A goal-driven modular neural network predicts parietofrontal neural dynamics during grasping

Reach-dependent reorientation of rotational dynamics in motor cortex

During reaching, neurons in motor cortex exhibit complex, time-varying activity patterns. Though single-neuron activity correlates with movement parameters, movement correlations explain neural activity only partially. Neural responses also reflect population-level dynamics thought to generate outputs. These dynamics have previously been described as "rotational," such that activity orbits in neural state space. Here, we reanalyze reaching datasets from male Rhesus macaques and find two essential features that cannot be accounted for with standard dynamics models. First, the planes in which rotations occur differ for different reaches. Second, this variation in planes reflects the overall location of activity in neural state space. Our "location-dependent rotations" model fits nearly all motor cortex activity during reaching, and high-quality decoding of reach kinematics reveals a quasilinear relationship with spiking. Varying rotational planes allows motor cortex to produce richer outputs than possible under previous models. Finally, our model links representational and dynamical ideas: representation is present in the state space location, which dynamics then convert into time-varying command signals.

Neural Manifold Constraint for Spike Prediction Models Under Behavioral Reinforcement

Spike prediction models effectively predict downstream spike trains from upstream neural activity for neural prostheses. Such prostheses could potentially restore damaged neural communication pathways using predicted patterns to guide electrical stimulations on downstream. Since the ground truth of downstream neural activity is unavailable for subjects with the damage, reinforcement learning (RL) with behavior-level rewards becomes necessary for model training. However, existing models do not involve any constraint on the generated firing patterns and neglect the correlations among neural activities. Thus, the model outputs can greatly deviate from the natural range of neural activities, causing concerns for clinical usage. This study proposes the neural manifold constraint to solve this problem, shaping RL-generated spike trains in the feature space. The constraint terms describe the first and second order statistics of the neural manifold estimated from neural recordings during subjects' freely moving period. Then, the models can be optimized within the neural manifold by behavioral reinforcement. We test the method to predict primary motor cortex (M1) spikes from medial prefrontal (mPFC) spikes when rats perform the two-lever discrimination task. Results show that the neural activity generated by constrained models resembles the real M1 recordings. Compared with models without constraints, our approach achieves similar behavioral success rates, but reduces the mean squared error of neural firing by 61%. The constraints also increase the model's robustness across data segments and induce realistic neural correlations. Our method provides a promising tool to restore transregional communication with high behavioral performance and more realistic microscopic patterns.

A virtual rodent predicts the structure of neural activity across behaviours

Animals have exquisite control of their bodies, allowing them to perform a diverse range of behaviours. How such control is implemented by the brain, however, remains unclear. Advancing our understanding requires models that can relate principles of control to the structure of neural activity in behaving animals. Here, to facilitate this, we built a 'virtual rodent', in which an artificial neural network actuates a biomechanically realistic model of the rat1 in a physics simulator2. We used deep reinforcement learning3-5 to train the virtual agent to imitate the behaviour of freely moving rats, thus allowing us to compare neural activity recorded in real rats to the network activity of a virtual rodent mimicking their behaviour. We found that neural activity in the sensorimotor striatum and motor cortex was better predicted by the virtual rodent's network activity than by any features of the real rat's movements, consistent with both regions implementing inverse dynamics6. Furthermore, the network's latent variability predicted the structure of neural variability across behaviours and afforded robustness in a way consistent with the minimal intervention principle of optimal feedback control7. These results demonstrate how physical simulation of biomechanically realistic virtual animals can help interpret the structure of neural activity across behaviour and relate it to theoretical principles of motor control.

MotorNet, a Python toolbox for controlling differentiable biomechanical effectors with artificial neural networks

Artificial neural networks (ANNs) are a powerful class of computational models for unravelling neural mechanisms of brain function. However, for neural control of movement, they currently must be integrated with software simulating biomechanical effectors, leading to limiting impracticalities: (1) researchers must rely on two different platforms and (2) biomechanical effectors are not generally differentiable, constraining researchers to reinforcement learning algorithms despite the existence and potential biological relevance of faster training methods. To address these limitations, we developed MotorNet, an open-source Python toolbox for creating arbitrarily complex, differentiable, and biomechanically realistic effectors that can be trained on user-defined motor tasks using ANNs. MotorNet is designed to meet several goals: ease of installation, ease of use, a high-level user-friendly application programming interface, and a modular architecture to allow for flexibility in model building. MotorNet requires no dependencies outside Python, making it easy to get started with. For instance, it allows training ANNs on typically used motor control models such as a two joint, six muscle, planar arm within minutes on a typical desktop computer. MotorNet is built on PyTorch and therefore can implement any network architecture that is possible using the PyTorch framework. Consequently, it will immediately benefit from advances in artificial intelligence through PyTorch updates. Finally, it is open source, enabling users to create and share their own improvements, such as new effector and network architectures or custom task designs. MotorNet's focus on higher-order model and task design will alleviate overhead cost to initiate computational projects for new researchers by providing a standalone, ready-to-go framework, and speed up efforts of established computational teams by enabling a focus on concepts and ideas over implementation.

