In the field of brain-computer interfaces (BCIs), a sufficiently high information transfer rate (ITR) serves as the prerequisite guarantee for the realization of practical application of electroencephalography (EEG)—based BCIs. Hybrid BCIs combining multiple biosignals with EEG can broaden the command set and improve ITR. We propose a hybrid BCI system combining surface electromyography (sEMG) and steady-state visual evoked potential (SSVEP), which utilizes 120 flicker frequencies and 4 gesture movements to encode 480 targets. Moreover, high-density electrodes are used to acquire more EEG information. In the online experiment, the average classification accuracy achieved 84.55 ± 7.23
Higher-order memristors offer a compact approach to representing multiple internal memory states in neuronal models. In this work, a discrete second-order locally active memristor (SO-LAM), featuring two internal states and a bounded periodic memductance, is integrated into the Aihara neuron map, yielding a memristive neuron with two history-dependent internal states. The membrane potential drives the evolution of both memristive states, while the two states jointly determine the memductance and feedback current. The discrete SO-LAM exhibits memristive characteristics, local activity, and edge-of-chaos regions. Through the combined effects of neuronal nonlinearity and memristive coupling, the SO-LAM-based neuron map generates regular and complex firing dynamics. The periodicity of the two memristive states gives rise to two-dimensional translational equivariance, indicating that the coexisting attractors in the grid-like array are spatially shifted copies of the same fundamental attractor rather than dynamically independent states. Representative hardware trajectories corresponding to the numerically identified hyperchaotic regime and symmetry-related attractor copies are reproduced using a field-programmable gate array (FPGA), demonstrating the feasibility of digital implementation. These results provide a compact framework for studying multistate history-dependent modulation and symmetry-induced coexistence in discrete memristive neurons.
Electrocardiogram (ECG) signal classification is a crucial task in automated cardiac monitoring and early disease detection. Conventional machine learning and deep learning methods often face challenges related to high computational costs, large data requirements and difficulty in handling real-time ECG analysis. In this work, we have proposed three memristor-based echo state network (MESN) models for ECG classification: (1) MESN, (2) Parallel MESNs (PMESN) with a classifier and (3) Attention-enhanced MESN (AMESN), where all three models employ artificial neural network (ANN) as a readout layer. These models leverage the property of an ESN to transform ECG signals into high-dimensional representations, thereby enhancing classification performance. The proposed models were evaluated on different benchmark ECG datasets: ECG200, ECG5000, ECGFiveDays, NonInvasiveFetalECGThorax1, NonInvasiveFetalECGThorax2, and TwoLeadECG. The results demonstrate that AMESN achieves the highest precision, recall, F1-score and overall accuracy particularly excelling in datasets with high intra-class variability. Whereas, PMESN improves robustness by leveraging multiple reservoir layers, while MESN provides a strong baseline with computational efficiency. These models are also compared with other state-of-the-art models and the analysis shows that AMESN achieves 1.92–13.39
Artificial agents trained on non-stationary task streams often acquire new categories at the cost of older representations. Systems-level accounts of hippocampo-cortical consolidation suggest that stored traces can support later learning in different ways and that consolidation can depend on both new learning and the vulnerability of established knowledge. Motivated by these functional principles, we propose dual-role memory consolidation learning (DRMCL) for class-incremental learning with a bounded replay buffer. DRMCL stores prototype-preserving core samples and decision-sensitive boundary samples, then combines both roles with replay, logit distillation, and task-level feedback control. On Split CIFAR-10 with a pretrained ResNet18 backbone, the clearest benefit occurs when memory is scarce. With 1000 stored samples, average accuracy increases from 0.4823 for Dark Experience Replay++ (DER++) to 0.6905, while forgetting decreases from 0.5734 to 0.2650. At medium and high memory, DRMCL usually lowers forgetting, although DER++ can achieve higher raw accuracy. Ablations show that feedback-regulated replay and distillation account for most of the retention effect, while the core/boundary organization makes the selected memory easier to inspect. An electroencephalography (EEG) brain-computer interface (BCI) feasibility study on BCI Competition IV 2a provides a cross-domain check. In a five-seed subject-1 experiment, DRMCL remains close to ER and DER++ rather than separating from them, and session spectral/coherence shifts vary across subjects. The results support DRMCL as a stability-oriented computational framework; they do not establish a circuit-level model of hippocampo-cortical consolidation or a new EEG decoding benchmark.
