搜索结果：Computational brain & behavior

The dorsal raphe nucleus (DRN) shapes behaviors including mood and motivation. The DRN contains molecularly distinct and topographically organized neurons that target specific forebrain regions. To understand how DRN neurons process sensory information, we investigated the spatiotemporal activity patterns of DRN neurons, DRN axons, and their forebrain targets in zebrafish. We found a remarkable topographic organization of ongoing activity and sensory-motor responses within the DRN. A subset of DRN neurons was driven by locomotion and sensory stimuli. Gad1-positive DRN neurons exhibited distinct activity during rest and sensory-motor stimulation. DRN axons in the forebrain showed topographically organized excitation and inhibition in response to sensory stimulation and locomotion. DRN axons covaried with forebrain neuronal activity. DRN ablation reduced the synchrony and sensory-motor responses of forebrain neurons and enhanced defensive behaviors. We revealed the functional diversity of DRN neurons and their role in transmitting sensory and locomotor signals via topographically organized forebrain projections.

Concurrent transcranial magnetic stimulation and functional magnetic resonance imaging (TMS-fMRI) provides a mechanism for assessing the acute effects of transcranial magnetic stimulation (TMS) on functional connectivity (FC), allowing a unique perspective of how TMS induces antidepressant effects over the course of treatment. The aim of this secondary analysis of clinical trial data was to interrogate the relevance of the triple network theory in low-frequency TMS to the right dorsolateral prefrontal cortex (DLPFC) (low-frequency repetitive TMS [LFR]) by assessing perturbations in salience (SN), control (CN), and default mode (DMN) networks during TMS-fMRI. A total of 38 subjects with treatment-resistant depression underwent one session of concurrent TMS-fMRI at 1 Hz to the right DLPFC (LFR), with resting-state scans acquired immediately before and after. Patients subsequently underwent a four-week treatment course using the same protocol. Whole-brain FC was computed, as well as within- and between- network FC for the SN, CN, and DMN for each scan. FC modulation scores were computed to capture changes between resting-state and TMS-fMRI and were used to test for relationships between acute changes in FC during a single repetitive TMS treatment and clinical outcomes after a course of treatment. Whole-brain FC decreased during the TMS-fMRI scan, as did within- and between-network FC for the SN, CN, and DMN. Resting-state scans acquired immediately before and after TMS-fMRI showed no differences in FC. After the four-week treatment course, eight subjects were classified as remitters (21%), eight subjects were responders but fell short of remission (21%), with the remaining 22 subjects (58%) showing nonresponse. FC modulation scores for the whole-brain were significantly associated with decreased depression scores at the end of treatment. Between-network FC modulation was initially correlated with clinical improvement, but these correlations did not persist when controlling for whole-brain FC modulation. When tested using predictive modeling, both whole-brain and network-level data significantly predicted treatment outcomes, with modulation involving SN predicting outcomes as effectively as whole-brain models. LFR acutely disrupts FC at a global, whole-brain level that encompasses the triple networks. Modulation of the SN-CN and SN-DMN connectivity hold predictive value for clinical improvement comparable with that of global, whole-brain connectivity. This widespread global disruption may be an important mechanism through which TMS exerts antidepressant effects. Limitations include the use of atlas-based network definitions, small sample, and generalizability limited to LFR protocols only.

