International journal of neural systems最新文献

Neural Architecture Search (NAS) automates the design of deep neural networks but remains computationally expensive, particularly in multi-objective settings. Existing predictor-assisted evolutionary NAS methods suffer from slow convergence and rank disorder, which undermines prediction accuracy. To overcome these limitations, we propose CHENAS: a Classifier-assisted multi-objective Hybrid Evolutionary NAS framework. CHENAS combines the global exploration of evolutionary algorithms with the local refinement of gradient-based optimization to accelerate convergence and enhance solution quality. A novel dominance classifier predicts Pareto dominance relationships among candidate architectures, reframing multi-objective optimization as a classification task and mitigating rank disorder. To further improve efficiency, we employ a contrastive learning-based autoencoder that maps architectures into a continuous, structured latent space tailored for dominance prediction. Experiments on several benchmark datasets demonstrate that CHENAS outperforms state-of-the-art NAS approaches in identifying high-performing architectures across multiple objectives. Future work will focus on improving the computational efficiency of the framework and extending it to other application domains.

Areas, such as the identification of human activity, have accelerated thanks to the immense development of artificial intelligence (AI). However, the lack of data is a major obstacle to even faster progress. This is particularly true in computer vision, where training a model typically requires at least tens of thousands of images. Moreover, when the activity a researcher is interested in is far from the usual, such as falls, it is difficult to have a sufficiently large dataset. An example of this could be the identification of people suffering from a heart attack. In this sense, this work proposes a novel approach that relies on generative models to extend image datasets, adapting them to generate more domain-relevant images. To this end, a refinement to stable diffusion models was performed using low-rank adaptation. A dataset of 100 images of individuals simulating infarct situations and neutral poses was created, annotated, and used. The images generated with the adapted models were evaluated using learned perceptual image patch similarity to test their closeness to the target scenario. The results obtained demonstrate the potential of synthetic datasets, and in particular the strategy proposed here, to overcome data sparsity in AI-based applications. This approach can not only be more cost-effective than building a dataset in the traditional way, but also reduces the ethical concerns of its applicability in smart environments, health monitoring, and anomaly detection. In fact, all data are owned by the researcher and can be added and modified at any time without requiring additional permissions, streamlining their research.

Speech Emotion Recognition (SER) is becoming a key element of speech-based human-computer interfaces, endowing them with some form of empathy towards the emotional status of the human. Transformers have become a central Deep Learning (DL) architecture in natural language processing and signal processing, recently including audio signals for Automatic Speech Recognition (ASR) and SER. A central question addressed in this paper is the achievement of speaker-independent SER systems, i.e. systems that perform independently of a specific training set, enabling their deployment in real-world situations by overcoming the typical limitations of laboratory environments. This paper presents a comprehensive performance evaluation review of transformer architectures that have been proposed to deal with the SER task, carrying out an independent validation at different levels over the most relevant publicly available datasets for validation of SER models. The comprehensive experimental design implemented in this paper provides an accurate picture of the performance achieved by current state-of-the-art transformer models in speaker-independent SER. We have found that most experimental instances reach accuracies below 40% when a model is trained on a dataset and tested on a different one. A speaker-independent evaluation combining up to five datasets and testing on a different one achieves up to 58.85% accuracy. In conclusion, the SER results improved with the aggregation of datasets, indicating that model generalization can be enhanced by extracting data from diverse datasets.

Accurate anomaly detection in brain Magnetic Resonance Imaging (MRI) is crucial for early diagnosis of neurological disorders, yet remains a significant challenge due to the high heterogeneity of brain abnormalities and the scarcity of annotated data. Traditional one-class classification models require extensive training on normal samples, limiting their adaptability to diverse clinical cases. In this work, we introduce MadIRC, an unsupervised anomaly detection framework that leverages Inter-Realization Channels (IRC) to construct a robust nominal model without any reliance on labeled data. We extensively evaluate MadIRC on brain MRI as the primary application domain, achieving a localization AUROC of 0.96 outperforming state-of-the-art supervised anomaly detection methods. Additionally, we further validate our approach on liver CT and retinal images to assess its generalizability across medical imaging modalities. Our results demonstrate that MadIRC provides a scalable, label-free solution for brain MRI anomaly detection, offering a promising avenue for integration into real-world clinical workflows.

