Neural Subspace Research Articles

Neural dynamics of reversal learning in the prefrontal cortex and recurrent neural networks.

In probabilistic reversal learning, the choice option yielding reward at higher probability switches at a random trial. To perform optimally in this task, one has to accumulate evidence across trials to infer the probability that a reversal has occurred. In this study, we investigated how this reversal probability is represented in cortical neurons by analyzing the neural activity in prefrontal cortex of monkeys and recurrent neural networks trained on the task. We found that neural trajectories encoding reversal probability had substantial dynamics associated with intervening behaviors necessary to perform the task. Furthermore, the neural trajectories were translated systematically in response to whether outcomes were rewarded, and their position in the neural subspace captured information about reward outcomes. These findings suggested that separable dynamic trajectories, instead of fixed points on a line attractor, provided a better description of neural representation of reversal probability. Near the behavioral reversal, in particular, the trajectories shifted monotonically across trials with stable ordering, representing varying estimates of reversal probability around the reversal point. Perturbing the neural trajectory of trained networks biased when the reversal trial occurred, showing the role of reversal probability activity in decision-making. In sum, our study shows that cortical neurons encode reversal probability in a family of dynamic neural trajectories that accommodate flexible behavior while maintaining separability to represent distinct probabilistic values.

Flexible control of sequence working memory in the macaque frontal cortex

To memorize a sequence, one must serially bind each item to its rank order. How the brain controls a given input to bind its associated order in sequence working memory (SWM) remains unexplored. Here, we investigated the neural representations underlying SWM control using electrophysiological recordings in the frontal cortex of macaque monkeys performing forward and backward SWM tasks. Separate and generalizable low-dimensional subspaces for sensory and memory information were found within the same frontal circuitry, and SWM control was reflected in these neural subspaces' organized dynamics. Each item at each rank was sequentially entered into a common sensory subspace and, depending on forward or backward task requirement, flexibly and timely sent into rank-selective SWM subspaces. Neural activity in these SWM subspaces faithfully predicted the recalled item and order information in single error trials. Thus, compositional neural population codes with well-orchestrated dynamics in frontal cortex support the flexible control of SWM.

Initial and corrective submovement encoding differences within primary motor cortex during precision reaching.

Precision reaching often requires corrective submovements to obtain the desired goal. Most studies of reaching have focused on single initial movements, and implied the cortical encoding model was the same for all submovements. However, corrective submovements may show different encoding patterns from the initial submovement with distinct patterns of activation across the population. Two rhesus macaques performed a precision center-out-task with small targets. Neural activity from single units in the primary motor cortex and associated behavioral data were recorded to evaluate movement characteristics. Neural population data and individual neuronal firing rates identified with a peak finding algorithm to identify peaks in hand speed were examined for encoding differences between initial and corrective submovements. Individual neurons were fitted with a regression model that included the reach vector, position, and speed to predict firing rate. For both initial and corrective submovements, the largest effect remained movement direction. We observed a large subset changed their preferred direction greater than 45° between initial and corrective submovements. Neuronal depth of modulation also showed considerable variation when adjusted for movement speed. By using principal component analysis, neural trajectories of initial and corrective submovements progressed through different neural subspaces. These findings all suggest that different neural encoding patterns exist for initial and corrective submovements within the cortex. We hypothesize that this variation in how neurons change to encode small, corrective submovements might allow for a larger portion of the neural space being used to encode a greater range of movements with varying amplitudes and levels of precision.NEW & NOTEWORTHY Neuronal recordings matched with kinematic behavior were collected in a precision center-out task that often required corrective movements. We reveal large differences in preferred direction and depth of modulation between initial and corrective submovements across the neural population. We then present a model of the neural population describing how these shifts in tuning create different subspaces for signaling initial and corrective movements likely to improve motor precision.

Rule-based modulation of a sensorimotor transformation across cortical areas.

