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.

Preserved neural dynamics across animals performing similar behaviour.

Animals of the same species exhibit similar behaviours that are advantageously adapted to their body and environment. These behaviours are shaped at the species level by selection pressures over evolutionary timescales. Yet, it remains unclear how these common behavioural adaptations emerge from the idiosyncratic neural circuitry of each individual. The overall organization of neural circuits is preserved across individuals1 because of their common evolutionarily specified developmental programme2-4. Such organization at the circuit level may constrain neural activity5-8, leading to low-dimensional latent dynamics across the neural population9-11. Accordingly, here we suggested that the shared circuit-level constraints within a species would lead to suitably preserved latent dynamics across individuals. We analysed recordings of neural populations from monkey and mouse motor cortex to demonstrate that neural dynamics in individuals from the same species are surprisingly preserved when they perform similar behaviour. Neural population dynamics were also preserved when animals consciously planned future movements without overt behaviour12 and enabled the decoding of planned and ongoing movement across different individuals. Furthermore, we found that preserved neural dynamics extend beyond cortical regions to the dorsal striatum, an evolutionarily older structure13,14. Finally, we used neural network models to demonstrate that behavioural similarity is necessary but not sufficient for this preservation. We posit that these emergent dynamics result from evolutionary constraints on brain development and thus reflect fundamental properties of the neural basis of behaviour.

A large-scale neural network training framework for generalized estimation of single-trial population dynamics.

Achieving state-of-the-art performance with deep neural population dynamics models requires extensive hyperparameter tuning for each dataset. AutoLFADS is a model-tuning framework that automatically produces high-performing autoencoding models on data from a variety of brain areas and tasks, without behavioral or task information. We demonstrate its broad applicability on several rhesus macaque datasets: from motor cortex during free-paced reaching, somatosensory cortex during reaching with perturbations, and dorsomedial frontal cortex during a cognitive timing task.

Sensory computations in the cuneate nucleus of macaques.

Tactile nerve fibers fall into a few classes that can be readily distinguished based on their spatiotemporal response properties. Because nerve fibers reflect local skin deformations, they individually carry ambiguous signals about object features. In contrast, cortical neurons exhibit heterogeneous response properties that reflect computations applied to convergent input from multiple classes of afferents, which confer to them a selectivity for behaviorally relevant features of objects. The conventional view is that these complex response properties arise within the cortex itself, implying that sensory signals are not processed to any significant extent in the two intervening structures-the cuneate nucleus (CN) and the thalamus. To test this hypothesis, we recorded the responses evoked in the CN to a battery of stimuli that have been extensively used to characterize tactile coding in both the periphery and cortex, including skin indentations, vibrations, random dot patterns, and scanned edges. We found that CN responses are more similar to their cortical counterparts than they are to their inputs: CN neurons receive input from multiple classes of nerve fibers, they have spatially complex receptive fields, and they exhibit selectivity for object features. Contrary to consensus, then, the CN plays a key role in processing tactile information.

Targeted Modulation of Human Brain Interregional Effective Connectivity With Spike-Timing Dependent Plasticity.

OBJECTIVE: The ability to selectively up- or downregulate interregional brain connectivity would be useful for research and clinical purposes. Toward this aim, cortico-cortical paired associative stimulation (ccPAS) protocols have been developed in which two areas are repeatedly stimulated with a millisecond-level asynchrony. However, ccPAS results in humans using bifocal transcranial magnetic stimulation (TMS) have been variable, and the mechanisms remain unproven. In this study, our goal was to test whether ccPAS mechanism is spike-timing-dependent plasticity (STDP). MATERIALS AND METHODS: Eleven healthy participants received ccPAS to the left primary motor cortex (M1) → right M1 with three different asynchronies (5 milliseconds shorter, equal to, or 5 milliseconds longer than the 9-millisecond transcallosal conduction delay) in separate sessions. To observe the neurophysiological effects, single-pulse TMS was delivered to the left M1 before and after ccPAS while cortico-cortical evoked responses were extracted from the contralateral M1 using source-resolved electroencephalography. RESULTS: Consistent with STDP mechanisms, the effects on synaptic strengths flipped depending on the asynchrony. Further implicating STDP, control experiments suggested that the effects were unidirectional and selective to the targeted connection. CONCLUSION: The results support the idea that ccPAS induces STDP and may selectively up- or downregulate effective connectivity between targeted regions in the human brain.

Estimating the dimensionality of the manifold underlying multi-electrode neural recordings.

