Emotional information is present in every spoken audio event that individuals frequently hear. As a result, Speech Emotion Recognition (SER) has gained widespread recognition. Over the past ten years, this has grown into a significant research topic. Through the use of human voices or everyday conversation, SER can detect people's emotional states. It is essential for developing Human-Computer Interaction (HCI) and signal processing systems. Emotions in humans also evolve with time. Therefore, to comprehend the dependencies in the speech sign over time, a strong model is required. In this work, Transformer-based method using the CNN model in parallel (TASER-Net) method is used for SER. With the use of a parallel CNN model, our novel approach to temporal emotion modelling for SER overcomes information loss from noise and bi-directional propagation while creating multi-scale contextual emotional representations across a range of time frames.
The increasing prevalence of malware propagation in IoT networks requires the development of efficient and adaptive mitigation strategies. Classical approaches based on optimal control theory, such as Pontryagin's Maximum Principle (PMP), provide mathematically optimal solutions, but require solving a complex two-point boundary value problem. Recent advances in machine learning have introduced Neural Ordinary Differential Equations (ODEs) as a new alternative, allowing learning of non-linear control policies through gradient-based optimization. In this work, we apply neural ODEs to solve an optimal control problem in a modified SIR model of malware propagation, incorporating two control functions: u(1)(t), which reduces transmission, and u(2)(t), which enhances recovery. We compare this approach with a PMP-based control solution obtained using the shooting method. Our results show that both methods effectively reduce the infection peak, but exhibit different behaviours in the timing and magnitude of interventions. This study highlights the advantages and limitations of both approaches, and provides a foundation for hybrid control strategies that combine machine learning with traditional optimal control techniques for malware mitigation.
Digital TV broadcasting networks rely on alarm systems to monitor equipment and service health. However, the large volume of alarms, many of which are non-critical, makes it difficult for Network Operation Center (NOC) operators to identify early signs of service-affecting incidents. This paper presents an automated incident prediction system based on machine learning techniques, developed with a Spanish digital TV operator, to address this challenge. The system is tested under three synthetic scenarios-Baseline, Degraded, and Alarm-storm-that simulate increasing levels of network degradation. Alarm data are encoded under four contextual settings: (i) without additional information; (ii) including pattern-based features to capture interactions with other network elements; (iii) incorporating time-based statistical features extracted from alarm activity; and (iv) combining both types of contextual information. Results show that incorporating temporal context significantly improves detection, especially in highly imbalanced conditions where incidents are rare. LightGBM stands out among the evaluated classifiers, achieving the highest balanced accuracy when time-based statistical features are used.
Incremental feature learning (IFL) is a supervised learning paradigm for neural networks (NNs), where the input layer is incrementally expanded over time. NNs dynamically expand the input layer with new features, while also reducing overfitting and model complexity. Feature ranking based on feature importance determines the order of feature integration. The incremental nature of IFL results in a dynamic optimization problem (DOP), where both the search space and its dimensionality changes over time. Particle swarm optimisation (PSO) has been extended to dynamic environments. This study adapts various dynamic PSO variants to train incrementally constructed NNs (INNs). The performance of INNs is compared to fully constructed NNs (FNNs) trained with BP and standard PSO on seven regression tasks. Results demonstrate that IFL effectively allows NNs to incorporate new features dynamically and acts as a regularisation technique.
The safe and efficient operation of hydrogen refueling stations is essential to support the global transition towards low-carbon energy systems. However, the scarcity of real-world operational data remains a major obstacle for advanced monitoring and anomaly detection. This study proposes a deep learning framework that combines LSTM-based synthetic data generation with unsupervised anomaly detection. A generative LSTM was used to simulate 756 realistic hydrogen refueling scenarios enriched with physically plausible anomalies. An LSTM-Autoencoder was subsequently trained to detect deviations in key process variables, achieving 92
The exponential growth of internet usage and social media has inundated us with an immense volume of unannotated textual content. Annotating such vast amounts of data requires considerable time and expertise, making it a daunting task. While handling massive datasets in an unsupervised manner offers certain advantages, the challenges inherent in effectively grouping this data underscore the importance of text clustering. To address this challenge, we introduce TextNet, an end-to-end deep learning model. Unlike traditional methods, TextNet does not rely on external tags or labels. Instead, it leverages the assumption that different clusters exhibit variations in their distributions–samples within a cluster share similar distributions, while samples from different clusters display distinct variations. The innovation of TextNet lies in its utilization of dual models: one processes the original text input, while the other handles augmented samples generated to resemble the input. Through unsupervised training, TextNet employs the Clussimloss loss function to iteratively refine the model, leading to more effective text clustering. We extensively evaluated TextNet on two widely used standard text datasets. The experimental results demonstrate that our proposed method surpasses existing models in terms of clustering accuracy and effectiveness.
