Recurrent Neural Networks (RNNs) are widely used to model neural activity in Computational Neuroscience. Here, we explore the mathematical foundations of three fundamental procedures that can be implemented: temporal rescaling, discretization, and linearization. These techniques provide crucial tools for characterizing the behavior of RNNs, offering insights into their temporal dynamics, facilitating practical computational implementation, and allowing for linear approximations for analysis. We discuss the flexible order in which these procedures can be applied, emphasizing their importance in modeling and analyzing RNNs for neuroscience and formally prove that these three operations commute pairwise. We also explicitly describe the conditions under which these procedures can be considered interchangeable. Our findings directly inform the design of biologically plausible RNN models for simulating neural dynamics observed in decision-making circuits and motor control, where temporal scaling and stability are critical for matching experimental recordings. Furthermore, we show that this exact commutativity guarantees the structural preservation of the network's controllability, preventing the emergence of inaccessible state-spaces under numerical discretization or temporal rescaling.
IntroductionIntracortical microstimulation (ICMS) of the primary somatosensory cortex can evoke localized tactile percepts, yet the spatial factors that influence perceptual discrimination remain poorly defined. In prior work, we showed that discrimination accuracy between behaviorally evaluated ICMS-evoked percepts declines as stimulation sites converge across cortical depths and adjacent cortical columns. Those results suggest that overlap in neuronal recruitment may constrain perceptual differentiation.MethodsHere, we combine simulated data from a biophysically realistic computational model of the somatosensory cortex with previously collected behavioral data from rats to quantify how overlap in ICMS-evoked activation volume relates to discrimination performance. Within the model, ICMS patterns investigated behaviorally were simulated, and activation volumes were estimated by fitting a range of 50–100% capture ellipsoids to the spatial distribution of activated somata. Overlap in activation volumes between pairs of ICMS patterns was then quantified using the intersection-over-union (IoU) metric.ResultsAcross both single- and four-shank microelectrode array configurations, we found that discrimination accuracy decreased in an exponential decay-like relationship (R2 = 0.88) as model-derived activation volume overlap increased. Independent of depth vs. lateral separation between ICMS pattern pairs, minimal overlap (IoU < 1%) was associated with high discrimination accuracy (>70%; average of 85%), whereas IoU values exceeding 20% corresponded to near-chance performance.DiscussionThese results suggest that ICMS-evoked activation volume overlap between stimulation sites may provide mechanistic insight into the spatial limits of perceptual discrimination in ICMS applications. More broadly, these findings may help guide future investigations aimed at determining appropriate electrode spacing and stimulation strategies for sensory neuroprosthetic design.
Heart rate variability (HRV) represents a rich physiological signal that captures the computational dynamics of autonomic regulation. Emerging evidence suggests that atypical autonomic control contributes to the neurocognitive phenotype of neurodevelopmental disorders (NDD), including autism spectrum disorder (ASD) and attention-deficit/hyperactivity disorder (ADHD). However, the characterization of HRV alterations in these populations remains incomplete. This review was motivated by the broader research context of the EMPOWER project, funded through the Horizon Europe program, which investigates technology-supported approaches for children with NDD, including smartwatch-based physiological sensing and machine-learning analysis of heart rate, HRV, skin temperature, and photoplethysmography signals during cognitive and relaxation tasks. The present manuscript does not report EMPOWER empirical data. Instead, it systematically synthesizes previously published studies on HRV in pediatric ASD and ADHD. The review examines how HRV has been acquired, processed, modeled, and interpreted in pediatric NDD research, with particular attention to nonlinear signal properties, computational approaches, and interactions between autonomic regulation and cognitive processes. Following PRISMA guidelines, the review surveyed publications over the past decade and included 24 empirical studies from five major databases. ASD studies most often indicated altered vagally mediated HRV and autonomic reactivity, whereas ADHD studies more often suggested task-dependent changes linked to attentional control. Cross-study comparability was constrained by substantial heterogeneity in preprocessing pipelines, feature extraction procedures, and analytical frameworks. Despite these methodological inconsistencies, current evidence supports HRV as a promising physiological signal marker of autonomic dysregulation in pediatric NDD. Nevertheless, the field requires standardized signal processing protocols and reproducible modeling approaches to elucidate the mechanistic and predictive relevance of HRV in ASD and ADHD.
