model. Based on similarities between this and other frameworks, we speculate that these results could apply to more general scenarios.

Differential encoding in prefrontal cortex projection neuron classes across cognitive tasks

Single-cell transcriptomics has been widely applied to classify neurons in the mammalian brain, while systems neuroscience has historically analyzed the encoding properties of cortical neurons without considering cell types. Here we examine how specific transcriptomic types of mouse prefrontal cortex (PFC) projection neurons relate to axonal projections and encoding properties across multiple cognitive tasks. We found that most types projected to multiple targets, and most targets received projections from multiple types, except PFC→PAG (periaqueductal gray). By comparing Ca2+ activity of the molecularly homogeneous PFC→PAG type against two heterogeneous classes in several two-alternative choice tasks in freely moving mice, we found that all task-related signals assayed were qualitatively present in all examined classes. However, PAG-projecting neurons most potently encoded choice in cued tasks, whereas contralateral PFC-projecting neurons most potently encoded reward context in an uncued task. Thus, task signals are organized redundantly, but with clear quantitative biases across cells of specific molecular-anatomical characteristics.

A Model of Online Temporal-Spatial Integration for Immediacy and Overrule in Discourse Comprehension

During discourse comprehension, information from prior processing is integrated and appears to be immediately accessible. This was remarkably demonstrated by an N400 for "salted" and not "in love" in response to "The peanut was salted/in love." Discourse overrule was induced by prior discourse featuring the peanut as an animate agent. Immediate discourse overrule requires a model that integrates information at two timescales. One is over the lifetime and includes event knowledge and word semantics. The second is over the discourse in an event context. We propose a model where both are accounted for by temporal-to-spatial integration of experience into distributed spatial representations, providing immediate access to experience accumulated over different timescales. For lexical semantics, this is modeled by a word embedding system trained by sequential exposure to the entire Wikipedia corpus. For discourse, this is modeled by a recurrent reservoir network trained to generate a discourse vector for input sequences of words. The N400 is modeled as the difference between the instantaneous discourse vector and the target word. We predict this model can account for semantic immediacy and discourse overrule. The model simulates lexical priming and discourse overrule in the "Peanut in love" discourse, and it demonstrates that an unexpected word elicits reduced N400 if it is generally related to the event described in prior discourse, and that this effect disappears when the discourse context is removed. This neurocomputational model is the first to simulate immediacy and overrule in discourse-modulated N400, and contributes to characterization of online integration processes in discourse.

A Diversity of Intrinsic Timescales Underlie Neural Computations

Neural processing occurs across a range of temporal scales. To facilitate this, the brain uses fast-changing representations reflecting momentary sensory input alongside more temporally extended representations, which integrate across both short and long temporal windows. The temporal flexibility of these representations allows animals to behave adaptively. Short temporal windows facilitate adaptive responding in dynamic environments, while longer temporal windows promote the gradual integration of information across time. In the cognitive and motor domains, the brain sets overarching goals to be achieved within a long temporal window, which must be broken down into sequences of actions and precise movement control processed across much shorter temporal windows. Previous human neuroimaging studies and large-scale artificial network models have ascribed different processing timescales to different cortical regions, linking this to each region's position in an anatomical hierarchy determined by patterns of inter-regional connectivity. However, even within cortical regions, there is variability in responses when studied with single-neuron electrophysiology. Here, we review a series of recent electrophysiology experiments that demonstrate the heterogeneity of temporal receptive fields at the level of single neurons within a cortical region. This heterogeneity appears functionally relevant for the computations that neurons perform during decision-making and working memory. We consider anatomical and biophysical mechanisms that may give rise to a heterogeneity of timescales, including recurrent connectivity, cortical layer distribution, and neurotransmitter receptor expression. Finally, we reflect on the computational relevance of each brain region possessing a heterogeneity of neuronal timescales. We argue that this architecture is of particular importance for sensory, motor, and cognitive computations.

Integration of value and action in medial prefrontal neural systems

The rodent medial prefrontal cortex (mPFC) plays a key role in regulating cognition, emotion, and behavior. mPFC neurons are activated in diverse experimental paradigms, raising the questions of whether there are specific task elements or dimensions encoded by mPFC neurons, and whether these encoded parameters are selective to neurons in particular mPFC subregions or networks. Here, we consider the role of mPFC neurons in processing appetitive and aversive cues, outcomes, and related behaviors. mPFC neurons are strongly activated in tasks probing value and outcome-associated actions, but these responses vary across experimental paradigms. Can we identify specific categories of responses (e.g., positive or negative value), or do mPFC neurons exhibit response properties that are too heterogeneous/complex to cluster into distinct conceptual groups? Based on a review of relevant studies, we consider what has been done and what needs to be further explored in order to address these questions.

Themes of advanced information processing in the primate brain

Here is a review of several empirical examples of information processing that occur in the primate cerebral cortex. These include visual processing, object identification and perception, information encoding, and memory. Also, there is a discussion of the higher scale neural organization, mainly theoretical, which suggests hypotheses on how the brain internally represents objects. Altogether they support the general attributes of the mechanisms of brain computation, such as efficiency, resiliency, data compression, and a modularization of neural function and their pathways. Moreover, the specific neural encoding schemes are expectedly stochastic, abstract and not easily decoded by theoretical or empirical approaches.

Approaches to inferring multi-regional interactions from simultaneous population recordings: Inferring multi-regional interactions from simultaneous population recordings

Most past studies of neural representations and dynamics have focused on recordings from single brain areas. However, growing evidence of brain-wide, parallel representations of cognitive variables suggests that analyzing neural representations and dynamics in individual brain areas can benefit from understanding the context of multi-regional interactions that support them. Moreover, perturbation experiments revealed that the manner in which these parallel representations interact with each other can differ dramatically across different pairs of brain areas. Recent advances in recording technology offer a potentially powerful substrate to study how multi-regional interactions coordinate neural representations in individual brain areas and dictate behavior on a single-trial basis through simultaneous recordings of multiple brain areas. We review pragmatic approaches to studying multi-regional interactions and illustrate them in the concrete context of a rodent delayed response task paradigm.