De novo motor learning creates structure in neural activity that shapes adaptation

Animals can quickly adapt learned movements to external perturbations, and their existing motor repertoire likely influences their ease of adaptation. Long-term learning causes lasting changes in neural connectivity, which shapes the activity patterns that can be produced during adaptation. Here, we examined how a neural population's existing activity patterns, acquired through de novo learning, affect subsequent adaptation by modeling motor cortical neural population dynamics with recurrent neural networks. We trained networks on different motor repertoires comprising varying numbers of movements, which they acquired following various learning experiences. Networks with multiple movements had more constrained and robust dynamics, which were associated with more defined neural 'structure'-organization in the available population activity patterns. This structure facilitated adaptation, but only when the changes imposed by the perturbation were congruent with the organization of the inputs and the structure in neural activity acquired during de novo learning. These results highlight trade-offs in skill acquisition and demonstrate how different learning experiences can shape the geometrical properties of neural population activity and subsequent adaptation.

Modeling proprioception with task-driven neural network models

In a recent issue of Cell, Vargas and colleagues1 demonstrate that task-driven neural network models are superior at predicting proprioceptive activity in the primate cuneate nucleus and sensorimotor cortex compared with other models. This provides valuable insights for better understanding the proprioceptive pathway.

Mapping model units to visual neurons reveals population code for social behaviour

The rich variety of behaviours observed in animals arises through the interplay between sensory processing and motor control. To understand these sensorimotor transformations, it is useful to build models that predict not only neural responses to sensory input1-5 but also how each neuron causally contributes to behaviour6,7. Here we demonstrate a novel modelling approach to identify a one-to-one mapping between internal units in a deep neural network and real neurons by predicting the behavioural changes that arise from systematic perturbations of more than a dozen neuronal cell types. A key ingredient that we introduce is 'knockout training', which involves perturbing the network during training to match the perturbations of the real neurons during behavioural experiments. We apply this approach to model the sensorimotor transformations of Drosophila melanogaster males during a complex, visually guided social behaviour8-11. The visual projection neurons at the interface between the optic lobe and central brain form a set of discrete channels12, and prior work indicates that each channel encodes a specific visual feature to drive a particular behaviour13,14. Our model reaches a different conclusion: combinations of visual projection neurons, including those involved in non-social behaviours, drive male interactions with the female, forming a rich population code for behaviour. Overall, our framework consolidates behavioural effects elicited from various neural perturbations into a single, unified model, providing a map from stimulus to neuronal cell type to behaviour, and enabling future incorporation of wiring diagrams of the brain15 into the model.

Emergent neural dynamics and geometry for generalization in a transitive inference task

Relational cognition-the ability to infer relationships that generalize to novel combinations of objects-is fundamental to human and animal intelligence. Despite this importance, it remains unclear how relational cognition is implemented in the brain due in part to a lack of hypotheses and predictions at the levels of collective neural activity and behavior. Here we discovered, analyzed, and experimentally tested neural networks (NNs) that perform transitive inference (TI), a classic relational task (if A > B and B > C, then A > C). We found NNs that (i) generalized perfectly, despite lacking overt transitive structure prior to training, (ii) generalized when the task required working memory (WM), a capacity thought to be essential to inference in the brain, (iii) emergently expressed behaviors long observed in living subjects, in addition to a novel order-dependent behavior, and (iv) expressed different task solutions yielding alternative behavioral and neural predictions. Further, in a large-scale experiment, we found that human subjects performing WM-based TI showed behavior inconsistent with a class of NNs that characteristically expressed an intuitive task solution. These findings provide neural insights into a classical relational ability, with wider implications for how the brain realizes relational cognition.

