This paper presents an in-depth evaluation of “MCsiNet,” an AI-based feedback channel state information compression and reconstruction model, in a 5G massive MIMO OFDM system, adhering to 3GPP standards using the “CDL-C” MIMO channel model. “M-CsiNet” demonstrates significant improvements over legacy schemes, such as Type-II and Enhanced Type-II. Furthermore, “M-CsiNet” delivers optimal performance in terms of link-level block error rate (BLER) and throughput with 10-15 dB lower SNR, along with a two-order reduction in overhead and a one-order reduction in complexity. These results position “M-CsiNet” as a highly efficient solution for capacity-limited scenarios in both rank-1 and rank-2 cases.

This paper introduces a groundbreaking computing paradigm based on neural oscillatory circuits, which harnesses the power of oscillatory dynamics found in the brain to overcome the fundamental limitations of traditional computing architectures. By encoding and processing information in the phase and amplitude dynamics of coupled oscillators, these circuits achieve unprecedented levels of parallelism, energy efficiency, and adaptability. We present a comprehensive mathematical framework for designing and analyzing oscillatory computing systems, drawing upon advanced concepts from dynamical systems theory, stochastic differential equations, and complex network analysis. Through extensive numerical simulations and physical prototypes, we demonstrate that oscillatory circuits can solve complex computational problems orders of magnitude faster and more efficiently than state-of-the-art algorithms, paving the way for transformative applications in domains ranging from neuromorphic computing and artificial intelligence to quantum computing and smart sensors. Our results provide compelling evidence that neural oscillatory circuits represent a paradigm shift in computing, with profound implications for the future of technology and our understanding of the computational principles underlying biological intelligence.

Alzheimer's Disease (AD) is a prevalent neurodegenerative condition where early detection is vital. Handwriting, often affected early in AD, offers a non-invasive and costeffective way to capture subtle motor changes. State-of-the-art research on handwriting, mostly online, based AD detection has predominantly relied on manually extracted features, fed as input to shallow machine learning models. Some recent works have proposed deep learning (DL)-based models, either 1D-CNN or 2D-CNN architectures, with performance comparing favorably to handcrafted schemes. These approaches, however, overlook the intrinsic relationship between the 2D spatial patterns of handwriting strokes and their 1D dynamic characteristics, thus limiting their capacity to capture the multimodal nature of handwriting data. Moreover, the application of Transformer models remains basically unexplored. To address these limitations, we propose a novel approach for AD detection, consisting of a learnable multimodal hybrid attention model that integrates simultaneously 2D handwriting images with 1D dynamic handwriting signals. Our model leverages a gated mechanism to combine similarity and difference attention, blending the two modalities and learning robust features by incorporating information at different scales. Our model achieved state-of-theart performance on the DARWIN dataset, with an F1-score of 90.32% and accuracy of 90.91% in Task 8 ('L' writing), surpassing the previous best by 4.61% and 6.06% respectively.

Large language models (LLMs) have received considerable interest recently due to their outstanding reasoning and comprehension capabilities. This work explores applying LLMs to vehicular networks, aiming to jointly optimize vehicleto-infrastructure (V2I) communications and autonomous driving (AD) policies. We deploy LLMs for AD decision-making to maximize traffic flow and avoid collisions for road safety, and a double deep Q-learning algorithm (DDQN) is used for V2I optimization to maximize the received data rate and reduce frequent handovers. In particular, for LLM-enabled AD, we employ the Euclidean distance to identify previously explored AD experiences, and then LLMs can learn from past good and bad decisions for further improvement. Then, LLM-based AD decisions will become part of states in V2I problems, and DDQN will optimize the V2I decisions accordingly. After that, the AD and V2I decisions are iteratively optimized until convergence. Such an iterative optimization approach can better explore the interactions between LLMs and conventional reinforcement learning techniques, revealing the potential of using LLMs for network optimization and management. Finally, the simulations demonstrate that our proposed hybrid LLM-DDQN approach outperforms the conventional DDQN algorithm, showing faster convergence and higher average rewards.