Flexible responses to sensory stimuli based on changing rules are critical for adapting to a dynamic environment. However, it remains unclear how the brain encodes and uses rule information to guide behavior. Here, we made single-unit recordings while head-fixed mice performed a cross-modal sensory selection task where they switched between two rules: licking in response to tactile stimuli while rejecting visual stimuli, or vice versa. Along a cortical sensorimotor processing stream including the primary (S1) and secondary (S2) somatosensory areas, and the medial (MM) and anterolateral (ALM) motor areas, single-neuron activity distinguished between the two rules both prior to and in response to the tactile stimulus. We hypothesized that neural populations in these areas would show rule-dependent preparatory states, which would shape the subsequent sensory processing and behavior. This hypothesis was supported for the motor cortical areas (MM and ALM) by findings that (1) the current task rule could be decoded from pre-stimulus population activity; (2) neural subspaces containing the population activity differed between the two rules; and (3) optogenetic disruption of pre-stimulus states impaired task performance. Our findings indicate that flexible action selection in response to sensory input can occur via configuration of preparatory states in the motor cortex.

Dimensionality reduction beyond neural subspaces with slice tensor component analysis

Recent work has argued that large-scale neural recordings are often well described by patterns of coactivation across neurons. Yet the view that neural variability is constrained to a fixed, low-dimensional subspace may overlook higher-dimensional structure, including stereotyped neural sequences or slowly evolving latent spaces. Here we argue that task-relevant variability in neural data can also cofluctuate over trials or time, defining distinct ‘covariability classes’ that may co-occur within the same dataset. To demix these covariability classes, we develop sliceTCA (slice tensor component analysis), a new unsupervised dimensionality reduction method for neural data tensors. In three example datasets, including motor cortical activity during a classic reaching task in primates and recent multiregion recordings in mice, we show that sliceTCA can capture more task-relevant structure in neural data using fewer components than traditional methods. Overall, our theoretical framework extends the classic view of low-dimensional population activity by incorporating additional classes of latent variables capturing higher-dimensional structure.

Neural Subspaces for Light Fields.

We introduce a framework for compactly representing light field content with the novel concept of neural subspaces. While the recently proposed neural light field representation achieves great compression results by encoding a light field into a single neural network, the unified design is not optimized for the composite structures exhibited in light fields. Moreover, encoding every part of the light field into one network is not ideal for applications that require rapid transmission and decoding. We recognize this problem's connection to subspace learning. We present a method that uses several small neural networks, specializing in learning the neural subspace for a particular light field segment. Moreover, we propose an adaptive weight sharing strategy among those small networks, improving parameter efficiency. In effect, this strategy enables a concerted way to track the similarity among nearby neural subspaces by leveraging the layered structure of neural networks. Furthermore, we develop a soft-classification technique to enhance the color prediction accuracy of neural representations. Our experimental results show that our method better reconstructs the light field than previous methods on various light field scenes. We further demonstrate its successful deployment on encoding light fields with irregular viewpoint layout and dynamic scene content.

Distinct neural signatures underlying information maintenance and manipulation in working memory.

Previous working memory research has demonstrated robust stimulus representations during memory maintenance in both voltage and alpha-band activity in electroencephalography. However, the exact functions of these 2 neural signatures have remained controversial. Here we systematically investigated their respective contributions to memory manipulation. Human participants either maintained a previously seen spatial location, or manipulated the location following a mental rotation cue over a delay. Using multivariate decoding, we observed robust location representations in low-frequency voltage and alpha-band oscillatory activity with distinct spatiotemporal dynamics: location representations were most evident in posterior channels in alpha-band activity, but were most prominent in the more anterior, central channels in voltage signals. Moreover, the temporal emergence of manipulated representation in central voltage preceded that in posterior alpha-band activity, suggesting that voltage might carry stimulus-specific source signals originated internally from anterior cortex, whereas alpha-band activity might reflect feedback signals in posterior cortex received from higher-order cortex. Lastly, while location representations in both signals were coded in a low-dimensional neural subspace, location representation in central voltage was higher-dimensional and underwent a representational transformation that exclusively predicted memory behavior. Together, these results highlight the crucial role of central voltage in working memory, and support functional distinctions between voltage and alpha-band activity.

