Short-term load forecasting (STLF) is essential for efficient power system operations, including unit commitment and energy market operations. Traditional deep learning models like Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks have shown promise but often struggle with capturing complex temporal dependencies and nonlinearities in electricity load data. This paper introduces SpikeNet, a novel neural network architecture incorporating integrate-and-fire (IF) spiking neurons with Recurrent Neural Network (RNN) and LSTM layers to improve temporal dynamics modelling in STLF. Spiking neurons process information through discrete events "spikes", offering enhanced efficiency and the ability to model temporal patterns more effectively. We evaluate SpikeNet's performance against state-of-the-art models on three diverse datasets: Islamabad Electric Supply Company (IESCO), the IEEE data portal, and Panama City. Our results demonstrate that SpikeNet outperforms traditional models, achieving up to 33% improvement in mean absolute percentage error (MAPE) and up to 42% reduction in inference time. These findings highlight SpikeNet's potential as a more accurate and efficient solution for STLF, with significant implications for real-world power system applications.

This paper presents a novel modeling approach for flying capacitor dynamics in boost-type multilevel converters (FCML-boost) controlled by Phase Shift Pulse Width Modulation (PSPWM). By explicitly taking into account the interaction between the inductor current and the flying capacitor voltage, the model is able to reveal an underlying resonance phenomenon and to predict its frequency at each operating point. Based on such model, whose derivation is explained in detail, both passive and active damping solutions are proposed, designed, and experimentally verified that significantly reduce the undesirable oscillations. The analytical results and the devised control solutions are tested on a 1kW, 4 level, boost DC-DC converter prototype employing 200 V, 48 A rated EPC2034C GaN devices.

In recent years, wireless sensing has become a significant research focus in the field of Internet of Things due to its advantages of wide sensing range, low deployment cost, and non-intrusive to user privacy. However, wireless signals are susceptible to environmental interference, and their sensing capability is highly coupled with the scenario, leading to limitations in robustness and scalability of existing systems. To address these challenges, one possible solution is to fully explore the collaborative sensing potentials of ubiquitous wireless sensing resources. To this end, we introduce the concept of "Crowd Wireless Sensing" for the first time, and propose a new wireless sensing framework named CrowdSignal, aiming to achieve robust and scalable wireless sensing based on the collaboration and complementarity of dynamic gathering sensing resources in real-world scenarios. Preliminary experimental results validate the importance and viability of the proposed framework. To further advance the research in this emerging field, we elaborate possible future directions at the end of the article.

This paper aims to address the typical needs met in designing Bio/CMOS interfaces for the remote monitoring of human metabolism. In particular, the design of robust and reliable interfaces for wearable, implantable, or point-of-care devices is considered as model application. Selectivity, sensitivity, and autonomy of such devices are introduced and discussed as the main features of such personal tools for health monitoring. The typical solutions presented in literature to assure the right selectivity, sensitivity, and autonomy are discussed in detail. Some of those solutions are on board in many technologies introduced into the market already, demonstrating the feasibility of such design approach obtained by integrating three system' layers: bio, nano, and CMOS (Complementary Metal-Oxide-Semiconductor). These largely spread technologies for personal health monitoring (e.g., the glucose auto-monitoring devices for chronic patients with diabetes) provide distributed diagnostics in term of billions of connected devices. Billions of connected personal devices worldwide distributed oblige to consider the several issues related to sustainability in term of minimal environmental impact. Therefore, the paper also introduces some new emerging design strategies by modular design, substrates made of recycled plastic or bioplastics by crop waste, and biodegradable materials.