Adaptive and adverse brain states are often assumed to lie on a shared molecular continuum, but this assumption has rarely been evaluated against explicit transcriptomic alternatives. This study aimed to compare two representations of cross-context brain transcriptomic organization: a transcriptome-wide global-axis model and a low-dimensional reciprocal model. We benchmarked these models across a curated cross-study brain cohort spanning exercise, alcohol-related adversity-like contexts, stress, aging, and neurodegeneration, using prespecified intervention-like and adversity-like directional contrast labels rather than assuming homogeneous biological states. We assessed the competing representations using signed-effect correlations, permutation analyses, non-linear fitting, and held-out reconstruction, and we then examined the resulting structure through region-specific human bulk evaluation and exploratory cellular, single-nucleus, spatial, and chromatin projection analyses. These downstream analyses were used to examine localization and biological interpretability and were not treated as independent evaluation of the module 1/module 2 (M1/M2) partition. The combined signed-effect statistics were interpreted as representation-level directional summaries rather than estimates of a homogeneous cross-study biological effect. The global-axis model received limited support: intervention-like and adversity-like signed-effect summaries were only weakly correlated, were not stronger than permutation null expectations, and were not improved by non-linear fitting. Within the selected reciprocal-gene space, a rank-1 latent profile reconstructed held-out genes more accurately than the hard M1/M2 partition, whereas the M1/M2 discretization provided a more interpretable but selection-conditioned directional summary. Human analyses yielded an asymmetric pattern: a significant M1 association was observed only in the hippocampal dataset, whereas M2, the reciprocal index, and the other examined brain regions showed no consistent corresponding effects; leave-one-stratum-out analyses indicated poor cross-stratum reproducibility of the exact gene-level partition. These findings motivate a low-dimensional reciprocal representation as an exploratory framework while emphasizing context dependence, cohort dependence, and heterogeneity.

In recent decades, neural energy-budget research has extensively quantified signaling costs-the "read" side of computation-including action potentials, synaptic transmission, and ionic homeostasis. Yet the metabolic cost of learning itself remains far less studied. The "write" side-encompassing synaptic plasticity (LTP, LTD, and homeostatic scaling), receptor trafficking, protein synthesis, spine remodeling, and long-term consolidation-is still poorly understood. We argue that these write costs, although smaller than signaling costs in their time-averaged share of the mature brain's energy budget, constitute a real peak-throughput bottleneck for any intelligent system that must learn continuously within fixed energy constraints. A durable synaptic update aggregates an order-of-magnitude estimate of ~ 10⁴-10⁶ ATP per synapse across multiple coincident processes (receptor trafficking, protein turnover, local translation, actin remodeling, Ca²⁺ handling, and concurrent ionic recovery), with the range reflecting both natural variability and a 1-2 order-of-magnitude uncertainty across independent estimates (Karbowski, 2019). This per-synapse multi-process sum is larger than the protein-synthesis component alone-a term that standard energy budgets typically fold into housekeeping rather than treat as a distinct plasticity cost. Dense simultaneous writes are therefore unlikely to be sustainable at scale, because active cortical and hippocampal tissue can operate close to its oxidative ceiling during high-demand states, suggesting that many mature cortical circuits operate with limited additional metabolic headroom during high-demand states, plausibly on the order of tens of percent rather than several-fold (Hyder et al., 2013; Yu et al., 2023; Watts et al., 2018). The brain manages this constraint across multiple regimes: awake activity is characterized by sparse population coding and asynchronous irregular firing that preserves metabolic headroom; durable plasticity also occurs during waking but is sparse, neuromodulator- and attention-gated, and supplemented by brief awake replay events; and sleep provides additional write windows-slow-wave sleep offers a low-cost offline window, while REM sleep provides a selective plasticity window in which restricted circuits sustain patterned activity supporting durable plasticity. We recast learning as the selective allocation of scarce writes-metabolic investments that reorganize circuit structure and can reduce future signaling costs at the population level. Gated plasticity, synaptic caching, and sleep-dependent consolidation are, under this view, write-management strategies. At the systems level, selective attention allows the brain to concentrate plasticity on the circuits engaged in a current task while other regions idle, further constraining writes in space as well as time. By contrast, backpropagation in modern machine learning tightly couples reads and writes, which increases update traffic, amplifies interference, and raises energy consumption. We argue that imposing biologically inspired write-cost constraints during training, rather than relying on post-hoc compression, offers a plausible direction for energy-efficient continual learning in machines. Our claim is not that write costs dominate the brain's average energy budget, but that durable plasticity imposes a locally clustered peak-throughput demand that must be scheduled within the limited metabolic headroom left by ongoing signaling.