Event-specific and persistent representations for contextual states in orbitofrontal neurons

Flexible and context-dependent behaviors require animals, including humans, to identify their current contextual state for proper rules to apply, especially when information that defines these states is partially observable. Depending on behavioral needs, contextual states usually persist for prolonged periods and across other events, including sensory stimuli, actions, and rewards, highlighting prominent challenges of holding a reliable state representation. The orbitofrontal cortex (OFC) is crucial in behaviors requiring the identification of the current context (e.g., reversal learning); however, how single units in the OFC accomplish this function has not been assessed. Do they maintain such information persistently, in separate populations from those responding phasically to events within a task, or is contextual information dynamic and embedded in these phasic responses? Here, we investigated this question by recording single units from OFC in rats performing a task that required them to identify the current contextual state related to estimated proximity to future reward with distracting olfactory cues. We found that while some OFC neurons encode contextual states, most change their selectivity upon the transition of task events. Nevertheless, despite dynamic activities in single neurons, the neural populations maintain persistent representations regarding current contextual states within particular neural subspaces.

Rapid Online Corrections for Proprioceptive and Visual Perturbations Recruit Similar Circuits in Primary Motor Cortex.

An important aspect of motor function is our ability to rapidly generate goal-directed corrections for disturbances to the limb or behavioral goal. The primary motor cortex (M1) is a key region involved in processing feedback for rapid motor corrections, yet we know little about how M1 circuits are recruited by different sources of sensory feedback to make rapid corrections. We trained two male monkeys (Macaca mulatta) to make goal-directed reaches and on random trials introduced different sensory errors by either jumping the visual location of the goal (goal jump), jumping the visual location of the hand (cursor jump), or applying a mechanical load to displace the hand (proprioceptive feedback). Sensory perturbations evoked a broad response in M1 with ∼73% of neurons (n = 257) responding to at least one of the sensory perturbations. Feedback responses were also similar as response ranges between the goal and cursor jumps were highly correlated (range of r = [0.91, 0.97]) as were the response ranges between the mechanical loads and the visual perturbations (range of r = [0.68, 0.86]). Lastly, we identified the neural subspace each perturbation response resided in and found a strong overlap between the two visual perturbations (range of overlap index, 0.73-0.89) and between the mechanical loads and visual perturbations (range of overlap index, 0.36-0.47) indicating each perturbation evoked similar structure of activity at the population level. Collectively, our results indicate rapid responses to errors from different sensory sources target similar overlapping circuits in M1.

Neural population dynamics of human working memory

The activity of neurons in macaque prefrontal cortex (PFC) persists during working memory (WM) delays, providing a mechanism for memory.1,2,3,4,5,6,7,8,9,10,11 Although theory,11,12 including formal network models,13,14 assumes that WM codes are stable over time, PFC neurons exhibit dynamics inconsistent with these assumptions.15,16,17,18,19 Recently, multivariate reanalyses revealed the coexistence of both stable and dynamic WM codes in macaque PFC.20,21,22,23 Human EEG studies also suggest that WM might contain dynamics.24,25 Nonetheless, how WM dynamics vary across the cortical hierarchy and which factors drive dynamics remain unknown. To elucidate WM dynamics in humans, we decoded WM content from fMRI responses across multiple cortical visual field maps.26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48 We found coexisting stable and dynamic neural representations of WM during a memory-guided saccade task. Geometric analyses of neural subspaces revealed that early visual cortex exhibited stronger dynamics than high-level visual and frontoparietal cortex. Leveraging models of population receptive fields, we visualized and made the neural dynamics interpretable. We found that during WM delays, V1 population initially encoded a narrowly tuned bump of activation centered on the peripheral memory target. Remarkably, this bump then spread inward toward foveal locations, forming a vector along the trajectory of the forthcoming memory-guided saccade. In other words, the neural code transformed into an abstraction of the stimulus more proximal to memory-guided behavior. Therefore, theories of WM must consider both sensory features and their task-relevant abstractions because changes in the format of memoranda naturally drive neural dynamics.