Interpersonal synchrony (IS), a crucial indicator of social interactions, is traditionally studied through video data and manual coding methods. Our study aims to develop automated coding and visualization tools for time series data collected by IMU sensors, offering a more efficient, objective, and scalable analysis of interaction dynamics. This study evaluates several time series similarity analyses, including Cross-correlation (CC), Dynamic Time Warping (DTW), and Cross-Wavelet Analysis, to predict interaction levels using regression and classification techniques. We conducted a comparative study using both simulated data and real-world data involving sessions between teachers and children with and without autism to optimize the parameters for each algorithm and investigate different combinations of time series analysis. The results demonstrate that combining different time series analyses is advantageous, with Cross-wavelet Analysis (CWA) being particularly effective in quantifying interpersonal synchrony due to its ability to analyze synchrony across different frequencies. Additionally, a visualization tool is designed to display the dynamics of interaction over time and pairwise interactions conveniently. This paper provides a step-by-step guideline for studying interpersonal synchrony. It establishes a robust automated algorithm as a tool to track, visualize and understand social interactions, which can be easily adapted to a broader range of motor-coordination-related applications.

The integration of wireless power (WP) and mobile edge computing (MEC) offers a promising solution to enhance both system performance and the computational capabilities of mobile devices. This paper investigates a WP-MEC system employing non-orthogonal multiple access (NOMA) to improve communication efficiency, particularly in user cooperation scenarios. Our objective is to maximize system throughput by optimizing time fraction allocation, transmit power, and task offloading decisions, while ensuring task queue stability constraints in a dynamic network environment characterized by time-varying wireless channels and stochastic task arrivals. To address this, we employ a non-linear utility function of time-averaged throughput based on the law of diminishing marginal utility in economics and formulate a stochastic optimization problem. By leveraging perturbed Lyapunov optimization theory and virtual queues method, we decompose the problem into tractable sub-problems for each time slot. For resource allocation, we apply variable substitution and convex optimization techniques to efficiently solve the inherently non-convex problem. We propose a low-complexity Dynamic User Cooperation Throughput Maximization (DUCTM) algorithm, which balances system throughput and stability. Our theoretical analysis shows that DUCTM achieves a trade-off between throughput and queue backlog characterized by [O(1/V ), O(V )]. Numerical results validate the superiority of our proposed algorithm over baseline schemes, demonstrating robustness and adaptability to time-varying and bursty task arrivals.

Deforestation monitoring in Brazil for Land use land cover application is the combination of up-to-date monitoring and accuracy. For detailed observation on time, need a Sentinel-2 multi-spectral satellite imagery which is a combination of multiple bands of different frequency for better analysis. Sentinel-2 Multispectral Images are the combination of high resolution and low-resolution images which create problem for classifying the object due to mixed pixel problem. Due to mixed pixel problem in optical satellite detection of the object is difficult which effects the accuracy. To identify the mixed pixel problem to combine multiple bands using Band Math to create a new band for detecting mixed pixel and to analyse the pixel using segmentation and clustering. To classify the Brazil Amazon Forest deforestation between 2019 and 2023 proposing a Satellite Image clustering Transpose Transformation Deep Neural Networking (SiCTT.net). To compare the CNN and Transpose CNN transformation with the help of accuracy and based on the results, proposed network gives better accuracy and helps to detect mixed pixel problem.

The increasing complexity of language tasks necessitates models capable of complex contextual understanding across multiple levels. Addressing this need, Adaptive Semantic Layering (ASL) introduces a hierarchical framework that structures semantic information within Large Language Models (LLMs), thereby enhancing their ability to process intricate linguistic patterns. Through the integration of ASL, LLMs exhibit improved performance in domain-specific terminology comprehension, response time efficiency, and adaptability to linguistic variations. Empirical evaluations demonstrate that ASL effectively addresses existing limitations in LLM architectures, offering a novel approach that contributes to the theoretical and practical advancement of natural language processing technologies.

This research bridges the gap between claimed and actual ethical behavior by analyzing subtle "hidden honest signals" in language using AI and Natural Language Processing (NLP). It categorizes individuals into four "tribes": "doves" (ethical, collaborative), "wolves" (loyal, diligent), "sharks" (competitive, aggressive), and "parasites" (self-serving, exploitative). Using data from migrant workers in Taiwan, the study examines how ethical behavior correlates with productivity in healthcare and manufacturing sectors. Results show that ethical behavior influences work outcomes, though not always in expected ways. The findings suggest that developing ethical training programs could improve productivity, particularly for migrant workers in these settings.