IntroductionTemporal melody similarity is a fundamental problem in music modelling, and current methods are mostly based on recurrent or attention-based architectures that implicitly learn sequential structures.MethodsThis work adopts an explicit feature-representation approach using temporal-style descriptors, including first-order deltas, relative ratios, polynomial interactions, and rolling statistical features. These features are represented in a model-agnostic manner to enable controlled comparison, interpretability, and reproducible evaluation across architectures. Seven model families were evaluated, including linear baselines, ensemble methods, recurrent networks, attention-based models, and dense fusion architectures, using a large-scale melody similarity dataset containing over 11,000 samples. Cross-validation and bootstrap-based uncertainty estimation were applied for performance evaluation.ResultsResults show that the engineered temporal feature space exhibits significant nonlinearity, with the highest performance achieved by Deep GRU (R2 = 0.863, MAE = 0.058), followed by XGBoost among non-recurrent models (R2 = 0.824, MAE = 0.091). Baseline models showed lower performance, demonstrating the nonlinear characteristics of the task. Feature attribution analysis identified temporal descriptors and embedding-based variables as the most influential features.DiscussionThe results demonstrate that TFR provides an interpretable and reproducible foundation for evaluating future sequence-learning architectures, particularly for data-constrained or deployment-oriented applications.
George A. Miller’s classic discussion of memory capacity and Sidney Smith’s recoding experiments demonstrated that cognitive limits depend more strongly on the number of active representational units (“chunks”) than on the total amount of raw information being processed. Here, we reinterpret Smith’s experiments from the perspective of modern computational neuroscience and representation learning. We argue that Smith’s recoding procedure illustrates a general principle of representational recoding, whereby learning increases the amount of information associated with each active representational unit without increasing the number of units available to the system. This principle provides conceptual links among classical theories of chunking, modern representation learning, vector symbolic architectures, and biological memory systems. The framework may offer useful guidance for future neuromorphic systems operating under strict energetic and structural constraints.
IntroductionPost-traumatic stress disorder (PTSD) is characterized by intrusive memories and an impaired resilience framework, which often leads to chronic psychological distress. Eye Movement Desensitization and Reprocessing (EMDR) is an emerging therapeutic approach targeting PTSD, yet the precise neurobiological mechanisms remain inadequately defined. Recent studies suggest immune modulation, mainly through molecules such as Interleukin-6 (IL-6), Interleukin-10 (IL-10), Neuropeptide Y (NPY), and Oxytocin (OXY), as well as additional proteins, may play a key role in longitudinal outcomes. This study investigated the roles of immune factors, systemic interactions, and neurobiological functions, including the Default Mode Network (DMN) and B-cell regulation, in PTSD remission and EMDR efficacy.MethodologyA systematic dry-lab analytical approach was employed using Artificial Neural Network (ANN) modeling and dynamic pathway analysis to reveal neurobiological factors, immune function, and neural network differences involved in PTSD remission and EMDR functionality. Publicly available gene expression datasets were analyzed, focusing on key pathways, interactions, and gene loci related to immune regulation, neurobiological modulation, and inflammatory response. Key molecules, including IL-6, IL-10, NPY, and OXY, were examined for potential contributions to neurobiological resilience, therapeutic response, and immune function in response to adverse stimuli.ResultsThe ANN analysis revealed that immunological function, notably through IL-10 and B-cell activity, plays a prominent role in PTSD remission, EMDR outcomes, and resilience. IL-10 emerged as central to B-cell differentiation and proliferation regarding trauma and resilience, indicating its potential as a biomarker for PTSD within the first year of symptom onset. Additional analyses implicated IL-6 and NPY as critical to longitudinal neurobiological resilience mechanisms. Interestingly, while OXY was initially classified as significant for social bonding and PTSD remission, analysis showed this gene only played a secondary mediating role, aligning with recent findings that OXY is not strictly necessary for prosocial outcomes such as PTSD remission.ConclusionThis study suggests that IL-10, IL-6, and NPY are key neuroimmune modulators in PTSD remission and may explain EMDR functionality. Immunologic activity may also explain variations in EMDR outcomes, such as therapeutic success, resistance, and relapse rates. These findings underscore the potential of targeted co-occurring immunotherapies, possible objective PTSD tests, potential objective measures for EMDR outcomes, and personalized approaches for enhanced treatment.