Task-driven neural network models predict neural dynamics of proprioception

Proprioception tells the brain the state of the body based on distributed sensory neurons. Yet, the principles that govern proprioceptive processing are poorly understood. Here, we employ a task-driven modeling approach to investigate the neural code of proprioceptive neurons in cuneate nucleus (CN) and somatosensory cortex area 2 (S1). We simulated muscle spindle signals through musculoskeletal modeling and generated a large-scale movement repertoire to train neural networks based on 16 hypotheses, each representing different computational goals. We found that the emerging, task-optimized internal representations generalize from synthetic data to predict neural dynamics in CN and S1 of primates. Computational tasks that aim to predict the limb position and velocity were the best at predicting the neural activity in both areas. Since task optimization develops representations that better predict neural activity during active than passive movements, we postulate that neural activity in the CN and S1 is top-down modulated during goal-directed movements.

μSim: A goal-driven framework for elucidating the neural control of movement through musculoskeletal modeling

How does the motor cortex (MC) produce purposeful and generalizable movements from the complex musculoskeletal system in a dynamic environment? To elucidate the underlying neural dynamics, we use a goal-driven approach to model MC by considering its goal as a controller driving the musculoskeletal system through desired states to achieve movement. Specifically, we formulate the MC as a recurrent neural network (RNN) controller producing muscle commands while receiving sensory feedback from biologically accurate musculoskeletal models. Given this real-time simulated feedback implemented in advanced physics simulation engines, we use deep reinforcement learning to train the RNN to achieve desired movements under specified neural and musculoskeletal constraints. Activity of the trained model can accurately decode experimentally recorded neural population dynamics and single-unit MC activity, while generalizing well to testing conditions significantly different from training. Simultaneous goal- and data- driven modeling in which we use the recorded neural activity as observed states of the MC further enhances direct and generalizable single-unit decoding. Finally, we show that this framework elucidates computational principles of how neural dynamics enable flexible control of movement and make this framework easy-to-use for future experiments.

Dynamical flexible inference of nonlinear latent factors and structures in neural population activity

Modelling the spatiotemporal dynamics in the activity of neural populations while also enabling their flexible inference is hindered by the complexity and noisiness of neural observations. Here we show that the lower-dimensional nonlinear latent factors and latent structures can be computationally modelled in a manner that allows for flexible inference causally, non-causally and in the presence of missing neural observations. To enable flexible inference, we developed a neural network that separates the model into jointly trained manifold and dynamic latent factors such that nonlinearity is captured through the manifold factors and the dynamics can be modelled in tractable linear form on this nonlinear manifold. We show that the model, which we named 'DFINE' (for 'dynamical flexible inference for nonlinear embeddings') achieves flexible inference in simulations of nonlinear dynamics and across neural datasets representing a diversity of brain regions and behaviours. Compared with earlier neural-network models, DFINE enables flexible inference, better predicts neural activity and behaviour, and better captures the latent neural manifold structure. DFINE may advance the development of neurotechnology and investigations in neuroscience.

AngoraPy: A Python toolkit for modeling anthropomorphic goal-driven sensorimotor systems

Goal-driven deep learning increasingly supplements classical modeling approaches in computational neuroscience. The strength of deep neural networks as models of the brain lies in their ability to autonomously learn the connectivity required to solve complex and ecologically valid tasks, obviating the need for hand-engineered or hypothesis-driven connectivity patterns. Consequently, goal-driven models can generate hypotheses about the neurocomputations underlying cortical processing that are grounded in macro- and mesoscopic anatomical properties of the network's biological counterpart. Whereas, goal-driven modeling is already becoming prevalent in the neuroscience of perception, its application to the sensorimotor domain is currently hampered by the complexity of the methods required to train models comprising the closed sensation-action loop. This paper describes AngoraPy, a Python library that mitigates this obstacle by providing researchers with the tools necessary to train complex recurrent convolutional neural networks that model the human sensorimotor system. To make the technical details of this toolkit more approachable, an illustrative example that trains a recurrent toy model on in-hand object manipulation accompanies the theoretical remarks. An extensive benchmark on various classical, 3D robotic, and anthropomorphic control tasks demonstrates AngoraPy's general applicability to a wide range of tasks. Together with its ability to adaptively handle custom architectures, the flexibility of this toolkit demonstrates its power for goal-driven sensorimotor modeling.