Cellular heterogeneity is one of the main hallmarks of cancer, referring to the coexistence of different phenotypes with very distinct biological behaviors in single isolates. Automatically detecting single-cell heterogeneity is therefore critical, and can provide important information on cancer initiation, tumour aggressiveness or drug-related resistance. Here we introduce He2Cl, a novel machinelearning algorithm able to detect single cell heterogeneity from the morpho-dynamic analysis of 2D cell cultures. Built on label-free quantitative time-lapse imaging, He2Cl performs a 2-step clustering algorithm for unsupervised classification of sub-phenotypes within a cell population. He2Cl outperforms state-of-the-art clustering models and allows the simultaneous classification of thousands of cell trajectories from their birth to their division, paving the way towards AI-enhanced microscopy for live-cell analysis.

The proliferation of large multi-antenna configurations operating in high frequency bands has recently challenged the conventional far-field, rich-scattering paradigm of wireless channels. Extra large antenna arrays must usually work in the near field and in the absence of multipath, which are far from traditional assumptions in conventional wireless communication systems. The present study proposes to analyze the spatial multiplexing capabilities of large multi-antenna configurations under line-of-sight, near field conditions by considering the use of multiple orthogonal diversities at both transmitter and receiver. The analysis is carried out using a holographic approximation to the problem, whereby the number of radiating elements is assumed to become large while their separation becomes asymptotically negligible. This emulates the operation of a continuous aperture of infinitesimal radiating elements, also recently known as holographic surfaces. The present study characterizes the asymptotic MIMO channel as seen by extra large uniform linear and planar arrays, as well as their associated achievable rates assuming access to perfect channel state information (CSI). It is shown, in particular, that for a given distance between the receiver and the center of the array and a given signal quality, there exists an optimum dimension of the multi-antenna surface that maximizes the spectral efficiency.

The rapid integration of the Internet of Things (IoT) into high-end equipment has dramatically transformed fault diagnosis processes, enabling real-time monitoring and predictive maintenance. This technological advancement, while enhancing operational efficiency, has also introduced critical security vulnerabilities, particularly within supply chains susceptible to cyberattacks. This paper investigates the intersection of IoT security and fault diagnosis in high-end equipment, proposing a systematic secure software engineering framework. Through an extensive literature review and a detailed analysis of existing datasets on supply chain attacks, an innovative secure software engineering approach is developed. The methodology is validated through a case study on IoT-Enabled Fault Diagnosis in a High-End Car Manufacturing Supply Chain, showcasing the practical application of the proposed framework. Business Process Modelling and Notation (BPMN) simulations of the car manufacturing process reveal that 1,000 cars can be assembled and repaired in 467.33 minutes (approximately 7.79 hours or onethird of a day) using this approach. Additionally, the study introduces key metrics for process quality, data quality, and predictive model accuracy, offering a quantitative assessment of the framework's effectiveness. The findings emphasize the critical need for robust security measures in IoT-enabled fault diagnosis systems to mitigate potential risks and ensure operational integrity.

The widespread use of lithium-ion batteries and the demand for high-performance battery packs have made battery thermal modelling a crucial research area. This field helps to understand and manage the heat generated and the temperature distribution during battery operation. Effective thermal management is essential to ensure safety, extend battery lifespan and optimise efficiency. In this work, heat generation is identified as the primary driver of temperature change and distribution within the cell. Various battery models are reviewed and classified, driving the selection of the right model according to the application. Several thermal characterisation methods are described in detail, with a focus on experimental techniques that can be applied in-situ. Finally, the challenges posed by temperature distribution are discussed extensively, with a focus on how these challenges could shape the future research and design of the next generation of lithium-ion cells and battery packs.  

