Teaching CORnet human fMRI representations for enhanced model-brain alignment

Deep convolutional neural networks (DCNNs) have demonstrated excellent performance in object recognition and have been found to share some similarities with brain visual processing. However, the substantial gap between DCNNs and human visual perception still exists. Functional magnetic resonance imaging (fMRI) as a widely used technique in cognitive neuroscience can record neural activation in the human visual cortex during the process of visual perception. Can we teach DCNNs human fMRI signals to achieve a more brain-like model? To answer this question, this study proposed ReAlnet-fMRI, a model based on the SOTA vision model CORnet but optimized using human fMRI data through a multi-layer encoding-based alignment framework. This framework has been shown to effectively enable the model to learn human brain representations. The fMRI-optimized ReAlnet-fMRI exhibited higher similarity to the human brain than both CORnet and the control model in within- and across-subject as well as within- and across-modality model-brain (fMRI and EEG) alignment evaluations. Additionally, we conducted an in-depth analysis to investigate how the internal representations of ReAlnet-fMRI differ from CORnet in encoding various object dimensions. These findings provide the possibility of enhancing the brain-likeness of visual models by integrating human neural data, helping to bridge the gap between computer vision and visual neuroscience.

Supplementary Information

The online version contains supplementary material available at 10.1007/s11571-025-10252-y.

Brain-region specific epileptic seizure detection through EEG dynamics: integrating spectral features, SMOTE and long short-term memory networks

Investigating neural dynamics through EEG signals offers valuable insights into brain activity, especially for automated seizure detection. The identification of epileptogenic zones is crucial for effective epilepsy treatment, particularly in surgical planning. This work introduces a novel method for seizure detection using EEG signals, designed to benefit clinicians by integrating spectral features with Long Short-Term Memory (LSTM) networks, enhanced by brain region-specific analysis. This research work captures critical frequency domain characteristics by extracting pivotal spectral features from EEG data, thereby improving the signal representation for LSTM networks. Additionally, this proposed work has employed the Synthetic Minority Over-sampling Technique (SMOTE) to handle the class imbalance problem. Furthermore, a comprehensive spatial analysis of EEG signals is performed to evaluate performance variations across distinct brain regions, enabling targeted region-wise analysis. This strategy effectively reduces the number of channels required, minimizing the need to process all 22 channels specified in the CHB-MIT dataset, thus significantly decreasing computational complexity while preserving high seizure detection performance. This work has obtained a mean value of accuracy of 95.43%, precision of 95.46%, sensitivity of 95.59%, F1-score of 95.48%, and specificity of 95.25% for the brain region providing the best performance for seizure discrimination. The results demonstrate that integrating spectral features and LSTM, augmented by spatial insights, enhances seizure detection performance and hence assists in identifying epileptogenic regions. This tool enhances clinical applications by improving diagnostic precision, personalized treatment strategies, and supporting precise surgical planning for epilepsy, ensuring safer resection and better outcomes.

Characteristics analysis of a single electromechanical arm driven by a functional neural circuit

From a biological viewpoint, the muscle tissue produces efficient gait behavior that can be adjusted by neural signals. From the physical viewpoint, the limb movement can be simulated by applying a neural circuit to control the artificial electromechanical arm (EA). In this paper, a functional neural circuit is used to excite a single EA, the load circuit attached to the moving beam is driven by a neural circuit, and the Ampere force is activated by the load circuit to control the artificial EA. The dynamic equations of the neural circuit are derived using Kirchhoff's theorem, while the energy and motion equations of the beam are computed through the application of mechanics and related theoretical principles. Furthermore, the dynamic characteristics of the functional neural circuit forced EA are analyzed. The results indicate that the beam movement can be controlled by the electrical activity of this functional neural circuit. This work will provide theoretical guidance to build the electromechanical device for complex gait behaviors.

Visual image reconstructed without semantics from human brain activity using linear image decoders and nonlinear noise suppression

In recent years, substantial strides have been made in the field of visual image reconstruction, particularly in its capacity to generate high-quality visual representations from human brain activity while considering semantic information. This advancement not only enables the recreation of visual content but also provides valuable insights into the intricate processes occurring within high-order functional brain regions, contributing to a deeper understanding of brain function. However, considering fusion semantics in reconstructing visual images from brain activity involves semantic-to-image guide reconstruction and may ignore underlying neural computational mechanisms, which does not represent true reconstruction from brain activity. In response to this limitation, our study introduces a novel approach that combines linear mapping with nonlinear noise suppression to reconstruct visual images perceived by subjects based on their brain activity patterns. The primary challenge associated with linear mapping lies in its susceptibility to noise interference. To address this issue, we leverage a flexible denoised deep convolutional neural network, which can suppress noise from linear mapping. Our investigation encompasses linear mapping as well as the training of shallow and deep autoencoder denoised neural networks, including a pre-trained, state-of-the-art denoised neural network. The outcome of our study reveals that combining linear image decoding with nonlinear noise reduction significantly enhances the quality of reconstructed images from human brain activity. This suggests that our methodology holds promise for decoding intricate perceptual experiences directly from brain activity patterns without semantic information. Moreover, the model has strong neural explanatory power because it shares structural and functional similarities with the visual brain.

Marques A, Mora-Jiménez I, Álvarez-Rodríguez J, Soguero-Ruiz C. Multimodal interpretable data-driven models for early prediction of multidrug resistance using multivariate time series

Electronic Health Records (EHRs) serve as a comprehensive repository of multimodal patient health data, combining static demographic attributes with dynamic, irregular Multivariate Time Series (MTS), characterized by varying lengths. While MTS provide critical insights for clinical predictions, their integration with static features enables a more nuanced understanding of patient health trajectories and enhances predictive accuracy. Deep Neural Networks (DNNs) have proven highly effective in capturing complex patterns in healthcare data, offering a framework for multimodal data fusion. However, their adoption in clinical practice is limited by a lack of interpretability, as transparency and explainability are essential for supporting informed medical decisions. This study presents interpretable multimodal DNN architectures for predicting and understanding the emergence of Multidrug Resistance (MDR) in Intensive Care Units (ICUs). The proposed models integrate static demographic data with temporal variables, providing a holistic view of baseline patient characteristics and health progression. To address predictive performance and interpretability challenges, we introduce a novel methodology combining feature selection techniques with attention mechanisms and post-hoc explainability tools. This approach not only reduces feature redundancy but also highlights key risk factors, thereby improving model accuracy and robustness. Experimental results demonstrate the effectiveness of the proposed framework, achieving a Receiver Operating Characteristic Area Under the Curve of 76.90 ± 3.10, a significant improvement over baseline models. Beyond MDR prediction, this methodology offers a scalable and interpretable framework for addressing various clinical challenges involving EHR data. By integrating predictive accuracy with explanatory insights-such as the identification of key risk factors-this work supports timely, evidence-based interventions to improve patient outcomes in ICU settings.

Unraveling the functional complexity of the locus coeruleus-norepinephrine system: insights from molecular anatomy to neurodynamic modeling

The locus coeruleus (LC), as the primary source of norepinephrine (NE) in the brain, is central to modulating cognitive and behavioral processes. This review synthesizes recent findings to provide a comprehensive understanding of the LC-NE system, highlighting its molecular diversity, neurophysiological properties, and role in various brain functions. We discuss the heterogeneity of LC neurons, their differential responses to sensory stimuli, and the impact of NE on cognitive processes such as attention and memory. Furthermore, we explore the system's involvement in stress responses and pain modulation, as well as its developmental changes and susceptibility to stressors. By integrating molecular, electrophysiological, and theoretical modeling approaches, we shed light on the LC-NE system's complex role in the brain's adaptability and its potential relevance to neurological and psychiatric disorders.

