Task representations in neural networks trained to perform many cognitive tasks

Distributed context-dependent choice information in mouse posterior cortex

Choice information appears in multi-area brain networks mixed with sensory, motor, and cognitive variables. In the posterior cortex-traditionally implicated in decision computations-the presence, strength, and area specificity of choice signals are highly variable, limiting a cohesive understanding of their computational significance. Examining the mesoscale activity in the mouse posterior cortex during a visual task, we found that choice signals defined a decision variable in a low-dimensional embedding space with a prominent contribution along the ventral visual stream. Their subspace was near-orthogonal to concurrently represented sensory and motor-related activations, with modulations by task difficulty and by the animals' attention state. A recurrent neural network trained with animals' choices revealed an equivalent decision variable whose context-dependent dynamics agreed with that of the neural data. Our results demonstrated an independent, multi-area decision variable in the posterior cortex, controlled by task features and cognitive demands, possibly linked to contextual inference computations in dynamic animal-environment interactions.

Cerebro-cerebellar networks facilitate learning through feedback decoupling

Behavioural feedback is critical for learning in the cerebral cortex. However, such feedback is often not readily available. How the cerebral cortex learns efficiently despite the sparse nature of feedback remains unclear. Inspired by recent deep learning algorithms, we introduce a systems-level computational model of cerebro-cerebellar interactions. In this model a cerebral recurrent network receives feedback predictions from a cerebellar network, thereby decoupling learning in cerebral networks from future feedback. When trained in a simple sensorimotor task the model shows faster learning and reduced dysmetria-like behaviours, in line with the widely observed functional impact of the cerebellum. Next, we demonstrate that these results generalise to more complex motor and cognitive tasks. Finally, the model makes several experimentally testable predictions regarding cerebro-cerebellar task-specific representations over learning, task-specific benefits of cerebellar predictions and the differential impact of cerebellar and inferior olive lesions. Overall, our work offers a theoretical framework of cerebro-cerebellar networks as feedback decoupling machines.

Emergent behaviour and neural dynamics in artificial agents tracking odour plumes

Tracking an odour plume to locate its source under variable wind and plume statistics is a complex task. Flying insects routinely accomplish such tracking, often over long distances, in pursuit of food or mates. Several aspects of this remarkable behaviour and its underlying neural circuitry have been studied experimentally. Here we take a complementary in silico approach to develop an integrated understanding of their behaviour and neural computations. Specifically, we train artificial recurrent neural network agents using deep reinforcement learning to locate the source of simulated odour plumes that mimic features of plumes in a turbulent flow. Interestingly, the agents' emergent behaviours resemble those of flying insects, and the recurrent neural networks learn to compute task-relevant variables with distinct dynamic structures in population activity. Our analyses put forward a testable behavioural hypothesis for tracking plumes in changing wind direction, and we provide key intuitions for memory requirements and neural dynamics in odour plume tracking.

Conserved structures of neural activity in sensorimotor cortex of freely moving rats allow cross-subject decoding

Our knowledge about neuronal activity in the sensorimotor cortex relies primarily on stereotyped movements that are strictly controlled in experimental settings. It remains unclear how results can be carried over to less constrained behavior like that of freely moving subjects. Toward this goal, we developed a self-paced behavioral paradigm that encouraged rats to engage in different movement types. We employed bilateral electrophysiological recordings across the entire sensorimotor cortex and simultaneous paw tracking. These techniques revealed behavioral coupling of neurons with lateralization and an anterior-posterior gradient from the premotor to the primary sensory cortex. The structure of population activity patterns was conserved across animals despite the severe under-sampling of the total number of neurons and variations in electrode positions across individuals. We demonstrated cross-subject and cross-session generalization in a decoding task through alignments of low-dimensional neural manifolds, providing evidence of a conserved neuronal code.

Episodic memory in aspects of brain information transfer by resting-state network topology

Cognitive functionality emerges due to neural interactions. The interregional signal interactions underlying episodic memory are a complex process. Thus, we need to quantify this process more accurately to understand how brain regions receive information from other regions. Studies suggest that resting-state functional connectivity (FC) conveys cognitive information; additionally, activity flow estimates the contribution of the source region to the activation pattern of the target region, thus decoding the cognitive information transfer. Therefore, we performed a combined analysis of task-evoked activation and resting-state FC voxel-wise by activity flow mapping to estimate the information transfer pattern of episodic memory. We found that the cinguloopercular (CON), frontoparietal (FPN) and default mode networks (DMNs) were the most recruited structures in information transfer. The patterns and functions of information transfer differed between encoding and retrieval. Furthermore, we found that information transfer was a better predictor of memory ability than previous methods. Additional analysis indicated that structural connectivity (SC) had a transportive role in information transfer. Finally, we present the information transfer mechanism of episodic memory from multiple neural perspectives. These findings suggest that information transfer is a better biological indicator that accurately describes signal communication in the brain and strongly influences the function of episodic memory.

Three aspects of representation in neuroscience

Neuroscientists often describe neural activity as a representation of something, or claim to have found evidence for a neural representation, but there is considerable ambiguity about what such claims entail. Here we develop a thorough account of what 'representation' does and should do for neuroscientists in terms of three key aspects of representation. (i) Correlation: a neural representation correlates to its represented content; (ii) causal role: the representation has a characteristic effect on behavior; and (iii) teleology: a goal or purpose served by the behavior and thus the representation. We draw broadly on literature in both neuroscience and philosophy to show how these three aspects are rooted in common approaches to understanding the brain and mind. We first describe different contexts that 'representation' has been closely linked to in neuroscience, then discuss each of the three aspects in detail.