Artificial Intelligence (AI) and Machine Learning (ML) have become pervasive technologies, raising complex ethical challenges. While prior discussions often addressed ethical themes in general, the rapid deployment of Large Language Models (LLMs) presents new, concrete dilemmas. This work critically examines key ethical dimensions, including data privacy, algorithmic bias, transparency, accountability, and sustainability, with a particular focus on their implications for public administration. The regulatory context of the European AI Act provides a reference framework, but significant ethical and governance gaps remain. We propose a structured analysis of the unique risks posed by LLMs in administrative decision-making, and outline actionable recommendations for responsible deployment.
This paper introduces hybrid AI models for structured mobility prediction in metropolitan areas, focusing on Vienna, to guide citizens toward greener transportation options. The AI-CENTIVE project explores how AI can identify effective incentives by forecasting future trips using a combination of traditional machine learning and modern deep learning architectures. Trained on a dataset of commuter trips from the Ummadum app, the models predict transport mode, time, origin, destination, distance, and duration. The most accurate predictions trigger notifications suggesting sustainable alternatives. The evaluation of various hybrid architectures revealed that a graph convolutional network that uses statistical patterns achieved the best performance on the analyzed dataset. The presented research contributes to leveraging AI to promote sustainable mobility through targeted incentivization.
We evaluate the feasibility of controlling an event-related potential (ERP) matrix speller using a commercially available, wearable, EEG headset with single-channel electrodes placed on the forehead in a headband configuration. Despite the suboptimal recording location for detecting typical visual ERP components, we conducted experiments with 11 right-handed healthy participants. Time-domain analysis revealed significant differences between attended and non-attended conditions in 6 participants, as determined by cluster-based permutation testing. A separate decoding analysis using linear discriminant analysis identified above-chance single-trial accuracy in a different subset of 6 participants. The highest decoding accuracy reached 52.5
Contextualizing museum objects into their historical scene is important, both for understanding how those items are genuinely used and for testing the capabilities of Stable Difussion (SD) methods to generate historical scenes with unusual, ancient items. For these aims, we propose two approaches based on SD: one that relies on inpainting and outpainting and one that uses LoRA model training and generation again with SD. We make a comparison between the two approaches. We evaluate how these methods excel and fail, respectively, with the help of two cultural heritage experts. They evaluate the produced images based on six criteria that take into account information like historical context alignment, coherence and aesthetic quality, object integration, proportionality, clothing accuracy and atmospheric appropriateness, respectively.
Modern power grids are becoming increasingly complex with the integration of heterogeneous distributed energy resources, underscoring the need for accurate and efficient Power Flow Analysis to ensure stability, reliability, and market operations. Existing methods generally rely on iterative numerical techniques (INT) or machine learning (ML). While INT is physically consistent and highly accurate, it can be computationally expensive and vulnerable to slow or non-convergence. ML methods offer faster solutions but often require extensive data, suffer from limited extrapolation capabilities, and lack physical consistency. Physics-informed ML (PIML) bridges these gaps by embedding domain knowledge before, during, and after training. However, current PIML approaches typically do not leverage this full range of opportunities. In this paper, we propose a novel PIML framework for Power Flow Analysis that integrates physical insights at all three stages (pre, in, and post-processing) to achieve superior accuracy and efficiency. Notably, we introduce a new post-processing technique that partitions the power network into its mesh and radial components: the mesh portion is handled via PIML, while the radial portion is efficiently solved with a convex optimization approach informed by the PIML outputs. This approach is efficient with radial topologies, especially in power distribution networks where the radial part is predominant. Experiments on realistic power networks demonstrate that our method outperforms state-of-the-art approaches in both accuracy and computational performance.