Training biologically plausible recurrent neural networks on cognitive tasks with long-term dependencies

Training recurrent neural networks (RNNs) has become a go-to approach for generating and evaluating mechanistic neural hypotheses for cognition. The ease and efficiency of training RNNs with backpropagation through time and the availability of robustly supported deep learning libraries has made RNN modeling more approachable and accessible to neuroscience. Yet, a major technical hindrance remains. Cognitive processes such as working memory and decision making involve neural population dynamics over a long period of time within a behavioral trial and across trials. It is difficult to train RNNs to accomplish tasks where neural representations and dynamics have long temporal dependencies without gating mechanisms such as LSTMs or GRUs which currently lack experimental support and prohibit direct comparison between RNNs and biological neural circuits. We tackled this problem based on the idea of specialized skip-connections through time to support the emergence of task-relevant dynamics, and subsequently reinstitute biological plausibility by reverting to the original architecture. We show that this approach enables RNNs to successfully learn cognitive tasks that prove impractical if not impossible to learn using conventional methods. Over numerous tasks considered here, we achieve less training steps and shorter wall-clock times, particularly in tasks that require learning long-term dependencies via temporal integration over long timescales or maintaining a memory of past events in hidden-states. Our methods expand the range of experimental tasks that biologically plausible RNN models can learn, thereby supporting the development of theory for the emergent neural mechanisms of computations involving long-term dependencies.

Mouse visual cortex as a limited resource system that self-learns an ecologically-general representation

Studies of the mouse visual system have revealed a variety of visual brain areas that are thought to support a multitude of behavioral capacities, ranging from stimulus-reward associations, to goal-directed navigation, and object-centric discriminations. However, an overall understanding of the mouse's visual cortex, and how it supports a range of behaviors, remains unknown. Here, we take a computational approach to help address these questions, providing a high-fidelity quantitative model of mouse visual cortex and identifying key structural and functional principles underlying that model's success. Structurally, we find that a comparatively shallow network structure with a low-resolution input is optimal for modeling mouse visual cortex. Our main finding is functional-that models trained with task-agnostic, self-supervised objective functions based on the concept of contrastive embeddings are much better matches to mouse cortex, than models trained on supervised objectives or alternative self-supervised methods. This result is very much unlike in primates where prior work showed that the two were roughly equivalent, naturally leading us to ask the question of why these self-supervised objectives are better matches than supervised ones in mouse. To this end, we show that the self-supervised, contrastive objective builds a general-purpose visual representation that enables the system to achieve better transfer on out-of-distribution visual scene understanding and reward-based navigation tasks. Our results suggest that mouse visual cortex is a low-resolution, shallow network that makes best use of the mouse's limited resources to create a light-weight, general-purpose visual system-in contrast to the deep, high-resolution, and more categorization-dominated visual system of primates.

Distinctive properties of biological neural networks and recent advances in bottom-up approaches toward a better biologically plausible neural network

Although it may appear infeasible and impractical, building artificial intelligence (AI) using a bottom-up approach based on the understanding of neuroscience is straightforward. The lack of a generalized governing principle for biological neural networks (BNNs) forces us to address this problem by converting piecemeal information on the diverse features of neurons, synapses, and neural circuits into AI. In this review, we described recent attempts to build a biologically plausible neural network by following neuroscientifically similar strategies of neural network optimization or by implanting the outcome of the optimization, such as the properties of single computational units and the characteristics of the network architecture. In addition, we proposed a formalism of the relationship between the set of objectives that neural networks attempt to achieve, and neural network classes categorized by how closely their architectural features resemble those of BNN. This formalism is expected to define the potential roles of top-down and bottom-up approaches for building a biologically plausible neural network and offer a map helping the navigation of the gap between neuroscience and AI engineering.

The neuroconnectionist research programme

Artificial neural networks (ANNs) inspired by biology are beginning to be widely used to model behavioural and neural data, an approach we call 'neuroconnectionism'. ANNs have been not only lauded as the current best models of information processing in the brain but also criticized for failing to account for basic cognitive functions. In this Perspective article, we propose that arguing about the successes and failures of a restricted set of current ANNs is the wrong approach to assess the promise of neuroconnectionism for brain science. Instead, we take inspiration from the philosophy of science, and in particular from Lakatos, who showed that the core of a scientific research programme is often not directly falsifiable but should be assessed by its capacity to generate novel insights. Following this view, we present neuroconnectionism as a general research programme centred around ANNs as a computational language for expressing falsifiable theories about brain computation. We describe the core of the programme, the underlying computational framework and its tools for testing specific neuroscientific hypotheses and deriving novel understanding. Taking a longitudinal view, we review past and present neuroconnectionist projects and their responses to challenges and argue that the research programme is highly progressive, generating new and otherwise unreachable insights into the workings of the brain.