It is generally accepted that the number of neurons in a given brain area far exceeds the number of neurons needed to carry any specific function controlled by that area. For example, motor areas of the human brain contain tens of millions of neurons that control the activation of tens or at most hundreds of muscles. This massive redundancy implies the covariation of many neurons, which constrains the population activity to a low-dimensional manifold within the space of all possible patterns of neural activity. To gain a conceptual understanding of the complexity of the neural activity within a manifold, it is useful to estimate its dimensionality, which quantifies the number of degrees of freedom required to describe the observed population activity without significant information loss. While there are many algorithms for dimensionality estimation, we do not know which are well suited for analyzing neural activity. The objective of this study was to evaluate the efficacy of several representative algorithms for estimating the dimensionality of linearly and nonlinearly embedded data. We generated synthetic neural recordings with known intrinsic dimensionality and used them to test the algorithms' accuracy and robustness. We emulated some of the important challenges associated with experimental data by adding noise, altering the nature of the embedding of the low-dimensional manifold within the high-dimensional recordings, varying the dimensionality of the manifold, and limiting the amount of available data. We demonstrated that linear algorithms overestimate the dimensionality of nonlinear, noise-free data. In cases of high noise, most algorithms overestimated the dimensionality. We thus developed a denoising algorithm based on deep learning, the "Joint Autoencoder", which significantly improved subsequent dimensionality estimation. Critically, we found that all algorithms failed when the intrinsic dimensionality was high (above 20) or when the amount of data used for estimation was low. Based on the challenges we observed, we formulated a pipeline for estimating the dimensionality of experimental neural data.

Encoding of limb state by single neurons in the cuneate nucleus of awake monkeys.

The cuneate nucleus (CN) is among the first sites along the neuraxis where proprioceptive signals can be integrated, transformed, and modulated. The objective of the study was to characterize the proprioceptive representations in CN. To this end, we recorded from single CN neurons in three monkeys during active reaching and passive limb perturbation. We found that many neurons exhibited responses that were tuned approximately sinusoidally to limb movement direction, as has been found for other sensorimotor neurons. The distribution of their preferred directions (PDs) was highly nonuniform and resembled that of muscle spindles within individual muscles, suggesting that CN neurons typically receive inputs from only a single muscle. We also found that the responses of proprioceptive CN neurons tended to be modestly amplified during active reaching movements compared to passive limb perturbations, in contrast to cutaneous CN neurons whose responses were not systematically different in the active and passive conditions. Somatosensory signals thus seem to be subject to a "spotlighting" of relevant sensory information rather than uniform suppression as has been suggested previously.NEW & NOTEWORTHY The cuneate nucleus (CN) is the somatosensory gateway into the brain, and only recently has it been possible to record these signals from an awake animal. We recorded single CN neurons in monkeys. Proprioceptive CN neurons appear to receive input from very few muscles, and their sensitivity to movement changes reliably during reaching relative to passive arm perturbations. Sensitivity is generally increased, but not exclusively so, as though CN "spotlights" critical proprioceptive information during reaching.

Latent Factors and Dynamics in Motor Cortex and Their Application to Brain-Machine Interfaces.

In the 1960s, Evarts first recorded the activity of single neurons in motor cortex of behaving monkeys (Evarts, 1968). In the 50 years since, great effort has been devoted to understanding how single neuron activity relates to movement. Yet these single neurons exist within a vast network, the nature of which has been largely inaccessible. With advances in recording technologies, algorithms, and computational power, the ability to study these networks is increasing exponentially. Recent experimental results suggest that the dynamical properties of these networks are critical to movement planning and execution. Here we discuss this dynamical systems perspective and how it is reshaping our understanding of the motor cortices. Following an overview of key studies in motor cortex, we discuss techniques to uncover the "latent factors" underlying observed neural population activity. Finally, we discuss efforts to use these factors to improve the performance of brain-machine interfaces, promising to make these findings broadly relevant to neuroengineering as well as systems neuroscience.

A muscle-activity-dependent gain between motor cortex and EMG.