Large-scale electrophysiology and deep learning reveal distorted neural signal dynamics after hearing loss.

Listeners with hearing loss often struggle to understand speech in noise, even with a hearing aid. To better understand the auditory processing deficits that underlie this problem, we made large-scale brain recordings from gerbils, a common animal model for human hearing, while presenting a large database of speech and noise sounds. We first used manifold learning to identify the neural subspace in which speech is encoded and found that it is low-dimensional and that the dynamics within it are profoundly distorted by hearing loss. We then trained a deep neural network (DNN) to replicate the neural coding of speech with and without hearing loss and analyzed the underlying network dynamics. We found that hearing loss primarily impacts spectral processing, creating nonlinear distortions in cross-frequency interactions that result in a hypersensitivity to background noise that persists even after amplification with a hearing aid. Our results identify a new focus for efforts to design improved hearing aids and demonstrate the power of DNNs as a tool for the study of central brain structures.

Neural Subspace Learning for Surface Defect Detection

Surface defect inspection is a key technique in industrial product assessments. Compared with other visual applications, industrial defect inspection suffers from a small sample problem and a lack of labeled data. Therefore, conventional deep-learning methods depending on huge supervised samples cannot be directly generalized to this task. To deal with the lack of labeled data, unsupervised subspace learning provides more clues for the task of defect inspection. However, conventional subspace learning methods focus on studying the linear subspace structure. In order to explore the nonlinear manifold structure, a novel neural subspace learning algorithm is proposed by substituting linear operators with nonlinear neural networks. The low-rank property of the latent space is approximated by limiting the dimensions of the encoded feature, and the sparse coding property is simulated by quantized autoencoding. To overcome the small sample problem, a novel data augmentation strategy called thin-plate-spline deformation is proposed. Compared with the rigid transformation methods used in previous literature, our strategy could generate more reliable training samples. Experiments on real-world datasets demonstrate that our method achieves state-of-the-art performance compared with unsupervised methods. More importantly, the proposed method is competitive and has a better generalization capability compared with supervised methods based on deep learning techniques.

Emergence of Distinct Neural Subspaces in Motor Cortical Dynamics during Volitional Adjustments of Ongoing Locomotion.

The ability to modulate ongoing walking gait with precise, voluntary adjustments is what allows animals to navigate complex terrains. However, how the nervous system generates the signals to precisely control the limbs while simultaneously maintaining locomotion is poorly understood. One potential strategy is to distribute the neural activity related to these two functions into distinct cortical activity coactivation subspaces so that both may be conducted simultaneously without disruptive interference. To investigate this hypothesis, we recorded the activity of primary motor cortex in male nonhuman primates during obstacle avoidance on a treadmill. We found that the same neural population was active during both basic unobstructed locomotion and volitional obstacle avoidance movements. We identified the neural modes spanning the subspace of the low-dimensional dynamics in primary motor cortex and found a subspace that consistently maintains the same cyclic activity throughout obstacle stepping, despite large changes in the movement itself. All of the variance corresponding to this large change in movement during the obstacle avoidance was confined to its own distinct subspace. Furthermore, neural decoders built for ongoing locomotion did not generalize to decoding obstacle avoidance during locomotion. Our findings suggest that separate underlying subspaces emerge during complex locomotion that coordinates ongoing locomotor-related neural dynamics with volitional gait adjustments. These findings may have important implications for the development of brain-machine interfaces.SIGNIFICANCE STATEMENT Locomotion and precise, goal-directed movements are two distinct movement modalities with known differing requirements of motor cortical input. Previous studies have characterized the cortical activity during obstacle avoidance while walking in rodents and felines, but, to date, no such studies have been completed in primates. Additionally, in any animal model, it is unknown how these two movements are represented in primary motor cortex (M1) low-dimensional dynamics when both activities are performed at the same time, such as during obstacle avoidance. We developed a novel obstacle avoidance paradigm in freely moving nonhuman primates and discovered that the rhythmic locomotion-related dynamics and the voluntary, gait-adjustment movement separate into distinct subspaces in M1 cortical activity. Our analysis of decoding generalization may also have important implications for the development of brain-machine interfaces.