Neocortical layer 5 subclasses: From cellular properties to roles in behavior

Layer 5 (L5) serves as the main output layer of cortical structures, where long-range projecting pyramidal neurons broadcast the columnar output to other cortical and extracortical regions of the brain. L5 pyramidal neurons are grouped into two subclasses based on their projection targets; while intratelencephalic (IT) neurons project to cortical areas and the striatum, extratelencephalic (ET) neurons project to subcortical areas such as the thalamus, midbrain, and brainstem. Each L5 subclass possesses distinct morphological and electrophysiological properties and is incorporated into a unique synaptic network. Thanks to recent advances in genetic tools and methodologies, it has now become possible to distinguish between the two subclasses in the living brain. There is increasing evidence indicating that each subclass plays a unique role in sensory processing, decision-making, and learning. This review first summarizes the anatomical and physiological properties as well as the neuromodulation of IT and ET neurons in the rodent neocortex, and then reviews recent literature on their roles in sensory processing and rodent behavior. Our ultimate goal is to provide a comprehensive understanding of the role of each subclass in cortical function by examining their operational regimes based on their cellular properties.

A Recurrent Neural Network Model Accounts for Both Timing and Working Memory Components of an Interval Discrimination Task

Interval discrimination is of fundamental importance to many forms of sensory processing, including speech and music. Standard interval discrimination tasks require comparing two intervals separated in time, and thus include both working memory (WM) and timing components. Models of interval discrimination invoke separate circuits for the timing and WM components. Here we examine if, in principle, the same recurrent neural network can implement both. Using human psychophysics, we first explored the role of the WM component by varying the interstimulus delay. Consistent with previous studies, discrimination was significantly worse for a 250 ms delay, compared to 750 and 1500 ms delays, suggesting that the first interval is stably stored in WM for longer delays. We next successfully trained a recurrent neural network (RNN) on the task, demonstrating that the same network can implement both the timing and WM components. Many units in the RNN were tuned to specific intervals during the sensory epoch, and others encoded the first interval during the delay period. Overall, the encoding strategy was consistent with the notion of mixed selectivity. Units generally encoded more interval information during the sensory epoch than in the delay period, reflecting categorical encoding of short versus long in WM, rather than encoding of the specific interval. Our results demonstrate that, in contrast to standard models of interval discrimination that invoke a separate memory module, the same network can, in principle, solve the timing, WM, and comparison components of an interval discrimination task.

Recurrent neural networks with explicit representation of dynamic latent variables can mimic behavioral patterns in a physical inference task

Primates can richly parse sensory inputs to infer latent information. This ability is hypothesized to rely on establishing mental models of the external world and running mental simulations of those models. However, evidence supporting this hypothesis is limited to behavioral models that do not emulate neural computations. Here, we test this hypothesis by directly comparing the behavior of primates (humans and monkeys) in a ball interception task to that of a large set of recurrent neural network (RNN) models with or without the capacity to dynamically track the underlying latent variables. Humans and monkeys exhibit similar behavioral patterns. This primate behavioral pattern is best captured by RNNs endowed with dynamic inference, consistent with the hypothesis that the primate brain uses dynamic inferences to support flexible physical predictions. Moreover, our work highlights a general strategy for using model neural systems to test computational hypotheses of higher brain function.

Functional coordinates: Modeling interactions between brain regions as points in a function space

Here, we propose a novel technique to investigate nonlinear interactions between brain regions that captures both the strength and type of the functional relationship. Inspired by the field of functional analysis, we propose that the relationship between activity in separate brain areas can be viewed as a point in function space, identified by coordinates along an infinite set of basis functions. Using Hermite polynomials as bases, we estimate a subset of these values that serve as "functional coordinates," characterizing the interaction between BOLD activity across brain areas. We provide a proof of the convergence of the estimates in the limit, and we validate the method with simulations in which the ground truth is known, additionally showing that functional coordinates detect statistical dependence even when correlations ("functional connectivity") approach zero. We then use functional coordinates to examine neural interactions with a chosen seed region: the fusiform face area (FFA). Using k-means clustering across each voxel's functional coordinates, we illustrate that adding nonlinear basis functions allows for the discrimination of interregional interactions that are otherwise grouped together when using only linear dependence. Finally, we show that regions in V5 and medial occipital and temporal lobes exhibit significant nonlinear interactions with the FFA.

Geometry of neural computation unifies working memory and planning

Real-world tasks require coordination of working memory, decision-making, and planning, yet these cognitive functions have disproportionately been studied as independent modular processes in the brain. Here, we propose that contingency representations, defined as mappings for how future behaviors depend on upcoming events, can unify working memory and planning computations. We designed a task capable of disambiguating distinct types of representations. In task-optimized recurrent neural networks, we investigated possible circuit mechanisms for contingency representations and found that these representations can explain neurophysiological observations from the prefrontal cortex during working memory tasks. Our experiments revealed that human behavior is consistent with contingency representations and not with traditional sensory models of working memory. Finally, we generated falsifiable predictions for neural data to identify contingency representations in neural data and to dissociate different models of working memory. Our findings characterize a neural representational strategy that can unify working memory, planning, and context-dependent decision-making.