Whether one is delicately placing a contact lens on the surface of the eye or lifting a heavy weight from the floor, the motor system must produce a wide range of forces under different dynamical loads. How does the motor cortex, with neurons that have a limited activity range, function effectively under these widely varying conditions? In this study, we explored the interaction of activity in primary motor cortex (M1) and muscles (electromyograms, EMGs) of two male rhesus monkeys for wrist movements made during three tasks requiring different dynamical loads and forces. Despite traditionally providing adequate predictions in single tasks, in our experiments, a single linear model failed to account for the relation between M1 activity and EMG across conditions. However, a model with a gain parameter that increased with the target force remained accurate across forces and dynamical loads. Surprisingly, this model showed that a greater proportion of EMG changes were explained by the nonlinear gain than the linear mapping from M1. In addition to its theoretical implications, the strength of this nonlinearity has important implications for brain-computer interfaces (BCIs). If BCI decoders are to be used to control movement dynamics (including interaction forces) directly, they will need to be nonlinear and include training data from broad data sets to function effectively across tasks. Our study reinforces the need to investigate neural control of movement across a wide range of conditions to understand its basic characteristics as well as translational implications. NEW & NOTEWORTHY We explored the motor cortex-to-electromyogram (EMG) mapping across a wide range of forces and loading conditions, which we found to be highly nonlinear. A greater proportion of EMG was explained by a nonlinear gain than a linear mapping. This nonlinearity allows motor cortex to control the wide range of forces encountered in the real world. These results unify earlier observations and inform the next-generation brain-computer interfaces that will control movement dynamics and interaction forces.

Restoring sensorimotor function through intracortical interfaces: progress and looming challenges.

The loss of a limb or paralysis resulting from spinal cord injury has devastating consequences on quality of life. One approach to restoring lost sensory and motor abilities in amputees and patients with tetraplegia is to supply them with implants that provide a direct interface with the CNS. Such brain-machine interfaces might enable a patient to exert voluntary control over a prosthetic or robotic limb or over the electrically induced contractions of paralysed muscles. A parallel interface could convey sensory information about the consequences of these movements back to the patient. Recent developments in the algorithms that decode motor intention from neuronal activity and in approaches to convey sensory feedback by electrically stimulating neurons, using biomimetic and adaptation-based approaches, have shown the promise of invasive interfaces with sensorimotor cortices, although substantial challenges remain.

Linear-nonlinear-time-warp-poisson models of neural activity.

Prominent models of spike trains assume only one source of variability - stochastic (Poisson) spiking - when stimuli and behavior are fixed. However, spike trains may also reflect variability due to internal processes such as planning. For example, we can plan a movement at one point in time and execute it at some arbitrary later time. Neurons involved in planning may thus share an underlying time course that is not precisely locked to the actual movement. Here we combine the standard Linear-Nonlinear-Poisson (LNP) model with Dynamic Time Warping (DTW) to account for shared temporal variability. When applied to recordings from macaque premotor cortex, we find that time warping considerably improves predictions of neural activity. We suggest that such temporal variability is a widespread phenomenon in the brain which should be modeled.

Musculoskeletal geometry accounts for apparent extrinsic representation of paw position in dorsal spinocerebellar tract.

Proprioception, the sense of limb position and motion, arises from individual muscle receptors. An important question is how and where in the neuroaxis our high level "extrinsic" sense of limb movement originates. In the 1990s, a series of papers detailed the properties of neurons in the dorsal spinocerebellar tract (DSCT) of the cat. Despite their direct projections from sensory receptors, it appeared that half of these neurons had consistent, high-level tuning to paw position rather than to joint angles (or muscle lengths). These results suggested that many DSCT neurons compute paw position from lower level sensory information. We examined the contribution of musculoskeletal geometry to this apparent extrinsic representation by simulating a three-joint hindlimb with mono- and biarticular muscles, each providing a muscle spindlelike signal, modulated by the muscle length. We simulated neurons driven by randomly weighted combinations of these signals and moved the paw to different positions under two joint-covariance conditions similar to the original experiments. Our results paralleled those experiments in a number of respects: 1) Many neurons were tuned to paw position relative to the hip under both conditions. 2) The distribution of tuning was strongly bimodal, with most neurons driven by whole-leg flexion or extension. 3) The change in tuning between conditions clustered around zero (median absolute change ~20°). These results indicate that, at least for these constraint conditions, extrinsic-like representation can be achieved simply through musculoskeletal geometry and convergent muscle length inputs. Consequently, they suggest a reinterpretation of the earlier results may be required.NEW & NOTEWORTHY A classic experiment concluding that many dorsal spinocerebellar tract neurons encode paw position rather than joint angles has been cited by many studies as evidence for high-level computation occurring within a single synapse of the sensors. However, our study provides evidence that such a computation is not required to explain the results. Using simulation, we replicated many of the original results with purely random connectivity, suggesting that a reinterpretation of the classic experiment is needed.

