Power delay profile (PDP) of a wireless channel has been widely studied using different approaches and measurements. In this paper we take a new approach by applying Friis transmission equation to derive the PDP in a Gaussian scattering scenario. Using this approach we notice that if the scattering spread is large or the distance between the base station and the mobile station is small, the PDP has a peak at a delay greater than the line of sight delay. The obtained PDP is able to explain the peak reported in measurements and is consistent with the delay probability density function. Further, the profile comes closer to the PDP reported in the literature when either the distance between the base station and mobile station is large or the scattering spread is small.

This paper presents a novel framework for real-time object detection in Wireless Mesh Networks (WMNs) using AI driven drones. The system integrates YOLOv5 for object detection, a hybrid greedy and reinforcement learning-based routing algorithm for efficient data communication, and a scalable multihop mesh topology. MATLAB-based simulations evaluate energy consumption, detection accuracy, and network connectivity, showcasing significant performance improvements. The results demonstrate the system's ability to detect objects accurately, maintain robust connectivity, and optimize energy consumption, even in dynamic environments. Future work includes expanding the system to integrate 5G and satellite networks and conducting field tests.

CHARGE (Charged Hyperspherical Autoencoder with Repulsive Gradient Encoding) is a novel autoencoder designed with a unique latent space that is hyperspherical and of n dimensions. The model leverages a specialized loss function based on Coulomb's Law or the Inverse Square Law, promoting a more effective and efficient latent space. This approach enables CHARGE to perform better on generative AI tasks compared to traditional Variational Autoencoders (VAEs), as it is deterministic rather than probabilistic. The results show that CHARGE provides superior performance in generating realistic data and encoding meaningful latent representations, making it a promising candidate for various generative tasks in AI.

Modern cybersecurity demands innovative approaches capable of addressing the rapidly evolving nature of sophisticated threats targeting critical systems. Temporal Entropy Fingerprinting introduces a novel framework for detecting malicious activities by leveraging the dynamic properties of entropy patterns over time. Through the systematic analysis of entropy fluctuations, the approach identifies behaviors associated with encryption, file manipulation, and anomalous system activity, key characteristics of ransomware attacks. The framework achieves high detection rates by integrating probabilistic models, dynamic thresholding, and advanced segmentation techniques to quantify anomalies across varying operational contexts. Experiments conducted with datasets encompassing contemporary ransomware variants, such as LockBit and BlackMatter, demonstrated consistent performance across diverse scenarios, ensuring both accuracy and scalability. The entropy-based architecture supports real-time anomaly detection while maintaining computational efficiency, addressing the limitations observed in traditional static and dynamic analysis methods. By enabling a granular understanding of entropy distributions during critical ransomware operations, the framework provides actionable insights that enhance situational awareness. Furthermore, modular design principles facilitate seamless integration with existing cybersecurity infrastructures, ensuring adaptability to resource-intensive environments. The results indicate significant improvements in detection precision and recall, particularly in identifying previously unseen ransomware families. Comparative analysis against baseline methods highlighted the framework's capability to maintain low false positive rates even under highvolume data processing conditions. With entropy metrics serving as a foundation, the proposed approach offers a resilient and effective solution for mitigating modern ransomware threats.

This work introduces a novel concept for the realization of biological artificial neural networks (ANNs), called shared volume computing (SVC). The central units in this concept are biological cells, that function as individual neurons of the ANN. These cells reside inside a shared volume, where they sense signaling molecules released by cells of the previous layer or a biological process. They respond by the production of a chemical that they release into the shared volume, thereby creating the input into the next layer. We introduce the SVC, present mathematical models of it and show, how these models can be used to design a biological ANN for a given task. Specifically, we discuss the design process on a first example: breast cancer detection from biomarkers. We show in simulation, that the proposed system is capable of solving this task with an accuracy of up to 75%, making it especially useful for rapid testing.