Small, correlated changes in synaptic connectivity may facilitate rapid motor learning

Animals rapidly adapt their movements to external perturbations, a process paralleled by changes in neural activity in the motor cortex. Experimental studies suggest that these changes originate from altered inputs (H_input) rather than from changes in local connectivity (H_local), as neural covariance is largely preserved during adaptation. Since measuring synaptic changes in vivo remains very challenging, we used a modular recurrent neural network to qualitatively test this interpretation. As expected, H_input resulted in small activity changes and largely preserved covariance. Surprisingly given the presumed dependence of stable covariance on preserved circuit connectivity, H_local led to only slightly larger changes in activity and covariance, still within the range of experimental recordings. This similarity is due to H_local only requiring small, correlated connectivity changes for successful adaptation. Simulations of tasks that impose increasingly larger behavioural changes revealed a growing difference between H_input and H_local, which could be exploited when designing future experiments.

Beneficial Perturbation Network for Designing General Adaptive Artificial Intelligence Systems

The human brain is the gold standard of adaptive learning. It not only can learn and benefit from experience, but also can adapt to new situations. In contrast, deep neural networks only learn one sophisticated but fixed mapping from inputs to outputs. This limits their applicability to more dynamic situations, where the input to output mapping may change with different contexts. A salient example is continual learning-learning new independent tasks sequentially without forgetting previous tasks. Continual learning of multiple tasks in artificial neural networks using gradient descent leads to catastrophic forgetting, whereby a previously learned mapping of an old task is erased when learning new mappings for new tasks. Herein, we propose a new biologically plausible type of deep neural network with extra, out-of-network, task-dependent biasing units to accommodate these dynamic situations. This allows, for the first time, a single network to learn potentially unlimited parallel input to output mappings, and to switch on the fly between them at runtime. Biasing units are programed by leveraging beneficial perturbations (opposite to well-known adversarial perturbations) for each task. Beneficial perturbations for a given task bias the network toward that task, essentially switching the network into a different mode to process that task. This largely eliminates catastrophic interference between tasks. Our approach is memory-efficient and parameter-efficient, can accommodate many tasks, and achieves the state-of-the-art performance across different tasks and domains.

Integrated neural dynamics for behavioural decisions and attentional competition in the prefrontal cortex

In the behaving monkey, complex neural dynamics in the prefrontal cortex contribute to context-dependent decisions and attentional competition. We used demixed principal component analysis to track prefrontal activity dynamics in a cued target detection task. In this task, the animal combined identity of a visual object with a prior instruction cue to determine a target/nontarget decision. From population activity, we extracted principal components for each task feature and examined their time course and sensitivity to stimulus and task variations. For displays containing a single choice object in left or right hemifield, object identity, cue identity and decision were all encoded in population activity, with different dynamics and lateralisation. Object information peaked at 100-200 ms from display onset and was largely confined to the contralateral hemisphere. Cue information was weaker and present even prior to display onset. Integrating information from cue and object, decision information arose more slowly and was bilateral. Individual neurons contributed independently to coding of the three task features. The analysis was then extended to displays with a target in one hemifield and a competing distractor in the other. In this case, the data suggest that each hemisphere initially encoded the identity of the contralateral object. The distractor representation was then rapidly suppressed, with the final target decision again encoded bilaterally. The results show how information is coded along task-related dimensions while competition is resolved and suggest how information flows within and across frontal lobes to implement a learned behavioural decision.

Neural Algorithms and Circuits for Motor Planning

The brain plans and executes volitional movements. The underlying patterns of neural population activity have been explored in the context of movements of the eyes, limbs, tongue, and head in nonhuman primates and rodents. How do networks of neurons produce the slow neural dynamics that prepare specific movements and the fast dynamics that ultimately initiate these movements? Recent work exploits rapid and calibrated perturbations of neural activity to test specific dynamical systems models that are capable of producing the observed neural activity. These joint experimental and computational studies show that cortical dynamics during motor planning reflect fixed points of neural activity (attractors). Subcortical control signals reshape and move attractors over multiple timescales, causing commitment to specific actions and rapid transitions to movement execution. Experiments in rodents are beginning to reveal how these algorithms are implemented at the level of brain-wide neural circuits.

The role of population structure in computations through neural dynamics

Neural computations are currently investigated using two separate approaches: sorting neurons into functional subpopulations or examining the low-dimensional dynamics of collective activity. Whether and how these two aspects interact to shape computations is currently unclear. Using a novel approach to extract computational mechanisms from networks trained on neuroscience tasks, here we show that the dimensionality of the dynamics and subpopulation structure play fundamentally complementary roles. Although various tasks can be implemented by increasing the dimensionality in networks with fully random population structure, flexible input-output mappings instead require a non-random population structure that can be described in terms of multiple subpopulations. Our analyses revealed that such a subpopulation structure enables flexible computations through a mechanism based on gain-controlled modulations that flexibly shape the collective dynamics. Our results lead to task-specific predictions for the structure of neural selectivity, for inactivation experiments and for the implication of different neurons in multi-tasking.

