Suspiciousness detection is crucial for anticipating potential security threats and facilitating timely intervention and risk mitigation. Enhanced by deep learning, vision-based systems can play a significant role in this area. This paper presents a computationally efficient vision-based suspiciousness estimation system cascading detection, classification, and analysis modules. It incorporates factors like suspicious objects, facial expressions, and abnormal body language. A suspicious object detector (SOD) is designed to precisely locate the objects invariant to the size, scale, rotation, translation, and occlusion. A deep convolutional neural network is implemented for body language and facial expression recognition with the image and landmark features as its input. To estimate the final suspiciousness, an algorithm (USE-riskometer) is proposed in this work by considering all the detected and analyzed factors. All the modules are separately trained with the corresponding object detection and facial expression datasets. In addition to this, a novel dataset for suspiciousness estimation is proposed in this work. This dataset includes various suspicious elements, such as weapons, fire, crowd presence, facial expressions, and body language in an uncontrolled environment. Each module, along with the dataset, is evaluated against state-of-the-art methods, demonstrating robustness. This work represents a significant advancement in preemptive security measures, leveraging advanced technologies for improved recognition of suspicious activities.

Recent findings reveal that a simple one-layer linear model can significantly outperform transformer-based models. To further enhance time series forecasting performance, we introduce a feedforward neural network (FNN) -based deep learning model, which surpassed most existing models. The novel FNN-based model presents three innovations: input and output structure, FNN division and combination, and an FNN sharing mechanism. The model utilizes three inputs: historical values, historical timestamps, and future timestamps, in order to predict the future demand as its output. FNN division means different inputs corresponding to different FNNs, and the outputs of these FNNs are combined to form the final prediction. The FNN sharing mechanism involves mapping each column (or row) of the matrix inputs to the column (or row) using a common FNN, termed FNN kernel. This FNN neural network framework reduces computational complexity and the number of weights significantly while still providing interpretability and strong performance. Our comparative experiments demonstrate that all three models outperform existing models on two real-world datasets.

We propose to estimate the region of attraction (ROA) for the stability of nonlinear systems from only system measurement data and without knowledge of the system model. The key to our result is the use of Koopman operator theory to approximate the nonlinear dynamics in linear coordinates. This approximation is typically more accurate than the traditional Jacobian-based linearization method. We then employ the Extended Dynamic Mode Decomposition (EDMD) method to estimate the linear approximation of the system through data. This is then used to construct a Lyapunov function that helps estimate the ROA. However, this estimate is typically very conservative. The trajectory reversing method is then used on the set of points that form this conservative estimate, to enlarge the ROA approximation. The output of EDMD is also utilized in the trajectory reversing method, keeping the entire analysis data-driven. Finally, an example is used to show the accuracy of this data-driven method, despite not knowing the system.

This paper delves into the integration of quantum technologies with semantic communications, addressing challenges related to transmitting knowledge graphs over conventional wireless channels. As traditional communications evolve towards the sharing and elaboration of knowledge between intelligent agents, both natural and artificial, they encounter critical challenges in terms of efficiency, accuracy, and the associated high complexity. To address these challenges, we propose leveraging quantum mechanical principles to communicate conceptual graph-based semantic representations of data between networked parties. The primary goal of the protocol is to transmit essential semantic features accurately while minimizing resource consumption and maintaining resilience against depolarizing errors. These include representing the semantic messages efficiently, using a quantum data encoding method, and employing remote state preparation to transmit the messages. Our work evaluates both the success probability of decoding semantic messages and investigates scalability analysis. Our results reveal that quantum semantic communication is robust against a noisy environment, and the evaluation of scalability analysis shows a trade-off regarding the complexity of communicating conceptual graphs.

The crucial element for ensuring secure optical communication lies in the consistent recovery of high-quality messages and validation needs to occur periodically. Semiconductor lasers have proven to be outstanding devices in the role of transceivers. This study delves into optical communication utilizing unidirectionally coupled multimode semiconductor lasers, specifically 3-mode lasers, as both transmitter and receiver. Two types of coupling are examined: global coupling, where the output of transmitter modes connects to corresponding receiver modes, and cross-modal coupling, which introduces a detuning equivalent to the frequency spacing of modes. The research is focused on investigating the impact of the transmission coefficient and bias current on the quality of the transmitted message. In the first coupling scenario, a high-quality message is successfully recovered, whereas the second coupling scheme demonstrates a neutral effect on message recovery quality. Message recovery is shown with an analog signal and a periodic bit sequence.

A document by André Costa Drummond. Click on the document to view its contents.