Reflection is widely recognized as a cornerstone of student development, fostering critical thinking, self-regulation, and deep conceptual understanding. Traditionally, reflective skills are cultivated through structured feedback, mentorship, and guided self-assessment. However, these approaches often face challenges such as limited scalability, difficulties in delivering individualized feedback, and a shortage of instructors proficient in facilitating meaningful reflection. This study pioneers the exploration of generative AI, specifically large language models (LLMs), as an innovative solution to these limitations. By leveraging the capacity of LLMs to provide personalized, context-sensitive feedback at scale, this research examines their potential to serve as effective facilitators of reflective exercises, maintaining the depth of engagement and promoting critical thinking. Through an in-depth analysis of prompt engineering strategies and the efficacy of LLMs in simulated multi-turn dialogues between tutors and students, this study demonstrates that, with pedagogically aligned prompts, LLMs can function as accessible and adaptive tools for automating reflective guidance and objectively assessing the performance of both tutors and students. This work also contributes to the evolving understanding of AI's role in reflective pedagogy and highlights new possibilities for AI-driven intelligent tutoring systems.

This paper analyses the limit cycle oscillation (LCO) phenomena that characterize some of the most commonly adopted digital algorithms for maximum power point tracking (MPPT) in photovoltaic (PV) generators. The paper identifies necessary conditions for the LCO existence and clarifies the relationship between its amplitude, and frequency, and some relevant system parameters, e.g. DC-DC converter control strategy, adopted MPPT algorithm and iteration frequency, A/D converter resolution. The analysis allows to devise an optimized design procedure, where the algorithm iteration frequency can be maximized without causing limit cycle amplification. The correctness of the results is verified performing computer simulations as well as experimental tests on a 1 kW peak photovoltaic (PV) generator.

This paper explores the rapidly evolving ecosystem of publicly available AI models, and their potential implications on the security and safety landscape. As AI models become increasingly prevalent, understanding their potential risks and vulnerabilities is crucial. We review the current security and safety scenarios while highlighting challenges such as tracking issues, remediation, and the apparent absence of AI model lifecycle and ownership processes. Comprehensive strategies to enhance security and safety for both model developers and end-users are proposed. This paper aims to provide some of the foundational pieces for more standardized security, safety, and transparency in the development and operation of AI models and the larger open ecosystems and communities forming around them.

Hierarchical Navigable Small World (HNSW) has demonstrated impressive accuracy and low latency for high dimensional nearest neighbor searches. However, its high computational demands and irregular, large-volume data access patterns present significant challenges to search efficiency. To address these challenges, we introduce P-HNSW, a software-hardware codesign solution for HNSW that integrates a Principal Component Analysis Filter (PCAF). On the software side, we apply PCAF to reduce dataset dimensionality, thereby lowering the volume of neighbor access and decreasing the computational load for distance calculations. On the hardware size, we design the P-HNSW processor with custom instructions to optimize search throughput and energy efficiency. Experimental results from RTL design synthesized using a 65nm technology node show that the proposed P-HNSW implementation achieves an 11.65× increase in Queries per Second (QPS) and a 56.8% reduction in energy consumption, compared to the original HNSW implementation

Dementia is a progressive condition that significantly impairs quality of life, with limited effective treatments available for managing its late stages. Music therapy has emerged as a promising intervention, potentially improving well-being in dementia patients through physiological and emotional engagement. This study explores the physiological responses of late-stage dementia patients undergoing a music therapy intervention, utilizing manifold learning to analyze data recorded over three sessions: a control session, after 1 week of intervention, and after 6 weeks. We focus on characterizing changes in heart rate (HR), electrodermal activity (EDA), movement (ACC), and skin temperature (ST) to identify patterns associated with potential improvements in well-being. Manifold learning was applied to these physiological metrics to discern nonlinear relationships that may highlight the intervention's effects. Statistical analyses further examine these differences, providing quantitative support for improvements in physiological responses, indicating enhanced well-being in response to prolonged music therapy interventions.