Deep neural networks and stochastic methods for cognitive modeling of rat behavioral dynamics in T -mazes

Modeling animal decision-making requires mathematical rigor and computational analysis to capture underlying cognitive mechanisms. This study presents a cognitive model for rat decision-making behavior in T -mazes by combining stochastic methods with deep neural architectures. The model adapts Wyckoff's stochastic framework, originally grounded in Bush's discrimination learning theory, to describe probabilistic transitions between directional choices under reinforcement contingencies. The existence and uniqueness of solutions are demonstrated via fixed-point theorems, ensuring the formulation is well-posed. The asymptotic properties of the system are examined under boundary conditions to understand the convergence behavior of decision probabilities across trials. Empirical validation is performed using Monte Carlo simulations to compare expected trajectories with the model's predictive output. The dataset comprises spatial trajectory recordings of rats navigating toward food rewards under controlled experimental protocols. Trajectories are preprocessed through statistical filtering, augmented to address data imbalance, and embedded using t-SNE to visualize separability across behavioral states. A hybrid convolutional-recurrent neural network (CNN-LSTM) is trained on these representations and achieves a classification accuracy of 82.24%, outperforming conventional machine learning models, including support vector machines and random forests. In addition to discrete choice prediction, the network reconstructs continuous paths, enabling full behavioral sequence modeling from partial observations. The integration of stochastic dynamics and deep learning develops a computational basis for analyzing spatial decision-making in animal behavior. The proposed approach contributes to computational models of cognition by linking observable behavior to internal processes in navigational tasks.

Deep learning methods for clinical workflow phase-based prediction of procedure duration: a benchmark study

This study evaluates the performance of deep learning models in the prediction of the end time of procedures performed in the cardiac catheterization laboratory (cath lab). We employed only the clinical phases derived from video analysis as input to the algorithms. Our results show that InceptionTime and LSTM-FCN yielded the most accurate predictions. InceptionTime achieves Mean Absolute Error (MAE) values below 5 min and Symmetric Mean Absolute Percentage Error (SMAPE) under 6% at 60-s sampling intervals. In contrast, LSTM with attention mechanism and standard LSTM models have higher error rates, indicating challenges in handling both long-term and short-term dependencies. CNN-based models, especially InceptionTime, excel at feature extraction across different scales, making them effective for time-series predictions. We also analyzed training and testing times. CNN models, despite higher computational costs, significantly reduce prediction errors. The Transformer model has the fastest inference time, making it ideal for real-time applications. An ensemble model derived by averaging the two best performing algorithms reported low MAE and SMAPE, although needing longer training. Future research should validate these findings across different procedural contexts and explore ways to optimize training times without losing accuracy. Integrating these models into clinical scheduling systems could improve efficiency in cath labs. Our research demonstrates that the models we implemented can form the basis of an automated tool, which predicts the optimal time to call the next patient with an average error of approximately 30 s. These findings show the effectiveness of deep learning models, especially CNN-based architectures, in accurately predicting procedure end times.

Dynamic modeling of astrocyte-neuron interactions under the influence of Aβ deposition

β-amyloid (Aβ) protein accumulation is recognized as a key factor in Alzheimer's disease (AD) pathogenesis. Its effects on astrocyte function appear primarily as disturbances to intracellular calcium signaling, which, in turn, affects neuronal excitability. We propose an innovative neuron-astrocyte interaction model to examine how Aβ accumulation influences astrocyte calcium oscillation and neuronal excitability, emphasizing its significance in AD pathogenesis. This comprehensive model describes not only the response of the astrocyte to presynaptic neuron stimulation but also the release of the downstream signaling glutamate and its consequential feedback on neurons. Our research concentrates on changes within two prominent pathways affected by Aβ: the creation of Aβ astrocyte membrane pores and the enhanced sensitivity of ryanodine receptors. By incorporating these adjustments into our astrocyte model, we can reproduce previous experimental findings regarding aberrant astrocyte calcium activity and neural behavior associated with Aβ from a neural computational viewpoint. Within a specified range of Aβ influence, our numerical analysis reveals that astrocyte cytoplasmic calcium rises, calcium oscillation frequency increases, and the time to the first calcium peak shortens, indicating the disrupted astrocyte calcium signaling. Simultaneously, the neuronal firing rate and cytosolic calcium concentration increase while the threshold current for initiating repetitive firing diminishes, implying heightened neuronal excitability. Given that increased neuronal excitability commonly occurs in early AD patients and correlates with cognitive decline, our findings may highlight the importance of Aβ accumulation in AD pathogenesis and provide a theoretical basis for identifying neuronal markers in the early stages of the disease.

Leveraging patients' longitudinal data to improve the Hospital One-year Mortality Risk

Purpose

Predicting medium-term survival after admission is necessary for identifying end-of-life patients who may benefit from goals of care (GOC) discussions. Considering that several patients have multiple hospital admissions, this study leverages patients' longitudinal data and information collected routinely at admission to predict the Hospital One-year Mortality Risk.

Methods

We propose the Ensemble Longitudinal Network (ELN) to predict one-year mortality using patients' longitudinal records. The model was evaluated: (i) with only predictors reported upon admission (AdmDemo); and (ii) also with diagnoses available later during patients' stay (AdmDemoDx). Using records of 123,646 patients with 250,812 hospitalizations from 2011 to 2021, our dataset was split into a learning set (2011-2017) to compare models with and without longitudinal information using nested cross-validation, and a holdout set (2017-2021) to assess clinical utility towards GOC discussions.

Results

The ELN achieved a significant increase in predictive performance using longitudinal information (p-value < 0.05) for both the AdmDemo and AdmDemoDx predictors. For randomly selected hospitalizations in the holdout set, the ELN showed: (i) AUROCs of 0.83 (AdmDemo) and 0.87 (AdmDemoDx); and (ii) superior decision-making properties, notably with an increase in precision from 0.25 for the standard process to 0.28 (AdmDemo) and 0.36 (AdmDemoDx). Feature importance analysis confirmed that the utility of the longitudinal information increases with the number of patient hospitalizations.

Conclusion

Integrating patients' longitudinal data provides better insights into the severity of illness and the overall patient condition, in particular when limited information is available during their stay.

Margin-aware optimized contrastive learning for enhanced self-supervised histopathological image classification

Histopathological images, characterized by their high resolution and intricate cellular structures, present unique challenges for automated analysis. Traditional supervised learning-based methods often rely on extensive labeled datasets, which are labour-intensive and expensive. In learning representations, self-supervised learning techniques have shown promising outcomes directly from raw image data without manual annotations. In this paper, we propose a novel margin-aware optimized contrastive learning approach to enhance representation learning from histopathological images using a self-supervised approach. The proposed approach optimizes contrastive learning with a margin-based strategy to effectively learn discriminative representations while enforcing a semantic similarity threshold. In the proposed loss function, a margin is used to enforce a certain level of similarity between positive pairs in the embedding space, and a scaling factor is introduced to adjust the sensitivity of the loss, thereby enhancing the discriminative capacity of the learned representations. Our approach demonstrates robust generalization in in- and out-domain settings through comprehensive experimental evaluations conducted on five distinct benchmark histopathological datasets belonging to three cancer types. The results obtained on different experimental settings show that the proposed approach outmatched the state-of-the-art approaches in cross-domain and cross-disease settings.