Over the past decade, comparisons between deep neural networks (DNNs) and the human brain have become central to cognitive neuroscience. Early work focused on vision, driven by the success of convolutional neural networks in object recognition, before such comparisons later gained traction in language with the rise of large-scale language models. These comparisons have validated existing hypotheses and generated new ones, challenging views of information processing, connectivity, and computational goals. Despite progress, debates persist over the interpretability and validity of mapping DNNs to brains, underscoring the need for more refined models and methods. Looking ahead, integrating cross-modal insights from vision and language, together with improved modeling and experimental frameworks, promises to advance the mechanistic understanding of cognition.

Accurate perception of subtle visuo-motor errors is essential for perceptual and sensorimotor learning, and supports timely corrective actions in precision-based task. However, conventional perceptual training, typically based on response-accuracy feedback, is limited in improving sensitivity to small, subtle errors. While prior approaches have focused on modulating sensory regions to enhance perceptual learning, we propose an alternative approach that targets a cognitive neural marker: the error positivity (Pe), a component of the error-related potential (ErrP) originating in the anterior cingulate cortex, a key decision-making region. We hypothesize that the Pe, which reflects conscious awareness of errors, serves as a modifiable neural correlate of error perception. In a five-day longitudinal study, we show that providing real-time feedback on the presence or absence of ErrPs during perceptual training accelerates perceptual learning at 3 ∘ $3^\circ$ errors and enhances perceptual performance at 6 ∘ $6^\circ$ errors without accelerating the learning rate, relative to behavioral training alone. These behavioral gains were accompanied by increase in Pe amplitude. Together, these findings offer new neurophysiological insights into the mechanisms of error perception, and establish ErrP-based brain-computer interface interventions as a promising approach for fostering perceptual learning in domains where detecting subtle errors is critical.

Altered sensitivity to environmental stimuli is a core feature of autism spectrum disorder (ASD) that may drive vulnerability to stress-related and anxiety disorders. To investigate this relationship, we evaluated autonomic and behavioral responses to tactile, nociceptive, and social stressors in juvenile Wistar rats prenatally exposed to valproic acid (VPA), a well-established model of ASD. VPA-exposed and saline-treated control (CTL) rats were subjected to a behavioral paradigm measuring defecation, freezing, and ultrasonic vocalizations (USVs) in response to handling, electro-tactile stimulation, classical fear conditioning, and an emotional contagion task. Compared to CTLs, VPA rats exhibited sustained hyperdefecation during handling and a higher prevalence of defecation during electro-tactile stimulation, independent of freezing changes. During fear conditioning, VPA rats demonstrated a delayed onset but subsequent enhancement of freezing responses, alongside disrupted temporal coordination between freezing and defecation behaviors. 22-kHz USVs positively correlated with freezing across both groups. In the emotional contagion task, observing a distressed conspecific increased freezing prevalence and suppressed vocalization rates in both cohorts; however, these effects were prolonged in VPA rats, which showed persistent freezing and an earlier reduction in vocalization rate. These findings indicate that VPA-treated rats display heightened stress reactivity, habituation deficit and suggest disrupted coordination of fear responses, supporting the VPA model as a relevant tool for investigating the neurobiological basis of stress vulnerability and social dysfunction in ASD.