De novo motor learning creates structure in neural activity space that shapes adaptation

Animals can quickly adapt learned movements in response to external perturbations. Motor adaptation is likely influenced by an animal's existing movement repertoire, but the nature of this influence is unclear. Long-term learning causes lasting changes in neural connectivity which determine the activity patterns that can be produced. Here, we sought to understand how a neural population's activity repertoire, acquired through long-term learning, affects short-term adaptation by modeling motor cortical neural population dynamics during de novo learning and subsequent adaptation using recurrent neural networks. We trained these networks on different motor repertoires comprising varying numbers of movements. Networks with multiple movements had more constrained and robust dynamics, which were associated with more defined neural 'structure'-organization created by the neural population activity patterns corresponding to each movement. This structure facilitated adaptation, but only when small changes in motor output were required, and when the structure of the network inputs, the neural activity space, and the perturbation were congruent. These results highlight trade-offs in skill acquisition and demonstrate how prior experience and external cues during learning can shape the geometrical properties of neural population activity as well as subsequent adaptation.

Linking cortex and contraction-Integrating models along the corticomuscular pathway

Computational models of the neuromusculoskeletal system provide a deterministic approach to investigate input-output relationships in the human motor system. Neuromusculoskeletal models are typically used to estimate muscle activations and forces that are consistent with observed motion under healthy and pathological conditions. However, many movement pathologies originate in the brain, including stroke, cerebral palsy, and Parkinson's disease, while most neuromusculoskeletal models deal exclusively with the peripheral nervous system and do not incorporate models of the motor cortex, cerebellum, or spinal cord. An integrated understanding of motor control is necessary to reveal underlying neural-input and motor-output relationships. To facilitate the development of integrated corticomuscular motor pathway models, we provide an overview of the neuromusculoskeletal modelling landscape with a focus on integrating computational models of the motor cortex, spinal cord circuitry, α-motoneurons  and skeletal muscle in regard to their role in generating voluntary muscle contraction. Further, we highlight the challenges and opportunities associated with an integrated corticomuscular pathway model, such as challenges in defining neuron connectivities, modelling standardisation, and opportunities in applying models to study emergent behaviour. Integrated corticomuscular pathway models have applications in brain-machine-interaction, education, and our understanding of neurological disease.

Dynamical flexible inference of nonlinear latent structures in neural population activity

Inferring complex spatiotemporal dynamics in neural population activity is critical for investigating neural mechanisms and developing neurotechnology. These activity patterns are noisy observations of lower-dimensional latent factors and their nonlinear dynamical structure. A major unaddressed challenge is to model this nonlinear structure, but in a manner that allows for flexible inference, whether causally, non-causally, or in the presence of missing neural observations. We address this challenge by developing DFINE, a new neural network that separates the model into dynamic and manifold latent factors, such that the dynamics can be modeled in tractable form. We show that DFINE achieves flexible nonlinear inference across diverse behaviors and brain regions. Further, despite enabling flexible inference unlike prior neural network models of population activity, DFINE also better predicts the behavior and neural activity, and better captures the latent neural manifold structure. DFINE can both enhance future neurotechnology and facilitate investigations across diverse domains of neuroscience.

Information-processing dynamics in neural networks of macaque cerebral cortex reflect cognitive state and behavior

One of the essential functions of biological neural networks is the processing of information. This includes everything from processing sensory information to perceive the environment, up to processing motor information to interact with the environment. Due to methodological limitations, it has been historically unclear how information processing changes during different cognitive or behavioral states and to what extent information is processed within or between the network of neurons in different brain areas. In this study, we leverage recent advances in the calculation of information dynamics to explore neural-level processing within and between the frontoparietal areas AIP, F5, and M1 during a delayed grasping task performed by three macaque monkeys. While information processing was high within all areas during all cognitive and behavioral states of the task, interareal processing varied widely: During visuomotor transformation, AIP and F5 formed a reciprocally connected processing unit, while no processing was present between areas during the memory period. Movement execution was processed globally across all areas with predominance of processing in the feedback direction. Furthermore, the fine-scale network structure reconfigured at the neuron level in response to different grasping conditions, despite no differences in the overall amount of information present. These results suggest that areas dynamically form higher-order processing units according to the cognitive or behavioral demand and that the information-processing network is hierarchically organized at the neuron level, with the coarse network structure determining the behavioral state and finer changes reflecting different conditions.

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.

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.