Artificial intelligence in kidney transplantation: a 30-year bibliometric analysis of research trends, innovations, and future directions

Kidney transplantation is the definitive treatment for end-stage renal disease (ESRD), yet challenges persist in optimizing donor-recipient matching, postoperative care, and immunosuppressive strategies. This study employs bibliometric analysis to evaluate 890 publications from 1993 to 2023, using tools such as CiteSpace and VOSviewer, to identify global trends, research hotspots, and future opportunities in applying artificial intelligence (AI) to kidney transplantation. Our analysis highlights the United States as the leading contributor to the field, with significant outputs from Mayo Clinic and leading authors like Cheungpasitporn W. Key research themes include AI-driven advancements in donor matching, deep learning for post-transplant monitoring, and machine learning algorithms for personalized immunosuppressive therapies. The findings underscore a rapid expansion in AI applications since 2017, with emerging trends in personalized medicine, multimodal data fusion, and telehealth. This bibliometric review provides a comprehensive resource for researchers and clinicians, offering insights into the evolution of AI in kidney transplantation and guiding future studies toward transformative applications in transplantation science.

Augmentation of PM1.0 measurements based on machine learning model and environmental factors

PM1.0, particulate matter with an aerodynamic diameter smaller than 1.0 µm, can adversely affect human health. However, fewer stations are capable of measuring PM1.0 concentrations than PM2.5 and PM10 concentrations in real time (i.e., only 9 locations for PM1.0 vs. 623 locations for PM2.5 or PM10) in South Korea, making it impossible to conduct a nationwide health risk analysis of PM1.0. Thus, this study aimed to develop a PM1.0 prediction model using a random forest algorithm based on PM1.0 data from the nine measurement stations and various environmental input factors. Cross validation, in which the model was trained in eight stations and tested in the remaining station, achieved an average R2 of 0.913. The high R2 value achieved under mutually exclusive training and test locations in the cross validation can be ascribed to the fact that all the locations had similar relationships between PM1.0 and the input factors, which were captured by our model. Moreover, results of feature importance analysis showed that PM2.5 and PM10 concentrations were the two most important input features in predicting PM1.0 concentration. Finally, the model was used to estimate the PM1.0 concentrations in 623 locations, where input factors such as PM2.5 and PM10 can be obtained. Based on the augmented profile, we identified Seoul and Ansan to be PM1.0 concentration hotspots. These regions are large cities or the center of anthropogenic and industrial activities. The proposed model and the augmented PM1.0 profiles can be used for large epidemiological studies to understand the health impacts of PM1.0.

Using longitudinal data and deep learning models to enhance resource allocation in home-based medical care

BACKGROUND

The aging population is driving increased healthcare demands and costs, prompting the need for effective home healthcare programs. Accurate patient assessment is essential for optimizing resource allocation and tailoring services.

OBJECTIVE

This retrospective study explores the application of artificial intelligence (AI) in predicting home medical care stages to enhance care delivery.

METHODS

Data from Taipei City Hospital (2015-2021) included inpatient, outpatient, and home medical care records. Three deep learning (DL) models-Transformer encoder-based, long short-term memory (LSTM), and gated recurrent unit (GRU)-were compared with three baseline machine learning (ML) models. Models were trained on 3, 5, and 10 consecutive visits for binary and multiclass classification. Performance was evaluated using accuracy, precision, recall, and the area under the receiver operating characteristic curve (AUC).

RESULTS

The study included 4,343 patients with a mean age of 85.04 ± 11.47 years. While models trained on 10 visits generally exhibited higher performance, data from 5 visits were sufficient for accurate predictions. With five visits, the LSTM model achieved the highest AUC (0.908) for distinguishing between the absence (S0) and presence (S1-S3) of home medical care. Meanwhile, the Transformer achieved the best AUC (0.86) for classifying S0-S3, with individual stage AUCs of 0.90, 0.82, 0.81, and 0.94 for S0, S1, S2, and S3, respectively.

CONCLUSIONS

AI deep learning models show strong potential for accurately predicting home medical care stages. The best-performing model could be a promising tool for healthcare professionals to optimize resource allocation in home medical care settings.

Neuronal correlates of sleep in honey bees

Honey bees Apis mellifera follow the day-night cycle for their foraging activity, entering rest periods during darkness. Despite considerable research on sleep behaviour in bees, its underlying neurophysiological mechanisms are not well understood, partly due to the lack of brain imaging data that allow for analysis from a network- or system-level perspective. This study aims to fill this gap by investigating whether neuronal activity during rest periods exhibits stereotypic patterns comparable to sleep signatures observed in vertebrates. Using two-photon calcium imaging of the antennal lobes (AL) in head-fixed bees, we analysed brain dynamics across motion and rest epochs during the nocturnal period. The recorded activity was computationally characterised, and machine learning was applied to determine whether a classifier could distinguish the two states after motion correction. Out-of-sample classification accuracy reached 93 %, and a feature importance analysis suggested network features to be decisive. Accordingly, the glomerular connectivity was found to be significantly increased in the rest-state patterns. A full simulation of the AL using a leaky spiking neural network revealed that such a transition in network connectivity could be achieved by weakly correlated input noise and a reduction of synaptic conductance of the inhibitive local neurons (LNs) which couple the AL network nodes. The difference in the AL response maps between awake- and sleep-like states generated by the simulation showed a decreased specificity of the odour code in the sleep state, suggesting reduced information processing during sleep. Since LNs in the bee brain are GABAergic, this suggests that the GABAergic system plays a central role in sleep regulation in bees as in many higher species including humans. Our findings support the theoretical view that sleep-related network modulation mechanisms are conserved throughout evolution, highlighting the bee's potential as an invertebrate model for studying sleep at the level of single neurons.

Modeling structured data learning with Restricted Boltzmann machines in the teacher-student setting

Restricted Boltzmann machines (RBM) are generative models capable to learn data with a rich underlying structure. We study the teacher-student setting where a student RBM learns structured data generated by a teacher RBM. The amount of structure in the data is controlled by adjusting the number of hidden units of the teacher and the correlations in the rows of the weights, a.k.a. patterns. In the absence of correlations, we validate the conjecture that the performance is independent of the number of teacher patterns and hidden units of the student RBMs, and we argue that the teacher-student setting can be used as a toy model for studying the lottery ticket hypothesis. Beyond this regime, we find that the critical amount of data required to learn the teacher patterns decreases with both their number and correlations. In both regimes, we find that, even with a relatively large dataset, it becomes impossible to learn the teacher patterns if the inference temperature used for regularization is kept too low. In our framework, the student can learn teacher patterns one-to-one or many-to-one, generalizing previous findings about the teacher-student setting with two hidden units to any arbitrary finite number of hidden units.

Memory Transmission Based Referring Video Object Segmentation

Referring Video Object Segmentation (RVOS) addresses the task of segmenting target objects described by textual descriptions from videos. In order to ensure the consistency of objects segmented from video frames, inter-frame modeling is adopted to capture the motion information of objects, which usually divides the video into several clips, and considers the association of video frames within each clip. However, the clip-level modeling cannot establish continuous motion changes of the object across the video. To address this issue, we suggest memory transmission based continuous inter-frame modeling, which uses the segmentation result of the previous frame to calculate a pseudo mask for the current frame. Based on the proposed continuous inter-frame modeling method, we propose Memory Transmission Based Referring Video Object Segmentation (MT-RVOS), which uses the transmitted pseudo mask to guide the segmentation mask inference for the current frame. Extensive experiments conducted on four referring video object segmentation benchmarks demonstrate that MT-RVOS achieves competitive performance.