Attention-Deficit/Hyperactivity Disorder (ADHD) is one of the most prevalent neurodevelopmental disorders, affecting approximately 5-7% of children and adolescents worldwide. Clinical diagnosis currently relies on behavioral assessments that are susceptible to subjectivity and inter-rater variability. Resting-state functional magnetic resonance imaging (rs-fMRI) offers a promising avenue for objective ADHD identification; however, most existing approaches depend on derivative feature representations, such as functional connectivity (FC), fractional amplitude of low-frequency fluctuations (fALFF), or regional homogeneity (ReHo), which substantially compress the original blood-oxygen-level-dependent (BOLD) signal prior to model training. Furthermore, limited cross-site generalizability and insufficient voxel-level interpretability remain barriers to clinical translation. We propose VoxSTNet (Voxel-level Spatiotemporal Network), an explainable and telepathology-ready framework that operates directly on four-dimensional rs-fMRI BOLD volumes. A two-stage processing pipeline preserves the complete raw BOLD signal while reducing computational burden through moderate compression. Subject-wise z-score normalization mitigates scanner-specific intensity variations without introducing fold leakage. A time-distributed three-dimensional convolutional neural network (3D-CNN) coupled with a gated recurrent unit (GRU) captures spatiotemporal representations, while HiResCAM provides voxel-level interpretability. Experiments were conducted on the ADHD-200 dataset comprising 760 subjects (300 ADHD and 460 controls) from six acquisition sites. Performance was evaluated using Leave-One-Site-Out (LOSO) cross-validation as the primary assessment and five-fold cross-validation as a secondary analysis. Five-fold cross-validation achieved an accuracy of 98.7 &#xb1; 0.4%, sensitivity of 98.2%, specificity of 99.1%, and area under the receiver operating characteristic curve (AUC) of 99.4% (95% confidence interval [CI]: 97.9-99.5%). Under the more stringent LOSO protocol, the model achieved a mean accuracy of 78.4% (95% CI: 75.1-81.7%). A controlled data-selection analysis demonstrated that retaining raw voxel-level information improved performance relative to derivative-feature baselines. HiResCAM saliency maps consistently highlighted the right caudate nucleus across validation subjects (mean Dice coefficient = 0.61 &#xb1; 0.08; Wilcoxon p < 0.001). VoxSTNet demonstrates that direct voxel-level modeling of rs-fMRI can achieve strong within-cohort performance while maintaining competitive cross-site generalizability. The identified saliency patterns align with established ADHD-related neurobiological findings, supporting the model's interpretability. Future work will focus on harmonization and domain-generalization strategies to further improve cross-site deployment performance.

One of the many goals of neuroscience is to understand how the brain encodes and transforms sensory information into behavior. These animal behaviors can be studied at the level of multi-limb poses or through the focused analysis of individual body parts. Techniques for tracking animal pose, such as DeepLabCut and SLEAP, enable detailed studies of large-scale multi-limb behaviors but show reduced accuracy when used for single-keypoint tracking, where insufficient spatial context leads to increased drift and instability in tracking (Arent I, Schmidt FP, Botsch M et al. Marker-less motion capture of insect locomotion with deep neural networks pre-trained on synthetic videos. Frontiers in Behavioral Neuroscience. Vol. 15. 2021. Tang G, Han Y, Sun X, et al. Anti-drift pose tracker (ADPT), a transformer-based network for robust animal pose estimation cross-species. eLife. Vol. 13. 2025). More general techniques, such as Faster Region-based Convolutional Neural Network (Faster R-CNN) and You Only Look Once (YOLO), have also been used to track location-based behaviors such as center-of-mass position and velocity. However, behaviors localized to a single body structure, such as the pharygeal pumping (i.e., feeding) in the microscopic roundworm Caenorhabditis elegans (C. elegans), are particularly sensitive to noise from moving non-target body parts. This limitation cannot be resolved by simply adding more training data, as doing so often leads to overfitting rather than improved robustness, and instead requires additional processing beyond existing object tracking packages. To address these challenges, we present a fast, automated method that reliably measures pumping in freely moving C. elegans by combining a state-of-the-art object detector (Faster R-CNN) with a tunable noise filter in a technique we call PumpKin. To validate its performance, we demonstrate both its speed (average of 0.4 seconds/frame) and its robust estimation capabilities through application to eight different experimental conditions that encompass both satiety and genetically-driven changes to feeding. PumpKin accurately estimates average pumping rates under eight different experimental conditions, which are positively correlated with the estimates of two expert annotators. Furthermore, PumpKin provides reliable estimates of the instantaneous pumping rate dynamics, achieving an average overlap that exceeds the human-human agreement measured via leave-one-out analysis. Applying PumpKin to conditions differing in satiety revealed a shared basal pumping rate of 0.5 Hz across all worm groups recorded off food, regardless of genetic background or satiety state. Together, these findings highlight PumpKin's ability to accurately isolate and estimate the motion of a single body part during locomotion. Although we present results specific to C. elegans, we anticipate that PumpKin will generalize to behaviors localized to a single body structure in other systems.