Graph theory has proven to be useful in studying brain dysfunction in Alzheimer's disease using MagnetoEncephaloGraphy (MEG) and fMRI signals. However, it has not yet been tested enough with reduced sets of electrodes, as in the 10-20 EEG. In this paper, we applied techniques from the Graph Spectral Analysis (GSA) derived from EEG signals of patients with Alzheimer, Frontotemporal Dementia and control subjects. A collection of global GSA metrics were computed, accounting for general properties of the adjacency or Laplacian matrices. Also, regional GSA metrics were calculated, disentangling centrality measures in five cortical regions (frontal, central, parietal, temporal and occipital). These two sort of measures were then utilized in a binary AD/controls classification problem to test their utility in AD diagnosis and identify most valuable parameters. The Theta band appeared as the most connected and synchronizable rhythm for all three groups. Also, it was the rhythm with most preserved connections among temporal electrodes, exhibiting the shortest average distances among [Formula: see text], [Formula: see text], [Formula: see text] and [Formula: see text]. In addition, Theta emerged as the rhythm with the highest classification performances based on regional parameters according to a [Formula: see text] cross-validation scheme (mean [Formula: see text], mean [Formula: see text] and mean F1-[Formula: see text]). In general, regional parameters produced better classification performances for most of the rhythms, encouraging further investigation into GSA parameters with refined spatial and functional specificity.

The reduction of dimensionality in machine learning and artificial intelligence problems constitutes a pivotal element in the simplification of models, significantly enhancing both their performance and execution time. This process enables the generation of results more rapidly while also facilitating the scalability and optimization of systems that rely on such models. Two primary approaches are commonly employed to achieve dimensionality reduction: feature selection-based methods and those grounded in feature extraction. In this paper, we propose a distance-correlation feature space, upon which we define a dimensionality reduction algorithm based on space transformations and graph embeddings. This methodology is applied in the context of dementia diagnosis through learning models, with the overarching objective of optimizing the diagnostic process.

Task-oriented rehabilitation is essential for hand function recovery in stroke patients, and recent advancements in BCI-controlled exoskeletons and neural biomarkers - such as post-movement beta rebound (PMBR) - offer new pathways to optimize these therapies. Movement-related EEG signals from the sensorimotor cortex, particularly PMBR (post-movement) and event-related desynchronization (ERD, during movement), exhibit high task specificity and correlate with stroke severity. This study evaluated PMBR in 34 chronic stroke patients across two cohorts, along with a control group of 16 healthy participants, during voluntary and exoskeleton-assisted movement tasks. Longitudinal tracking in the second cohort enabled the analysis of PMBR changes, with EEG recordings acquired at three timepoints over a 30-session rehabilitation program. Findings revealed significant PMBR alterations in both passive and active movement tasks: patients with severe impairment lacked a PMBR dipole in the ipsilesional hemisphere, while moderately impaired patients showed a diminished response. The marked differences in PMBR patterns between stroke patients and controls highlight the extent of sensorimotor cortex disruption due to stroke. ERD showed minimal task-specific variation, underscoring PMBR as a more reliable biomarker of motor function impairment. These findings support the use of PMBR, particularly the PMBR/ERD ratio, as a biomarker for EEG-guided monitoring of motor recovery over time during exoskeleton-assisted rehabilitation.