Application of machine learning approaches for estimating carbon dioxide absorption capacity of a variety of blended imidazolium-based ionic liquids

Ionic liquids (ILs) have gained attention in recent times as potentially effective absorbents for CO2 emissions owing to the number of their notable attributes, including reduced volatility, enhanced thermal consistency etc. Due to the number of challenges of thermodynamic models in forecasting CO2 solubility in ILs under a variety of operating conditions, machine learning (ML) approaches have been developed as a result of the necessity for an alternate solution. Nevertheless, there are currently quite a few of forecasting techniques available for evaluating the solubility of CO2, specifically in combinations of imidazolium-based ILs. For this reason, the present study focuses on the utilization of molecular structure-based descriptors as an alternative chemistry concept for predicting the CO2 solubility in an imidazolium-based ILs mixture. This research utilized and contrasted 6 sophisticated machine learning models (AdaBoost-SVR, Extra trees, DT, CatBoost, LightGBM, XGBoost) to determine the most effective method for target parameter estimation. The study employed an exclusive and all-encompassing databank consisting of 43 imidazolium-based ILs, 26 input variables, and 4397 experimental data points in total. The remarkable 90 % overall accuracy consistently surpassed by all models serves as evidence of the ML methodologies' robustness and efficacy. The highest-performing approaches, XGBoost, exhibited a remarkable precision level of R2 being equal to 0.999 and RMSE of 0.0077. A comprehensive trend analysis was performed to assess the XGBoost model's performance across different operational scenarios such as molecular weight, temperature, water content, and pressure. The developed model proved to be capable of accurately detecting patterns in various operating conditions. By employing sensitivity analysis with SHAP values, it was observed that pressure, temperature, and molecular weight were the most impactful factors influencing the XGBoost model's predictions.

Visual reasoning in object-centric deep neural networks: A comparative cognition approach

Achieving visual reasoning is a long-term goal of artificial intelligence. In the last decade, several studies have applied deep neural networks (DNNs) to the task of learning visual relations from images, with modest results in terms of generalization of the relations learned. However, in recent years, object-centric representation learning has been put forward as a way to achieve visual reasoning within the deep learning framework. Object-centric models attempt to model input scenes as compositions of objects and relations between them. To this end, these models use several kinds of attention mechanisms to segregate the individual objects in a scene from the background and from other objects. In this work we tested relation learning and generalization in several object-centric models, as well as a ResNet-50 baseline. In contrast to previous research, which has focused heavily in the same-different task in order to asses relational reasoning in DNNs, we use a set of tasks - with varying degrees of complexity - derived from the comparative cognition literature. Our results show that object-centric models are able to segregate the different objects in a scene, even in many out-of-distribution cases. In our simpler tasks, this improves their capacity to learn and generalize visual relations in comparison to the ResNet-50 baseline. However, object-centric models still struggle in our more difficult tasks and conditions. We conclude that abstract visual reasoning remains an open challenge for DNNs, including object-centric models.

Automatic and precise identification of volatile organic compounds from gas chromatography in prolonged atmospheric monitoring

Long-term continuous monitoring of volatile organic compounds (VOCs) is pivotal for climate change research, air quality assessment, pollution source identification, and public health early warning systems. Prolonged VOC monitoring is routinely implemented by gas chromatographs. However, accurate identification of target contaminants heavily relies on time-consuming and error-prone manual processes conducted by professional personnel due to complex chromatograms and anomalous patterns. This study proposes an artificial intelligence-based model, ResGRU, for the automated and precise identification of VOCs in a chromatograph. By taking real data from a monitoring site in Shanghai, the model achieved a mean absolute error of 0.0144 min for retention time localization, which is 2.76 to 38.19 times smaller compared to conventional machine learning or deep learning models by previous reports. Moreover, it achieves precise recognition of subtle chromatographic peaks and exceptional adaptability to abnormal chromatograms. Notably, the vast majority of these weak peaks are attributed to olefinic compounds, which exhibit exceptionally high ozone formation potential. In addition, cross-transfer verification of data from four monitoring sites in Shanghai, Hubei, and Jiangsu, China further proved the robust transferability of this model. This work provides a novel methodology for precise analysis of GC data, enabling deeper exploration of the mechanisms behind VOCs pollution over extended temporal scales.

SSE-Net: A novel network based on sequence spatial equation for Camellia sinensis lysine acetylation identification

Lysine acetylation (Kace) is one of the most important post-translational modifications. It is key to identify Kace sites for understanding regulation mechanisms in Camellia sinensis. In this study, we defined a mathematical formula, named sequence spatial equation (SSE), which could give each amino acid coordinate in 3-D space by rotating and translating. Based on SSE, an optional network SSE-Net was constructed for representing spatial structure information. Centrality metrics of SSE-Net were used to design structure feature vectors for reflecting the importance of sites. The optimal features were fed into classifier to construct model SSE-ET. The results showed that SSE-ET outperformed the other classifiers. Meanwhile, all MCC results were higher than 0.7 for different machine learning, which indicated that SSE-Net was effective for representing Kace sites in Camellia sinensis. Moreover, we implemented the other models on our dataset. The results of comparison showed that SSE-ET was much more powerful than the others. Specifically, the result of SN was nearly 20 % higher than the other models. These results showed that the proposed SSE was a valuable mathematics concept for reflecting 3-D space Kace site information in Camellia sinensis, and SSE-Net may be an essential complementary for biology and bioinformatics research.

Fixed-time synchronization of proportional delay memristive complex-valued competitive neural networks

The fixed-time synchronization (FXS) is considered for memristive complex-valued competitive neural networks (MCVCNNs) with proportional delays. Two less conservative criteria supporting the FXS of MCVCNNs are founded by involving Lyapunov method and inequality techniques. Suitable switch controllers are designed by defining different norms of complex numbers instead of treating complex-valued neural networks as two real-valued systems. Furthermore, the settling time (ST) has been approximated. Finally, two simulations are shown to confirm the effectiveness of criteria in this paper and the outcomes of practical application in image protection.

An in vitro platform for characterizing axonal electrophysiology of individual human iPSC-derived nociceptors

Neuropathic pain is characterized by aberrant activity of specific nociceptor populations, as demonstrated through functional assessments such as microneurography. Current treatments against severe forms of neuropathic pain demonstrate insufficient efficacy or lead to unwanted side effects as they fail to specifically target the affected nociceptors. Tools that can recapitulate aspects of microneurography in vitro would enable a more targeted compound screening. Therefore, we developed an in vitro platform combining a CMOS-based high-density microelectrode array with a polydimethylsiloxane (PDMS) guiding microstructure that captures the electrophysiological responses of individual axons. Human induced pluripotent stem cell-derived (hiPSC) sensory neurons were cultured in a way that allowed axons to be distributed through parallel 4 ×10μm microchannels exiting the seeding well before converging to a bigger axon-collecting channel. This configuration allowed the measurement of stimulation-induced responses of individual axons. Sensory neurons were found to exhibit a great diversity of electrophysiological response profiles that can be classified into different functional archetypes. Moreover, we show that some responses are affected by applying the TRPV1 agonist capsaicin. Overall, results using our platform demonstrate that we were able to distinguish individual axon responses, making the platform a promising tool for testing therapeutic candidates targeting particular sensory neuron subtypes.

Grain protein function prediction based on improved FCN and bidirectional LSTM

With the development of high-throughput sequencing technologies, predicting grain protein function from amino acid sequences based on intelligent model has become one of the significant tasks in bioinformatics. The soybean, maize, indica, and japonica are selected as grain dataset from the UniProtKB. Aiming at the problem of neglecting the sequence order of amino acids and the long-term dependence between amino acids, the PBiLSTM-FCN model is proposed for predicting grain protein function in this paper. The sequence of amino acid sequences is considered in the Fully Convolutional Networks (FCN), and the long-term dependence between amino acids is addressed by the bidirectional Long Short-Term Memory network (BiLSTM). The experimental results show that the PBiLSTM-FCN model is superior to existing models, and can predict more accurately by solving the problem of capturing long-range dependencies and the order of amino acid sequences. Finally, the interpretability analyses are performed by the actual protein function compared with the predicted protein function which proves the effectiveness of the PBiLSTM-FCN model in predicting grain protein function.