Frontal eye field neurons selectively signal the reward value of prior actions

The consequences of individual actions are typically unknown until well after they are executed. This fact necessitates a mechanism that bridges delays between specific actions and reward outcomes. We looked for the presence of such a mechanism in the post-movement activity of neurons in the frontal eye field (FEF), a visuomotor area in prefrontal cortex. Monkeys performed an oculomotor gamble task in which they made eye movements to different locations associated with dynamically varying reward outcomes. Behavioral data showed that monkeys tracked reward history and made choices according to their own risk preferences. Consistent with previous studies, we observed that the activity of FEF neurons is correlated with the expected reward value of different eye movements before a target appears. Moreover, we observed that the activity of FEF neurons continued to signal the direction of eye movements, the expected reward value, and their interaction well after the movements were completed and when targets were no longer within the neuronal response field. In addition, this post-movement information was also observed in local field potentials, particularly in low-frequency bands. These results show that neural signals of prior actions and expected reward value persist across delays between those actions and their experienced outcomes. These memory traces may serve a role in reward-based learning in which subjects need to learn actions predicting delayed reward.

If deep learning is the answer, what is the question?

Neuroscience research is undergoing a minor revolution. Recent advances in machine learning and artificial intelligence research have opened up new ways of thinking about neural computation. Many researchers are excited by the possibility that deep neural networks may offer theories of perception, cognition and action for biological brains. This approach has the potential to radically reshape our approach to understanding neural systems, because the computations performed by deep networks are learned from experience, and not endowed by the researcher. If so, how can neuroscientists use deep networks to model and understand biological brains? What is the outlook for neuroscientists who seek to characterize computations or neural codes, or who wish to understand perception, attention, memory and executive functions? In this Perspective, our goal is to offer a road map for systems neuroscience research in the age of deep learning. We discuss the conceptual and methodological challenges of comparing behaviour, learning dynamics and neural representations in artificial and biological systems, and we highlight new research questions that have emerged for neuroscience as a direct consequence of recent advances in machine learning.

The Geometry of Abstraction in the Hippocampus and Prefrontal Cortex

The curse of dimensionality plagues models of reinforcement learning and decision making. The process of abstraction solves this by constructing variables describing features shared by different instances, reducing dimensionality and enabling generalization in novel situations. Here, we characterized neural representations in monkeys performing a task described by different hidden and explicit variables. Abstraction was defined operationally using the generalization performance of neural decoders across task conditions not used for training, which requires a particular geometry of neural representations. Neural ensembles in prefrontal cortex, hippocampus, and simulated neural networks simultaneously represented multiple variables in a geometry reflecting abstraction but that still allowed a linear classifier to decode a large number of other variables (high shattering dimensionality). Furthermore, this geometry changed in relation to task events and performance. These findings elucidate how the brain and artificial systems represent variables in an abstract format while preserving the advantages conferred by high shattering dimensionality.

Designing spontaneous behavioral switching via chaotic itinerancy

Chaotic itinerancy is a frequently observed phenomenon in high-dimensional nonlinear dynamical systems and is characterized by itinerant transitions among multiple quasi-attractors. Several studies have pointed out that high-dimensional activity in animal brains can be observed to exhibit chaotic itinerancy, which is considered to play a critical role in the spontaneous behavior generation of animals. Thus, how to design desired chaotic itinerancy is a topic of great interest, particularly for neurorobotics researchers who wish to understand and implement autonomous behavioral controls. However, it is generally difficult to gain control over high-dimensional nonlinear dynamical systems. In this study, we propose a method for implementing chaotic itinerancy reproducibly in a high-dimensional chaotic neural network. We demonstrate that our method enables us to easily design both the trajectories of quasi-attractors and the transition rules among them simply by adjusting the limited number of system parameters and by using the intrinsic high-dimensional chaos.

Biased Neural Representation of Feature-Based Attention in the Human Frontoparietal Network

Selective attention is a core cognitive function for efficient processing of information. Although it is well known that attention can modulate neural responses in many brain areas, the computational principles underlying attentional modulation remain unclear. Contrary to the prevailing view of a high-dimensional, distributed neural representation, here we show a surprisingly simple, biased neural representation for feature-based attention in a large dataset including five human fMRI studies. We found that when human participants (both sexes) selected one feature from a compound stimulus, voxels in many cortical areas responded consistently higher to one attended feature over the other. This univariate bias was consistent across brain areas within individual subjects. Importantly, this univariate bias showed a progressively stronger magnitude along the cortical hierarchy. In frontoparietal areas, the bias was strongest and contributed largely to pattern-based decoding, whereas early visual areas lacked such a bias. These findings suggest a gradual transition from a more analog to a more abstract representation of attentional priority along the cortical hierarchy. Biased neural responses in high-level areas likely reflect a low-dimensional neural code that can facilitate a robust representation and simple readout of cognitive variables.SIGNIFICANCE STATEMENT It is typically assumed that cognitive variables are represented by distributed population activities. Although this view is rooted in decades of work in the sensory system, it has not been rigorously tested at different levels of cortical hierarchy. Here we show a novel, low-dimensional coding scheme that dominated the representation of feature-based attention in frontoparietal areas. The simplicity of such a biased code may confer a robust representation of cognitive variables, such as attentional selection, working memory, and decision-making.

Behavioral state coding by molecularly defined paraventricular hypothalamic cell type ensembles

Brains encode behaviors using neurons amenable to systematic classification by gene expression. The contribution of molecular identity to neural coding is not understood because of the challenges involved with measuring neural dynamics and molecular information from the same cells. We developed CaRMA (calcium and RNA multiplexed activity) imaging based on recording in vivo single-neuron calcium dynamics followed by gene expression analysis. We simultaneously monitored activity in hundreds of neurons in mouse paraventricular hypothalamus (PVH). Combinations of cell-type marker genes had predictive power for neuronal responses across 11 behavioral states. The PVH uses combinatorial assemblies of molecularly defined neuron populations for grouped-ensemble coding of survival behaviors. The neuropeptide receptor neuropeptide Y receptor type 1 (Npy1r) amalgamated multiple cell types with similar responses. Our results show that molecularly defined neurons are important processing units for brain function.