Priority-based transformations of stimulus representation in visual working memory

How does the brain prioritize among the contents of working memory (WM) to appropriately guide behavior? Previous work, employing inverted encoding modeling (IEM) of electroencephalography (EEG) and functional magnetic resonance imaging (fMRI) datasets, has shown that unprioritized memory items (UMI) are actively represented in the brain, but in a “flipped”, or opposite, format compared to prioritized memory items (PMI). To acquire independent evidence for such a priority-based representational transformation, and to explore underlying mechanisms, we trained recurrent neural networks (RNNs) with a long short-term memory (LSTM) architecture to perform a 2-back WM task. Visualization of LSTM hidden layer activity using Principal Component Analysis (PCA) confirmed that stimulus representations undergo a representational transformation–consistent with a flip—while transitioning from the functional status of UMI to PMI. Demixed (d)PCA of the same data identified two representational trajectories, one each within a UMI subspace and a PMI subspace, both undergoing a reversal of stimulus coding axes. dPCA of data from an EEG dataset also provided evidence for priority-based transformations of the representational code, albeit with some differences. This type of transformation could allow for retention of unprioritized information in WM while preventing it from interfering with concurrent behavior. The results from this initial exploration suggest that the algorithmic details of how this transformation is carried out by RNNs, versus by the human brain, may differ.

Social impact and governance of AI and neurotechnologies

Advances in artificial intelligence (AI) and brain science are going to have a huge impact on society. While technologies based on those advances can provide enormous social benefits, adoption of new technologies poses various risks. This article first reviews the co-evolution of AI and brain science and the benefits of brain-inspired AI in sustainability, healthcare, and scientific discoveries. We then consider possible risks from those technologies, including intentional abuse, autonomous weapons, cognitive enhancement by brain-computer interfaces, insidious effects of social media, inequity, and enfeeblement. We also discuss practical ways to bring ethical principles into practice. One proposal is to stop giving explicit goals to AI agents and to enable them to keep learning human preferences. Another is to learn from democratic mechanisms that evolved in human society to avoid over-consolidation of power. Finally, we emphasize the importance of open discussions not only by experts, but also including a diverse array of lay opinions.

The Spatiotemporal Neural Dynamics of Intersensory Attention Capture of Salient Stimuli: A Large-Scale Auditory-Visual Modeling Study

The spatiotemporal dynamics of the neural mechanisms underlying endogenous (top-down) and exogenous (bottom-up) attention, and how attention is controlled or allocated in intersensory perception are not fully understood. We investigated these issues using a biologically realistic large-scale neural network model of visual-auditory object processing of short-term memory. We modeled and incorporated into our visual-auditory object-processing model the temporally changing neuronal mechanisms for the control of endogenous and exogenous attention. The model successfully performed various bimodal working memory tasks, and produced simulated behavioral and neural results that are consistent with experimental findings. Simulated fMRI data were generated that constitute predictions that human experiments could test. Furthermore, in our visual-auditory bimodality simulations, we found that increased working memory load in one modality would reduce the distraction from the other modality, and a possible network mediating this effect is proposed based on our model.

The geometry of domain-general performance monitoring in the human medial frontal cortex

Controlling behavior to flexibly achieve desired goals depends on the ability to monitor one's own performance. It is unknown how performance monitoring can be both flexible, to support different tasks, and specialized, to perform each task well. We recorded single neurons in the human medial frontal cortex while subjects performed two tasks that involve three types of cognitive conflict. Neurons encoding conflict probability, conflict, and error in one or both tasks were intermixed, forming a representational geometry that simultaneously allowed task specialization and generalization. Neurons encoding conflict retrospectively served to update internal estimates of conflict probability. Population representations of conflict were compositional. These findings reveal how representations of evaluative signals can be both abstract and task-specific and suggest a neuronal mechanism for estimating control demand.

Minimal cross-trial generalization in learning the representation of an odor-guided choice task

There is no single way to represent a task. Indeed, despite experiencing the same task events and contingencies, different subjects may form distinct task representations. As experimenters, we often assume that subjects represent the task as we envision it. However, such a representation cannot be taken for granted, especially in animal experiments where we cannot deliver explicit instruction regarding the structure of the task. Here, we tested how rats represent an odor-guided choice task in which two odor cues indicated which of two responses would lead to reward, whereas a third odor indicated free choice among the two responses. A parsimonious task representation would allow animals to learn from the forced trials what is the better option to choose in the free-choice trials. However, animals may not necessarily generalize across odors in this way. We fit reinforcement-learning models that use different task representations to trial-by-trial choice behavior of individual rats performing this task, and quantified the degree to which each animal used the more parsimonious representation, generalizing across trial types. Model comparison revealed that most rats did not acquire this representation despite extensive experience. Our results demonstrate the importance of formally testing possible task representations that can afford the observed behavior, rather than assuming that animals’ task representations abide by the generative task structure that governs the experimental design.

To study how animals learn and make decisions, scientists design experiments, train animals to perform them, and observe how they behave. During this process, an important but rarely asked question is how animals understand the experiment. Merely through observing animals’ behavior in a task, it is often hard to determine if they understand the task in the same way as the experimenter expects. Assuming that animals represent tasks differently than they actually do may lead to incorrect interpretations of behavioral or neural results. Here, we compared different possible representations for a simple reward-learning task in terms of how well these alternative models explain animal’s choice behavior. We found that rats did not represent the task in the most parsimonious way, thereby failing to learn from forced-choice trials what rewards are available on free-choice trials, despite extensive training on the task. These results caution against simply assuming that animals’ understanding of a task corresponds to the way the task was designed.

Brain-like functional specialization emerges spontaneously in deep neural networks

The human brain contains multiple regions with distinct, often highly specialized functions, from recognizing faces to understanding language to thinking about what others are thinking. However, it remains unclear why the cortex exhibits this high degree of functional specialization in the first place. Here, we consider the case of face perception using artificial neural networks to test the hypothesis that functional segregation of face recognition in the brain reflects a computational optimization for the broader problem of visual recognition of faces and other visual categories. We find that networks trained on object recognition perform poorly on face recognition and vice versa and that networks optimized for both tasks spontaneously segregate themselves into separate systems for faces and objects. We then show functional segregation to varying degrees for other visual categories, revealing a widespread tendency for optimization (without built-in task-specific inductive biases) to lead to functional specialization in machines and, we conjecture, also brains.