This paper proposes WaveDepth, a novel method for enabling accurate water depth estimation by combining water semantic segmentation models with depth estimation models. By masking out the water portions in images before inputting them into the depth estimation model, WaveDepth significantly improves the accuracy and detail representation of depth estimation for water targets. To overcome the lack of water depth estimation datasets, a dedicated dataset was constructed using multiple devices and WaveShot water photography drones. Multiple depth estimation models were trained and tested on this dataset to select an optimized model that can estimate depths in water scenes without additional image preprocessing. This work contributes an innovative water depth estimation method, a purpose-built water depth estimation dataset, and an optimized model for practical application.

E-learning is one of the most prominent phenomena that has accompanied the world community since the emergence of the digital and technological revolution. The need for these digital technologies has increased with the increasing cases of urgent need for them according to the realities of the situation. For example, the world's need for e-learning technologies emerged after the COVID-19 pandemic, which proved that the advantages of e-learning are no less important than face-to-face education, but rather that e-learning has exceeded the importance that face-to-face education has enjoyed, according to what this type of education (distance learning) has provided from features and data that have affected the virtual educational reality and led to the emergence of a distinguished group of professors who are proficient in dealing with computers and electronics. This in itself formed a positive point among a series of positives that came out of this education.

Speech command recognition (SCR) is the task of detecting predefined voice commands in an audio stream. SCR is used widely in voice assistants, augmented reality, and robotics. Nowadays, Google Speech Commands (GSC) is a benchmark dataset for training and testing SCR models for English. In this work, we developed an open-source speech commands dataset for the low-resourced Tatar language. The dataset employs 35 commands in the GSC dataset and contains 3,547 utterances. To prove the efficacy of our dataset, we trained and evaluated the Keyword-MLP model on our dataset. The model achieved an accuracy of 98.28% on the test set. We optimized the model by converting it to the ONNX Runtime format and deployed it on an NVIDIA Jetson Orin NX 16GB. The optimized model showed an average inference time of 0.003 s on the device, making it suitable for real-time IoT and robotics applications. We have made the dataset, source code, and pretrained models publicly available at https://github.com/IS2AI/TatarSCR to promote the development of voice-controlled systems for the Tatar language.

Human pose estimation has applications in numerous fields, including action recognition, human-robot interaction, motion capture, augmented reality, sports analytics, and healthcare. Many datasets and deep learning models are available for human pose estimation within the visible domain. However, challenges such as poor lighting and privacy issues persist. These challenges can be addressed using thermal cameras; nonetheless, only a few annotated thermal human pose datasets are available for training deep learning-based human pose estimation models. In this regard, we extend our previously released OpenThermalPose dataset with more data, human instances, and poses. In the new OpenThermalPose2 dataset, the number of thermal images increased from 6,090 to 11,391, and the number of labeled humans rose from 14,315 to 21,125. The annotations include bounding boxes and 17 anatomical key points. To show the efficacy of the dataset, we trained and evaluated the YOLOv8-pose and YOLO11-pose models. The experimental results showed that the models trained on the extended dataset outperformed the previous models. We have made the dataset, source code, and pre-trained models publicly available at https://github.com/IS2AI/OpenThermalPose to bolster research in this field.

Ransomware continues to pose a significant threat to the security of Windows systems, as its rapidly evolving nature enables attackers to bypass traditional signature-based detection methods. The use of opcode-based features, combined with K-Nearest Neighbors (KNN) classification, offers a novel and effective approach to identifying ransomware by analyzing the low-level behavior of executables rather than relying solely on external characteristics. Through extracting opcode sequences from both ransomware and benign Windows executables, and converting them into high-dimensional feature vectors, the method captures behavioral distinctions that facilitate accurate classification. The study demonstrates that feature selection techniques, such as Chi-Square and Information Gain, not only improve computational efficiency but also maintain the model's detection accuracy, significantly enhancing the practicality of deploying machine learning models in real-world cybersecurity environments. Comparative evaluations show that KNN performs competitively with other machine learning approaches, making it a suitable choice for detecting ransomware while balancing efficiency and accuracy. The results emphasize the importance of opcode-based analysis in providing deeper insights into executable behavior, offering robust defense mechanisms against ransomware attacks.