Buice M. MouseNet: A biologically constrained convolutional neural network model for the mouse visual cortex

Convolutional neural networks trained on object recognition derive inspiration from the neural architecture of the visual system in mammals, and have been used as models of the feedforward computation performed in the primate ventral stream. In contrast to the deep hierarchical organization of primates, the visual system of the mouse has a shallower arrangement. Since mice and primates are both capable of visually guided behavior, this raises questions about the role of architecture in neural computation. In this work, we introduce a novel framework for building a biologically constrained convolutional neural network model of the mouse visual cortex. The architecture and structural parameters of the network are derived from experimental measurements, specifically the 100-micrometer resolution interareal connectome, the estimates of numbers of neurons in each area and cortical layer, and the statistics of connections between cortical layers. This network is constructed to support detailed task-optimized models of mouse visual cortex, with neural populations that can be compared to specific corresponding populations in the mouse brain. Using a well-studied image classification task as our working example, we demonstrate the computational capability of this mouse-sized network. Given its relatively small size, MouseNet achieves roughly 2/3rds the performance level on ImageNet as VGG16. In combination with the large scale Allen Brain Observatory Visual Coding dataset, we use representational similarity analysis to quantify the extent to which MouseNet recapitulates the neural representation in mouse visual cortex. Importantly, we provide evidence that optimizing for task performance does not improve similarity to the corresponding biological system beyond a certain point. We demonstrate that the distributions of some physiological quantities are closer to the observed distributions in the mouse brain after task training. We encourage the use of the MouseNet architecture by making the code freely available.

Modulation of neural co-firing to enhance network transmission and improve motor function after stroke

Stroke is a leading cause of disability. While neurotechnology has shown promise for improving upper limb recovery after stroke, efficacy in clinical trials has been variable. Our central thesis is that to improve clinical translation, we need to develop a common neurophysiological framework for understanding how neurotechnology alters network activity. Our perspective discusses principles for how motor networks, both healthy and those recovering from stroke, subserve reach-to-grasp movements. We focus on neural processing at the resolution of single movements, the timescale at which neurotechnologies are applied, and discuss how this activity might drive long-term plasticity. We propose that future studies should focus on cross-area communication and bridging our understanding of timescales ranging from single trials within a session to across multiple sessions. We hope that this perspective establishes a combined path forward for preclinical and clinical research with the goal of more robust clinical translation of neurotechnology.

Recurrent neural networks controlling musculoskeletal models predict motor cortex activity during novel limb movements

Goal-driven networks trained to perform a task analogous to that performed by biological neural populations are being increasingly utilized as insightful computational models of motor control. The resulting dynamics of the trained networks are then analyzed to uncover the neural strategies employed by the motor cortex to produce movements. However, these networks do not take into account the role of sensory feedback in producing movement, nor do they consider the complex biophysical underpinnings of the underlying musculoskeletal system. Moreover, these models can not be used in context of predictive neuromechanical simulations for hypothesis generation and prediction of neural strategies during novel movements. In this research, we adapt state-of-the-art deep reinforcement learning (DRL) algorithms to train a controller to drive a developed anatomically accurate monkey arm model to track experimentally recorded kinematics. We validate that the trained controller mimics biologically observed neural strategies to produce movement. The trained controller generalizes well to unobserved conditions as well as to perturbation analyses. The recorded firing rates of motor cortex neurons can be predicted from the controller activity with high accuracy even on unseen conditions. Finally, we validate that the trained controller outperforms existing goal-driven and representational models of motor cortex in single neuron decoding accuracy, thus showing the utility of the complex underpinnings of anatomically accurate models in shaping motor cortex neural activity during limb movements. The learned controller can be used for hypothesis generation and prediction of neural strategies during novel movements and unobserved conditions.

Decoding sensorimotor information from superior parietal lobule of macaque via Convolutional Neural Networks

Despite the well-recognized role of the posterior parietal cortex (PPC) in processing sensory information to guide action, the differential encoding properties of this dynamic processing, as operated by different PPC brain areas, are scarcely known. Within the monkey's PPC, the superior parietal lobule hosts areas V6A, PEc, and PE included in the dorso-medial visual stream that is specialized in planning and guiding reaching movements. Here, a Convolutional Neural Network (CNN) approach is used to investigate how the information is processed in these areas. We trained two macaque monkeys to perform a delayed reaching task towards 9 positions (distributed on 3 different depth and direction levels) in the 3D peripersonal space. The activity of single cells was recorded from V6A, PEc, PE and fed to convolutional neural networks that were designed and trained to exploit the temporal structure of neuronal activation patterns, to decode the target positions reached by the monkey. Bayesian Optimization was used to define the main CNN hyper-parameters. In addition to discrete positions in space, we used the same network architecture to decode plausible reaching trajectories. We found that data from the most caudal V6A and PEc areas outperformed PE area in the spatial position decoding. In all areas, decoding accuracies started to increase at the time the target to reach was instructed to the monkey, and reached a plateau at movement onset. The results support a dynamic encoding of the different phases and properties of the reaching movement differentially distributed over a network of interconnected areas. This study highlights the usefulness of neurons' firing rate decoding via CNNs to improve our understanding of how sensorimotor information is encoded in PPC to perform reaching movements. The obtained results may have implications in the perspective of novel neuroprosthetic devices based on the decoding of these rich signals for faithfully carrying out patient's intentions.