Electroencephalography (EEG)-based depression classification requires interpretable machine-learning approaches and validation strategies that avoid subject-level information leakage. This single-center pilot study developed and internally evaluated a marker-based interpretable machine-learning framework using eight-channel resting-state EEG. After quality control, 48 participants were included, comprising 23 clinician-diagnosed major depressive disorder (MDD) patients recruited at Zhejiang Provincial Tongde Hospital and 25 healthy controls (HCs). Subject-level spectral, entropy/complexity, asymmetry, and coherence-based connectivity features were extracted from pre-processed EEG epochs. Exploratory feature analysis was used to define a three-component EEG marker comprising fronto-posterior beta- and gamma-band coherence heterogeneity and F8-F7 beta-band asymmetry variability. The resulting marker was evaluated using classical classifiers under strict subject-wise leave-one-subject-out validation and compared with EEGNet and 1D-CNN baselines under the same subject-wise protocol. Among the evaluated classical models, RBF-SVM achieved the highest subject-level discrimination in the primary native-reference 2-s analysis, with an AUC of 0.910, accuracy of 85.42%, sensitivity of 82.61%, and specificity of 88.00%. The primary result used the native A1/A2 acquisition reference and non-overlapping 2-s epochs, consistent with segmentation used in previous resting-state EEG depression studies. SHAP analysis indicated that fronto-posterior gamma-band coherence heterogeneity and F8-F7 beta-band asymmetry variability were the dominant contributors to the RBF-SVM decision function. These findings support the exploratory value of compact EEG markers for interpretable internal discrimination in data-limited eight-channel settings and motivate validation in independent external cohorts.
Biological sex is increasingly recognized as a fundamental dimension of brain organization, yet how it influences early neurodevelopment, particularly autism spectrum disorder (ASD), remains understudied. Using data from the International Infant EEG Data Integration Platform (EEG-IP), a multi-site longitudinal cohort of 179 infants (91 males, 88 females) at elevated (ELA) and typical (TLA) likelihood of ASD, we conducted a series of re-analyses of previously published EEG studies to examine whether key EEG metrics differ by biological sex and whether biological sex can moderate the relationship between early brain profiles and language development from 6 to 36 months. Across all EEG metrics, we found that females displayed significantly higher left-hemisphere functional connectivity within speech-related regions at 6 months of age compared to males. Females also demonstrated increased posterior theta consistency during face processing as well as steeper expressive language trajectories from 6 to 36 months. Speech connectivity at 6 months also moderated receptive language growth, but not expressive language growth in females, suggesting sex-specific relationships between brain dynamics and language. These findings show the importance of including biological sex as a primary variable in infant developmental studies and provide support for the study of sex-differentiated biomarkers in prospective ASD studies.
The capacity of long-term memory seems to be extremely large, capable of storing information spanning almost a lifetime. Why does it have such a vast capacity? Why are some memories so enduring? What is the actual physical form of long-term memory? In the movie Inside Out, it is depicted as individual orbs containing information. Is that really the case? Simply explaining this by saying that the cortex has many neurons, numerous neural connections, and complex electrochemical activity between them is not sufficient to answer these fundamental questions. We need to uncover the theory hidden behind these phenomena. In essence, a neural network is equivalent to a very large directed graph, with a massive number of nodes and directed connections. This paper posits that the physical form of long-term memory is a connected subgraph within this complex directed graph. This subgraph is capable of linking together the disparate fragments of the same event, spread across different sensory cortices, to form associations. This provides a physical realization of the engram theory. The robustness of the connected subgraph and the resources it consumes can explain various memory behaviors. Based on anatomical, brain imaging, and electrophysiological evidence, this paper constructs a probabilistic connectivity model and uses theorems from graph theory to prove the ease of constructing connected subgraphs. Finally, it explains why the potential capacity for memory is immense.