On the use of a Transformer Neural Network to deconvolve ultrasonic signals

Pulse-echo ultrasonic techniques play a crucial role in assessing wall thickness deterioration in safety-critical industries. Current approaches face limitations with low signal-to-noise ratios, weak echoes, or vague echo patterns typical of heavily corroded profiles. This study proposes a novel combination of Convolution Neural Networks (CNN) and Transformer Neural Networks (TNN) to improve thickness gauging accuracy for complex geometries and echo patterns. Recognizing the strength of TNN in language processing and speech recognition, the proposed network comprises three modules: 1. pre-processing CNN, 2. a Transformer model and 3. a post-processing CNN. Two datasets, one being simulation-generated, and the other, experimentally gathered from a corroded carbon steel staircase specimen, support the training and testing processes. Results indicate that the proposed model outperforms other AI architectures and traditional methods, providing a 5.45% improvement over CNN architectures from NDE literature, a 1.81% improvement over ResNet-50, and a 17.5% improvement compared to conventional thresholding techniques in accurately detecting depths with a precision under 0.5λ.

Robust one-class support vector machine

One-Class Support Vector Machine (OCSVM) is an effective algorithm in one-class classification task. However, it exhibits sensitivity to noise and outliers. Current solutions often employ bounded loss functions that impose finite but relatively large penalties on noise or outliers, and these loss functions suffer from limitations of discontinuity and non-differentiability. To address these issues, this paper introduces a novel, continuous, smooth, and differentiable loss function, namely Quadratic Type Squared Error Loss Function (QTSELF), and proposes a more robust OCSVM (Q-OCSVM). Q-OCSVM not only differentiates samples based on their positions and applies distinct treatments accordingly but also enhances model robustness by imposing minimal penalties on noise and outliers. Moreover, the elegant mathematical properties of the loss function facilitate model optimization. Theoretical analysis utilizes Rademacher complexity theory to conduct the generalization error bound of the model. Momentum method is used to optimize Q-OCSVM. Extensive experiments convincingly demonstrate that Q-OCSVM outperforms the benchmark techniques.

The body intervenes: How active inference explains depression's clinical presentation

The low mood that characterises depression is accompanied by changes in bodily processes, manifested in symptoms such as insomnia, reduced appetite and fatigue. The active inference framework provides an explanation as to how mood-related symptoms are linked. It suggests that affective experiences arise from predictions about interoceptive states and their corresponding prediction errors, with the relative influence of each modified by precision weighting. Moods reflect long-term predictions about the state of the body, incorporating parameters related to sleep, appetite and energy levels. Depression emerges from the interplay between reduced confidence in long-term prospects and heightened expectation of shorter-term negative affect, which sees a re-weighting of the precision of interoceptive prediction errors. The ensuing bodily changes contribute to the emergence of depressed mood; and underpin disturbances in shorter-term interoceptive predictions and the experience of emotions such as anxiety and irritability. This framework details how interoceptive processes shape the phenomenological and symptomatic experience of depression, helping us to understand the disorder's multifaceted and often idiosyncratic clinical presentation, and with implications for the way we understand and treat depression and its co-morbidities.

Estimating uncertainty from feed-forward network based sensing using quasi-linear approximation

A fundamental problem in neural network theory is the quantification of uncertainty as it propagates through these constructs. Such quantification is crucial as neural networks become integrated into broader engineered systems that render decisions based on their outputs. In this paper, we engage the problem of estimating uncertainty in feedforward neural network constructs. Mathematically, the problem, in essence, amounts to understanding how the moments of an input distribution become modifies as they move through network layers. Despite its straightforward formulation, the nonlinear nature of modern feedforward architectures makes this is a mathematically challenging problem. Most contemporary approaches rely on some form of Monte Carlo sampling to construct inter-laminar distributions. Here, we borrow an approach from the control systems community known as quasilinear approximation, to enable a more analytical approach to the uncertainty quantification problem in this setting. Specifically, by using quasilinear approximation, nonlinearities are linearized in terms of the expectation of their gain in an input-output sense. We derive these expectations for several commonly used nonlinearities, under the assumption of Gaussian inputs. We then establish that the ensuing approximation is accurate relative to traditional linearization. Furthermore, we provide a rigorous example how this method can enable formal estimation of uncertainty in latent variables upstream of the network, within a target-tracking case study.

Saccade and purify: Task adapted multi-view feature calibration network for few shot learning

Current few-shot image classification methods encounter challenges in extracting multi-view features that can complement each other and selecting optimal features for classification in a specific task. To address this problem, we propose a novel Task-adapted Multi-view feature Calibration Network (TMCN) inspired by the different saccade patterns observed in the human visual system. The TMCN is designed to "saccade" for extracting complementary multi-view features and "purify" multi-view features in a task-adapted manner. To capture more representative features, we propose a multi-view feature extraction method that simulates the voluntary saccades and scanning saccades in the human visual system, which generates global, local grid, and randomly sampled multi-view features. To purify and obtain the most appropriate features, we employ a global local feature calibration module to calibrate global and local grid features for achieving more stable non-local image features. Furthermore, a sampling feature fusion method is proposed to fuse the randomly sampled features from classes to obtain better prototypes, and a multi-view feature calibrating module is proposed to adaptively fuse purified multi-view features based on the task information obtained from the task feature extracting module. Extensive experiments conducted on three widely used public datasets prove that our proposed TMCN can achieve excellent performance and surpass state-of-the-art methods. The code is available at the following address: https://github.com/huyunzuo/TMCN.

SpikeCLIP: A contrastive language-image pretrained spiking neural network

Spiking Neural Networks (SNNs) have emerged as a promising alternative to conventional Artificial Neural Networks (ANNs), demonstrating comparable performance in both visual and linguistic tasks while offering the advantage of improved energy efficiency. Despite these advancements, the integration of linguistic and visual features into a unified representation through spike trains poses a significant challenge, and the application of SNNs to multimodal scenarios remains largely unexplored. This paper presents SpikeCLIP, a novel framework designed to bridge the modality gap in spike-based computation. Our approach employs a two-step recipe: an "alignment pre-training" to align features across modalities, followed by a "dual-loss fine-tuning" to refine the model's performance. Extensive experiments reveal that SNNs achieve results on par with ANNs while substantially reducing energy consumption across various datasets commonly used for multimodal model evaluation. Furthermore, SpikeCLIP maintains robust image classification capabilities, even when dealing with classes that fall outside predefined categories. This study marks a significant advancement in the development of energy-efficient and biologically plausible multimodal learning systems.

Opposing-through crash risk forecasting using artificial intelligence-based video analytics for real-time application: integrating generalized extreme value theory and time series forecasting models

Recent advancements in artificial intelligence (AI) and traffic sensing technologies provide significant opportunities for real-time crash risk forecasting. While forecasting based on historical crash data yields macroscopic insights into future crash risks, such information is often insufficient for real-time applications. In contrast, traffic conflict techniques (TCTs) leveraged by extreme value theory (EVT) and AI-based video analytics have enabled crash risk estimation to a granular level, presenting a promising potential for real-time applications. This study develops a unified framework of integrating generalized extreme value (GEV) theory with parametric and non-parametric forecasting models to predict opposing-through crash risks at signalized intersections. A deep neural network-based computer vision technique was employed to extract post encroachment time (PET) traffic conflicts from 97 h of video footage. Crash risks were estimated using a non-stationary GEV model, incorporating traffic conflict counts, speed variations, and signal timing characteristics. These risk estimates were then forecasted using autoregressive integrated moving average (ARIMA), gated recurrent unit (GRU), and long short-term memory (LSTM) models to analyze short-term crash trends. Results show that the mean crash frequency estimates fell within the 95 % confidence limits of observed crashes and confirm the adequacy of the developed EVT model in estimating opposing-through crashes. The autoregressive and recurrent neural network models exhibit similar forecasting accuracy for crash risk forecasting, with reliable predictions extending up to 11 future signal cycles. The proposed real-time crash risk forecasting framework can be a crucial component of an intelligent transport system, leading to proactive safety management for signalized intersections.