A Standardized Nonvisual Behavioral Event Is Broadcasted Homogeneously across Cortical Visual Areas without Modulating Visual Responses.

Multiple recent studies have shown that motor activity greatly impacts the activity of primary sensory areas like V1. Yet, the role of this motor related activity in sensory processing is still unclear. Here, we dissect how these behavior signals are broadcast to different layers and areas of the visual cortex. To do so, we leveraged a standardized and spontaneous behavioral fidget event in passively viewing mice. Importantly, this behavior event had no relevance to any ongoing task allowing us to compare its neuronal correlates with visually relevant behaviors (e.g., running). A large two-photon Ca2+ imaging database of neuronal responses uncovered four neural response types during fidgets that were consistent in their proportion and response patterns across all visual areas and layers of the visual cortex. Indeed, the layer and area identity could not be decoded above chance level based only on neuronal recordings. In contrast to running behavior, fidget evoked neural responses that were independent to visual processing. The broad availability of visually orthogonal standardized behavior signals could be a key component in how the cortex selects, learns and binds local sensory information with motor outputs. Contrary to behaviorally relevant motor outputs, irrelevant motor signals could project to separate local neural subspaces.

Invariant neural subspaces maintained by feedback modulation.

Sensory systems reliably process incoming stimuli in spite of changes in context. Most recent models accredit this context invariance to an extraction of increasingly complex sensory features in hierarchical feedforward networks. Here, we study how context-invariant representations can be established by feedback rather than feedforward processing. We show that feedforward neural networks modulated by feedback can dynamically generate invariant sensory representations. The required feedback can be implemented as a slow and spatially diffuse gain modulation. The invariance is not present on the level of individual neurons, but emerges only on the population level. Mechanistically, the feedback modulation dynamically reorients the manifold of neural activity and thereby maintains an invariant neural subspace in spite of contextual variations. Our results highlight the importance of population-level analyses for understanding the role of feedback in flexible sensory processing.

Priority coding in the visual system.

Although we are continuously bombarded with visual input, only a fraction of incoming visual events is perceived, remembered or acted on. The neural underpinnings of various forms of visual priority coding, including perceptual expertise, goal-directed attention, visual salience, image memorability and preferential looking, have been studied. Here, we synthesize information from these different examples to review recent developments in our understanding of visual priority coding and its neural correlates, with a focus on the role of behaviour to evaluate candidate correlates. We propose that the brain combines different types of priority into a unified priority signal while also retaining the ability to differentiate between them, and that this happens by leveraging partially overlapping low-dimensional neural subspaces for each type of priority that are shared with the downstream neural populations involved in decision-making. Finally, we describe the gulfs in understanding that have resulted from different research approaches, and we point towards future directions that will lead to fundamental insights about neural coding and how prioritization influences visually guided behaviours.

Cortical preparatory activity indexes learned motor memories.