Neural Mechanisms of Working Memory Accuracy Revealed by Recurrent Neural Networks

Understanding the neural mechanisms of working memory has been a long-standing Neuroscience goal. Bump attractor models have been used to simulate persistent activity generated in the prefrontal cortex during working memory tasks and to study the relationship between activity and behavior. How realistic the assumptions of these models are has been a matter of debate. Here, we relied on an alternative strategy to gain insights into the computational principles behind the generation of persistent activity and on whether current models capture some universal computational principles. We trained Recurrent Neural Networks (RNNs) to perform spatial working memory tasks and examined what aspects of RNN activity accounted for working memory performance. Furthermore, we compared activity in fully trained networks and immature networks, achieving only imperfect performance. We thus examined the relationship between the trial-to-trial variability of responses simulated by the network and different aspects of unit activity as a way of identifying the critical parameters of memory maintenance. Properties that spontaneously emerged in the artificial network strongly resembled persistent activity of prefrontal neurons. Most importantly, these included drift of network activity during the course of a trial that was causal to the behavior of the network. As a consequence, delay period firing rate and behavior were positively correlated, in strong analogy to experimental results from the prefrontal cortex. These findings reveal that delay period activity is computationally efficient in maintaining working memory, as evidenced by unbiased optimization of parameters in artificial neural networks, oblivious to the properties of prefrontal neurons.

Constructing neural network models from brain data reveals representational transformations linked to adaptive behavior

The human ability to adaptively implement a wide variety of tasks is thought to emerge from the dynamic transformation of cognitive information. We hypothesized that these transformations are implemented via conjunctive activations in "conjunction hubs"-brain regions that selectively integrate sensory, cognitive, and motor activations. We used recent advances in using functional connectivity to map the flow of activity between brain regions to construct a task-performing neural network model from fMRI data during a cognitive control task. We verified the importance of conjunction hubs in cognitive computations by simulating neural activity flow over this empirically-estimated functional connectivity model. These empirically-specified simulations produced above-chance task performance (motor responses) by integrating sensory and task rule activations in conjunction hubs. These findings reveal the role of conjunction hubs in supporting flexible cognitive computations, while demonstrating the feasibility of using empirically-estimated neural network models to gain insight into cognitive computations in the human brain.

Thunderstruck: The ACDC model of flexible sequences and rhythms in recurrent neural circuits

Adaptive sequential behavior is a hallmark of human cognition. In particular, humans can learn to produce precise spatiotemporal sequences given a certain context. For instance, musicians can not only reproduce learned action sequences in a context-dependent manner, they can also quickly and flexibly reapply them in any desired tempo or rhythm without overwriting previous learning. Existing neural network models fail to account for these properties. We argue that this limitation emerges from the fact that sequence information (i.e., the position of the action) and timing (i.e., the moment of response execution) are typically stored in the same neural network weights. Here, we augment a biologically plausible recurrent neural network of cortical dynamics to include a basal ganglia-thalamic module which uses reinforcement learning to dynamically modulate action. This “associative cluster-dependent chain” (ACDC) model modularly stores sequence and timing information in distinct loci of the network. This feature increases computational power and allows ACDC to display a wide range of temporal properties (e.g., multiple sequences, temporal shifting, rescaling, and compositionality), while still accounting for several behavioral and neurophysiological empirical observations. Finally, we apply this ACDC network to show how it can learn the famous “Thunderstruck” song intro and then flexibly play it in a “bossa nova” rhythm without further training.

How do humans flexibly adapt action sequences? For instance, musicians can learn a song and quickly speed up or slow down the tempo, or even play the song following a completely different rhythm (e.g., a rock song using a bossa nova rhythm). In this work, we build a biologically plausible network of cortico-basal ganglia interactions that explains how this temporal flexibility may emerge in the brain. Crucially, our model factorizes sequence order and action timing, respectively represented in cortical and basal ganglia dynamics. This factorization allows full temporal flexibility, i.e. the timing of a learned action sequence can be recomposed without interfering with the order of the sequence. As such, our model is capable of learning asynchronous action sequences, and flexibly shift, rescale, and recompose them, while accounting for biological data.

Advances in modeling learning and decision-making in neuroscience

An organism's survival depends on its ability to learn about its environment and to make adaptive decisions in the service of achieving the best possible outcomes in that environment. To study the neural circuits that support these functions, researchers have increasingly relied on models that formalize the computations required to carry them out. Here, we review the recent history of computational modeling of learning and decision-making, and how these models have been used to advance understanding of prefrontal cortex function. We discuss how such models have advanced from their origins in basic algorithms of updating and action selection to increasingly account for complexities in the cognitive processes required for learning and decision-making, and the representations over which they operate. We further discuss how a deeper understanding of the real-world complexities in these computations has shed light on the fundamental constraints on optimal behavior, and on the complex interactions between corticostriatal pathways to determine such behavior. The continuing and rapid development of these models holds great promise for understanding the mechanisms by which animals adapt to their environments, and what leads to maladaptive forms of learning and decision-making within clinical populations.