Unmanned Aerial Vehicles (UAV) play a critical role in the low-altitude economy; however, they face significant challenges in network coverage during transit. This paper investigates the characteristics and challenges of communication networks for UAV. First, we delineate typical scenarios, traffic characteristics, and the dual-layer heterogeneous network topology necessary for three-dimensional continuous coverage and network coexistence. Additionally, the paper addresses the channel characteristics and specific challenges faced by UAV Integrated Sensing and Communication (ISAC) networks, including the limitations of traditional APAA-based networks in terms of cost, complexity, and site selection. Subsequently, we present Reconfigurable Intelligent Surface (RIS)-assisted networks as a viable solution for enhancing UAV signal coverage. Key technical characteristics of RIS are discussed, along with design principles, antenna tilt design, and the application of new beam types. We also explore the implications of high-frequency bands and their absorption peaks on signal attenuation, as well as network architecture designs aimed at improving UAV signal coverage and facilitating network coexistence. Furthermore, the importance of standardization in RIS-based UAV networks is emphasized. We also present field trial results evaluating the effectiveness of RIS in improving UAV coverage. Finally, we outline future technology trends, highlighting potential advancements that could further optimize UAV communication systems.

We propose a universal approach for modeling complex integrated photonic resonators based on the scattering matrix method. By dividing devices into basic elements including directional couplers and connecting waveguides, our approach can be used to model integrated photonic resonators with both unidirectional and bidirectional light propagation, with the simulated spectral response showing good agreement with experimental results. A simplified form of our approach, which divides devices into several independent submodules such as microring resonators and Sagnac interferometers, is also introduced to streamline the calculation of spectral transfer functions. Finally, we discuss the deviations introduced by approximations in our modeling, along with strategies for improving modeling accuracy. Our approach is universal across different integrated platforms, providing a useful tool for designing and optimizing integrated photonic devices with complex configurations.

The rapid expansion of multimodal data, encompassing text, images, and audio, presents significant challenges in efficiently processing and interpreting incomplete or partiallydefined information. The proposed framework introduces a novel approach to token propagation, addressing the inherent difficulties of grounding partially-defined tokens across multiple modalities without human intervention. Through an open-source LLM model, Llama, the study successfully demonstrated how cross-modal interactions can resolve ambiguities in token definitions, particularly in text and image data, while also identifying the limitations associated with audio processing. The methodology showed that contextual data from various modalities can be effectively integrated to refine token representations, leading to improved accuracy in multimodal tasks. The comprehensive evaluation of token grounding performance highlights the potential for further extending LLM capabilities in real-world applications, where heterogeneous data sources are increasingly prevalent.

The performance of model-based battery algorithms depends on the fidelity of the adopted battery model including the open-circuit voltage (OCV) versus state-of-charge (SoC) relationship. The OCV-SoC model is usually extracted using offline experiments at the beginning-of-life (BoL) and is often assumed not to change over the battery lifetime. However, as shown in recent literature and this article, the battery OCV-SoC curve will change because of cyclic and calendar aging. Thus, the effect of aging on the OCV-SoC curve should be compensated to maintain accuracy in different battery algorithms. To address this gap, this article puts forward a reinforcement learning (RL) framework to train an artificial intelligence agent that can adjust and compensate for the changes of the OCV-SoC curve in different temperature and aging conditions. First, parameter and state estimators are designed to estimate the OCV and SoC terms and the feedback is then provided to train the RL agent used for prediction of the parameters of the OCV curve model. The effectiveness of the proposed method is demonstrated in SoC estimation using actual battery performance and lifetime data related to four high-capacity lithium-ion cells. The proposed RL agent yields an average RMSE of about 10 mV and improves SoC estimation accuracy up to 7% when the agent is applied to compensate OCV-SoC curve of the aged cells.

Understanding the limitations of Large Language Models (LLMs) in code generation is crucial for optimizing their use in software development. This paper explores these limitations, discusses their impact on coding practices, and oers insights into potential solutions and areas for further research.