Integrated Gradients is a well-known technique for explaining deep learning models. It calculates feature importance scores by employing a gradient based approach computing gradients of the model output with respect to input features and accumulating them along a linear path. While this works well for continuous features spaces, it may not be the most optimal way to deal with discrete spaces like word embeddings. For interpreting LLMs (Large Language Models), there exists a need for a non-linear path where intermediate points, whose gradients are to be computed, lie close to actual words in the embedding space. In this paper, we propose a method called Uniform Discretized Integrated Gradients (UDIG) based on a new interpolation strategy where we choose a favorable nonlinear path for computing attribution scores suitable for predictive language models. We evaluate our method on two types of NLP tasks-Sentiment Classification and Question Answering and show that our method yields better results than existing approaches.

In recent years, WiFi-based gesture recognition has garnered increasing attention due to its convenience and widespread applicability. However, achieving generalized crossdomain gesture recognition remains a formidable challenge. A prevalent approach to addressing this issue involves using domain labels to help gesture recognition systems acquire domainagnostic features. Nevertheless, we propose that the subjective nature of current domain dividing methods, which are based on human cognition, limits the generalization capabilities of existing methods in WiFi sensing. To overcome this limitation, we propose GESFI, an objective WiFi-based gesture recognition framework. Unlike traditional approaches, GESFI abandons the biased, subjectively defined domain labels and instead identifies latent domains within the WiFi sensing data objectively. The process begins by roughly training a gesture recognition model to understand the data distribution. Then, using a worst-case criterion, latent domains are mined from the model's perspective. Finally, GESFI minimizes the discrepancies between these latent domains to effectively learn domain-agnostic features, enabling generalized gesture recognition. Extensive experiments demonstrate that GESFI significantly outperforms existing state-of-theart methods across various scenarios. The code is available at: https://github.com/purpleleaves007/GesFi.

Recent years have witnessed a growing interest in Wi-Fi-based gesture recognition. However, existing works have predominantly focused on closed-set paradigms, where all testing gestures are predefined during training. This poses a significant challenge in real-world applications, as unseen gestures might be misclassified as known classes during testing. To address this issue, we propose WiOpen, a robust Wi-Fi-based Open-Set Gesture Recognition (OSGR) framework. Implementing OSGR requires addressing challenges caused by the unique uncertainty in Wi-Fi sensing. This uncertainty, resulting from noise and domains, leads to widely scattered and irregular data distributions in collected Wi-Fi sensing data. Consequently, data ambiguity between classes and challenges in defining appropriate decision boundaries to identify unknowns arise. To tackle these challenges, WiOpen adopts a twofold approach to eliminate uncertainty and define precise decision boundaries. Initially, it addresses uncertainty induced by noise during data preprocessing by utilizing the CSI ratio. Next, it designs the OSGR network based on an uncertainty quantification method. Throughout the learning process, this network effectively mitigates uncertainty stemming from domains. Ultimately, the network leverages relationships among samples' neighbors to dynamically define open-set decision boundaries, successfully realizing OSGR. Comprehensive experiments on publicly accessible datasets confirm WiOpen's effectiveness. Code is available at https://github.com/purpleleaves007/WiOpen.

