Frontiers in Computational Neuroscience最新文献

We present a Spiking Neural Network (SNN) model that incorporates learnable synaptic delays through two approaches: per-synapse delay learning via Dilated Convolutions with Learnable Spacings (DCLS) and a dynamic pruning strategy that also serves as a form of delay learning. In the latter approach, the network dynamically selects and prunes connections, optimizing the delays in sparse connectivity settings. We evaluate both approaches on the Raw Heidelberg Digits keyword spotting benchmark using Backpropagation Through Time with surrogate gradients. Our analysis of the spatio-temporal structure of synaptic interactions reveals that, after training, excitation and inhibition group together in space and time. Notably, the dynamic pruning approach, which employs DEEP R for connection removal and RigL for reconnection, not only preserves these spatio-temporal patterns but outperforms per-synapse delay learning in sparse networks. Our results demonstrate the potential of combining delay learning with dynamic pruning to develop efficient SNN models for temporal data processing. Moreover, the preservation of spatio-temporal dynamics throughout pruning and rewiring highlights the robustness of these features, providing a solid foundation for future neuromorphic computing applications.

It is a universal phenomenon for patients who do not know which clinical department to register in large general hospitals. Although triage nurses can help patients, due to the larger number of patients, they have to stand in a queue for minutes to consult. Recently, there have already been some efforts to devote deep-learning techniques or pre-trained language models (PLMs) to triage recommendations. However, these methods may suffer two main limitations: (1) These methods typically require a certain amount of labeled or unlabeled data for model training, which are not always accessible and costly to acquire. (2) These methods have not taken into account the distortion of semantic feature structure and the loss of category discriminability in the model training. To overcome these limitations, in this study, we propose a cross-domain text classification method based on prompt-tuning, which can classify patients' questions or texts about their symptoms into several given categories to give suggestions on which kind of consulting room patients could choose. Specifically, first, different prompt templates are manually crafted based on various data contents, embedding source domain information into the prompt templates to generate another text with similar semantic feature structures for performing classification tasks. Then, five different strategies are employed to expand the label word space for modifying prompts, and the integration of these strategies is used as the final verbalizer. The extensive experiments on Chinese Triage datasets demonstrate that our method achieved state-of-the-art performance.

Background: Automatic sleep staging is essential for assessing sleep quality and diagnosing sleep disorders. While previous research has achieved high classification performance, most current sleep staging networks have only been validated in healthy populations, ignoring the impact of Obstructive Sleep Apnea (OSA) on sleep stage classification. In addition, it remains challenging to effectively improve the fine-grained detection of polysomnography (PSG) and capture multi-scale transitions between sleep stages. Therefore, a more widely applicable network is needed for sleep staging.

Methods: This paper introduces MSDC-SSNet, a novel deep learning network for automatic sleep stage classification. MSDC-SSNet transforms two channels of electroencephalogram (EEG) and one channel of electrooculogram (EOG) signals into time-frequency representations to obtain feature sequences at different temporal and frequency scales. An improved Transformer encoder architecture ensures temporal consistency and effectively captures long-term dependencies in EEG and EOG signals. The Multi-Scale Feature Extraction Module (MFEM) employs convolutional layers with varying dilation rates to capture spatial patterns from fine to coarse granularity. It adaptively fuses the weights of features to enhance the robustness of the model. Finally, multiple channel data are integrated to address the heterogeneity between different modalities effectively and alleviate the impact of OSA on sleep stages.

Results: We evaluated MSDC-SSNet on three public datasets and our collection of PSG records of 17 OSA patients. It achieved an accuracy of 80.4% on the OSA dataset. It also outperformed the state-of-the-art methods in terms of accuracy, F1 score, and Cohen's Kappa coefficient on the remaining three datasets.

Conclusion: The MSDC-SSRNet multi-channel sleep staging architecture proposed in this study enhances widespread system applicability by supplementing inter-channel features. It employs multi-scale attention to extract transition rules between sleep stages and effectively integrates multimodal information. Our method address the limitations of single-channel approaches, enhancing interpretability for clinical applications.