Spiking recurrent neural networks (SRNNs) rival gated recurrent neural networks (RNNs) on various tasks, yet they still lack several hallmarks of biological neural networks. We introduce a biologically grounded SRNN that implements Dale's law with conductance-based stands for a-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid (AMPA) and gamma-aminobutyric acid (GABA) reversal potentials. These reversal potentials modulate synaptic gain as a function of the postsynaptic membrane potential, and we derive theoretically how they make each neuron's effective dynamics and subthreshold resonance input-dependent. We trained SRNNs on the Spiking Heidelberg Digits (SHD) dataset and show that SRNNs with reversal potentials reduce spike energy by up to 3 × , while maintaining, or increasing, task accuracy. This leads to high-performing Dalean SRNNs that substantially improve on Dalean networks without reversal potentials. SRNNs with reversal potentials exhibited spike-train statistics closer to Poisson statistics, similar to biological neurons, and showed a substantial reduction in oscillatory activity, leading to increased heterogeneity in response properties. Thus, Dale's law with reversal potentials, a core feature of biological neural networks, can render SRNNs more accurate and energy-efficient.
Along the lines of neuronal global workspace theories, the paper hypothesizes that active neural signal regeneration in recurrent, re-entrant circuits could constitute an organizational basis for states of conscious awareness. In this view, brains are self-production systems that regenerate their own informational, signal states (“neural autopoiesis”). States of awareness themselves depend critically on sustained regeneration of sets of neural signals in local and global circuits. The set of regenerated signals at each moment determines the specific contents of consciousness. Two hypotheses are proposed. H1: Organizational closure and with it, awareness, is achieved when regenerative, self-production loops are completed, such that a stable set of signal productions is sustained. The organization of informational processes stabilizes and closes on itself. H2: Neural coding is critical for signal regeneration. Neural signals must be properly encoded in order for them to be regenerated and thereby affect the contents of awareness. In addition to simple signal suppressions at their points of origin, one means by which general anesthetic agents may abolish awareness is by scrambling neural signals. Rendering internal control signals incoherent by altering within- and across-neuron rate profiles and/or temporal patternings may disrupt signal regeneration in local and global circuits. The two hypotheses are agnostic with respect to neural coding. Our own signal-centric time-domain (TD) brain theory framework is presented to illustrate how signal regeneration could be mediated by local and global temporal codes and neural temporal processing networks. Similarities and differences between TD, global neuronal workspaces (GNW), recurrent processing (RP), integrated information (IIT) and predictive processing (PP) are discussed. Empirical testing will necessitate solving the neural coding problem at multiple system levels. This will involve investigating correlations and causal linkages between regeneration of tracked neural signals and states of awareness as well as between alternative candidate neural codes and the contents of awareness. Rhythm-tagged stimuli and high temporal resolution neural recordings are proposed to enable tracking of specific neural signals throughout the brain under conditions of waking vs. sleep vs. anesthesia, with masking and unmasking signal/noise levels. Some philosophical comments regarding organization as a potential basis for consciousness are made.
Neurons near a subcritical Hopf bifurcation are distinguished by their ability to produce subthreshold oscillations, maintain bistability between rest and repetitive spiking, and fire preferentially in response to inputs near their intrinsic resonant frequency. Simulating these properties in large neural networks using continuous-time models such as the Hodgkin-Huxley formalism is computationally prohibitive, motivating the search for reduced representations that retain the essential dynamics. Here we derive a two-dimensional discrete-time map from the continuous INa, p+IK model by exploiting the time-scale separation between the membrane potential and the potassium gating variable, and using nullcline geometry to approximate the slow drift along each branch of the cubic v-nullcline. The resulting map consists of an outer piecewise-linear loop that produces realistic action potential waveforms and an inner switching region that reproduces focus-like subthreshold oscillations, with a slow recovery variable governing the transition between the two. The injected current modulates the size of the inner region, naturally encoding the bifurcation structure of the original system. Numerical simulations confirm that the discrete model reproduces the main dynamical signatures of bistability, hysteresis under a slowly varying current ramp, and frequency-selective firing in response to periodic burst stimulation—all without solving a single differential equation. Crucially, the map retains selected physically interpretable parameters of the continuous model while replacing continuous-time integration with a lightweight algebraic update. Benchmarking against explicit Euler and fourth-order Runge–Kutta integration shows that the optimized map has a per-update cost comparable to Euler and substantially lower than Runge–Kutta, supporting its use as a compact, interpretable discrete representation of subcritical-Hopf dynamics rather than as a direct numerical integrator of the continuous model.