Excitation-inhibition (E–I) imbalance is a core pathological mechanism underlying many psychiatric disorders. In this study, we integrate structural and functional magnetic resonance imaging data to construct a large-scale brain network model, aiming to investigate how perturbed structural connectivity drive brain network reorganization and how neurotransmitters regulate E–I balance. The results show that the multiscale dynamic mean-field (MDMF) model identifies steady-state regimes of neurotransmitter-related parameters that yield relatively high correspondence between simulated and empirical resting-state functional connectivity. Furthermore, synaptic pathology is simulated by altering structural connectivity, and the resulting functional connectivity reproduces network topological features observed in neurological and psychiatric disorders. Finally, the E–I ratio is quantified at both the neurotransmitter level and the neuronal firing-rate level. By modulating inhibitory neurotransmitter-related parameters, selected topological properties of the simulated FC networks shift closer to the empirical condition. These findings establish a link between E–I imbalance, neurotransmitter regulation, and structural abnormalities, providing new insights into the pathogenesis and potential treatment strategies for psychiatric disorders.
Working memory (WM) depends on coordinated interactions among distributed brain systems. Although sex differences in WM have been widely studied, most previous work has focused on group-level activation differences. As a result, it remains unclear whether males and females can be characterized by distinct functional connectivity-based models of WM performance. We analyzed 622 participants (311 females and 311 males) from the Human Connectome Project who were retained after data quality and completeness screening, with the male and female groups matched on relevant covariates and showing no significant between-group differences. Using n-back task functional magnetic resonance imaging data and corresponding behavioral measures, we constructed connectome-based predictive models of WM performance separately in females and males. WM performance could be predicted in both sexes. However, cross-sex validation showed that the female-trained model significantly predicted male WM performance, whereas the male-trained model did not significantly predict female WM performance. To examine factors associated with this asymmetry, we analyzed inter-subject consistency within each sex group. Females showed greater within-sex neural dynamic consistency than males across multiple brain regions, and subsampling analyses further linked higher inter-subject consistency to greater predictive feature stability and better cross-sex prediction performance. Importantly, sex differences extended beyond model performance to the predictive connectivity patterns themselves, mainly involving higher-order cognitive control and lower-level visual systems. Functional connectivity gradient analyses provided convergent support for this pattern. Together, these findings deepen our understanding of WM network mechanisms and provide a basis for understanding sex-related differences in WM impairment and for exploring potential interventions.
This paper is concerned with global polynomial state estimation issue of memristive neural networks which take quaternion-valued parameters and fuzzy terms into account. First, via drawing support from quaternion-valued norm, an easily analyzable estimation error model is established, which overcomes the complexity brought by the system parameters. Then, a simple feedback controller is designed aiming to obtain the polynomial stability conditions. It is worth noting that, several algebraic forms of polynomial stability criteria for the error system proposed are achieved by applying the quaternion-valued norm, each of these criteria is represented by algebraic inequality, which facilitates validation. Ultimately, illustrative examples are given to show the effectiveness of the theoretical results.