Methodological considerations for a chronic neural interface with the cuneate nucleus of macaques.

While the response properties of neurons in the somatosensory nerves and anterior parietal cortex have been extensively studied, little is known about the encoding of tactile and proprioceptive information in the cuneate nucleus (CN) or external cuneate nucleus (ECN), the first recipients of upper limb somatosensory afferent signals. The major challenge in characterizing neural coding in CN/ECN has been to record from these tiny, difficult-to-access brain stem structures. Most previous investigations of CN response properties have been carried out in decerebrate or anesthetized animals, thereby eliminating the well-documented top-down signals from cortex, which likely exert a strong influence on CN responses. Seeking to fill this gap in our understanding of somatosensory processing, we describe an approach to chronically implanting arrays of electrodes in the upper limb representation in the brain stem in primates. First, we describe the topography of CN/ECN in rhesus macaques, including its somatotopic organization and the layout of its submodalities (touch and proprioception). Second, we describe the design of electrode arrays and the implantation strategy to obtain stable recordings. Third, we show sample responses of CN/ECN neurons in brain stem obtained from awake, behaving monkeys. With this method, we are in a position to characterize, for the first time, somatosensory representations in CN and ECN of primates.NEW & NOTEWORTHY In primates, the neural basis of touch and of our sense of limb posture and movements has been studied in the peripheral nerves and in somatosensory cortex, but coding in the cuneate and external cuneate nuclei, the first processing stage for these signals in the central nervous system, remains an enigma. We have developed a method to record from these nuclei, thereby paving the way to studying how sensory information from the limb is encoded there.

Altered tuning in primary motor cortex does not account for behavioral adaptation during force field learning.

Although primary motor cortex (M1) is intimately involved in the dynamics of limb movement, its inputs may be more closely related to higher-order aspects of movement and multi-modal sensory feedback. Motor learning is thought to result from the adaption of internal models that compute transformations between these representations. While the psychophysics of motor learning has been studied in many experiments, the particular role of M1 in the process remains the subject of debate. Studies of learning-related changes in the spatial tuning of M1 neurons have yielded conflicting results. To resolve the discrepancies, we recorded from M1 during curl field adaptation in a reaching task. Our results suggest that aside from the addition of the load itself, the relation of M1 to movement dynamics remains unchanged as monkeys adapt behaviorally. Accordingly, we implemented a musculoskeletal model to generate synthetic neural activity having a fixed dynamical relation to movement and showed that these simulated neurons reproduced the observed behavior of the recorded M1 neurons. The stable representation of movement dynamics in M1 suggests that behavioral changes are mediated through progressively altered recruitment of M1 neurons, while the output effect of those neurons remained largely unchanged.

Toward a Proprioceptive Neural Interface that Mimics Natural Cortical Activity.

The dramatic advances in efferent neural interfaces over the past decade are remarkable, with cortical signals used to allow paralyzed patients to control the movement of a prosthetic limb or even their own hand. However, this success has thrown into relief, the relative lack of progress in our ability to restore somatosensation to these same patients. Somatosensation, including proprioception, the sense of limb position and movement, plays a crucial role in even basic motor tasks like reaching and walking. Its loss results in crippling deficits. Historical work dating back decades and even centuries has demonstrated that modality-specific sensations can be elicited by activating the central nervous system electrically. Recent work has focused on the challenge of refining these sensations by stimulating the somatosensory cortex (S1) directly. Animals are able to detect particular patterns of stimulation and even associate those patterns with particular sensory cues. Most of this work has involved areas of the somatosensory cortex that mediate the sense of touch. Very little corresponding work has been done for proprioception. Here we describe the effort to develop afferent neural interfaces through spatiotemporally precise intracortical microstimulation (ICMS). We review what is known of the cortical representation of proprioception, and describe recent work in our lab that demonstrates for the first time, that sensations like those of natural proprioception may be evoked by ICMS in S1. These preliminary findings are an important first step to the development of an afferent cortical interface to restore proprioception.

Brain-controlled muscle stimulation for the restoration of motor function.