IntroductionThis study proposes a neural-network-inspired computational framework for hierarchical, noise-tolerant coordination in distributed agent networks. The framework is evaluated in a cyber-physical demand-side management testbed in which autonomous home energy management systems perform local scheduling under a shared global price signal.MethodsPrivacy-Preserving Federated Congestion-Signal Coordination separates local appliance scheduling from global adaptive coordination. Each home energy management system uses a genetic algorithm with a population of 100 over 50 generations. A federated coordinator aggregates clipped congestion-gradient updates and applies Gaussian differential privacy with a noise multiplier of 1.1 and a clipping norm of 1.0. The main evaluation was conducted over 30 federated rounds using 50 home energy management systems, EirGrid demand traces, SEM-O market prices, and 10 independent simulation runs. A complementary noise-location analysis used 20 home energy management systems and five independent seeds.ResultsThe proposed method reduced the peak-to-average ratio from 1.468 to 1.276, corresponding to a statistically significant 13.1% improvement. Its normalized energy cost remained within 0.04% of centralized coordination. Energy cost varied by only 0.016% across eight evaluated privacy budgets, and the framework transmitted 200 times fewer numerical values than centralized coordination at 50 agents. In the complementary noise-location experiment, aggregation-stage noise was better tolerated than local-encoding noise at four of eight evaluated noise levels.DiscussionThe results show that hierarchical feedback, bounded unit influence, stochastic aggregation, and compressed message passing can support stable privacy-preserving distributed coordination. The framework is not a biological neural-circuit model but provides a reproducible testbed for studying neural-network-inspired principles of noise-tolerant collective computation.
Waves are fundamental. In our view, waves in the brain may constitute and drive organized neural activity patterns on individual neural and population levels. Their interactions follow basic physical principles. Taking a comprehensive, temporal and spatiotemporal perspective, we endeavor to explain multiple brain functions and behaviors with a unified mechanistic approach. Starting with neural architectures, traveling waves, and spikes as the basic signals of the system, our multidisciplinary theory proposes that precise temporal phase-locking codes, spike coincidences, phase relationships, recurrent networks, temporal and spatiotemporal population patterns, pattern correlations, their interactions, synchronizations and couplings play an essential role in determining brain dynamics at multiple processing scales. Analogously to optical holography, it posits that traveling brain waves convey spike timing information, interact to form distinctive time/phase interference patterns and spatially distribute these widely. These in turn can interact with other traveling waves producing yet new spike patterns. Traveling waves also serve to selectively reactivate/refresh existing patterns. Spatially distributed temporal spike patterns and representations derived from these, are used to code, process, synchronize, integrate, and decode objects (e.g., sensorimotor events, concepts, etc.). We apply established physical principles (e.g., wave dynamics, holography) and signal processing principles (e.g., linear additive operations, and nonlinear, multiplicative frequency mixing). This theory proposes that oscillations may serve as signal carriers for communications in transmitting progressively processed signals through an emergent cascade of neuron mixing stages. Such cascades closely correspond to intermediate frequency (IF) processing stages in broadly used radio communications, specifically superheterodyned Single Sideband Suppressed Carrier (SSBSC = SSB) communications technology. This is illustrated with a numerical speech/language hierarchy oscillatory cascade model. This model correlates signal processing stages with both canonical oscillation bands and associated cognitive stages. Plausible biophysical mechanisms are proposed to realize these processes. Many neurophysiological observations consistent with these proposed mechanisms are referenced. This proposal is novel in suggesting conceptual integrations and coordinations of multiple disciplines that mechanistically trace informational neural signals from inception to conception. Some suggestions for empirically testing these hypotheses are presented.
Human facial emotion recognition (FER) is a vibrant research field. This research proposes a novel, biologically inspired hybrid FER framework that uniquely connects event-driven Spiking Neural Networks (SNNs) with deep learning, specifically a Spike-based Support Vector Machine (S-SVM), which is designed for its event-driven processing and energy efficiency. The article proposes a novel SNN-based framework for FER that extracts robust image features using a Vision Transformer (ViT). Spikes are generated from the features using the rate-and-threshold encoding technique. The core algorithmic novelty lies in the integration of Leaky Integrate-and-Fire (LIF) and Quadratic Integrate-and-Fire (QIF) neurons, which are used in S-SVM, creating a highly optimized decision boundary in the spike domain. The goal is to determine which neuron performs best, as confirmed by the CK+ dataset. The same technique was validated on RAF-CE, a unprocessed dataset derived from real-world events and movie scenes. The approach was validated on the CK+ dataset, achieving 99.14% and 99.94% accuracy for the QIF and LIF neurons, respectively. For the compound emotions in the unprocessed dataset, which are hard to distinguish, the accuracy rates achieved with QIF and LIF neurons are 78.87% and 98.91%, respectively. In the FER system, the LIF neuron with an S-SVM classifier performs better than the LIF neuron. This hybrid design's capacity to provide cutting-edge FER accuracy while also enabling previously unheard-of energy efficiency gains of 99.52% for CK+ and 99.60% for RAF-CE when compared to traditional deep learning models is its primary significance.