Interactivity is crucial for enabling models to adjust and optimize based on user feedback, thereby enhancing overall performance. However, existing electroencephalogram (EEG)-based emotion recognition models rely on static training paradigms, lack interactivity, and struggle to effectively handle uncertainty in predictions. To address this issue, we propose a novel paradigm for interactive emotion recognition based on incremental Gaussian processes (GP). Unlike existing methods, our approach introduces an expert interaction mechanism to correct samples with high predictive uncertainty and incrementally update the model accordingly, thereby optimizing its performance. First, we model the emotion recognition task as a GP-based framework, utilizing the variance of the GP to quantify the model's uncertainty, thereby guiding experts in targeted interactions. Second, within the GP framework, we propose a novel incremental update strategy that allows the GP to incrementally update prediction results and uncertainties based only on new data obtained through expert interactions, without reprocessing all existing data. This effectively overcomes the shortcomings of traditional GP in updating efficiency. Third, to address the high computational complexity of GP, we use a sparse approximation strategy, selecting inducing points and performing variational inference to efficiently approximate the GP posterior, thereby reducing computational complexity. Subject-dependent and subject-independent experiments conducted on the DEAP and DREAMER datasets demonstrate that the proposed method exhibits significant advantages over state-of-the-art (SOTA) methods. In subject-dependent experiments, our method achieved the highest improvement (1.73%) in the Dominance dimension on the DREAMER dataset. In subject-independent experiments, it attained the largest performance improvement (2.96%) in the Arousal dimension on the DEAP dataset. These results further validate the proposed method's effectiveness.

Due to the differences in size, shape, and location of brain tumors, brain tumor segmentation differs greatly from that of other organs. The purpose of brain tumor segmentation is to accurately locate and segment tumors from MRI images to assist doctors in diagnosis, treatment planning and surgical navigation. NSNP-like convolutional model is a new neural-like convolutional model inspired by nonlinear spiking mechanism of nonlinear spiking neural P (NSNP) systems. Therefore, this paper proposes a global-local feature fusion network based on NSNP-like convolutional model for MRI brain tumor segmentation. To this end, we have designed three characteristic modules that take full advantage of the NSNP-like convolution model: dilated SNP module (DSNP), multi-path dilated SNP pooling module (MDSP) and Poolformer module. The DSNP and MDSP modules are employed to construct the encoders. These modules help address the issue of feature loss and enable the fusion of more high-level features. On the other hand, the Poolformer module is used in the decoder. It processes features that contain global context information and facilitates the interaction between local and global features. In addition, channel spatial attention (CSA) module is designed at the skip connection between encoder and decoder to establish the long-range dependence between the same layers, thereby enhancing the relationship between channels and making the model have global modeling capabilities. In the experiments, our model achieves Dice coefficients of 85.71[Formula: see text], 92.32[Formula: see text], 87.75[Formula: see text] for ET, WT, and TC, respectively, on the N-BraTS2021 dataset. Moreover, our model achieves Dice coefficients of 83.91[Formula: see text], 91.96[Formula: see text], 90.14[Formula: see text] and 85.05[Formula: see text], 92.30[Formula: see text], 90.31[Formula: see text] on the BraTS2018 and BraTS2019 datasets respectively. Experimental results also indicate that our model not only achieves good brain tumor segmentation performance, but also has good generalization ability. The code is already available on GitHub: https://github.com/Li-JJ-1/NSNP-brain-tumor-segmentation.

Palmprint recognition is a pivotal biometric modality, renowned for its numerous advantages and applications in the field of biometrics. The Gabor filter is a classic and efficient texture feature extractor abstracted from the nervous system. The existing palmprint texture coding methods only focus on first-order texture features (1TFs), while neglecting discriminative second-order texture features (2TFs). Therefore, this paper proposes multi-order extensions for state-of-the-art (SOTA) palmprint texture coding methods, which makes full usage of 1TFs and 2TFs. A filter is used to extract 1TFs from the palmprint image, and the same filter is applied to extract 2TFs from 1TFs. Here, different methods employ various filters to extract diverse textures. Due to the simultaneous participations of 1TFs and 2TFs in multi-order extension codes, more discriminative features are extracted and fused. The experimental results on three public databases, including contact, noncontact and multispectral acquisition types, show that the accuracies of all the palmprint texture coding methods are remarkably improved by multi-order extension, establishing it as a general framework extendable to other texture-based recognition tasks.