The evolution of language models has achieved remarkable breakthroughs in understanding and generating human-like text, yet a gap persists in managing recursive semantic dependencies that require dynamically layered contextual interpretation. Addressing this limitation, the proposed Dynamic Recursive Framework (DRF) introduces an innovative recursive processing mechanism that adapts context dynamically across layered dependencies, enhancing language models' abilities to handle complex, recursive structures in text. Unlike conventional architectures that rely on static attention mechanisms, the DRF utilizes recursive feedback loops to achieve complex semantic alignment, particularly valuable for tasks requiring the interpretation of long-range dependencies, polysemous disambiguation, and layered meanings. Experimental evaluations reveal that the DRF improves semantic alignment accuracy and contextual consistency, while only minimally increasing computational demands, showing its potential as a significant advancement for tasks demanding recursive contextual comprehension. Implemented on an open-source language model, the DRF's contributions to both performance and interpretive depth illustrate its transformative impact on the structural capabilities of language models, paving the way for models that can process the complexity of human language with enhanced precision and adaptability.

The IETF Manufacturer Usage Description (MUD) standard seeks to enhance IoT security by restricting device communication to specified network flows. However, while MUD profiles define permitted flow destinations, they lack visibility into the content exchanged, leaving devices vulnerable to unintended payloads. This paper introduces a systematic approach to categorize MUD flows by their payload types-transparent, encoded, or encrypted-providing insights crucial for cyber-risk assessment, anomaly detection, and selective packet inspection. Our contributions are threefold: (1) We collect a dataset consisting of over 123000 network flows involving 24 different protocols, label each record with a ground-truth label of payload type, and analyze them to identify 17 attributes predictive of three classes-we make our data publicly available; (2) We train and evaluate a baseline classifier with over 90% accuracy for detecting flow payload types, and show encoded and encrypted flows demand higher confidence scores to pass a practical threshold 0.8; (3) We demonstrate how detailed payload characterization in representative MUD flows significantly improves security profiling for IoT devices at computing costs higher than our baseline. We develop an inference pipeline with three specialized classifiers that are dynamically and progressively applied to the payload of input MUD flows. Our evaluation results show that highly (more than 99%) accurate and confident inferences can be made across three payload classes by spending 44% more computing resources.

A document by Jishnudeep Kar. Click on the document to view its contents.

Sequential Model-Based Optimization (SMBO) is a highly effective strategy for hyperparameter search in machine learning. It utilizes a surrogate model that fits previous trials and approximates hyperparameters' response surface (performance). This surrogate model primarily guides the decision-making process for selecting the next set of hyperparameters. Existing classic surrogates, such as Gaussian processes and random forests, focus solely on the current task of interest and lack the capability to incorporate trials from historical tasks. This limitation hinders their efficacy in various applications. Inspired by the state-of-the-art convolutional neural process, this paper proposes a novel meta-learning-based surrogate model for efficient and effective hyperparameter optimization. Our surrogate is trained on the meta-knowledge from a range of historical tasks, enabling it to accurately predict the hyperparameter response surface even with a limited number of trials on a new task. We tested our approach on the hyperparameter selection problem for the well-known support vector machine (SVM) and residual neural network (ResNet) across hundreds of real-world classification datasets. The empirical results demonstrate its superiority over existing surrogate models, highlighting the effectiveness of metalearning in hyperparameter optimization.

Realistic leukemia analysis on microscopic slides requires a system that localizes, classifies, and performs WBC morphological analysis. While deep learning has advanced medical imaging, leukemia analysis is hindered by the lack of a large, diverse dataset. Our contribution here is twofold. First, we present a large-scale Leukemia dataset collected through Peripheral Blood Films (PBF) from many patients, through multiple microscopes, multi-cameras, and multi-magnification. To enhance explainability and medical expert acceptance, each leukemia cell is annotated with 7 morphological attributes, ranging from Cell Size to Nuclear Shape. Specifically, we used two microscopes from two different cost spectrums (high-cost: HCM and low-cost: LCM) to capture the dataset at three magnifications (100x, 40x,10x) through different sensors (high-end camera for HCM, middle-level camera for LCM, and mobile phone camera for both). Our dataset has 2.4k annotated images per resolution (in both HCM and LCM), with 10.3K cells annotated for 14 WBC types and 7 morphological properties. Secondly, hematologists typically annotate only well-spread, non-overlapping cell regions within small fields of view (FoV), leaving larger FoVs unannotated. Models trained solely on these annotated areas underutilize available information. To overcome this limitation, we present a novel method for training object detectors using sparse annotations. Along with this, we present a second dataset Leukemia-Sparse-Attri dataset. Curated at 40x FoV, this dataset consists of 1546 sparsely annotated microscopic images, and 185 fully annotated images (used for testing). From explainability to overcoming domain-shift challenges, presented datasets could be used for many challenging aspects of microscopic image analysis.