Navigation in complex environments is a key challenge in robotics and autonomous systems, traditionally tackled with pathfinding algorithms like A* and its variants. These methods discretise the domain into a uniform grid, enabling movement between adjacent nodes. While computationally efficient, these approaches compromise path smoothness and optimality due to the limited number of movement directions. Alternatively, Any-Angle path planning methods address this by using visibility graphs, allowing direct connections between nodes when there is no obstacle between them. This improves trajectory flexibility but significantly increases computational costs, limiting scalability. To overcome these challenges, this paper introduces a Deep Learning-accelerated approach for efficient navigation in obstacle maps. A customised Conditioned U-Net, trained on a solution database, provides heuristic estimates to guide the search. Combined with a k-Non Aligned Nearest Neighbours (k-NANN) graph structure, this method ensures smooth, optimal trajectories while reducing computational overhead.
Knee Osteoarthritis (KOA) is a progressive joint disease characterised by stiffness and pain, among others. It is generally diagnosed by evaluating physical symptoms, medical history, and screening techniques. However, conventional methods are often subjective, posing a significant challenge to the early grading of disease progression. To address this issue and support clinical decision-making, we propose an ordinal deep learning framework to study the optimal combination of loss functions, and output methodologies with soft labelling approaches, for automatic KOA severity grading based on Kellgren and Lawrence scores from X-ray images. A total of 20 combinations (2 loss functions × 2 output methodologies × 5 soft labelling approaches) are compared in this study, using a public dataset. The optimal configuration uses the categorical cross entropy loss, a cumulative link model as output, and a beta distribution for soft labelling. The results achieved demonstrate the efficacy of these ordinal classification approaches.
This paper presents a new method for detecting defects in composite materials examined by non-destructive testing using active thermography. The proposed method includes a fusion stage where the scores from multiple classifiers are fused under the mean-square error optimization criterion using alpha integration method. The goal is to improve the performance of individual classifiers based on different and sometimes complementary principles and that fusion can be used to exploit such a complementarity in both accuracy and variance. Several time-domain, frequency-domain, and statistics features were extracted from a dataset of thermography signals measured in composite material specimens. Seven individual classifiers were implemented. The results of fusion based on alpha integration were compared to the ones of the individual classifiers and the fusion by the mean showing the superiority of the proposed method in terms of several indices such as receiver operating characteristic and precision-recall curves.
Tensor decomposition methods constitute an alternative way to analyze multi-dimensional data, with advantages with respect to classical linear techniques such as PCA (Principal Component Analysis). This is especially useful in the context of phase-amplitude coupling data, where these are computed for different band combinations for each channel and for each subject. Unlike PCA, which assumes that the data matrix is linear and typically focuses on variance maximization, tensor decomposition methods can handle multi-dimensional data (e.g., EEG signals across different frequency bands and channels) in a more flexible way by decomposing it into components that are not only spatially but also temporally and spectrally meaningful. As a result, it enables to model interactions across multiple modes, allowing a more accurate representation of the brain activity by capturing non-linear patterns. Moreover, it provides a more interpretable, robust and unique solution. In this work, we propose a method to study cerebral lateralization by means of tensor decomposition using the PARAFAC (Parallel Factor Analysis) model to identify spatial patterns along with bands contribution. This provides insights into hemispheric differences in cognitive processing, being able to detect abnormal patterns linked to language impairments.