Spatial organization of functional clusters representing reward and movement information in the striatal direct and indirect pathways

To obtain insights into striatal neural processes underlying reward-based learning and movement control, we examined spatial organizations of striatal neurons related to movement and reward-based learning. For this, we recorded the activity of direct- and indirect-pathway neurons (D1 and A2a receptor-expressing neurons, respectively) in mice engaged in probabilistic classical conditioning and open-field free exploration. We found broadly organized functional clusters of striatal neurons in the direct as well as indirect pathways for both movement- and reward-related variables. Functional clusters for different variables were partially overlapping in both pathways, but the overlap between outcome- and value-related functional clusters was greater in the indirect than direct pathway. Also, value-related spatial clusters were progressively refined during classical conditioning. Our study shows the broad and learning-dependent spatial organization of functional clusters of dorsal striatal neurons in the direct and indirect pathways. These findings further argue against the classic model of the basal ganglia and support the importance of spatiotemporal patterns of striatal neuronal ensemble activity in the control of behavior.

A shared neural substrate for action verbs and observed actions in human posterior parietal cortex

High-level sensory and motor cortical areas are activated when processing the meaning of language, but it is unknown whether, and how, words share a neural substrate with corresponding sensorimotor representations. We recorded from single neurons in human posterior parietal cortex (PPC) while participants viewed action verbs and corresponding action videos from multiple views. We find that PPC neurons exhibit a common neural substrate for action verbs and observed actions. Further, videos were encoded with mixtures of invariant and idiosyncratic responses across views. Action verbs elicited selective responses from a fraction of these invariant and idiosyncratic neurons, without preference, thus associating with a statistical sampling of the diverse sensory representations related to the corresponding action concept. Controls indicated that the results are not the product of visual imagery or arbitrary learned associations. Our results suggest that language may activate the consolidated visual experience of the reader.

Mixed Selectivity in Macaque Medial Parietal Cortex during Eye-Hand Reaching

The activity of neurons of the medial posterior parietal area V6A in macaque monkeys is modulated by many aspects of reach task. In the past, research was mostly focused on modulating the effect of single parameters upon the activity of V6A cells. Here, we used Generalized Linear Models (GLMs) to simultaneously test the contribution of several factors upon V6A cells during a fix-to-reach task. This approach resulted in the definition of a representative “functional fingerprint” for each neuron. We first studied how the features are distributed in the population. Our analysis highlighted the virtual absence of units strictly selective for only one factor and revealed that most cells are characterized by “mixed selectivity.” Then, exploiting our GLM framework, we investigated the dynamics of spatial parameters encoded within V6A. We found that the tuning is not static, but changed along the trial, indicating the sequential occurrence of visuospatial transformations helpful to guide arm movement.

•

The parietal cortex integrates a variety of sensorimotor inputs to guide reaching

•

GLM disentangled the effect of various reaching parameters upon cell activity

•

V6A neurons were not functionally clustered, but characterized by mixed selectivity

•

Spatial selectivity was dynamic and reached its peak during the movement phase

Theoretical distinction between functional states in working memory and their corresponding neural states

Working memory (WM) is important for guiding behaviour, but not always for the next possible action. Here we define a WM item that is currently relevant for guiding behaviour as the functionally "active" item; whereas items maintained in WM, but not immediately relevant to behaviour, are defined as functionally "latent". Traditional neurophysiological theories of WM proposed that content is maintained via persistent neural activity (e.g., stable attractors); however, more recent theories have highlighted the potential role for "activity-silent" mechanisms (e.g., short-term synaptic plasticity). Given these somewhat parallel dichotomies, functionally active and latent cognitive states of WM have been associated with storage based on persistent-activity and activity-silent neural mechanisms, respectively. However, in this article we caution against a one-to-one correspondence between functional and activity states. We argue that the principal theoretical requirement for active and latent WM is that the corresponding neural states play qualitatively different functional roles. We consider a number of candidate solutions, and conclude that the neurophysiological mechanisms for functionally active and latent WM items are theoretically independent of the distinction between persistent activity-based and activity-silent forms of WM storage.

Phase of firing coding of learning variables across the fronto-striatal network during feature-based learning

The prefrontal cortex and striatum form a recurrent network whose spiking activity encodes multiple types of learning-relevant information. This spike-encoded information is evident in average firing rates, but finer temporal coding might allow multiplexing and enhanced readout across the connected network. We tested this hypothesis in the fronto-striatal network of nonhuman primates during reversal learning of feature values. We found that populations of neurons encoding choice outcomes, outcome prediction errors, and outcome history in their firing rates also carry significant information in their phase-of-firing at a 10–25 Hz band-limited beta frequency at which they synchronize across lateral prefrontal cortex, anterior cingulate cortex and anterior striatum when outcomes were processed. The phase-of-firing code exceeds information that can be obtained from firing rates alone and is evident for inter-areal connections between anterior cingulate cortex, lateral prefrontal cortex and anterior striatum. For the majority of connections, the phase-of-firing information gain is maximal at phases of the beta cycle that were offset from the preferred spiking phase of neurons. Taken together, these findings document enhanced information of three important learning variables at specific phases of firing in the beta cycle at an inter-areally shared beta oscillation frequency during goal-directed behavior.

The average spiking frequency in the fronto-striatal network encodes multiple types of learning-relevant information. Here, the authors show that populations of neurons in non-human primates also carry significant information in their phase-of-firing when learning-relevant outcomes are processed.