Reinforcement-learning in fronto-striatal circuits

We review the current state of knowledge on the computational and neural mechanisms of reinforcement-learning with a particular focus on fronto-striatal circuits. We divide the literature in this area into five broad research themes: the target of the learning-whether it be learning about the value of stimuli or about the value of actions; the nature and complexity of the algorithm used to drive the learning and inference process; how learned values get converted into choices and associated actions; the nature of state representations, and of other cognitive machinery that support the implementation of various reinforcement-learning operations. An emerging fifth area focuses on how the brain allocates or arbitrates control over different reinforcement-learning sub-systems or "experts". We will outline what is known about the role of the prefrontal cortex and striatum in implementing each of these functions. We then conclude by arguing that it will be necessary to build bridges from algorithmic level descriptions of computational reinforcement-learning to implementational level models to better understand how reinforcement-learning emerges from multiple distributed neural networks in the brain.

Using top-down modulation to optimally balance shared versus separated task representations

Human adaptive behavior requires continually learning and performing a wide variety of tasks, often with very little practice. To accomplish this, it is crucial to separate neural representations of different tasks in order to avoid interference. At the same time, sharing neural representations supports generalization and allows faster learning. Therefore, a crucial challenge is to find an optimal balance between shared versus separated representations. Typically, models of human cognition employ top-down modulatory signals to separate task representations, but there exist surprisingly little systematic computational investigations of how such modulation is best implemented. We identify and systematically evaluate two crucial features of modulatory signals. First, top-down input can be processed in an additive or multiplicative manner. Second, the modulatory signals can be adaptive (learned) or non-adaptive (random). We cross these two features, resulting in four modulation networks which are tested on a variety of input datasets and tasks with different degrees of stimulus-action mapping overlap. The multiplicative adaptive modulation network outperforms all other networks in terms of accuracy. Moreover, this network develops hidden units that optimally share representations between tasks. Specifically, different than the binary approach of currently popular latent state models, it exploits partial overlap between tasks.

Computational mechanisms of distributed value representations and mixed learning strategies

Learning appropriate representations of the reward environment is challenging in the real world where there are many options, each with multiple attributes or features. Despite existence of alternative solutions for this challenge, neural mechanisms underlying emergence and adoption of value representations and learning strategies remain unknown. To address this, we measure learning and choice during a multi-dimensional probabilistic learning task in humans and trained recurrent neural networks (RNNs) to capture our experimental observations. We find that human participants estimate stimulus-outcome associations by learning and combining estimates of reward probabilities associated with the informative feature followed by those of informative conjunctions. Through analyzing representations, connectivity, and lesioning of the RNNs, we demonstrate this mixed learning strategy relies on a distributed neural code and opponency between excitatory and inhibitory neurons through value-dependent disinhibition. Together, our results suggest computational and neural mechanisms underlying emergence of complex learning strategies in naturalistic settings.

Gated recurrence enables simple and accurate sequence prediction in stochastic, changing, and structured environments

From decision making to perception to language, predicting what is coming next is crucial. It is also challenging in stochastic, changing, and structured environments; yet the brain makes accurate predictions in many situations. What computational architecture could enable this feat? Bayesian inference makes optimal predictions but is prohibitively difficult to compute. Here, we show that a specific recurrent neural network architecture enables simple and accurate solutions in several environments. This architecture relies on three mechanisms: gating, lateral connections, and recurrent weight training. Like the optimal solution and the human brain, such networks develop internal representations of their changing environment (including estimates of the environment’s latent variables and the precision of these estimates), leverage multiple levels of latent structure, and adapt their effective learning rate to changes without changing their connection weights. Being ubiquitous in the brain, gated recurrence could therefore serve as a generic building block to predict in real-life environments.

Similarity of computations across domains does not imply shared implementation: The case of language comprehension

Understanding language requires applying cognitive operations (e.g., memory retrieval, prediction, structure building)-relevant across many cognitive domains-to specialized knowledge structures (a particular language's phonology, lexicon, and syntax). Are these computations carried out by domain-general circuits or by circuits that store domain-specific representations? Recent work has characterized the roles in language comprehension of the language-selective network and the multiple demand (MD) network, which has been implicated in executive functions and linked to fluid intelligence, making it a prime candidate for implementing computations that support information processing across domains. The language network responds robustly to diverse aspects of comprehension, but the MD network shows no sensitivity to linguistic variables. We therefore argue that the MD network does not play a core role in language comprehension, and that past claims to the contrary are likely due to methodological artifacts. Although future studies may discover some aspects of language that require the MD network, evidence to date suggests that those will not be related to core linguistic processes like lexical access or composition. The finding that the circuits that store linguistic knowledge carry out computations on those representations aligns with general arguments against the separation between memory and computation in the mind and brain.

Evolving the olfactory system with machine learning

The convergent evolution of the fly and mouse olfactory system led us to ask whether the anatomic connectivity and functional logic of olfactory circuits would evolve in artificial neural networks trained to perform olfactory tasks. Artificial networks trained to classify odor identity recapitulate the connectivity inherent in the olfactory system. Input units are driven by a single receptor type, and units driven by the same receptor converge to form a glomerulus. Glomeruli exhibit sparse, unstructured connectivity onto a larger expansion layer of Kenyon cells. When trained to both classify odor identity and to impart innate valence onto odors, the network develops independent pathways for identity and valence classification. Thus, the defining features of fly and mouse olfactory systems also evolved in artificial neural networks trained to perform olfactory tasks. This implies that convergent evolution reflects an underlying logic rather than shared developmental principles.