A deeper understanding of brain function requires resolving the intricate networks formed by neurons at the microscopic scale. Imaging the connecting nerve fibers remains a significant challenge, particularly due to the difficulty of resolving crossing fibers using conventional optical imaging techniques. Computational Scattered Light Imaging (ComSLI) addresses this by using obliquely incident light to reconstruct the in-plane orientations of nerve fibers based on their scattering profiles, enabling the resolution of fiber crossings. One approach is to use an LED display as a light source, which allows for illuminating the sample by arbitrary patterns and measuring full scattering patterns as well as angular scattering profiles. However, when using an LED display instead of a high-intensity LED spot, ComSLI is limited by a low signal and acquisition times of several seconds per image. To overcome these limitations, this work introduces Hadamard basis sampling for the angular illumination patterns, allowing an increase in illumination intensity and a corresponding reduction in measurement time. Compared to standard sampling approaches, this method yields significantly sharper defined scattering peaks, resulting in enhanced angular resolution of the scattering profiles. The Hadamard-based illumination enhances the reconstruction of cortical fiber organization, overcoming a key limitation in challenging ComSLI applications and neuroscience.

The Adolescent Brain Cognitive Development (ABCD) Study has substantially advanced developmental neuroscience through its large scale and open-science framework. This review synthesizes the study's significant statistical and methodological contributions over its first ten years, organized around the pillars of population neuroscience, longitudinal modeling, and causal inference. We first examine how ABCD's population-based design has prompted a reconsideration of how effect sizes are interpreted, helping to establish new benchmarks for distinguishing stable, biologically relevant signals from trivial associations in large-N contexts. We detail the computational innovations required to process high-dimensional data at scale, specifically highlighting new analytic tools like the Fast and Efficient Mixed Effects Algorithm (FEMA) framework for mass-univariate modeling and advanced strategies for managing selective attrition and missing data in large-scale longitudinal cohorts. In the domain of longitudinal and multilevel modeling, we discuss the transition from traditional cross-lagged designs to sophisticated frameworks - such as random-intercept cross-lagged panel models, latent growth curves, and parallel process models - that disentangle within-person developmental trajectories from stable between-person traits. We further highlight the study's role in advancing causal inference in observational research through "G-methods", marginal structural models, and quasi-experimental family-based designs. Finally, we explore how ABCD serves as a critical bridge for cross-cohort generalizability and lifespan validation using datasets like the UK Biobank. By contributing to new standards for reproducibility and methodological rigor, the ABCD Study has helped move neuroscience toward a "big data" era, providing a comprehensive statistical foundation for understanding the complex interplay between biology and environment during the transition to adulthood.

The ventromedial prefrontal cortex (vmPFC) encodes subjective value across monetary and social outcomes, but whether it tracks individual differences in fairness preferences remains unclear. We applied computational modeling to an openly available fMRI dataset in which participants completed an Ultimatum Game with a computer, an ingroup peer, and an outgroup peer. Individual disadvantageous inequality aversion parameters (&#x3b1;) were estimated from behavioral data using a one-parameter Fehr-Schmidt model, and vmPFC activation was extracted trial-by-trial using a coordinate-based region of interest. Offer size predicted vmPFC activation across proposer types, consistent with a value-coding role. &#x3b1; moderated vmPFC responses for ingroup offers: participants with stronger inequality aversion showed a weaker vmPFC signal when receiving offers from an ingroup peer. A direct contrast of the ingroup and outgroup slopes did not reach significance, and the corresponding effect in a ventral striatum control region was null. At the subject level, &#x3b1; did not predict overall vmPFC engagement. These results indicate that vmPFC value signals are associated with individual fairness preferences within the ingroup context, with only weak evidence that this association differs across proposer types. The findings extend prior work on group-modulated fairness processing.