Low-dimensional dynamics for working memory and time encoding

Our decisions often depend on multiple sensory experiences separated by time delays. The brain can remember these experiences and, simultaneously, estimate the timing between events. To understand the mechanisms underlying working memory and time encoding, we analyze neural activity recorded during delays in four experiments on nonhuman primates. To disambiguate potential mechanisms, we propose two analyses, namely, decoding the passage of time from neural data and computing the cumulative dimensionality of the neural trajectory over time. Time can be decoded with high precision in tasks where timing information is relevant and with lower precision when irrelevant for performing the task. Neural trajectories are always observed to be low-dimensional. In addition, our results further constrain the mechanisms underlying time encoding as we find that the linear "ramping" component of each neuron's firing rate strongly contributes to the slow timescale variations that make decoding time possible. These constraints rule out working memory models that rely on constant, sustained activity and neural networks with high-dimensional trajectories, like reservoir networks. Instead, recurrent networks trained with backpropagation capture the time-encoding properties and the dimensionality observed in the data.

Representation of the observer's predicted outcome value in mirror and nonmirror neurons of macaque F5 ventral premotor cortex

In the search for the function of mirror neurons, a previous study reported that F5 mirror neuron responses are modulated by the value that the observing monkey associates with the grasped object. Yet we do not know whether mirror neurons are modulated by the expected reward value for the observer or also by other variables, which are causally dependent on value (e.g., motivation, attention directed at the observed action, arousal). To clarify this, we trained two rhesus macaques to observe a grasping action on an object kept constant, followed by four fully predictable outcomes of different values (2 outcomes with positive and 2 with negative emotional valence). We found a consistent order in population activity of both mirror and nonmirror neurons that matches the order of the value of this predicted outcome but that does not match the order of the above-mentioned value-dependent variables. These variables were inferred from the probability not to abort a trial, saccade latency, modulation of eye position during action observation, heart rate, and pupil size. Moreover, we found subpopulations of neurons tuned to each of the four predicted outcome values. Multidimensional scaling revealed equal normalized distances of 0.25 between the two positive and between the two negative outcomes suggesting the representation of a relative value, scaled to the task setting. We conclude that F5 mirror neurons and nonmirror neurons represent the observer's predicted outcome value, which in the case of mirror neurons may be transferred to the observed object or action.NEW & NOTEWORTHY Both the populations of F5 mirror neurons and nonmirror neurons represent the predicted value of an outcome resulting from the observation of a grasping action. Value-dependent motivation, arousal, and attention directed at the observed action do not provide a better explanation for this representation. The population activity's metric suggests an optimal scaling of value representation to task setting.

Multidimensional processing in the amygdala

Brain-wide circuits that coordinate affective and social behaviours intersect in the amygdala. Consequently, amygdala lesions cause a heterogeneous array of social and non-social deficits. Social behaviours are not localized to subdivisions of the amygdala even though the inputs and outputs that carry social signals are anatomically restricted to distinct subnuclear regions. This observation may be explained by the multidimensional response properties of the component neurons. Indeed, the multitudes of circuits that converge in the amygdala enlist the same subset of neurons into different ensembles that combine social and non-social elements into high-dimensional representations. These representations may enable flexible, context-dependent social decisions. As such, multidimensional processing may operate in parallel with subcircuits of genetically identical neurons that serve specialized and functionally dissociable functions. When combined, the activity of specialized circuits may grant specificity to social behaviours, whereas multidimensional processing facilitates the flexibility and nuance needed for complex social behaviour.

Different Multidimensional Representations across the Amygdalo-Prefrontal Network during an Approach-Avoidance Task

The prelimbic (PL) area and basolateral amygdala (lateral [LA] and basolateral [BL] nuclei) have closely related functions and similar extrinsic connectivity. Reasoning that the computational advantage of such redundancy should be reflected in differences in how these structures represent information, we compared the coding properties of PL and amygdala neurons during a task that requires rats to produce different conditioned defensive or appetitive behaviors. Rather than unambiguous regional differences in the identities of the variables encoded, we found gradients in how the same variables are represented. Whereas PL and BL neurons represented many different parameters through minor variations in firing rates, LA cells coded fewer task features with stronger changes in activity. At the population level, whereas valence could be easily distinguished from amygdala activity, PL neurons could distinguish both valence and trial identity as well as or better than amygdala neurons. Thus, PL has greater representational capacity.

A Distributed Neural Code in the Dentate Gyrus and in CA1

Neurons are often considered specialized functional units that encode a single variable. However, many neurons are observed to respond to a mix of disparate sensory, cognitive, and behavioral variables. For such representations, information is distributed across multiple neurons. Here we find this distributed code in the dentate gyrus and CA1 subregions of the hippocampus. Using calcium imaging in freely moving mice, we decoded an animal's position, direction of motion, and speed from the activity of hundreds of cells. The response properties of individual neurons were only partially predictive of their importance for encoding position. Non-place cells encoded position and contributed to position encoding when combined with other cells. Indeed, disrupting the correlations between neural activities decreased decoding performance, mostly in CA1. Our analysis indicates that population methods rather than classical analyses based on single-cell response properties may more accurately characterize the neural code in the hippocampus.

Decoding of standard and non-standard visuomotor associations from parietal cortex

OBJECTIVE

Neural signals can be decoded and used to move neural prostheses with the purpose of restoring motor function in patients with mobility impairments. Such patients typically have intact eye movement control and visual function, suggesting that cortical visuospatial signals could be used to guide external devices. Neurons in parietal cortex mediate sensory-motor transformations, encode the spatial coordinates for reaching goals, hand position and movements, and other spatial variables. We studied how spatial information is represented at the population level, and the possibility to decode not only the position of visual targets and the plans to reach them, but also conditional, non-spatial motor responses.

APPROACH

The animals first fixated one of nine targets in 3D space and then, after the target changed color, either reached toward it, or performed a non-spatial motor response (lift hand from a button). Spiking activity of parietal neurons was recorded in monkeys during two tasks. We then decoded different task related parameters.

MAIN RESULTS

We first show that a maximum-likelihood estimation (MLE) algorithm trained separately in each task transformed neural activity into accurate metric predictions of target location. Furthermore, by combining MLE with a Naïve Bayes classifier, we decoded the monkey's motor intention (reach or hand lift) and the different phases of the tasks. These results show that, although V6A encodes the spatial location of a target during a delay period, the signals they carry are updated around the movement execution in an intention/motor specific way.