TSCMamba: Mamba Meets Multi-View Learning for Time Series Classification

Multivariate time series classification (TSC) is critical for various applications in fields such as healthcare and finance. While various approaches for TSC have been explored, important properties of time series, such as shift equivariance and inversion invariance, are largely underexplored by existing works. To fill this gap, we propose a novel multi-view approach to capture patterns with properties like shift equivariance. Our method integrates diverse features, including spectral, temporal, local, and global features, to obtain rich, complementary contexts for TSC. We use continuous wavelet transform to capture time-frequency features that remain consistent even when the input is shifted in time. These features are fused with temporal convolutional or multilayer perceptron features to provide complex local and global contextual information. We utilize the Mamba state space model for efficient and scalable sequence modeling and capturing long-range dependencies in time series. Moreover, we introduce a new scanning scheme for Mamba, called tango scanning, to effectively model sequence relationships and leverage inversion invariance, thereby enhancing our model's generalization and robustness. Experiments on two sets of benchmark datasets (10+20 datasets) demonstrate our approach's effectiveness, achieving average accuracy improvements of 4.01-6.45% and 7.93% respectively, over leading TSC models such as TimesNet and TSLANet. The code is available at: https://drive.google.com/file/d/1fScmALgreb_sE9_P2kIsQCmt9SNxp7GP/view?usp=sharing.

TCH: A novel multi-view dimensionality reduction method based on triple contrastive heads

Multi-view dimensionality reduction (MvDR) is a potent approach for addressing the high-dimensional challenges in multi-view data. Recently, contrastive learning (CL) has gained considerable attention due to its superior performance. However, most CL-based methods focus on promoting consistency between any two cross views from the perspective of subspace samples, which extract features containing redundant information and fail to capture view-specific discriminative information. In this study, we propose feature- and recovery-level contrastive losses to eliminate redundant information and capture view-specific discriminative information, respectively. Based on this, we construct a novel MvDR method based on triple contrastive heads (TCH). This method combines sample-, feature-, and recovery-level contrastive losses to extract sufficient yet minimal subspace discriminative information in accordance with the information bottleneck principle. Furthermore, the relationship between TCH and mutual information is revealed, which provides the theoretical support for the outstanding performance of our method. Our experiments on five real-world datasets show that the proposed method outperforms existing methods.

Linking cellular-level phenomena to brain architecture: the case of spiking cerebellar controllers

Linking cellular-level phenomena to brain architecture and behavior is a holy grail for theoretical and computational neuroscience. Advances in neuroinformatics have recently allowed scientists to embed spiking neural networks of the cerebellum with realistic neuron models and multiple synaptic plasticity rules into sensorimotor controllers. By minimizing the distance (error) between the desired and the actual sensory state, and exploiting the sensory prediction, the cerebellar network acquires knowledge about the body-environment interaction and generates corrective signals. In doing so, the cerebellum implements a generalized computational algorithm, allowing it "to learn to predict the timing between correlated events" in a rich set of behavioral contexts. Plastic changes evolve trial by trial and are distributed over multiple synapses, regulating the timing of neuronal discharge and fine-tuning high-speed movements on the millisecond timescale. Thus, spiking cerebellar built-in controllers, among various computational approaches to studying cerebellar function, are helping to reveal the cellular-level substrates of network learning and signal coding, opening new frontiers for predictive computing and autonomous learning in robots.

Universality of reservoir systems with recurrent neural networks

Approximation capability of reservoir systems whose reservoir is a recurrent neural network (RNN) is discussed. We show what we call uniform strong universality of RNN reservoir systems for a certain class of dynamical systems. This means that, given an approximation error to be achieved, one can construct an RNN reservoir system that approximates each target dynamical system in the class just via adjusting its linear readout. To show the universality, we construct an RNN reservoir system via parallel concatenation that has an upper bound of approximation error independent of each target in the class.

A fractional-order multi-delayed bicyclic crossed neural network: Stability, bifurcation, and numerical solution

In this paper, we propose a fractional-order bicyclic crossed neural network (NN) with multiple time delays consisting of two sharing neurons between rings. The given fractional-order NN is defined in terms of the Caputo fractional derivatives. We prove boundedness and the existence of a unique solution for the proposed NN. We do the stability and the onset of Hopf bifurcation analyses by converting the proposed multiple-delayed NN into a single-delay NN. Later, we numerically solve the proposed NN with the help of the L1 predictor-corrector algorithm and justify the theoretical results with graphical simulations. We explore that the time delay and the order of the derivative both influence the stability and bifurcation of the fractional-order NN. The proposed fractional-order NN is a unique multi-delayed bicyclic crossover NN that has two sharing neurons between rings. Such ring structure appropriately mimics the information transmission process within intricate NNs.

Broad learning system based on fractional order optimization

Due to its efficient incremental learning performance, the broad learning system (BLS) has received widespread attention in the field of machine learning. Scholars have found in algorithm research that using the maximum correntropy criterion (MCC) can further improves the performance of broad learning in handling outliers. Recent studies have shown that differential equations can be used to represent the forward propagation of deep learning. The BLS based on MCC uses differentiation to optimize parameters, which indicates that differential methods can also be used for BLS optimization. But general methods use integer order differential equations, ignoring system information between integer orders. Due to the long-term memory property of fractional differential equations, this paper innovatively introduces fractional order optimization into the BLS, called FOBLS, to better enhance the data processing capability of the BLS. Firstly, a BLS is constructed using fractional order, incorporating long-term memory characteristics into the weight optimization process. In addition, constructing a dynamic incremental learning system based on fractional order further enhances the ability of network optimization. The experimental results demonstrate the excellent performance of the method proposed in this paper.

S4-KD: A single step spiking SiamFC+ + for object tracking with knowledge distillation

Spiking neural networks (SNNs), which transmit information through binary spikes, have the advantages of high efficiency and low energy consumption. At present, the multiple time steps of SNNs can lead to increased latency and power consumption. To this end, we propose Single Step Spiking SiamFC+ + (S4), an improved single-step end-to-end direct training target tracking framework that compresses the time step to 1 by temporal pruning, using AlexNet as the backbone network. Experimental results show that, even when only a single time step is used, the tracking performance of the proposed S4 is still comparable to the original Spiking SiamFC+ +. Furthermore, we introduce the knowledge distillation to improve the performance of the proposed S4, which is called S4-KD for clarity. Three kinds of distillation loss functions are designed for the S4-KD. An artificial neural network model based on the AlexNet network serves as the teacher model, while the temporal-pruned S4 model acts as the student model for retraining. Experimental results show that the S4-KD tracker achieves higher performance on several tracking benchmarks. More specifically, on the OTB100 dataset, Precision and Success are 0.871 and 0.657 respectively, on the UAV123 dataset, Precision and Success are 0.766 and 0.603 respectively, and on the VOT2018 dataset, A, R, and EAO are 0.582, 0.370, and 0.278 respectively. In addition, the estimated energy consumption of the S4-KD is only 34.6 % of that of the original Spiking SiamFC+ +. To the best of our knowledge, the proposed S4-KD tracker surpasses all the existing SNN-based object tracking methods, achieving state-of-the-art performance. Our codes will be available at https://github.com/PSNN-xd/S4-KD.