This study aims to enhance the classification accuracy of adverse events associated with the da Vinci surgical robot through advanced natural language processing techniques, thereby ensuring medical device safety and protecting patient health. Addressing the issues of incomplete and inconsistent adverse event records, we employed a deep learning model that combines BERT and BiLSTM to predict whether adverse event reports resulted in patient harm. We developed the Bert-BiLSTM-Att_dropout model specifically for text classification tasks with small datasets, optimizing the model's generalization ability and key information capture through the integration of dropout and attention mechanisms. Our model demonstrated exceptional performance on a dataset comprising 4,568 da Vinci surgical robot adverse event reports collected from 2013 to 2023, achieving an average F1 score of 90.15%, significantly surpassing baseline models such as GRU, LSTM, BiLSTM-Attention, and BERT. This achievement not only validates the model's effectiveness in text classification within this specific domain but also substantially improves the usability and accuracy of adverse event reporting, contributing to the prevention of medical incidents and reduction of patient harm. Furthermore, our research experimentally confirmed the model's performance, alleviating the data classification and analysis burden for healthcare professionals. Through comparative analysis, we highlighted the potential of combining BERT and BiLSTM in text classification tasks, particularly for small datasets in the medical field. Our findings advance the development of adverse event monitoring technologies for medical devices and provide critical insights for future research and enhancements.

Facial emotion recognition (FER) can serve as a valuable tool for assessing emotional states, which are often linked to mental health. However, mental health encompasses a broad range of factors that go beyond facial expressions. While FER provides insights into certain aspects of emotional well-being, it can be used in conjunction with other assessments to form a more comprehensive understanding of an individual's mental health. This research work proposes a framework for human FER using UNet image segmentation and transfer learning with the EfficientNetB4 model (called FacialNet). The proposed model demonstrates promising results, achieving an accuracy of 90% for six emotion classes (happy, sad, fear, pain, anger, and disgust) and 96.39% for binary classification (happy and sad). The significance of FacialNet is judged by extensive experiments conducted against various machine learning and deep learning models, as well as state-of-the-art previous research works in FER. The significance of FacialNet is further validated using a cross-validation technique, ensuring reliable performance across different data splits. The findings highlight the effectiveness of leveraging UNet image segmentation and EfficientNetB4 transfer learning for accurate and efficient human facial emotion recognition, offering promising avenues for real-world applications in emotion-aware systems and effective computing platforms. Experimental findings reveal that the proposed approach performs substantially better than existing works with an improved accuracy of 96.39% compared to existing 94.26%.

Inspired by animal navigation strategies, we introduce a novel computational model to navigate and map a space rooted in biologically inspired principles. Animals exhibit extraordinary navigation prowess, harnessing memory, imagination, and strategic decision-making to traverse complex and aliased environments adeptly. Our model aims to replicate these capabilities by incorporating a dynamically expanding cognitive map over predicted poses within an active inference framework, enhancing our agent's generative model plasticity to novelty and environmental changes. Through structure learning and active inference navigation, our model demonstrates efficient exploration and exploitation, dynamically expanding its model capacity in response to anticipated novel un-visited locations and updating the map given new evidence contradicting previous beliefs. Comparative analyses in mini-grid environments with the clone-structured cognitive graph model (CSCG), which shares similar objectives, highlight our model's ability to rapidly learn environmental structures within a single episode, with minimal navigation overlap. Our model achieves this without prior knowledge of observation and world dimensions, underscoring its robustness and efficacy in navigating intricate environments.

We adapt non-linear optimal control theory (OCT) to control oscillations and network synchrony and apply it to models of neural population dynamics. OCT is a mathematical framework to compute an efficient stimulation for dynamical systems. In its standard formulation, it requires a well-defined reference trajectory as target state. This requirement, however, may be overly restrictive for oscillatory targets, where the exact trajectory shape might not be relevant. To overcome this limitation, we introduce three alternative cost functionals to target oscillations and synchrony without specification of a reference trajectory. We successfully apply these cost functionals to single-node and network models of neural populations, in which each node is described by either the Wilson-Cowan model or a biophysically realistic high-dimensional mean-field model of exponential integrate-and-fire neurons. We compute efficient control strategies for four different control tasks. First, we drive oscillations from a stable stationary state at a particular frequency. Second, we switch between stationary and oscillatory stable states and find a translational invariance of the state-switching control signals. Third, we switch between in-phase and out-of-phase oscillations in a two-node network, where all cost functionals lead to identical OC signals in the minimum-energy limit. Finally, we (de-) synchronize an (a-) synchronously oscillating six-node network. In this setup, for the desynchronization task, we find very different control strategies for the three cost functionals. The suggested methods represent a toolbox that enables to include oscillatory phenomena into the framework of non-linear OCT without specification of an exact reference trajectory. However, task-specific adjustments of the optimization parameters have to be performed to obtain informative results.