Emotion recognition in videos is still difficult due to the fact that affective cues are spread throughout facial expressions, body movements and scene context. Face only approaches tend to perform poorly in most cases because of being affected by such factors as varying illumination, occlusion, camera motion, dynamic backgrounds, and multiple interacting subjects. To overcome these drawbacks, a novel and flexible model called ContextFusion-EmoNet (CFEN) is proposed in this paper to utilize appearance, motion and relational information for robust emotion recognition from video. CFEN uses an Average Contextual Loss (ACL) guided key-frame selection method, incorporating the differences of the VGG16 features and Farneback optical flow weight to select 4 key frames that are rich in motion and extremely informative per video. The selected images are then fed into ResNet18 feature extractor, a temporal context modeler, which is a Transformer encoder, and a deterministic Graph Convolutional Network (GCN) module for modeling the relation between frame tokens. Good intra-dataset performance is shown through extensive experimentation on the four datasets, CAER, CK + , DFEW, and FERV39k, with accuracy rates of 97.71
Paroxysmal kinesigenic dyskinesia is a rare neurological disorder characterized by brief, recurrent motor attacks that significantly impair quality of life. Prior studies have largely relied on unimodal data, which offer partial insights into neural regulation but are constrained by trade-offs between temporal and spatial resolution. To address this limitation, we developed a multimodal recognition and tracing framework integrating electroencephalography and functional magnetic resonance imaging. We propose GTBL-AF, a deep multimodal neural architecture that captures spatial connectivity and temporal dynamics of brain function through graph attention, Transformers, and bidirectional long short-term memory networks, with cross-attention enabling modality-level fusion. GTBL-AF achieved 94.2
Cognitive impairment is one of the most functionally debilitating non-motor symptoms in Parkinson's disease (PD). Yet, current diagnostic and clinical practices rely heavily on subjective assessments and burdensome behavioral testing, which lack the temporal resolution to detect subtle deficits. While EEG and pupillometry offer promising non-invasive insights into cognitive processing, more scalable and objective methods are needed to detect subtle and often elusive dysfunction in early PD. Machine learning (ML) leverages patterns in neurophysiological signals recorded during cognitive tasks to identify subtle impairments that may not be evident in standard evaluations. We recorded EEG and pupillometry data from 68 participants (35 PD, 33 healthy controls (HC)) during a visual Change Detection working memory task. Using our custom, standardized feature-extraction toolbox, we extracted 108 features and incorporated the resulting feature matrices into an ML pipeline. We applied SMOTE to balance the classes, then used principal component analysis (PCA) to reduce dimensionality. An automated elbow method identified optimal cutoffs for principal components (PCs) and original features, which guided subsequent recursive elimination (RE) for computational efficiency. We trained a support vector machine with a radial basis function kernel (SVM-RBF) to classify PD and evaluated the final model's performance on a hold-out dataset. The automated elbow method identified optimal cutoffs at 20 PCs and 22 original features. The RE revealed that a model using 14 PCs and the top 7 PCA-weighted features achieved the highest performance, with 71
Understanding how cognition unfolds from neurophysiological signals presents a promising direction for cognitive science studies and wearable-enabled human–robot interaction applications. However, uncovering latent neurodynamic geometry and temporal progression remains challenging and underexplored due to the lack of observable temporal organization for annotation and, consequently, the difficulty of training models in a supervised manner. This study proposes a representational learning method for this segmentation problem that shifts the solution away from statistical change-point detection methods and Hidden Markov Models. Our method employs self-supervised learning to discover emergent properties of the underlying temporal organization directly from the neurodynamic data itself. Four objectives are introduced and jointly optimized, including within-stage temporal predictability, boundary contrast, cross-trial alignment, and sparse stage-specific feature weights. Population-based evolutionary search was adopted to explore the multiple-basins-of-attraction landscape, where mutation and crossover govern the convergence process. We validated the framework on EEG recordings collected from participants performing an embodied road-crossing decision-making task, which simulates a typical cognitive processing transition from perceptual assessment to risk evaluation and decision commitment. Results showed that our method achieves an order-of-magnitude improvement in boundary contrast of the discovered stages, indicating that the learning behavior fundamentally changes the working principle from seeking local statistical consistency to capturing higher-order global temporal organization. This inter-stage divergence serves as the driving force for latent regime discovery while preserving local temporal continuity and coherence. Ablation and sensitivity studies demonstrate that the model performance is robust in identifying cross-trial transferable state geometry and handling data variability introduced by subject and stimulus heterogeneity. The reconstructed cognitive stages are also behaviorally plausible, and the dimensions attended by the model are well aligned with the neurophysiological underpinnings governing critical cognitive activities underlying each stage.