SIGNIFICANCE

These findings show the presence of multiple levels of information in parietal cortex that could be decoded and used in brain machine interfaces to control both goal-directed movements and more cognitive visuomotor associations.

Sparse coding with a somato-dendritic rule

Cortical neurons are silent most of the time: sparse activity enables low-energy computation in the brain, and promises to do the same in neuromorphic hardware. Beyond power efficiency, sparse codes have favourable properties for associative learning, as they can store more information than local codes but are easier to read out than dense codes. Auto-encoders with a sparse constraint can learn sparse codes, and so can single-layer networks that combine recurrent inhibition with unsupervised Hebbian learning. But the latter usually require fast homeostatic plasticity, which could lead to catastrophic forgetting in embodied agents that learn continuously. Here we set out to explore whether plasticity at recurrent inhibitory synapses could take up that role instead, regulating both the population sparseness and the firing rates of individual neurons. We put the idea to the test in a network that employs compartmentalised inputs to solve the task: rate-based dendritic compartments integrate the feedforward input, while spiking integrate-and-fire somas compete through recurrent inhibition. A somato-dendritic learning rule allows somatic inhibition to modulate nonlinear Hebbian learning in the dendrites. Trained on MNIST digits and natural images, the network discovers independent components that form a sparse encoding of the input and support linear decoding. These findings confirm that intrinsic homeostatic plasticity is not strictly required for regulating sparseness: inhibitory synaptic plasticity can have the same effect. Our work illustrates the usefulness of compartmentalised inputs, and makes the case for moving beyond point neuron models in artificial spiking neural networks.

Value and choice as separable and stable representations in orbitofrontal cortex

Value-based decision-making requires different variables-including offer value, choice, expected outcome, and recent history-at different times in the decision process. Orbitofrontal cortex (OFC) is implicated in value-based decision-making, but it is unclear how downstream circuits read out complex OFC responses into separate representations of the relevant variables to support distinct functions at specific times. We recorded from single OFC neurons while macaque monkeys made cost-benefit decisions. Using a novel analysis, we find separable neural dimensions that selectively represent the value, choice, and expected reward of the present and previous offers. The representations are generally stable during periods of behavioral relevance, then transition abruptly at key task events and between trials. Applying new statistical methods, we show that the sensitivity, specificity and stability of the representations are greater than expected from the population's low-level features-dimensionality and temporal smoothness-alone. The separability and stability suggest a mechanism-linear summation over static synaptic weights-by which downstream circuits can select for specific variables at specific times.

Stable and dynamic representations of value in the prefrontal cortex

Optimal decision-making requires that stimulus-value associations are kept up to date by constantly comparing the expected value of a stimulus with its experienced outcome. To do this, value information must be held in mind when a stimulus and outcome are separated in time. However, little is known about the neural mechanisms of working memory (WM) for value. Contradicting theories have suggested WM requires either persistent or transient neuronal activity, with stable or dynamic representations, respectively. To test these hypotheses, we recorded neuronal activity in the orbitofrontal and anterior cingulate cortex of two monkeys performing a valuation task. We found that features of all hypotheses were simultaneously present in prefrontal activity, and no single hypothesis was exclusively supported. Instead, mixed dynamics supported robust, time invariant value representations while also encoding the information in a temporally specific manner. We suggest that this hybrid coding is a critical mechanism supporting flexible cognitive abilities.

Chemogenetic Modulation and Single-Photon Calcium Imaging in Anterior Cingulate Cortex Reveal a Mechanism for Effort-Based Decisions

The ACC is implicated in effort exertion and choices based on effort cost, but it is still unclear how it mediates this cost-benefit evaluation. Here, male rats were trained to exert effort for a high-value reward (sucrose pellets) in a progressive ratio lever-pressing task. Trained rats were then tested in two conditions: a no-choice condition where lever-pressing for sucrose was the only available food option, and a choice condition where a low-value reward (lab chow) was freely available as an alternative to pressing for sucrose. Disruption of ACC, via either chemogenetic inhibition or excitation, reduced lever-pressing in the choice, but not in the no-choice, condition. We next looked for value coding cells in ACC during effortful behavior and reward consumption phases during choice and no-choice conditions. For this, we used in vivo miniaturized fluorescence microscopy to reliably track responses of the same cells and compare how ACC neurons respond during the same effortful behavior where there was a choice versus when there was no-choice. We found that lever-press and sucrose-evoked responses were significantly weaker during choice compared with no-choice sessions, which may have rendered them more susceptible to chemogenetic disruption. Together, findings from our interference experiments and neural recordings suggest that a mechanism by which ACC mediates effortful decisions is in the discrimination of the utility of available options. ACC regulates these choices by providing a stable population code for the relative value of different options.

SIGNIFICANCE STATEMENT The ACC is implicated in effort-based decision-making. Here, we used chemogenetics and in vivo calcium imaging to explore its mechanism. Rats were trained to lever press for a high-value reward and tested in two conditions: a no-choice condition where lever-pressing for the high-value reward was the only option, and a choice condition where a low-value reward was also available. Inhibition or excitation of ACC reduced effort toward the high-value option, but only in the choice condition. Neural responses in ACC were weaker in the choice compared with the no-choice condition. A mechanism by which ACC regulates effortful decisions is in providing a stable population code for the discrimination of the utility of available options.

Recurrent circuit dynamics underlie persistent activity in the macaque frontoparietal network

During delayed oculomotor response tasks, neurons in the lateral intraparietal area (LIP) and the frontal eye fields (FEF) exhibit persistent activity that reflects the active maintenance of behaviorally relevant information. Despite many computational models of the mechanisms of persistent activity, there is a lack of circuit-level data from the primate to inform the theories. To fill this gap, we simultaneously recorded ensembles of neurons in both LIP and FEF while macaques performed a memory-guided saccade task. A population encoding model revealed strong and symmetric long-timescale recurrent excitation between LIP and FEF. Unexpectedly, LIP exhibited stronger local functional connectivity than FEF, and many neurons in LIP had longer network and intrinsic timescales. The differences in connectivity could be explained by the strength of recurrent dynamics in attractor networks. These findings reveal reciprocal multi-area circuit dynamics in the frontoparietal network during persistent activity and lay the groundwork for quantitative comparisons to theoretical models.