Leaving flatland: Advances in 3D behavioral measurement

Animals move in three dimensions (3D). Thus, 3D measurement is necessary to report the true kinematics of animal movement. Existing 3D measurement techniques draw on specialized hardware, such as motion capture or depth cameras, as well as deep multi-view and monocular computer vision. Continued advances at the intersection of deep learning and computer vision will facilitate 3D tracking across more anatomical features, with less training data, in additional species, and within more natural, occlusive environments. 3D behavioral measurement enables unique applications in phenotyping, investigating the neural basis of behavior, and designing artificial agents capable of imitating animal behavior.

Prediction and action in cortical pain processing

Predicting that a stimulus is painful facilitates action to avoid harm. But how distinct are the neural processes underlying the prediction of upcoming painful events vis-à-vis those taking action to avoid them? Here, we investigated brain activity as a function of current and predicted painful or nonpainful thermal stimulation, as well as the ability of voluntary action to affect the duration of upcoming stimulation. Participants performed a task which involved the administration of a painful or nonpainful stimulus (S1), which predicted an immediately subsequent very painful or nonpainful stimulus (S2). Pressing a response button within a specified time window during S1 either reduced or did not reduce the duration of the upcoming stimulation. Predicted pain increased activation in several regions, including anterior cingulate cortex (ACC), midcingulate cortex (MCC), and insula; however, activation in ACC and MCC depended on whether a meaningful action was performed, with MCC activation showing a direct relationship with motor output. Insula's responses for predicted pain were also modulated by potential action consequences, albeit without a direct relationship with motor output. These findings suggest that cortical pain processing is not specifically tied to the sensory stimulus, but instead, depends on the consequences of that stimulus for sensorimotor control of behavior.

Classification of amyotrophic lateral sclerosis by brain volume, connectivity, and network dynamics

Emerging studies corroborate the importance of neuroimaging biomarkers and machine learning to improve diagnostic classification of amyotrophic lateral sclerosis (ALS). While most studies focus on structural data, recent studies assessing functional connectivity between brain regions by linear methods highlight the role of brain function. These studies have yet to be combined with brain structure and nonlinear functional features. We investigate the role of linear and nonlinear functional brain features, and the benefit of combining brain structure and function for ALS classification. ALS patients (N = 97) and healthy controls (N = 59) underwent structural and functional resting state magnetic resonance imaging. Based on key hubs of resting state networks, we defined three feature sets comprising brain volume, resting state functional connectivity (rsFC), as well as (nonlinear) resting state dynamics assessed via recurrent neural networks. Unimodal and multimodal random forest classifiers were built to classify ALS. Out-of-sample prediction errors were assessed via five-fold cross-validation. Unimodal classifiers achieved a classification accuracy of 56.35-61.66%. Multimodal classifiers outperformed unimodal classifiers achieving accuracies of 62.85-66.82%. Evaluating the ranking of individual features' importance scores across all classifiers revealed that rsFC features were most dominant in classification. While univariate analyses revealed reduced rsFC in ALS patients, functional features more generally indicated deficits in information integration across resting state brain networks in ALS. The present work undermines that combining brain structure and function provides an additional benefit to diagnostic classification, as indicated by multimodal classifiers, while emphasizing the importance of capturing both linear and nonlinear functional brain properties to identify discriminative biomarkers of ALS.

Compartmentalized dynamics within a common multi-area mesoscale manifold represent a repertoire of human hand movements

The human hand is unique in the animal kingdom for unparalleled dexterity, ranging from complex prehension to fine finger individuation. How does the brain represent such a diverse repertoire of movements? We evaluated mesoscale neural dynamics across the human "grasp network," using electrocorticography and dimensionality reduction methods, for a repertoire of hand movements. Strikingly, we found that the grasp network represented both finger and grasping movements alike. Specifically, the manifold characterizing the multi-areal neural covariance structure was preserved during all movements across this distributed network. In contrast, latent neural dynamics within this manifold were surprisingly specific to movement type. Aligning latent activity to kinematics further uncovered distinct submanifolds despite similarities in synergistic coupling of joints between movements. We thus find that despite preserved neural covariance at the distributed network level, mesoscale dynamics are compartmentalized into movement-specific submanifolds; this mesoscale organization may allow flexible switching between a repertoire of hand movements.