Severe and enduring psychiatric illnesses, including melancholia and bipolar illness, show prolonged deviations in motivation and mood yet lack a unifying account of their long-term dynamics. Solomon’s opponent-process theory provides a qualitative framework for short-timescale affective responses, composed of a fast stimulus-locked a-process opposed by a slower adaptive b-process. Here we evaluated whether these dynamics can be implemented as a computational homeostatic controller, and whether distinct clinical trajectories emerge as canonical failure modes of the same system. We formulated a tractable control-systems model with feedforward a- and b-processes, and examined its behaviour across minute-scale and month-scale regimes. Under “healthy” parameters, the model reproduced classical opponent-process responses. Altering only opponent-process gain and decay generated a prolonged downward drift, matching the clinical timescale and asymmetry of melancholia. Reducing damping within the same controller produced an endogenous underdamped oscillation with long period and phase asymmetry characteristic of bipolar illness. Together, these proof-of-principle simulations suggest that severe severe affective illnesses may be expressed as distinct dynamical regimes of a single motivational homeostat. This framework generates testable predictions and may facilitate experimental quantification of opponent-process recovery, damping, and gain as mechanistic markers of severe affective illnesses.
Recently, several studies concerning the richness of electroencephalogram (EEG) signals’ content to detect pain and classify pain severity have been made. In this study, a novel band-wise adaptive classifier is proposed based on successive brain graphs through five levels of pain. The EEG data were acquired from 44 healthy subjects while putting their hand in the cold water to feel pain continuously and increasingly over time till the intolerable stage, that they withdrew their right hand. During the cold pressor test, participants reported their pain at five different levels while their EEG signals were captured by 32 silver electrodes. The data was decomposed into five frequency bands, and for each band-wise channel, discriminative EEG features were calculated in two different feature sets. Afterward, three connectivity estimators were applied to determine the adjacency matrices for each frequency band and construct the corresponding brain graphs. The constructed graphs, along with feature nodes, were fed into an adaptive graph convolutional neural network (GCNN) based hierarchical classifier, which selects the most discriminative graphs at each node. The discriminability of each estimator was tested by the Kruskal–Wallis test. In each stage of the proposed adaptive hierarchical tree, the most discriminative band-wise GCNN was selected and applied. Based on the proposed classifier, the highest accuracy reached 88.5
Epileptic seizures are serious neurological events that significantly affect patients’ health and quality of life. Accurate seizure prediction is essential for enabling early intervention and improving clinical outcomes. Most existing prediction systems are patient-specific, requiring large amounts of individualized data and exhibiting poor generalizability across subjects. A major challenge in this field lies in the limited availability of preictal EEG segments, recorded shortly before a seizure onset, compared to the more abundant interictal segments, which represent normal brain activity between seizures. This study proposes a hybrid deep learning architecture for patient-independent epileptic seizure identification. The framework is designed to perform robustly across multiple patients without the need for subject-specific calibration. A data augmentation technique based on a random walk algorithm was adopted from the literature to address the scarcity of preictal EEG segments. Then the power spectral density (PSD) was used to extract features, capturing important frequency-domain characteristics of brain activity. The proposed hybrid architecture integrates a Hierarchical Temporal Separable Convolutional Network (HTSCN), a Dual-Stage Bidirectional Recurrent Neural Network (DS-Bi-RNN), and a Multi-Head Attention Mechanism, enabling effective extraction of spatial, temporal, and contextual features from non-stationary EEG signals while addressing inter-patient variability. Experimental evaluations conducted on the CHB-MIT and Siena datasets demonstrated the strong discriminative performance of the proposed model in classifying preictal and interictal EEG states, achieving test accuracies of 98.74
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.