Towards the next generation of recurrent network models for cognitive neuroscience

Recurrent neural networks (RNNs) trained with machine learning techniques on cognitive tasks have become a widely accepted tool for neuroscientists. In this short opinion piece, we discuss fundamental challenges faced by the early work of this approach and recent steps to overcome such challenges and build next-generation RNN models for cognition. We propose several essential questions that practitioners of this approach should address to continue to build future generations of RNN models.

Emergence of prefrontal neuron maturation properties by training recurrent neural networks in cognitive tasks

Working memory and response inhibition are functions that mature relatively late in life, after adolescence, paralleling the maturation of the prefrontal cortex. The link between behavioral and neural maturation is not obvious, however, making it challenging to understand how neural activity underlies the maturation of cognitive function. To gain insights into the nature of observed changes in prefrontal activity between adolescence and adulthood, we investigated the progressive changes in unit activity of recurrent neural networks as they were trained to perform working memory and response inhibition tasks. These included increased delay period activity during working memory tasks and increased activation in antisaccade tasks. These findings reveal universal properties underlying the neuronal computations behind cognitive tasks and explicate the nature of changes that occur as the result of developmental maturation.

•

Properties of RNN networks during training offer insights in prefrontal maturation

•

Fully trained networks exhibit higher levels of activity in working memory tasks

•

Trained networks also exhibit higher activation in antisaccade tasks

•

Partially trained RNNs can generate accurate predictions of immature PFC activity

50 years of mnemonic persistent activity: quo vadis?

Half a century ago persistent spiking activity in the neocortex was discovered to be a neural substrate of working memory. Since then scientists have sought to understand this core cognitive function across biological and computational levels. Studies are reviewed here that cumulatively lend support to a synaptic theory of recurrent circuits for mnemonic persistent activity that depends on various cellular and network substrates and is mathematically described by a multiple-attractor network model. Crucially, a mnemonic attractor state of the brain is consistent with temporal variations and heterogeneity across neurons in a subspace of population activity. Persistent activity should be broadly understood as a contrast to decaying transients. Mechanisms in the absence of neural firing ('activity-silent state') are suitable for passive short-term memory but not for working memory - which is characterized by executive control for filtering out distractors, limited capacity, and internal manipulation of information.

Natural and Artificial Intelligence: A brief introduction to the interplay between AI and neuroscience research

Neuroscience and artificial intelligence (AI) share a long history of collaboration. Advances in neuroscience, alongside huge leaps in computer processing power over the last few decades, have given rise to a new generation of in silico neural networks inspired by the architecture of the brain. These AI systems are now capable of many of the advanced perceptual and cognitive abilities of biological systems, including object recognition and decision making. Moreover, AI is now increasingly being employed as a tool for neuroscience research and is transforming our understanding of brain functions. In particular, deep learning has been used to model how convolutional layers and recurrent connections in the brain's cerebral cortex control important functions, including visual processing, memory, and motor control. Excitingly, the use of neuroscience-inspired AI also holds great promise for understanding how changes in brain networks result in psychopathologies, and could even be utilized in treatment regimes. Here we discuss recent advancements in four areas in which the relationship between neuroscience and AI has led to major advancements in the field; (1) AI models of working memory, (2) AI visual processing, (3) AI analysis of big neuroscience datasets, and (4) computational psychiatry.

The population doctrine in cognitive neuroscience

A major shift is happening within neurophysiology: a population doctrine is drawing level with the single-neuron doctrine that has long dominated the field. Population-level ideas have so far had their greatest impact in motor neuroscience, but they hold great promise for resolving open questions in cognition as well. Here, we codify the population doctrine and survey recent work that leverages this view to specifically probe cognition. Our discussion is organized around five core concepts that provide a foundation for population-level thinking: (1) state spaces, (2) manifolds, (3) coding dimensions, (4) subspaces, and (5) dynamics. The work we review illustrates the progress and promise that population-level thinking holds for cognitive neuroscience-for delivering new insight into attention, working memory, decision-making, executive function, learning, and reward processing.

Modulation of the dynamical state in cortical network models

Cortical neural responses can be modulated by various factors, such as stimulus inputs and the behavior state of the animal. Understanding the circuit mechanisms underlying modulations of network dynamics is important to understand the flexibility of circuit computations. Identifying the dynamical state of a network is an important first step to predict network responses to external stimulus and top-down modulatory inputs. Models in stable or unstable dynamical regimes require different analytic tools to estimate the network responses to inputs and the structure of neural variability. In this article, I review recent cortical models of state-dependent responses and their predictions about the underlying modulatory mechanisms.

Interpreting neural computations by examining intrinsic and embedding dimensionality of neural activity

The ongoing exponential rise in recording capacity calls for new approaches for analysing and interpreting neural data. Effective dimensionality has emerged as an important property of neural activity across populations of neurons, yet different studies rely on different definitions and interpretations of this quantity. Here, we focus on intrinsic and embedding dimensionality, and discuss how they might reveal computational principles from data. Reviewing recent works, we propose that the intrinsic dimensionality reflects information about the latent variables encoded in collective activity while embedding dimensionality reveals the manner in which this information is processed. We conclude by highlighting the role of network models as an ideal substrate for testing more specifically various hypotheses on the computational principles reflected through intrinsic and embedding dimensionality.