The Dentate Gyrus Classifies Cortical Representations of Learned Stimuli

Animals must discern important stimuli and place them onto their cognitive map of their environment. The neocortex conveys general representations of sensory events to the hippocampus, and the hippocampus is thought to classify and sharpen the distinctions between these events. We recorded populations of dentate gyrus granule cells (DG GCs) and lateral entorhinal cortex (LEC) neurons across days to understand how sensory representations are modified by experience. We found representations of odors in DG GCs that required synaptic input from the LEC. Odor classification accuracy in DG GCs correlated with future behavioral discrimination. In associative learning, DG GCs, more so than LEC neurons, changed their responses to odor stimuli, increasing the distance in neural representations between stimuli, responding more to the conditioned and less to the unconditioned odorant. Thus, with learning, DG GCs amplify the decodability of cortical representations of important stimuli, which may facilitate information storage to guide behavior.

Contextual Codes in the Hippocampus

The hippocampus is thought to support memory and decisions by binding relevant aspects of experiences within a context. A recent paper by Gulli et al. studies how activity in the macaque hippocampus varies according to different contextual requirements in the same space. This study demonstrates how a hippocampal cognitive map can flexibly reflect both spatial and nonspatial task demands.

Conjunctive representations that integrate stimuli, responses, and rules are critical for action selection

A tennis player planning the next stroke has to consider various pieces of information: the type of stroke, the trajectory of the incoming ball, and where to hit the ball. Here, by decoding electroencephalography signals while participants executed simple, rule-based actions, we found that such different, actionrelevant aspects are integrated within a unified, conjunctive representation rather than being processed in a piecemeal manner. Furthermore, the strength of conjunctive representations is a highly robust predictor of how quickly actions are executed. Human-level theories and recent single-cell evidence from animal models suggest that conjunctive representations are a necessary condition for successful action selection. Our results provide direct evidence in humans that is consistent with this hypothesis.

People can use abstract rules to flexibly configure and select actions for specific situations, yet how exactly rules shape actions toward specific sensory and/or motor requirements remains unclear. Both research from animal models and human-level theories of action control point to the role of highly integrated, conjunctive representations, sometimes referred to as event files. These representations are thought to combine rules with other, goal-relevant sensory and motor features in a nonlinear manner and represent a necessary condition for action selection. However, so far, no methods exist to track such representations in humans during action selection with adequate temporal resolution. Here, we applied time-resolved representational similarity analysis to the spectral-temporal profiles of electroencephalography signals while participants performed a cued, rule-based action selection task. In two experiments, we found that conjunctive representations were active throughout the entire selection period and were functionally dissociable from the representation of constituent features. Specifically, the strength of conjunctions was a highly robust predictor of trial-by-trial variability in response times and was selectively related to an important behavioral indicator of conjunctive representations, the so-called partial-overlap priming pattern. These results provide direct evidence for conjunctive representations as critical precursors of action selection in humans.

A Domain-General Cognitive Core Defined in Multimodally Parcellated Human Cortex

Numerous brain imaging studies identified a domain-general or "multiple-demand" (MD) activation pattern accompanying many tasks and may play a core role in cognitive control. Though this finding is well established, the limited spatial localization provided by traditional imaging methods precluded a consensus regarding the precise anatomy, functional differentiation, and connectivity of the MD system. To address these limitations, we used data from 449 subjects from the Human Connectome Project, with the cortex of each individual parcellated using neurobiologically grounded multimodal MRI features. The conjunction of three cognitive contrasts reveals a core of 10 widely distributed MD parcels per hemisphere that are most strongly activated and functionally interconnected, surrounded by a penumbra of 17 additional areas. Outside cerebral cortex, MD activation is most prominent in the caudate and cerebellum. Comparison with canonical resting-state networks shows MD regions concentrated in the fronto-parietal network but also engaging three other networks. MD activations show modest relative task preferences accompanying strong co-recruitment. With distributed anatomical organization, mosaic functional preferences, and strong interconnectivity, we suggest MD regions are well positioned to integrate and assemble the diverse components of cognitive operations. Our precise delineation of MD regions provides a basis for refined analyses of their functions.

Temporal chunking as a mechanism for unsupervised learning of task-sets

Depending on environmental demands, humans can learn and exploit multiple concurrent sets of stimulus-response associations. Mechanisms underlying the learning of such task-sets remain unknown. Here we investigate the hypothesis that task-set learning relies on unsupervised chunking of stimulus-response associations that occur in temporal proximity. We examine behavioral and neural data from a task-set learning experiment using a network model. We first show that task-set learning can be achieved provided the timescale of chunking is slower than the timescale of stimulus-response learning. Fitting the model to behavioral data on a subject-by-subject basis confirmed this expectation and led to specific predictions linking chunking and task-set retrieval that were borne out by behavioral performance and reaction times. Comparing the model activity with BOLD signal allowed us to identify neural correlates of task-set retrieval in a functional network involving ventral and dorsal prefrontal cortex, with the dorsal system preferentially engaged when retrievals are used to improve performance.

The what, where and how of delay activity

Working memory is characterized by neural activity that persists during the retention interval of delay tasks. Despite the ubiquity of this delay activity across tasks, species and experimental techniques, our understanding of this phenomenon remains incomplete. Although initially there was a narrow focus on sustained activation in a small number of brain regions, methodological and analytical advances have allowed researchers to uncover previously unobserved forms of delay activity various parts of the brain. In light of these new findings, this Review reconsiders what delay activity is, where in the brain it is found, what roles it serves and how it may be generated.