Linking behavioral deficits with underlying neural property changes in amblyopia

While it is widely accepted that abnormal visual experience during critical period can lead to significant functional deficits and altered neural property, the quantitative link between behavioral visual losses and underlying neural changes remains elusive. To address this gap, we systematically varied stimulus orientation and contrast to measure 2D psychometric functions of amblyopic and normally sighted participants at two different spatial frequencies. A biologically-interpretable neural population model explicitly incorporated with neural contrast response function (CRF) and orientation tuning property accounted for the complex performance data for both groups. Our results revealed that the poor performance in the amblyopic group can be excellently explained by a rightward-shifted CRF at higher spatial frequency and reduced population Fisher information for coding orientation. Moreover, regression analysis revealed that the behavior contrast threshold from an independent measurement significantly depended on the neural properties estimated by the model. This study demonstrates the potential of biologically-interpretable models to quantitatively bridge the gap between behavioral deficits and underlying neural changes, offering a promising tool for understanding normal and abnormal visual systems.

Artificial intelligence in bacterial diagnostics and antimicrobial susceptibility testing: Current advances and future prospects

Recently, artificial intelligence (AI) has emerged as a transformative tool, enhancing the speed, accuracy, and scalability of bacterial diagnostics. This review explores the role of AI in revolutionizing bacterial detection and antimicrobial susceptibility testing (AST) by leveraging machine learning models, including Random Forest, Support Vector Machines (SVM), and deep learning architectures such as Convolutional Neural Networks (CNNs) and transformers. The integration of AI into these methods promises to address the current limitations of traditional techniques, offering a path toward more efficient, accessible, and reliable diagnostic solutions. In particular, AI-based approaches have demonstrated significant potential in resource-limited settings by enabling cost-effective and portable diagnostic solutions, reducing dependency on specialized infrastructure, and facilitating remote bacterial detection through smartphone-integrated platforms and telemedicine applications. This review highlights AI's transformative role in automating data analysis, minimizing human error, and delivering real-time diagnostic results, ultimately improving patient outcomes and optimizing healthcare efficiency. In addition, we not only examine the current advances in machine learning and deep learning but also review their applications in plate counting, mass spectrometry, morphology-based and motion-based microscopic detection, holographic microscopy, colorimetric and fluorescence detection, electrochemical sensors, Raman and Surface-Enhanced Raman Spectroscopy (SERS), and Atomic Force Microscopy (AFM) for bacterial diagnostics and AST. Finally, we discuss the future directions and potential advancements in AI-driven bacterial diagnostics.

Physiological noise: a comprehensive review on informative randomness in neural systems

Noise is often regarded as mere interference in the analysis of biomedical signals. Nonetheless, stochasticity plays a critical and informative role in the dynamics of complex systems, particularly in neurocardiovascular and neural systems. This review provides a comprehensive exploration on informative randomness in physiological contexts, tracing the evolution of noise research from its foundations on Brownian motion to its applications in neural systems, including the neuroautonomic regulation of cardiovascular dynamics. Key distinctions are made between output (measurement) noise and dynamic (intrinsic) noise, which directly influence the system behaviors at various levels. Several physiological noise identification techniques, such as stochastic differential equations, Bayesian methods, and Kalman filters, are evaluated in real-world scenarios. Special emphasis is placed on the role of physiological noise in multiscale neural systems, such as brain dynamics, neuronal communication, and heart-brain interactions, highlighting how it shapes complex functions. Furthermore, physiological noise is presented as a potential clinical biomarker, offering insights into the underlying structure and health of neural systems. Future research is encouraged to investigate multivariate noise estimation methods and their implications for understanding causality and systemic interactions in neurocardiovascular networks.

Chirp sensitivity and vowel coding in the inferior colliculus

The inferior colliculus (IC) is an important brain region to understand neural encoding of complex sounds due to its diverse sound-feature sensitivities, including features that are affected by peripheral nonlinearities. Recent physiological studies in rabbit IC demonstrate that IC neurons are sensitive to chirp direction and velocity. Fast spectrotemporal changes, known as chirps, are contained within pitch-periods of natural vowels. Here, we use a combination of physiological and modeling strategies to assess the impact of chirp-sensitivity on vowel coding. Neural responses to vowel stimuli were recorded and vowel-token identification was evaluated based on average-rate and spike-timing metrics. Response timing was found to result in higher identification accuracy than rate. Additionally, rate bias towards low-velocity chirps, independent of chirp direction, was shown to correlate with higher vowel-identification accuracy based on timing. Also, direction bias in response to chirps of high velocity was shown to correlate with vowel-identification accuracy based on both rate and timing. Responses to natural-vowel tokens of individual neurons were simulated using an IC model with controllable chirp sensitivity. Responses of upward-biased, downward-biased, and non-selective model neurons were generated. Manipulating chirp sensitivity influenced response profiles across natural vowel tokens and vowel discrimination based on model-neuron responses. More work is needed to match all features of model responses to those of physiological recordings.

Functional brain network dynamics of brooding in depression: Insights from real-time fMRI neurofeedback

BACKGROUND

Brooding is a critical symptom and prognostic factor of major depressive disorder (MDD), which involves passively dwelling on self-referential dysphoria and related abstractions. The neurobiology of brooding remains under characterized. We aimed to elucidate neural dynamics underlying brooding, and explore their responses to neurofeedback intervention in MDD.

METHODS

We investigated functional MRI (fMRI) dynamic functional network connectivity (dFNC) in 36 MDD subjects and 26 healthy controls (HCs) during rest and brooding. Rest was measured before and after fMRI neurofeedback (MDD-active/sham: n = 18/18, HC-active/sham: n = 13/13). Baseline brooding severity was recorded using Ruminative Response Scale - Brooding subscale (RRS-B).

RESULTS

Four recurrent dFNC states were identified. Measures of time spent were not significantly different between MDD and HC for any of these states during brooding or rest. RRS-B scores in MDD showed significant negative correlation with measures of time spent in dFNC state 3 during brooding (r = -0.4, p = 0.002, FDR-significant). This state comprises strong connections spanning several brain systems involved in sensory, attentional and cognitive processing. Time spent in this anti-brooding dFNC state significantly increased following neurofeedback only in the MDD active group (z = -2.09, FWE-p = 0.034).

LIMITATIONS

The sample size was small and imbalanced between groups. Brooding condition was not examined post-neurofeedback.

CONCLUSION

We identified a densely connected anti-brooding dFNC brain state in MDD. MDD subjects spent significantly longer time in this state after active neurofeedback intervention, highlighting neurofeedback's potential for modulating dysfunctional brain dynamics to treat MDD.

A multi-task neural network for full waveform ultrasonic bone imaging

BACKGROUND AND OBJECTIVE

It is a challenging task to use ultrasound for bone imaging, as the bone tissue has a complex structure with high acoustic impedance and speed-of-sound (SOS). Recently, full waveform inversion (FWI) has shown promising imaging for musculoskeletal tissues. However, the FWI showed a limited ability and tended to produce artifacts in bone imaging because the inversion process would be more easily trapped in local minimum for bone tissue with a large discrepancy in SOS distribution between bony and soft tissues. In addition, the application of FWI required a high computational burden and relatively long iterations. The objective of this study was to achieve high-resolution ultrasonic imaging of bone using a deep learning-based FWI approach.

METHOD

In this paper, we proposed a novel network named CEDD-Unet. The CEDD-Unet adopts a Dual-Decoder architecture, with the first decoder tasked with reconstructing the SOS model, and the second decoder tasked with finding the main boundaries between bony and soft tissues. To effectively capture multi-scale spatial-temporal features from ultrasound radio frequency (RF) signals, we integrated a Convolutional LSTM (ConvLSTM) module. Additionally, an Efficient Multi-scale Attention (EMA) module was incorporated into the encoder to enhance feature representation and improve reconstruction accuracy.