Working Memory: From Neural Activity to the Sentient Mind

Working memory (WM) is the ability to maintain and manipulate information in the conscious mind over a timescale of seconds. This ability is thought to be maintained through the persistent discharges of neurons in a network of brain areas centered on the prefrontal cortex, as evidenced by neurophysiological recordings in nonhuman primates, though both the localization and the neural basis of WM has been a matter of debate in recent years. Neural correlates of WM are evident in species other than primates, including rodents and corvids. A specialized network of excitatory and inhibitory neurons, aided by neuromodulatory influences of dopamine, is critical for the maintenance of neuronal activity. Limitations in WM capacity and duration, as well as its enhancement during development, can be attributed to properties of neural activity and circuits. Changes in these factors can be observed through training-induced improvements and in pathological impairments. WM thus provides a prototypical cognitive function whose properties can be tied to the spiking activity of brain neurons. © 2021 American Physiological Society. Compr Physiol 11:1-41, 2021.

Slow manifolds within network dynamics encode working memory efficiently and robustly

Working memory is a cognitive function involving the storage and manipulation of latent information over brief intervals of time, thus making it crucial for context-dependent computation. Here, we use a top-down modeling approach to examine network-level mechanisms of working memory, an enigmatic issue and central topic of study in neuroscience. We optimize thousands of recurrent rate-based neural networks on a working memory task and then perform dynamical systems analysis on the ensuing optimized networks, wherein we find that four distinct dynamical mechanisms can emerge. In particular, we show the prevalence of a mechanism in which memories are encoded along slow stable manifolds in the network state space, leading to a phasic neuronal activation profile during memory periods. In contrast to mechanisms in which memories are directly encoded at stable attractors, these networks naturally forget stimuli over time. Despite this seeming functional disadvantage, they are more efficient in terms of how they leverage their attractor landscape and paradoxically, are considerably more robust to noise. Our results provide new hypotheses regarding how working memory function may be encoded within the dynamics of neural circuits.

Stimulus-Driven and Spontaneous Dynamics in Excitatory-Inhibitory Recurrent Neural Networks for Sequence Representation

Recurrent neural networks (RNNs) have been widely used to model sequential neural dynamics ("neural sequences") of cortical circuits in cognitive and motor tasks. Efforts to incorporate biological constraints and Dale's principle will help elucidate the neural representations and mechanisms of underlying circuits. We trained an excitatory-inhibitory RNN to learn neural sequences in a supervised manner and studied the representations and dynamic attractors of the trained network. The trained RNN was robust to trigger the sequence in response to various input signals and interpolated a time-warped input for sequence representation. Interestingly, a learned sequence can repeat periodically when the RNN evolved beyond the duration of a single sequence. The eigenspectrum of the learned recurrent connectivity matrix with growing or damping modes, together with the RNN's nonlinearity, were adequate to generate a limit cycle attractor. We further examined the stability of dynamic attractors while training the RNN to learn two sequences. Together, our results provide a general framework for understanding neural sequence representation in the excitatory-inhibitory RNN.

Primate neuronal connections are sparse in cortex as compared to mouse

Detailing how primate and mouse neurons differ is critical for creating generalized models of how neurons process information. We reconstruct 15,748 synapses in adult Rhesus macaques and mice and ask how connectivity differs on identified cell types in layer 2/3 of primary visual cortex. Primate excitatory and inhibitory neurons receive 2-5 times fewer excitatory and inhibitory synapses than similar mouse neurons. Primate excitatory neurons have lower excitatory-to-inhibitory (E/I) ratios than mouse but similar E/I ratios in inhibitory neurons. In both species, properties of inhibitory axons such as synapse size and frequency are unchanged, and inhibitory innervation of excitatory neurons is local and specific. Using artificial recurrent neural networks (RNNs) optimized for different cognitive tasks, we find that penalizing networks for creating and maintaining synapses, as opposed to neuronal firing, reduces the number of connections per node as the number of nodes increases, similar to primate neurons compared with mice.

Distributed functions of prefrontal and parietal cortices during sequential categorical decisions

Comparing sequential stimuli is crucial for guiding complex behaviors. To understand mechanisms underlying sequential decisions, we compared neuronal responses in the prefrontal cortex (PFC), the lateral intraparietal (LIP), and medial intraparietal (MIP) areas in monkeys trained to decide whether sequentially presented stimuli were from matching (M) or nonmatching (NM) categories. We found that PFC leads M/NM decisions, whereas LIP and MIP appear more involved in stimulus evaluation and motor planning, respectively. Compared to LIP, PFC showed greater nonlinear integration of currently visible and remembered stimuli, which correlated with the monkeys' M/NM decisions. Furthermore, multi-module recurrent networks trained on the same task exhibited key features of PFC and LIP encoding, including nonlinear integration in the PFC-like module, which was causally involved in the networks' decisions. Network analysis found that nonlinear units have stronger and more widespread connections with input, output, and within-area units, indicating putative circuit-level mechanisms for sequential decisions.

Rationalizing constraints on the capacity for cognitive control

Humans are remarkably limited in: (i) how many control-dependent tasks they can execute simultaneously, and (ii) how intensely they can focus on a single task. These limitations are universal assumptions of most theories of cognition. Yet, a rationale for why humans are subject to these constraints remains elusive. This feature review draws on recent insights from psychology, neuroscience, and machine learning, to suggest that constraints on cognitive control may result from a rational adaptation to fundamental, computational dilemmas in neural architectures. The reviewed literature implies that limitations in multitasking may result from a trade-off between learning efficacy and processing efficiency and that limitations in the intensity of commitment to a single task may reflect a trade-off between cognitive stability and flexibility.