Visual cue-related activity of cells in the medial entorhinal cortex during navigation in virtual reality

During spatial navigation, animals use self-motion to estimate positions through path integration. However, estimation errors accumulate over time and it is unclear how they are corrected. Here we report a new cell class (‘cue cell’) encoding visual cues that could be used to correct errors in path integration in mouse medial entorhinal cortex (MEC). During virtual navigation, individual cue cells exhibited firing fields only near visual cues and their population response formed sequences repeated at each cue. These cells consistently responded to cues across multiple environments. On a track with cues on left and right sides, most cue cells only responded to cues on one side. During navigation in a real arena, they showed spatially stable activity and accounted for 32% of unidentified, spatially stable MEC cells. These cue cell properties demonstrate that the MEC contains a code representing spatial landmarks, which could be important for error correction during path integration.

Preservation of Partially Mixed Selectivity in Human Posterior Parietal Cortex across Changes in Task Context

Recent studies in posterior parietal cortex (PPC) have found multiple effectors and cognitive strategies represented within a shared neural substrate in a structure termed "partially mixed selectivity" (Zhang et al., 2017). In this study, we examine whether the structure of these representations is preserved across changes in task context and is thus a robust and generalizable property of the neural population. Specifically, we test whether the structure is conserved from an open-loop motor imagery task (training) to a closed-loop cortical control task (online), a change that has led to substantial changes in neural behavior in prior studies in motor cortex. Recording from a 4 × 4 mm electrode array implanted in PPC of a human tetraplegic patient participating in a brain-machine interface (BMI) clinical trial, we studied the representations of imagined/attempted movements of the left/right hand and compare their individual BMI control performance using a one-dimensional cursor control task. We found that the structure of the representations is largely maintained between training and online control. Our results demonstrate for the first time that the structure observed in the context of an open-loop motor imagery task is maintained and accessible in the context of closed-loop BMI control. These results indicate that it is possible to decode the mixed variables found from a small patch of cortex in PPC and use them individually for BMI control. Furthermore, they show that the structure of the mixed representations is maintained and robust across changes in task context.

Nonlinear mixed selectivity supports reliable neural computation

Neuronal activity in the brain is variable, yet both perception and behavior are generally reliable. How does the brain achieve this? Here, we show that the conjunctive coding of multiple stimulus features, commonly known as nonlinear mixed selectivity, may be used by the brain to support reliable information transmission using unreliable neurons. Nonlinearly mixed feature representations have been observed throughout primary sensory, decision-making, and motor brain areas. In these areas, different features are almost always nonlinearly mixed to some degree, rather than represented separately or with only additive (linear) mixing, which we refer to as pure selectivity. Mixed selectivity has been previously shown to support flexible linear decoding for complex behavioral tasks. Here, we show that it has another important benefit: in many cases, it makes orders of magnitude fewer decoding errors than pure selectivity even when both forms of selectivity use the same number of spikes. This benefit holds for sensory, motor, and more abstract, cognitive representations. Further, we show experimental evidence that mixed selectivity exists in the brain even when it does not enable behaviorally useful linear decoding. This suggests that nonlinear mixed selectivity may be a general coding scheme exploited by the brain for reliable and efficient neural computation.

Neurons in the brain are unreliable, while both perception and behavior are generally reliable. In this work, we study how the neural population response to sensory, motor, and cognitive features can produce this reliability. Across the brain, single neurons have been shown to respond to particular conjunctions of multiple features, termed nonlinear mixed selectivity. In this work, we show that populations of these mixed selective neurons lead to many fewer decoding errors than populations without mixed selectivity, even when both neural codes are given the same number of spikes. We show that the reliability benefits from mixed selectivity are quite general, holding under different assumptions about metabolic costs and neural noise as well as for both categorical and sensory errors. Further, previous theoretical work has shown that mixed selectivity enables the learning of complex behaviors with simple decoders. Through the analysis of neural data, we show that the brain implements mixed selectivity even when it would not serve this purpose. Thus, we argue that the brain also implements mixed selectivity to exploit its general benefits for reliable and efficient neural computation.

From thought to action: The brain-machine interface in posterior parietal cortex

A dramatic example of translational monkey research is the development of neural prosthetics for assisting paralyzed patients. A neuroprosthesis consists of implanted electrodes that can record the intended movement of a paralyzed part of the body, a computer algorithm that decodes the intended movement, and an assistive device such as a robot limb or computer that is controlled by these intended movement signals. This type of neuroprosthetic system is also referred to as a brain-machine interface (BMI) since it interfaces the brain with an external machine. In this review, we will concentrate on BMIs in which microelectrode recording arrays are implanted in the posterior parietal cortex (PPC), a high-level cortical area in both humans and monkeys that represents intentions to move. This review will first discuss the basic science research performed in healthy monkeys that established PPC as a good source of intention signals. Next, it will describe the first PPC implants in human patients with tetraplegia from spinal cord injury. From these patients the goals of movements could be quickly decoded, and the rich number of action variables found in PPC indicates that it is an appropriate BMI site for a very wide range of neuroprosthetic applications. We will discuss research on learning to use BMIs in monkeys and humans and the advances that are still needed, requiring both monkey and human research to enable BMIs to be readily available in the clinic.

Context-dependent representations of objects and space in the primate hippocampus during virtual navigation

The hippocampus is implicated in associative memory and spatial navigation. To investigate how these functions are mixed in the hippocampus, we recorded from single hippocampal neurons in macaque monkeys navigating a virtual maze during a foraging task and a context-object associative memory task. During both tasks, single neurons encoded information about spatial position; a linear classifier also decoded position. However, the population code for space did not generalize across tasks, particularly where stimuli relevant to the associative memory task appeared. Single-neuron and population-level analyses revealed that cross-task changes were due to selectivity for nonspatial features of the associative memory task when they were visually available (perceptual coding) and following their disappearance (mnemonic coding). Our results show that neurons in the primate hippocampus nonlinearly mix information about space and nonspatial elements of the environment in a task-dependent manner; this efficient code flexibly represents unique perceptual experiences and correspondent memories.