Artificial neural networks constitute simplified computational models of neural circuits that might help understand how the biological brain solves and represents complex tasks. Previous research revealed that recurrent neural networks (RNNs) with 48 hidden units show human-level performance in restless four-armed bandit tasks but differ from humans with respect to the task strategy employed. Here we systematically examined the impact of network capacity (no. of hidden units) on computational mechanisms and performance. Computational modeling was applied to investigate and compare network behavior between capacity levels as well as between RNNs and human learners. Using a task frequently employed in human cognitive neuroscience work as well as in animal systems neuroscience work, we show that high-capacity networks displayed increased directed exploration and attenuated random exploration relative to low-capacity networks. RNNs with 576 hidden units approached "human-like" exploration strategies, but the overall switch rate and the level of perseveration still deviated from human learners. In the context of the resource-rational framework, which posits a trade-off between reward and policy complexity, human learners may devote more resources to solving the task, albeit without performance benefits over RNNs. Taken together, this work reveals the importance of network capacity on exploration strategies during reinforcement learning and therefore contributes to the goal of building neural networks that behave "human-like" to possibly gain insights into computational mechanisms in human brains. The online version contains supplementary material available at 10.1007/s42113-025-00258-4.

The reliability of parameter estimation is crucial in using computational models for choice data in decision-making tasks, especially so since the cognitive meaningful parameters within these models often are leveraged for further analysis. Typically, model-fitting involves using the model log-likelihood as the loss function to quantify discrepancies between model predictions and observed data. However, outlier data in choice datasets can bias parameter estimation when using log-likelihood. Alternative loss functions that are less sensitive to outliers are available. In this study, we compared a total of 3 such outlier-insensitive loss functions with the log-likelihood function in terms of parameter recovery. We compared their performance in both a reinforcement learning model in a learning paradigm and a hyperbolic model in an intertemporal choice paradigm, in both systematically varying the presence of outliers (ranging from no outliers to 25% of the data being outliers). Our parameter recovery results show that even a small proportion of outlier data can substantially impair parameter identification when using the log-likelihood function, especially for the choice consistency/explore-exploit trade-off parameter. In contrast, outlier-insensitive loss functions markedly improve the recovery of computational model parameters. Moreover, our power analysis further suggests that even a small proportion of outlier trials (e.g., 5%) can potentially undermine the statistical power to detect condition differences, underscoring the importance of accounting for outliers when using cognitive models as measurement tools. Based on our results, we recommend using the outlier-insensitive loss functions for non-hierarchical model estimation as it performed well across both the learning and the intertemporal choice paradigms and under varying degrees of outlier presence.

The Brain-Computer Interface (BCI) is a system that enables communication between the brain and external devices by translating brain activity into commands. Electroencephalography (EEG) is a commonly used modality for measuring brain activity. However, its low signal-to-noise ratio (SNR) and electrode reference problems lead to poor spatial resolution. As a result, EEG signals are often contaminated with physiological artifacts such as muscle movements. Therefore, this study used novel tripolar concentric ring electrodes (TCREs) to record brain signals related to overt and covert speech. Brain signals associated with overt and covert speech were recorded using TCRE and disc electrodes. Classification algorithms, including K-Nearest Neighbors (KNN), Fully Connected Neural Networks (FCNN), and Convolutional Neural Networks (CNN), were used to classify the TCRE and conventional EEG signals. The data were collected from 16 healthy participants, consisting of 10 males and 6 females. The experimental results demonstrate that TCREs provide superior performance compared to conventional disc electrodes. In addition, the 0.5-1.2s interval, corresponding to the peak stimulus window, exhibits a maximum power of 250&#x3bc;V. The average accuracy achieved during this peak epoch was 86.25%, whereas the remaining epoch shows an accuracy of 83.5% using TCREs.

Cortical folds encode the architecture of human cognition, yet the mechanisms that transform the smooth fetal cortex into its convoluted geometry remain elusive. Biophysical modeling enables mechanistic insight into cortical morphogenesis, but existing models often lack anatomical realism and fail to capture key hallmarks and morphometrics of dynamic cortical folding in the developing human brain. Here, we introduce a whole-brain developmental framework that integrates region-specific, data-driven growth laws with anatomically realistic cortical geometry to enable biologically interpretable modeling of cortical morphogenesis during gestation. Growth fields derived from large-scale prenatal magnetic resonance imaging data capture spatiotemporal variations in cortical expansion and thickness across parcellated regions. Incorporating heterogeneous growth yields folding patterns that match key anatomical landmarks and quantitative morphometrics from human imaging. Systematic perturbations of geometry and growth attributes delineate control parameters that produce realistic morphological variability and replicate clinically atypical brain phenotypes consistent with lissencephaly, pachygyria, and polymicrogyria. This framework provides a quantitative foundation for elucidating the mechanisms of typical and atypical fetal brain development.