Distributed yet compartmentalized neural dynamics of hand actions

In this issue of Neuron,Natraj et al. (2021) demonstrate that finger and hand grasping movements are represented in the human fronto-parietal grasp network in a compartmentalized fashion. The movements are encoded in a distributed network that is preserved across various hand actions. The neural dynamics are specific to particular hand movements, leading to movement-specific submanifolds in the network.

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.

Multivariate Analysis of Electrophysiological Signals Reveals the Time Course of Precision Grasps Programs: Evidence for Nonhierarchical Evolution of Grasp Control

Current understanding of the neural processes underlying human grasping suggests that grasp computations involve gradients of higher to lower level representations and, relatedly, visual to motor processes. However, it is unclear whether these processes evolve in a strictly canonical manner from higher to intermediate and to lower levels given that this knowledge importantly relies on functional imaging, which lacks temporal resolution. To examine grasping in fine temporal detail here we used multivariate EEG analysis. We asked participants to grasp objects while controlling the time at which crucial elements of grasp programs were specified. We first specified the orientation with which participants should grasp objects, and only after a delay we instructed participants about which effector to use to grasp, either the right or the left hand. We also asked participants to grasp with both hands because bimanual and left-hand grasping share intermediate-level grasp representations. We observed that grasp programs evolved in a canonical manner from visual representations, which were independent of effectors to motor representations that distinguished between effectors. However, we found that intermediate representations of effectors that partially distinguished between effectors arose after representations that distinguished among all effector types. Our results show that grasp computations do not proceed in a strictly hierarchically canonical fashion, highlighting the importance of the fine temporal resolution of EEG for a comprehensive understanding of human grasp control.SIGNIFICANCE STATEMENT A long-standing assumption of the grasp computations is that grasp representations progress from higher to lower level control in a regular, or canonical, fashion. Here, we combined EEG and multivariate pattern analysis to characterize the temporal dynamics of grasp representations while participants viewed objects and were subsequently cued to execute an unimanual or bimanual grasp. Interrogation of the temporal dynamics revealed that lower level effector representations emerged before intermediate levels of grasp representations, thereby suggesting a partially noncanonical progression from higher to lower and then to intermediate level grasp control.

Completing the puzzle: Why studies in non-human primates are needed to better understand the effects of non-invasive brain stimulation

Brain stimulation is a core method in neuroscience. Numerous non-invasive brain stimulation (NIBS) techniques are currently in use in basic and clinical research, and recent advances promise the ability to non-invasively access deep brain structures. While encouraging, there is a surprising gap in our understanding of precisely how NIBS perturbs neural activity throughout an interconnected network, and how such perturbed neural activity ultimately links to behaviour. In this review, we will consider why non-human primate (NHP) models of NIBS are ideally situated to address this gap in knowledge, and why the oculomotor network that moves our line of sight offers a particularly valuable platform in which to empirically test hypothesis regarding NIBS-induced changes in brain and behaviour. NHP models of NIBS will enable investigation of the complex, dynamic effects of brain stimulation across multiple hierarchically interconnected brain areas, networks, and effectors. By establishing such links between brain and behavioural output, work in NHPs can help optimize experimental and therapeutic approaches, improve NIBS efficacy, and reduce side-effects of NIBS.

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

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.

From Observed Action Identity to Social Affordances

Others' observed actions cause continuously changing retinal images, making it challenging to build neural representations of action identity. The monkey anterior intraparietal area (AIP) and its putative human homologue (phAIP) host neurons selective for observed manipulative actions (OMAs). The neuronal activity of both AIP and phAIP allows a stable readout of OMA identity across visual formats, but human neurons exhibit greater invariance and generalize from observed actions to action verbs. These properties stem from the convergence in AIP of superior temporal signals concerning: (i) observed body movements; and (ii) the changes in the body-object relationship. We propose that evolutionarily preserved mechanisms underlie the specification of observed-actions identity and the selection of motor responses afforded by them, thereby promoting social behavior.

A Network Perspective on Sensorimotor Learning

What happens in the brain when we learn? Ever since the foundational work of Cajal, the field has made numerous discoveries as to how experience could change the structure and function of individual synapses. However, more recent advances have highlighted the need for understanding learning in terms of complex interactions between populations of neurons and synapses. How should one think about learning at such a macroscopic level? Here, we develop a conceptual framework to bridge the gap between the different scales at which learning operates, from synapses to neurons to behavior. Using this framework, we explore the principles that guide sensorimotor learning across these scales, and set the stage for future experimental and theoretical work in the field.