Loss of the ability to move, as a consequence of spinal cord injury or neuromuscular disorder, has devastating consequences for the paralyzed individual, and great economic consequences for society. Functional electrical stimulation (FES) offers one means to restore some mobility to these individuals, improving not only their autonomy, but potentially their general health and well-being as well. FES uses electrical stimulation to cause the paralyzed muscles to contract. Existing clinical systems require the stimulation to be preprogrammed, with the patient typically using residual voluntary movement of another body part to trigger and control the patterned stimulation. The rapid development of neural interfacing in the past decade offers the promise of dramatically improved control for these patients, potentially allowing continuous control of FES through signals recorded from motor cortex, as the patient attempts to control the paralyzed body part. While application of these 'brain-machine interfaces' (BMIs) has undergone dramatic development for control of computer cursors and even robotic limbs, their use as an interface for FES has been much more limited. In this review, we consider both FES and BMI technologies and discuss the prospect for combining the two to provide important new options for paralyzed individuals.

Less is more: selection from a small set of options improves BCI velocity control.

We designed the discrete direction selection (DDS) decoder for intracortical brain computer interface (iBCI) cursor control and showed that it outperformed currently used decoders in a human-operated real-time iBCI simulator and in monkey iBCI use. Unlike virtually all existing decoders that map between neural activity and continuous velocity commands, DDS uses neural activity to select among a small menu of preset cursor velocities. We compared closed-loop cursor control across four visits by each of 48 naïve, able-bodied human subjects using either DDS or one of three common continuous velocity decoders: direct regression with assist (an affine map from neural activity to cursor velocity), ReFIT, and the velocity Kalman Filter. DDS outperformed all three by a substantial margin. Subsequently, a monkey using an iBCI also had substantially better performance with DDS than with the Wiener filter decoder (direct regression decoder that includes time history). Discretizing the decoded velocity with DDS effectively traded high resolution velocity commands for less tortuous and lower noise trajectories, highlighting the potential benefits of simplifying online iBCI control.

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.

BRAND: a platform for closed-loop experiments with deep network models.

Objective.Artificial neural networks (ANNs) are state-of-the-art tools for modeling and decoding neural activity, but deploying them in closed-loop experiments with tight timing constraints is challenging due to their limited support in existing real-time frameworks. Researchers need a platform that fully supports high-level languages for running ANNs (e.g. Python and Julia) while maintaining support for languages that are critical for low-latency data acquisition and processing (e.g. C and C++).Approach.To address these needs, we introduce the Backend for Realtime Asynchronous Neural Decoding (BRAND). BRAND comprises Linux processes, termednodes, which communicate with each other in agraphvia streams of data. Its asynchronous design allows for acquisition, control, and analysis to be executed in parallel on streams of data that may operate at different timescales. BRAND uses Redis, an in-memory database, to send data between nodes, which enables fast inter-process communication and supports 54 different programming languages. Thus, developers can easily deploy existing ANN models in BRAND with minimal implementation changes.Main results.In our tests, BRAND achieved <600 microsecond latency between processes when sending large quantities of data (1024 channels of 30 kHz neural data in 1 ms chunks). BRAND runs a brain-computer interface with a recurrent neural network (RNN) decoder with less than 8 ms of latency from neural data input to decoder prediction. In a real-world demonstration of the system, participant T11 in the BrainGate2 clinical trial (ClinicalTrials.gov Identifier: NCT00912041) performed a standard cursor control task, in which 30 kHz signal processing, RNN decoding, task control, and graphics were all executed in BRAND. This system also supports real-time inference with complex latent variable models like Latent Factor Analysis via Dynamical Systems.Significance.By providing a framework that is fast, modular, and language-agnostic, BRAND lowers the barriers to integrating the latest tools in neuroscience and machine learning into closed-loop experiments.

Responses of somatosensory area 2 neurons to actively and passively generated limb movements.

Control of reaching movements requires an accurate estimate of the state of the limb, yet sensory signals are inherently noisy, because of both noise at the receptors themselves and the stochastic nature of the information representation by neural discharge. One way to derive an accurate representation from noisy sensor data is to combine it with the output of a forward model that considers both the previous state estimate and the noisy input. We recorded from primary somatosensory cortex (S1) in macaques (Macaca mulatta) during both active and passive movements to investigate how the proprioceptive representation of movement in S1 may be modified by the motor command (through efference copy). We found neurons in S1 that respond to one or both movement types covering a broad distribution from active movement only, to both, to passive movement only. Those neurons that responded to both active and passive movements responded with similar directional tuning. Confirming earlier results, some, but not all, neurons responded before the onset of volitional movements, possibly as a result of efference copy. Consequently, many of the features necessary to combine the forward model with proprioceptive feedback appear to be present in S1. These features would not be expected from combinations of afferent receptor responses alone.