Background:Obsessive-compulsive disorder (OCD) is characterized by intrusive thoughts and repetitive behaviors, with incomplete response to current treatments suggesting limitations in prevailing neurotransmitter-focused models. Epidemiological data indicate lifetime DSM-IV OCD in approximately 2.3% and 12-month OCD in approximately 1.2% of US adults, while subthreshold obsessions or compulsions are substantially more common. One mechanistic possibility is that abnormal synaptic pruning contributes to persistent cortico-striato-thalamo-cortical circuit rigidity. This idea remains indirect in OCD, but is supported by convergent synaptic-marker findings in OCD and by complement-linked pruning mechanisms shown most clearly in schizophrenia. Methods:We developed a modular gated recurrent unit network approximating cortico-striato-thalamo-cortical dynamics, trained on a rule-switching task sensitive to perseveration. The architecture included a recurrent cortical integration layer, a schematic striatal "habit" module, and a thalamic confidence-gating scalar. Excessive pruning was implemented as 60% activity-dependent pruning, using a composite of low recent gradient-based usage and low weight magnitude, with partial protection of recurrent and habit-related weights. From identical pruned baselines, three mechanistically distinct interventions were simulated: rapid gradient-guided structural reopening, treated as a ketamine-like synaptogenesis motif, prolonged low-learning-rate adaptation with stress-noise annealing (SSRI-like), and tonic inhibitory scaling (neurosteroid-like). An iso-dose pipeline recorded L1 and L2 weight-change norms, synaptic turnover, and change in sparsity; linear interpolation was used when bracketing sweep points existed, and residual mismatch was reported otherwise. Multi-seed statistics and sensitivity analyses were performed. Results:At 60% activity-dependent pruning, the untreated network showed impaired accuracy (0.4972) and elevated perseveration (0.5247). In the fixed-parameter comparison, ketamine-like structural repair reduced perseveration to 0.2582, SSRI-like adaptation to 0.3209, and neurosteroid-like inhibition to 0.2654. Relapse vulnerability differed by mechanism: cumulative relapse increased perseveration by +0.1086 after ketamine-like repair, +0.0259 after SSRI-like adaptation, and -0.0017 after neurosteroid-like inhibition in the representative seed. Across five seeds, best acute perseveration was lowest for ketamine-like repair (0.2330 ± 0.0065), followed by neurosteroid-like inhibition (0.2413 ± 0.0021) and SSRI-like adaptation (0.2831 ± 0.0099). SSRI-like adaptation showed the highest mean efficiency because it produced smaller absolute weight changes. Sensitivity analyses showed that untreated perseveration increased with pruning severity, from 0.2627 at 40% pruning to 0.6218 at 70% pruning, while ketamine-like repair remained relatively stable across the same range. Conclusion:These findings support excessive synaptic pruning as a plausible contributor to OCD-like cognitive inflexibility and illustrate that structural and functional interventions offer different trade-offs within a highly abstract computational model. Structural repair produced the most robust acute rescue and remained resilient across pruning severities, whereas functional mechanisms showed advantages in dose efficiency or relapse stability. The results are hypothesis-generating only and should not be read as clinical evidence for treatment ranking.