Neural ensemble dynamics in dorsal motor cortex during speech in people with paralysis

Speaking is a sensorimotor behavior whose neural basis is difficult to study with single neuron resolution due to the scarcity of human intracortical measurements. We used electrode arrays to record from the motor cortex 'hand knob' in two people with tetraplegia, an area not previously implicated in speech. Neurons modulated during speaking and during non-speaking movements of the tongue, lips, and jaw. This challenges whether the conventional model of a 'motor homunculus' division by major body regions extends to the single-neuron scale. Spoken words and syllables could be decoded from single trials, demonstrating the potential of intracortical recordings for brain-computer interfaces to restore speech. Two neural population dynamics features previously reported for arm movements were also present during speaking: a component that was mostly invariant across initiating different words, followed by rotatory dynamics during speaking. This suggests that common neural dynamical motifs may underlie movement of arm and speech articulators.

Neocortex-Cerebellum Circuits for Cognitive Processing

Although classically thought of as a motor circuit, the cerebellum is now understood to contribute to a wide variety of cognitive functions through its dense interconnections with the neocortex, the center of brain cognition. Recent investigations have shed light on the nature of cerebellar cognitive processing and information exchange with the neocortex. We review findings that demonstrate widespread reward-related cognitive input to the cerebellum, as well as new studies that have characterized the codependence of processing in the neocortex and cerebellum. Together, these data support a view of the neocortex-cerebellum circuit as a joint dynamic system both in classical sensorimotor contexts and reward-related, cognitive processing. These studies have also expanded classical theory on the computations performed by the cerebellar circuit.

Oscillatory Control over Representational States in Working Memory

In the visual world, attention is guided by perceptual goals activated in visual working memory (VWM). However, planning multiple-task sequences also requires VWM to store representations for future goals. These future goals need to be prevented from interfering with the current perceptual task. Recent findings have implicated neural oscillations as a control mechanism serving the implementation and switching of different states of prioritization of VWM representations. We review recent evidence that posterior alpha-band oscillations underlie the flexible activation and deactivation of VWM representations and that frontal delta-to-theta-band oscillations play a role in the executive control of this process. That is, frontal delta-to-theta appears to orchestrate posterior alpha through long-range oscillatory networks to flexibly set up and change VWM states during multitask sequences.

Multidimensional Neural Selectivity in the Primate Amygdala

The amygdala contributes to multiple functions including attention allocation, sensory processing, decision-making, and the elaboration of emotional behaviors. The diversity of functions attributed to the amygdala is reflected in the response selectivity of its component neurons. Previous work claimed that subsets of neurons differentiate between broad categories of stimuli (e.g., objects vs faces, rewards vs punishment), while other subsets are narrowly specialized to respond to individual faces or facial features (e.g., eyes). Here we explored the extent to which the same neurons contribute to more than one neural subpopulation in a task that activated multiple functions of the amygdala. The subjects (Macaca mulatta) watched videos depicting conspecifics or inanimate objects, and learned by trial and error to choose the individuals or objects associated with the highest rewards. We found that the same neurons responded selectively to two or more of the following task events or stimulus features: (1) alerting, task-related stimuli (fixation icon, video start, and video end); (2) reward magnitude; (3) stimulus categories (social vs nonsocial); and (4) stimulus-unique features (faces, eyes). A disproportionate number of neurons showed selectivity for all of the examined stimulus features and task events. These results suggest that neurons that appear specialized and uniquely tuned to specific stimuli (e.g., face cells, eye cells) are likely to respond to multiple other types of stimuli or behavioral events, if/when these become behaviorally relevant in the context of a complex task. This multidimensional selectivity supports a flexible, context-dependent evaluation of inputs and subsequent decision making based on the activity of the same neural ensemble.

Categorical encoding of decision variables in orbitofrontal cortex

A fundamental and recurrent question in systems neuroscience is that of assessing what variables are encoded by a given population of neurons. Such assessments are often challenging because neurons in one brain area may encode multiple variables, and because neuronal representations might be categorical or non-categorical. These issues are particularly pertinent to the representation of decision variables in the orbitofrontal cortex (OFC)-an area implicated in economic choices. Here we present a new algorithm to assess whether a neuronal representation is categorical or non-categorical, and to identify the encoded variables if the representation is indeed categorical. The algorithm is based on two clustering procedures, one variable-independent and the other variable-based. The two partitions are then compared through adjusted mutual information. The present algorithm overcomes limitations of previous approaches and is widely applicable. We tested the algorithm on synthetic data and then used it to examine neuronal data recorded in the primate OFC during economic decisions. Confirming previous assessments, we found the neuronal representation in OFC to be categorical in nature. We also found that neurons in this area encode the value of individual offers, the binary choice outcome and the chosen value. In other words, during economic choice, neurons in the primate OFC encode decision variables in a categorical way.

Decoding and encoding (de)mixed population responses

A central tenet of neuroscience is that the brain works through large populations of interacting neurons. With recent advances in recording techniques, the inner working of these populations has come into full view. Analyzing the resulting large-scale data sets is challenging because of the often complex and 'mixed' dependency of neural activities on experimental parameters, such as stimuli, decisions, or motor responses. Here we review recent insights gained from analyzing these data with dimensionality reduction methods that 'demix' these dependencies. We demonstrate that the mappings from (carefully chosen) experimental parameters to population activities appear to be typical and stable across tasks, brain areas, and animals, and are often identifiable by linear methods. By considering when and why dimensionality reduction and demixing work well, we argue for a view of population coding in which populations represent (demixed) latent signals, corresponding to stimuli, decisions, motor responses, and so on. These latent signals are encoded into neural population activity via non-linear mappings and decoded via linear readouts. We explain how such a scheme can facilitate the propagation of information across cortical areas, and we review neural network architectures that can reproduce the encoding and decoding of latent signals in population activities. These architectures promise a link from the biophysics of single neurons to the activities of neural populations.