Musical performances, particularly in drumming, are characterized not only by their structured rhythmic patterns but also by the subtle variations in timing and amplitude series that create expressive complexity. This study proposes a neural-inspired computational model to investigate how the brain might learn and internalize such complex rhythms. Inspired by the established roles of the cerebellum and basal ganglia in production of rhythms and timings, we utilize an oscillation-driven reservoir computer, a recurrent neural network model for temporal learning, to simulate the generation of human-like expressive drumming performances. First, the model was trained to replicate Jeff Porcaro's distinctive hi-hat patterns. Analyses revealed that the outputs of the model incorporating high-frequency oscillators ([50, 100] Hz), closely matched the original drumming, reproducing its characteristic fluctuations and patterns in inter-beat timings (microtiming) and amplitudes. Next, the model was trained to generate multidimensional drum kit performances for various genres (funk, jazz, samba, and rock). The model's outputs exhibited timing deviation and audio features characteristic of the original performances. Our findings demonstrate that oscillation-driven reservoir computing can replicate the rhythmic complexity of professional drumming, suggesting it as a potential computational principle for motor timing and rhythm generation. This approach provides a powerful framework for understanding how the brain generates and processes intricate rhythmic patterns. The online version contains supplementary material available at 10.1007/s11571-026-10512-5.

Cognitive and computational modeling has been used as a method to understand the processes underlying behavior in humans and other animals. A common approach in this field involves the use of theoretically constructed cognitive models, such as reinforcement learning models. However, human and animal decision-making often deviates from the predictions of these theoretical models. To capture characteristics that these cognitive models fail to account for, recurrent neural networks (RNNs) have been increasingly used to model choice behavior involving reinforcement learning. RNNs can capture how choice probabilities change depending on past experience. In this work, we demonstrate that RNNs can improve future choice predictions by capturing individual differences on the basis of past behavior, even when a single model is fit across the entire population. We refer to this capacity as the individual difference tracking (IDT) property. While the IDT property might be useful for prediction, it may introduce excessive flexibility when RNNs are used as benchmarks for predictive accuracy. We investigate the nature of the IDT property through simulation studies and examine how it affects the interpretation of predictive accuracy when RNNs are used as benchmarks for cognitive models. We also present examples using real-world data. Through these analyses, we discuss practical considerations and limitations in using RNNs as benchmarks for cognitive models. The online version contains supplementary material available at 10.1007/s42113-025-00254-8.

Millions of people worldwide are living with movement and sensory impairments owing to spinal cord injury, stroke and other neurological conditions. Here we report a double neural bypass (DNB), a hybrid neuroprosthetic system designed to restore both immediate and lasting gains in movement and sensation after a severe, complete spinal cord injury. The DNB links an intracortical brain-computer interface with targeted and patterned neuromodulation of the spinal cord and cortex. This allows brain signals associated with movement intention to directly control the movement of the user's own hand in real time while also promoting long-term sensorimotor recovery-even after the system is turned off. The DNB system uses recurrent artificial neural networks and reinforcement learning for fine grasp control, together with patterned spinal cord stimulation and activity-informed intracortical microstimulation ('cortical mirroring') to promote neuroplasticity and durable recovery of function. In a participant with chronic C4 sensory/C5 motor complete tetraplegia, this hybrid approach enabled recovery of functional abilities including self-feeding and manipulation of delicate objects, while also producing significant and persistent improvements in elbow flexion and wrist tactile sensation. These findings demonstrate the potential of combining a sensorimotor neuroprosthesis with targeted brain and spinal neuromodulation to restore clinically relevant function in severe paralysis.