Contextual Cascade Representations (CCR) have been introduced as a transformative framework designed to optimize the scalability and efficiency of large-scale language model architectures. The methodology employs hierarchical parameter activation, dynamically adjusting computational pathways to align with the contextual demands of specific tasks. Experimental findings demonstrated substantial reductions in energy consumption, inference latency, and parameter count, highlighting the framework's potential to achieve computational efficiency without sacrificing accuracy. Through a combination of sequentially weighted pruning mechanisms and dynamic thresholding, the approach maintained linguistic coherence across diverse tasks while ensuring adaptability to domain-specific challenges. Ablation studies demonstrated the synergistic contributions of individual components, including cross-layer coherence mechanisms, which preserved semantic alignment during the pruning process. Results from domain-specific evaluations revealed consistent improvements in accuracy and resource utilization, particularly within medical, legal, and financial datasets, where specialized linguistic structures posed significant challenges. A comprehensive analysis of error distributions further indicated that CCR excelled in mitigating inaccuracies in syntactically and semantically complex tasks. The proposed methodology leveraged context-sensitive parameter selection, enabling robust performance even in high-sparsity regimes. Quantitative metrics, including perplexity, BLEU scores, and classification accuracy, affirmed the advantages of CCR over conventional pruning and compression techniques. Comparisons with baseline architectures highlighted its ability to balance model size and performance more effectively. Additionally, insights from energy consumption analysis emphasized its contribution to environmentally sustainable computational practices. Contextual Cascade Representations exemplify a significant advancement in addressing the challenges of efficiency, adaptability, and performance retention in contemporary language model research.

Human Activity Recognition (HAR) is a field that automatically recognizes and categorizes human activities based on data collected from wearables, smart homes, and other IoT devices. Recognizing human activity is not an easy task because an action can occur periodically, and two activities can have almost similar properties in terms of signals received from sensors. Also, how an activity is performed may differ from person to person. This study describes how to identify a specific type of human physical activity using human pose detection. Human pose detection involves detecting critical key points that can describe the orientation or movement of the human body by correlating them with critical points of the body to detect body parts of interest. Therefore, proposing a method to label and mark activity transitions between two consecutive activities using body key detection and machine learning classifiers as the fundamental human activity recognition in the learning and classification module can improve the prediction probability of Human Activity Recognition. The combination of these two techniques improves the performance of the Human Activity Recognition system, making it more accurate in recognizing complex and varied human activities.

The classical definition of a charge is limited to assigning opposite polarities to it, which is used to design electric current circuits whose charges are sustained by an energy from a potential (voltage) difference for example. However, recent research and reviews of physical plasma chaotic behaviour revealed that the energy acquired by the charges can be studied closely in space and time domains using novel attributions such as charge (dimness) or (brightness). This is both qualitatively and computationally significant especially when building display screens. After the decline of physical plasma screens technology more than10 years ago, it has been recognized again as it provides unique displays for a variety of media channels. The research presented here concerns working on the so - called mixed modes pixels screens technology, where each screen pixel is allowed to house light emitting diodes elements in addition to making room for a physical plasma compartment to house an electric discharge. Switching between, or mixing, the output from these elements is what defines the overall viewer experience. Moreover, it will be shown that recent research on tracking charge transport across a display screen, allows for a selective capture of images from the sender and receiver at the screen surface, thereby enabling a possible two way televised transmission.

On-chip integration of 2D materials with exceptional optical properties provides an attractive solution for next-generation photonic integrated circuits to address the limitations of conventional bulk integrated platforms. Over the past two decades, significant advancements have been made in the interdisciplinary field of 2D material integrated photonics, greatly narrowing the gap between laboratory research and industrial applications. In this paper, we provide a perspective on the developments of this field towards industrial manufacturing and commercialization. First, we review recent progress towards commercialization. Next, we provide an overview of cuttingedge fabrication techniques, which are categorized into large-scale integration, precise patterning, dynamic tuning, and device packaging. Both the advantages and limitations of these techniques are discussed in relation to industrial manufacturing. Finally, we highlight some important issues related to commercialization, including fabrication standards, recycling, service life, and environmental implications.

Facial landmark detection plays a pivotal role in computer vision applications such as facial recognition, emotion analysis, and augmented reality. This study proposes an improved strategy for detecting landmarks from facial images using efficient machine learning algorithms. By leveraging Histogram of Oriented Gradients (HOG) for feature extraction and Support Vector Machines (SVM) for classification, the proposed methodology demonstrates enhanced accuracy and computational efficiency. Extensive experiments conducted on the iBUG-300W dataset show a significant improvement over existing methods. The results highlight the suitability of the approach for real-time applications while maintaining robustness to variations in pose, expression, and illumination.