Diabetes mellitus (DM) poses a critical global health challenge, with type 1 diabetes (T1D) patients presenting unique difficulties in maintaining a safe blood glucose level (BGL). This work demonstrates that evaluating BGL prediction models without considering different BGL ranges, hypoglycemia, hyperglycemia, and normoglycemia, introduces bias in assessing the prediction results. Data are obtained from the T1DiabetesGranada dataset, comprising over 22.5 million measured BGL values recorded at 15-min intervals, and are preprocessed into a uniform format for supervised learning. Time series are segmented into windows with a 2-h history length and prediction horizons of 30 and 60 min. An LSTM architecture is used to predict BGL values due to its ability to capture temporal dependencies. The evaluation combines traditional non-clinical metrics (RMSE, MAE, MAPE) with clinical metrics derived from the Clarke Error Grid. The newly proposed evaluation strategy assesses BGL prediction models performance not only across the entire BGL range but also within different BGL ranges. Results indicate that evaluation metrics computed using the entire BGL range may suggest satisfactory BGL prediction model performance. However, significant deficiencies emerge in hypoglycemic ranges, implying that conventional evaluation strategies may overestimate BGL prediction models capabilities. These findings highlight the need for a comprehensive evaluation strategy in different BGL ranges to avoid bias, especially while evaluating clinically critical regions.
5G technology, increasingly present in sectors such as industrial automation and smart cities, represents a major advance in connectivity, speed, and latency. However, these environments can also be affected by interference, such as jamming attacks, which can physically deny service by overwhelming the radio spectrum. This work presents a labeled dataset generated in a controlled private 5G SA (standalone) network environment, including both legitimate traffic and traffic affected by jamming. To this end, an experimental scenario was prepared using real devices and a dedicated system to perform the interference attack. The analysis focuses on key protocols such as NGAP and GTP, with the aim of providing a dataset for the development of AI-based intrusion detection systems in 5G environments, adapted to the specific characteristics of these networks.
Continuous monitoring of neonatal behavior in the Neonatal Intensive Care Unit (NICU) is essential for early detection of neurological disorders. Among behavioral indicators, eye state (open vs. closed) serves as a clinically relevant marker for alertness, sedation, and responsiveness. This study presents a deep learning-based system for automated eye state detection in NICU video recordings. Using a manually labeled dataset of 7,388 facial frames extracted from 154 clinical videos, we trained and evaluated binary classifiers based on VGG16 and VGG19 convolutional neural network architectures. A five-fold cross-validation scheme was implemented to ensure subject-independent evaluation. The models achieved mean frame-level accuracies above 0.86 and AUC-PR values of 0.97. Additionally, video-level evaluation under realistic conditions yielded up to 0.79 accuracy and 0.84 AUC-PR. These results support the feasibility of integrating eye state detection into broader AI frameworks for neonatal monitoring and early neurological assessment.
This paper presents a new application of a hybrid dropout technique for Ordinal Classification (OC), based on a novel regularisation method. Unlike standard dropout, which ignores class ordering, this hybrid dropout integrates ordinal information by adjusting neurons dropout probabilities based on their correlation with target labels. We evaluate its effectiveness using a ResNet18 architecture over three new OC datasets and compare it with the standard dropout approach and with an architecture with no dropout. Results show that the hybrid dropout consistently achieves the best performance across multiple well-known metrics (1-off, QWK, MAE, AMAE, and RPS), while also reducing prediction variability. Statistical analysis using the Wilcoxon signe-drank test confirms its robustness, obtaining 21 significant wins out of 30 comparisons, with no losses. These results highlight the importance of designing regularisation strategies that consider the problems ordinal structure, demonstrating that hybrid dropout effectively enhances generalisation and predictive accuracy.
Natural Language Processing (NLP) is the foundation that supports the technology around us today: from search engines to automated customer service. As these systems gain an increasing influence on social and economic outcomes, however, the question of bias in NLP has become hugely important. In this paper, we provide a comprehensive review of bias in NLP, from its sources, and societal impacts to the current approaches to mitigating it. We look at recent studies of data and algorithmic biases that persist and have a disproportionate impact on marginalized communities. Our results stress the necessity of interdisciplinary approaches to these challenges by merging the insights of computer science, linguistics, and ethical and social sciences. To this end, we develop a framework for building fairer, more inclusive NLP systems that leverage diverse data in combination with state-of-the-art debiasing methods and ethical AI guidelines. This work adds to an old debate about making ethical AI while also suggesting where to direct future efforts on the creation of fairer NLP technologies.