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.
Spiking neural networks (SNN) provide superior potential for neuromorphic architecture implementation due to its similarity to biological brain structures and exceptional computing efficiency. Neurons are the fundamental elements of SNN, and the incorporation of frequency adaptation enhances the performance of SNN significantly. The study implements an adaptive leaky-integrate-and-fire (LIF) neuron utilizing volatile and non-volatile memristors. The design offers control of adaptive response via inter-pulse interval of the input and certain circuit characteristics, including membrane capacitance and the initial resistance state of the volatile memristor. This work presents a novel SNN crossbar circuit of dimensions 2x2 and 5x5, offering several advantages including bio-realistic spike generation, reduce energy per spike consumption, no requirement of additional neuron reset circuitry, improving scalability and integration. The Cadence Virtuoso 180 nm simulation environment has been utilized to demonstrate firing dynamics of the adaptive SNN. The study emphasizes potential of adaptive spiking neural network circuits in facilitating efficient neuromorphic applications in forthcoming research.
Emotions are constantly generated in daily activities. They not only control people's behavioral patterns and thinking decisions, but also affect physical and mental health. Therefore, the emotion recognition technology based on electroencephalogram (EEG) signals has broad application prospects in multiple fields such as human-computer interaction, medical health, and intelligent driving. To address the issues of insufficient labeled data in EEG and significant differences in EEG data among different subjects and at different time periods, this paper introduces the domain adaptation (DA) technique to solve the task of cross-domain EEG emotion recognition under unsupervised conditions. Aiming at the problem that the existing domain adaptation methods ignore the different weights of the global domain and subdomains when reducing domain differences, this paper proposes a dynamic bi-domain discriminator adversarial network (DBDAN). A feature extractor is constructed to extract the low-level domain invariant features of the EEG signals, and a multi-branch domain-specific feature extractor is used to generate domain-specific features. In order to reduce domain differences, adversarial learning for the global domain and subdomains is achieved through a dual-domain discriminator. Meanwhile, dynamic factors are introduced to dynamically adjust adversarial learning, enabling the model to strike a balance between coarse-grained adversarial and fine-grained adversarial and achieve domain adaptation more precisely. The cross-subject accuracy rates on SEED, SEED-IV and DEAP were 89.43%, 75.09% and 63.42%, respectively. These results indicate that our method achieves competitive performance among the evaluated EEG transfer-learning baselines and suggest that dynamic global-subdomain alignment is beneficial for cross-domain EEG emotion recognition.
The human ability to smell functions as a critical cognitive function because it enables people to detect their surroundings while experiencing feelings and recalling memories and making choices. Researchers face difficulties when they use electroencephalography (EEG) to study how the brain responds to smells because olfactory brain signals produce low signal-to-noise ratios and different people show different response patterns and researchers lack established olfactory EEG databases for their studies. The study proposes a simulation-based framework which enables researchers to study olfactory EEG signals through power spectral density (PSD) analysis. The research team created a simulated olfactory EEG dataset which simulated the responses of fifty virtual participants who experienced two distinct odor categories of pleasant rose and unpleasant rotten at three different concentration levels of low medium and high to create six separate olfactory conditions. The simulated EEG signals included 45 channels which recorded data at a 256 Hz sampling rate. Welch’s method estimated PSD features for five canonical EEG frequency bands which included delta theta alpha beta and gamma after the data underwent band-pass filtering at the 0.5–70 Hz range. The researchers used Stratified 10-fold cross-validation to evaluate the band’s characteristics which they had developed as training data for their multiclass support vector machine (SVM) classification model. The PSD-based features demonstrated their ability to distinguish between different olfactory conditions in controlled tests which showed the system’s classification accuracy of 99.67