The brain's remarkable ability to learn and execute various motor behaviours harnesses the capacity of neural populations to generate a variety of activity patterns. Here we explore systematic changes in preparatory activity in motor cortex that accompany motor learning. We trained rhesus monkeys to learn an arm-reaching task1 in a curl force field that elicited new muscle forces for some, but not all, movement directions2,3. We found that in a neural subspace predictive of hand forces, changes in preparatory activity tracked the learned behavioural modifications and reassociated4 existing activity patterns with updated movements. Along a neural population dimension orthogonal to the force-predictive subspace, we discovered that preparatory activity shifted uniformly for all movement directions, including those unaltered by learning. During a washout period when the curl field was removed, preparatory activity gradually reverted in the force-predictive subspace, but the uniform shift persisted. These persistent preparatory activity patterns may retain a motor memory of the learned field5,6 and support accelerated relearning of the same curl field. When a set of distinct curl fields was learned in sequence, we observed a corresponding set of field-specific uniform shifts which separated the associated motor memories in the neural state space7-9. The precise geometry of these uniform shifts in preparatory activity could serve to index motor memories, facilitating the acquisition, retention and retrieval of a broad motor repertoire.

Hybrid dedicated and distributed coding in PMd/M1 provides separation and interaction of bilateral arm signals.

Pronounced activity is observed in both hemispheres of the motor cortex during preparation and execution of unimanual movements. The organizational principles of bi-hemispheric signals and the functions they serve throughout motor planning remain unclear. Using an instructed-delay reaching task in monkeys, we identified two components in population responses spanning PMd and M1. A "dedicated" component, which segregated activity at the level of individual units, emerged in PMd during preparation. It was most prominent following movement when M1 became strongly engaged, and principally involved the contralateral hemisphere. In contrast to recent reports, these dedicated signals solely accounted for divergence of arm-specific neural subspaces. The other "distributed" component mixed signals for each arm within units, and the subspace containing it did not discriminate between arms at any stage. The statistics of the population response suggest two functional aspects of the cortical network: one that spans both hemispheres for supporting preparatory and ongoing processes, and another that is predominantly housed in the contralateral hemisphere and specifies unilateral output.

Low-Dimensional Dynamics of Brain Activity Associated with Manual Acupuncture in Healthy Subjects.

Acupuncture is one of the oldest traditional medical treatments in Asian countries. However, the scientific explanation regarding the therapeutic effect of acupuncture is still unknown. The much-discussed hypothesis it that acupuncture’s effects are mediated via autonomic neural networks; nevertheless, dynamic brain activity involved in the acupuncture response has still not been elicited. In this work, we hypothesized that there exists a lower-dimensional subspace of dynamic brain activity across subjects, underpinning the brain’s response to manual acupuncture stimulation. To this end, we employed a variational auto-encoder to probe the latent variables from multichannel EEG signals associated with acupuncture stimulation at the ST36 acupoint. The experimental results demonstrate that manual acupuncture stimuli can reduce the dimensionality of brain activity, which results from the enhancement of oscillatory activity in the delta and alpha frequency bands induced by acupuncture. Moreover, it was found that large-scale brain activity could be constrained within a low-dimensional neural subspace, which is spanned by the “acupuncture mode”. In each neural subspace, the steady dynamics of the brain in response to acupuncture stimuli converge to topologically similar elliptic-shaped attractors across different subjects. The attractor morphology is closely related to the frequency of the acupuncture stimulation. These results shed light on probing the large-scale brain response to manual acupuncture stimuli.

Dynamic coordination of the perirhinal cortical neurons supports coherent representations between task epochs

Cortical neurons show distinct firing patterns across multiple task epochs characterized by different computations. Recent studies suggest that such distinct patterns underlie dynamic population code achieving computational flexibility, whereas neurons in some cortical areas often show coherent firing patterns across epochs. To understand how coherent single-neuron code contributes to dynamic population code, we analyzed neural responses in the rat perirhinal cortex (PRC) during cue and reward epochs of a two-alternative forced-choice task. We found that the PRC neurons often encoded the opposite choice directions between those epochs. By using principal component analysis as a population-level analysis, we identified neural subspaces associated with each epoch, which reflected coordination across the neurons. The cue and reward epochs shared neural dimensions where the choice directions were consistently discriminated. Interestingly, those dimensions were supported by dynamically changing contributions of the individual neurons. These results demonstrated heterogeneity of coherent single-neuron representations in their contributions to population code.