In this work, a modeling methodology is proposed for building up a quasi-analytical equivalent circuit of semiconductor diodes. In particular, it is demonstrated how this nonlinear intrinsic model can be directly extracted from port current and voltage waveform responses under single-tone injections through a formulation which relates the spectral coefficients of the state-functions involved. The extraction technique also includes a multi-bias S-parameters-based flow for characterizing the extrinsic linear network whose proper estimation is essential in high-frequency applications. Another strength of the proposed solution is that the quasi-static version can be extended to a first-order nonquasi-static approach for the modeling of typical relaxation times in high-frequency electron devices. These models have been analytically formulated and implemented in a Computer-Aided Design tool for a commercial Schottky diode and a varactor. Experimental tests for both devices at microwave frequencies prove how the nonquasi-static equivalent diode successfully responds when frequency grows up. In addition, extracted circuits are able to extrapolate in frequency under several bias and power conditions. This strategy provides a solution to overcome some foundry models' limitations under modern power and frequency operating conditions and enhances reconfigurable prototypes design.

Device-to-Device (D2D) communication, a state-ofthe-art for next Generation communication paradigm, facilitates direct file transfer among neighboring nodes, without involving cellular base station. In this paper, we consider a purchase model related to content distribution. In this context, we propose a novel D2D content sharing scheme using the frameworks of Evolutionary Minority Game(EMG) and Multiple Choice Minority Game(MCMG). The performance of our content purchase model, in terms of its convergence , content downloading efficiency, etc., has been studied through extensive simulation. In particular, we studied the system performance by varying different system parameters, such as pricing parameters, cellular data rates, and cluster size. We demonstrate that our self-organizing game model achieves near-optimal performance, without requiring excessive coordination and signaling overhead.

This systematic literature review (SLR) explores the application of Artificial Intelligence (AI) and deep learning methods in emotion recognition and patient-centric healthcare. The objective is to evaluate the effectiveness of these technologies in enhancing patient care through emotional analysis. We examine research studies utilizing AI to advance sentiment analysis, categorize emotions, and predict patient outcomes using textual data from clinical narratives, patient feedback, and online health discussions. The review highlights progress in emotion recognition algorithms, integration of cognitive and emotional analysis techniques, and creation of AI-powered systems for clinical decisionmaking. AI applications have shown significant improvements in personalizing therapy plans, facilitating early detection of mental health disorders, and providing robust support for clinical interventions. Despite advancements, challenges remain in ensuring ethical AI application, maintaining patient confidentiality, and addressing algorithmic biases. Nevertheless, AI's potential to transform healthcare practices is clear, pointing towards more informed, efficient, empathetic, and patient-centered healthcare. This study emphasizes AI's transformative impact on healthcare by providing a thorough understanding of its role in analyzing emotional content in healthcare texts and identifying emerging trends and future research directions. The findings advocate for harmonizing AI's analytical prowess with the human aspects of healthcare, ensuring technological progress supports patients' emotional well-being.

Artificial intelligence systems increasingly require robust evaluation techniques to ensure their logical reasoning capabilities align with practical applications. This research introduces a novel approach to empirically evaluate symmetrical reasoning in leading language models, specifically ChatGPT, Gemini, and Claude, utilizing automated interaction methods. The evaluation, encompassing metrics such as accuracy, consistency, and logical coherence, revealed distinct strengths and weaknesses among the models, with ChatGPT excelling in accuracy and logical coherence, Gemini demonstrating superior consistency, and Claude highlighting areas for further improvement. By employing a diverse dataset of symmetrical reasoning tasks and sophisticated automated scoring algorithms, the study provides an objective and comprehensive assessment, significantly contributing to the understanding of logical reasoning in AI. The findings underscore the critical importance of continuous model refinement and the integration of ethical considerations in developing advanced AI systems.

The escalating threat of climate change necessitates urgent and comprehensive action to mitigate its impacts. Renewable energy technologies, such as solar, wind, and hydroelectric power, offer sustainable alternatives to fossil fuels, significantly reducing greenhouse gas emissions. Concurrently, the adoption of EVs presents a viable solution to curbing emissions from the transportation sector, one of the largest contributors to global warming. However, the deployment of renewables and EVs, especially in electrical distribution systems, poses significant challenges - particularly in terms of loading and voltage violations, as well as of power quality deterioration - representing one of the major limitations to their large-scale deployment. Without careful planning and control, such deployment may be technically unfeasible or necessitate extensive and costly upgrades to the distribution infrastructure, potentially affecting up to 95% of existing assets. By analyzing a case study in Seychelles the paper provide a framework for sizing EV, battery storage systems (BESS) and photovoltaic (PV) with a clear focus in eliminating the needs of costly grid reinforcements accelerating a substantial reduction in carbon footprint. Specifically, our approach strikes a balance between grid impact and performance, resulting in a sustainable solution for both customers and system operators. The findings also underscore the necessity for policies to support the proposed solutions.