RESULTS

Using the ultrasonic imaging modality with a ring array transducer, the performance of CEDD-Unet was tested on the SOS model datasets from human bones (noted as Dataset1) and mouse bones (noted as Dataset2), and compared with three classic reconstruction architectures (Unet, Unet++, and Att-Unet), four state-of-the-art architecture (InversionNet, DD-Net, UPFWI, and DEFE-Unet). Experiments showed that CEDD-Unet outperforms all competing methods, achieving the lowest MAE of 23.30 on Dataset1 and 25.29 on Dataset2, the highest SSIM of 0.9702 on Dataset1 and 0.9550 on Dataset2, and the highest PSNR of 30.60 dB on Dataset1 and 32.87 dB on Dataset2. Our method demonstrated superior reconstruction quality, with clearer bone boundaries, reduced artifacts, and improved consistency with ground truth. Moreover, CEDD-Unet surpasses traditional FWI by producing sharper skeletal SOS reconstructions, reducing computational cost, and eliminating the reliance for an initial model. Ablation studies further confirm the effectiveness of each network component.

CONCLUSION

The results suggest that CEDD-Unet is a promising deep learning-based FWI method for high-resolution bone imaging, with the potential to reconstruct accurate and sharp-edged skeletal SOS models.

Machine learning-based prediction of ambient CO2 and CH4 concentrations with high temporal resolution in Seoul metropolitan area

Machine learning has the potential to support the growing need for high-resolution greenhouse gas monitoring in urban and industrial environments, where deploying extensive sensor networks is often limited by cost and operational challenges. This study presents a novel approach for estimating greenhouse gas (GHG) concentrations using routinely collected air quality and meteorological data from existing monitoring stations. Focusing on the Seoul metropolitan area in the Republic of Korea, we developed and evaluated three machine learning models - Random Forest, Long Short-Term Memory (LSTM), and an ensemble learning approach - to predict CO2 and CH4 concentrations without relying on additional GHG monitoring equipment. Among these, the ensemble learning model outperformed the individual models, consistently achieving lower error metrics, even in data-limited scenarios. Feature importance analysis identifies NO2, CO, O3, and temperature as key predictors of CO2 and CH4 level variations, highlighting the influence of combustion-related pollutants and photochemical processes. Cross-validation results confirm the model's out-of-sample capabilities; however, local factors, such as traffic density, industrial activities, and meteorology, can affect performance. Consequently, model retraining or transfer learning may be required when applying the model to new locations with comparable emission profiles or atmospheric conditions. These findings emphasize the importance of localized context in model application while also demonstrating the broader applicability of the approach. By utilizing data already available through urban monitoring networks, this study offers a scalable and cost-effective strategy to support high-resolution GHG monitoring and inform targeted climate policies in complex urban-industrial regions.

Mnemonic brain state engagement is diminished in healthy aging

Healthy older adults typically show impaired episodic memory - memory for when and where an event occurred. This selective episodic memory deficit may arise from differential engagement in the retrieval state, a brain state in which attention is focused internally in an attempt to access prior knowledge, and the encoding state, a brain state which supports the formation of new memories and that trades off with the retrieval state. We hypothesize that older adults are biased toward a retrieval state. We recorded scalp electroencephalography while young, middle-aged and older adults performed a memory task in which they were explicitly directed to either encode or retrieve on a given trial. We used multivariate pattern analysis of spectral activity to decode retrieval vs. encoding state engagement. We find that whereas all age groups can follow task demands to selectively engage in encoding or retrieval, mnemonic brain state engagement is diminished for older adults relative to young and middle-aged adults. These findings suggest that differential mnemonic state engagement may underlie age-related memory changes.

Dataset-free weight-initialization on restricted Boltzmann machine

In feed-forward neural networks, dataset-free weight-initialization methods such as LeCun, Xavier (or Glorot), and He initializations have been developed. These methods randomly determine the initial values of weight parameters based on specific distributions (e.g., Gaussian or uniform distributions) without using training datasets. To the best of the authors' knowledge, such a dataset-free weight-initialization method is yet to be developed for restricted Boltzmann machines (RBMs), which are probabilistic neural networks consisting of two layers. In this study, we derive a dataset-free weight-initialization method for Bernoulli-Bernoulli RBMs based on statistical mechanical analysis. In the proposed weight-initialization method, the weight parameters are drawn from a Gaussian distribution with zero mean. The standard deviation of the Gaussian distribution is optimized based on our hypothesis that a standard deviation providing a larger layer correlation (LC) between the two layers improves the learning efficiency. The expression of the LC is derived based on a statistical mechanical analysis. The optimal value of the standard deviation corresponds to the maximum point of the LC. The proposed weight-initialization method is identical to Xavier initialization in a specific case (i.e., when the sizes of the two layers are the same, the random variables of the layers are {-1,1}-binary, and all bias parameters are zero). The validity of the proposed weight-initialization method is demonstrated in numerical experiments using a toy dataset and real-world datasets.

Entrainment by transcranial alternating current stimulation: Insights from models of cortical oscillations and dynamical systems theory

Signature of neuronal oscillations can be found in nearly every brain function. However, abnormal oscillatory activity is linked with several brain disorders. Transcranial alternating current stimulation (tACS) is a non-invasive brain stimulation technique that can potentially modulate neuronal oscillations and influence behavior both in health and disease. Yet, a complete understanding of how interacting networks of neurons are affected by tACS remains elusive. Entrainment effects by which tACS synchronizes neuronal oscillations is one of the main hypothesized mechanisms, as evidenced in animals and humans. Computational models of cortical oscillations may shed light on the entrainment effects of tACS, but current modeling studies lack specific guidelines to inform experimental investigations. This study addresses the existing gap in understanding the mechanisms of tACS effects on rhythmogenesis within the brain by providing a comprehensive overview of both theoretical and experimental perspectives. We explore the intricate interactions between oscillators and periodic stimulation through the lens of dynamical systems theory. Subsequently, we present a synthesis of experimental findings that demonstrate the effects of tACS on both individual neurons and collective oscillatory patterns in animal models and humans. Our review extends to computational investigations that elucidate the interplay between tACS and neuronal dynamics across diverse cortical network models. To illustrate these concepts, we conclude with a simple oscillatory neuron model, showcasing how fundamental theories of oscillatory behavior derived from dynamical systems, such as phase response of neurons to external perturbation, can account for the entrainment effects observed with tACS. Studies reviewed here render the necessity of integrated experimental and computational approaches for effective neuromodulation by tACS in health and disease.

A novel STA-EEGNet combined with channel selection for classification of EEG evoked in 2D and 3D virtual reality

Virtual reality (VR), particularly through 3D presentations, significantly boosts user engagement and task efficiency in fields such as gaming, education, and healthcare, offering more immersive and interactive experiences than traditional 2D formats. This study investigates EEG classification in response to 2D and 3D VR stimuli to deepen our understanding of the neural mechanisms driving VR interactions, with implications for brain-computer interfaces (BCIs). We introduce STA-EEGNet, an innovative model that enhances EEGNet by incorporating spatial-temporal attention (STA), improving EEG signal classification from VR environments. A one-way analysis of variance (ANOVA) was utilized to optimize channel selection, enhancing model accuracy. Comparative experiments showed that STA-EEGNet surpassed traditional EEGNet, achieving a peak accuracy of 99.78 % with channel selection. These findings highlight the benefits of spatial-temporal attention and optimal channel selection in classifying VR-evoked EEG data. This study underscores the importance of integrating spatial-temporal attention with compact convolutional neural networks like EEGNet, not only improving EEG signal classification but also advancing neural decoding and optimizing BCI applications.