Artificial Intelligence (AI) is rapidly transforming educational management, presenting innovative opportunities to foster inclusive leadership within multicultural and globalized contexts. This qualitative study explores the potential of AI to enhance inclusive leadership by enabling real-time responsiveness, data-driven decision-making, and improved cultural competency. Utilizing content and Hybrid Thematic Content Analysis (HTCA) based on a comprehensive literature review, this research compares traditional leadership approaches with AI-integrated models, focusing on their impact on practice, theory, and policy within educational institutions. Key findings indicate that AI promotes inclusivity through proactive, culturally aware interventions and continuous tracking of inclusion metrics, which advance accountability and strategic planning. This study provides critical insights into the ethical considerations and operational benefits of AI in educational leadership, underscoring the need for balanced, human-centered AI applications that align with institutional values and support sustainable educational transformation.

The emergence of Large Language Models (LLMs) transformed the domain of Natural Language Processing (NLP) by enhancing the capability of machines to effectively comprehend and generate natural language. These LLMs are a type of generative AI and the underlying AI model that work behind the scenes of most modern AI chatbots. Generative AI is an umbrella term for all AI technologies which can generate original contents. These LLMs are pre-trained on a large amount of textual data and billions of parameters. This pre-training is normally unsupervised learning, meaning that it processes the unlabelled data for a comprehensive understanding of context, semantics, and grammar of natural language to generate coherent, context-relevant and credible text. In view of the significance of LLMs, this paper aims to perform a comprehensive study of LLMs, which will elucidate the historical journey of language processing and modelling, evolution and types of language models, and working, benefits and limitations of LLMs.

The rise of creative machines is attributed to generative AI which enabled machines to create new contents. Wherein the introduction of the advanced neural network architecture known as a transformer revolutionized the landscape of generative AI. A transformer transforms one sequence into another sequence, and is primarily used in natural language processing and computer vision tasks. Which determines the relationship between tokens or words in a sequence to understand the context, while processing these tokens or words simultaneously. Transformers were built to resolve various issues of its previous neural networks, such as a Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM); and are now the brains behind majority of generative AI models, for example, most Large Language Models (LLMs) and Large Multimodal Models (LMMs). This paper will explain about the transformer and its architectural components and working. Subsequently, it will illustrate the decoder-only transformer architecture and its components and working, including the reason why this type of transformer architecture is used in most generative AI models such as majority of LLMs and LMMs.

This paper presents a novel methodology for design space exploration and resource optimisation of three-dimensional Networks-on-Chip (3D NoC) architectures using hypergraph modelling and genetic algorithms. The proposed approach combines the mathematical rigour of hypergraph theory with the evolutionary search capabilities of genetic algorithms to efficiently explore the vast design space of 3D NoC configurations. Unlike existing vendor-specific tools, our framework provides a vendor-independent solution for NoC architects and designers. The key contribution is the development of Performance-Cost-Ratio (PCR) functions that enable quantitative evaluation of different topologies and routing algorithms. We validate our framework through two compute-intensive use cases: double SHA256 cryptographic operations and real-time facial recognition. Experimental results demonstrate significant improvements in both applications, with the optimised architectures achieving up to 33% reduction in latency, 40% increase in throughput, and 30% reduction in power consumption compared to baseline implementations. The insights and techniques presented have farreaching implications for developing efficient and scalable NoC solutions as processor designs advance towards kilo-core scales and beyond.