To enhance the dynamic adaptability and robustness of neuromuscular signal-driven exoskeleton robots in complex environments. This study develops a Deep Reinforcement Learning (DRL)-based exoskeleton control framework. A module for neuromuscular signal processing and motion intention modeling is designed, encompassing band-pass filtering, local normalization, time-frequency feature extraction, and Bidirectional Long Short-Term Memory (BiLSTM) time-series encoding. Additionally, a DRL-based dynamically adaptive control framework is established. Trajectory tracking error, torque smoothness, energy consumption agent, joint safety constraint and intention consistency are comprehensively introduced into the reward function to achieve the multi-objective balance of "accuracy-comfort-energy consumption-safety." Comparative experiments demonstrate that the joint trajectory tracking errors of the proposed optimized system are 2.37°, 2.64°, and 2.92°, respectively. These values are significantly lower than those of comparative systems, including the Deep Reinforcement Learning-based Robust Controller for Lower-Limb Rehabilitation Exoskeletons (DRL-RC-LLRE) and the Learning-in-Simulation Exoskeleton Assistance Framework (LiS-EXO). The corresponding torque smoothness indicators are 0.62, 0.68, and 0.71. The mechanical energy consumption proxy indicators are 10.48, 11.39, and 11.67, indicating that while improving the tracking accuracy, the torque oscillation is effectively suppressed and the energy consumption is reduced. The ablation experiments further show that when any of the neuromuscular intention embedding, intention consistency reward or safety constraint modules are removed, the average trajectory error, torque smoothness and energy consumption indicators all deteriorate to varying degrees. The complete model achieves 2.63°, 0.66, and 11.17, respectively, in the three indicators, verifying the collaborative contribution of the three key modules to the overall performance. Overall, this study has constructed and verified a reproducible "neuromuscular signal-motion intention-reinforcement learning control" integrated framework, and thus having certain contributions to the field of intelligent exoskeleton control.
IntroductionAccurately identifying student behavior in classroom environments is essential for the development of intelligent education systems. However, existing methods often exhibit high false-positive and false-negative rates in complex backgrounds or densely populated scenes, which degrades detection accuracy.MethodsTo overcome these limitations, we propose StudentNet, a novel framework incorporating the Behavior-Enhanced Fusion Super-Resolution Network (BEF-SRNet), the Direction-Sensitive Adaptive Convolution module (DSAC), and the Adaptive Spatial Feature Fusion (ASFF) strategy. BEF-SRNet separates regions of interest from background clutter to improve the discriminability of behavioral features. DSAC uses multi-directional branches and channel attention to extract structural behavioral cues while suppressing noise. ASFF enhances scale-invariant feature representation to improve semantic consistency and detection performance.ResultsExperiments on SCB-Dataset3 show that StudentNet improves mAP@50 and F1-score by 6.3 and 6.4 percentage points, respectively, compared with the YOLOv8n baseline. Deployment on the Core-3399Pro-JD4 edge platform further shows that the INT8-quantized model achieves 38.4 FPS with acceptable accuracy degradation.DiscussionThese results indicate that StudentNet improves student behavior detection in complex classroom scenes and is suitable for real-time smart classroom applications on edge devices.
Computational neuron models commonly reduce subthreshold membrane dynamics to a leaky RC integrator, confining computation to the spike and treating the inter-spike trajectory as passive decay. Yet impedance measurements and channel-specific analyses show that many excitable membranes have band-pass, inductance-like impedance profiles with a tuneable resonant peak — dynamics a first-order RC model cannot represent. This paper develops a spike-as-perturbation framework in which spikes act as impulse perturbations that launch regime-dependent transient trajectories in an equivalent parallel RLC membrane. We define the biological grounding and domain of validity of the reduction, show that post-perturbation RLC ringdowns carry circuit-identity and perturbation-timing state variables not available to a single matched first-order RC element without added delays, recurrence, or extra state variables, and demonstrate a concrete primitive — phase-based temporal discrimination — together with a spike-timing readout that makes it network-visible. A closed-form comparison (Q = 2, τRC = 25 ms, f0 = 10 Hz) gives an RLC sensitivity half-life of 44 ms versus 17 ms for a matched RC decay. We identify the quality factor Q — strongly shaped by Ih and related slow conductances — as a candidate neuromodulatory control variable that fixes the membrane pole radius and hence its transient (short-term) memory horizon, the biological analogue of selectivity in machine state-space models. We do not claim a universal in vivo phase code: a simulation under high-conductance bombardment shows that the readout is reliable in quiescent, high input-resistance states and is suppressed as conductance loading drives the effective Q below the underdamped threshold. We derive four falsifiable predictions spanning cellular, network, decoding, and population levels, and present the argument in three tiers of decreasing evidential support.