Line spectra is an important phenomenon in many signal processing application including radar, sonar, communications, and spectroscopy. The estimation of a frequency component in a noisy sinusoidal signal for source localization and Doppler estimation is often required for practical systems such as cooperative communication, and vehicular communications. Here, we propose a quantum framework for the line spectrum estimation in the presence of observation noise and quantum error. We propose quantum algorithm, namely quantum line spectrum estimation (QLSE) to estimate the frequency of a sinusoidal signal from a quantum measurement probability histogram. Considering complex sinusoidal waveform, we show how to estimate single tone and multiple tone signals from the quantum circuit. We demonstrate a fully quantum computational model implemented on an IBM quantum machine and discuss quantum complexity advantage and error performance of the proposed algorithm in comparison with classical method. The experimental study shows promising result for super-resolution frequency estimation application with the proposed QLSE algorithm. The numerical results are given to support the proposition in the quantum regime. This manuscript is submitted to a journal with adequate modifications. The copyright may be transferred without notice, and this preprint may not be available after publication. 

We introduce a Neural Network optimization model inspired by the renowned Proportional-Integral-Derivative (PID) controller. This method leverages the principles of the PID controller to effectively reduce loss errors during training. By applying the theory behind PID control, our approach aims to achieve higher accuracy across diverse objectives and problems involving large datasets. Additionally, it is well-suited for non-stationary objectives with noisy or sparse gradients. The hyperparameters of our PID-inspired optimizer have intuitive interpretations. The connections between these hyperparameters and the PID controller are thoroughly discussed. We tested the PID optimizer across various scenarios, demonstrating its superior performance and robustness.

Due to the time-varying nature of EMG signals, designing a practical EMG-based motor decoder for robotic prosthetic hands is still a challenge. Contribution: To address this issue, we present a novel personalized and adaptive motor decoding framework, considering the main challenges in practice. The proposed framework operates based on initial model calibration and unsupervised domain adaptation during daily usage to attain optimal motion detection rate and speed by minimizing hardware requirements. Methods: Twelve subjects (8 male and 4 female) participated in this study, where we recorded EMG signals from three of their forearm muscles while performing 20 repetitions of 9 different hand motions. Results: Ignoring the sequence of data, it was observed that the motion classifier trained on data from other subjects has a significantly lower classification rate (28±0.8%) compared to a personalized model (94±4.7%). We also demonstrated that data from the same subject on a different day is not comparable (37±4.6%) and can be considered as data from another subject. In addition to the inter-subject and inter-session variability of EMG signals, our observations also indicated significant intra-session variability. This necessitates the adaptation method to be conducted within a session while preserving the sequence of data for analysis, a consideration overlooked by previous studies. According to our results, at the start of usage, the accuracy of non-adaptive and proposed adaptive models is almost identical, with an accuracy of 96.3±4.4%. However, after only 17 repetitions, the adaptive model's accuracy (88.8±7.9%) is on average 4.86% higher than the non-adaptive model (84.02±8.38%); t-test result: p* = 0.031.

Education is a basic right, but not everyone gets the same opportunity to learn. Certain minorities and people in impoverished areas face hardships accessing basic informational and educational resources. Our solution provides them with educational materials such as open-source articles, lecture videos, notes etc. in myriad subjects. It also allows the users to contribute the abovementioned educational resources to the platform, according to their capacity, promulgating a smooth process of reciprocity. The main idea behind this is to make the resources accessible offline so that the users living in remote locations can download the content they choose and study it at their own pace, wherever they want, whenever they want.This case study posits a newfangled example of mobile learning application (MLA) development. While content delivery remains a core function, we argue that incorporating user interface and user experience design principles is more than just visual appeal. User preferences and accessibility considerations must be prioritized within MLAs. Learning environments must be dynamic; their content delivery must be effective. Further, they must cultivate engagement, optimize learning outcomes, and contribute to a more sustainable learning experience for users.

The advent of sophisticated language generation technologies has revolutionized various domains, from automated customer service to advanced research aids. However, a significant challenge persists in the form of contextual hallucinations, where models generate text that is factually incorrect or contextually inappropriate. The novel approach of dynamically adjusting attention weights within the Llama model addresses this issue by enhancing the relevance and accuracy of the generated content. The methodology involved collecting a diverse dataset, fine-tuning the model, and implementing attention map modifications to steer the model's focus towards pertinent information, thereby reducing hallucinations. Empirical evaluations demonstrated substantial improvements in factual consistency, context relevance, and semantic coherence scores, validating the efficacy of the proposed modifications. Comparative analysis with baseline models further highlighted the superior performance of the modified Llama model, revealing its potential for reliable application in critical tasks. Insights gained from attention pattern analysis provided a deeper understanding of hallucination mechanisms, informing future strategies for model refinement. The study's implications extend to the broader field of language model reliability, offering a robust framework for enhancing text generation quality across various applications.

A document by Mudi Zhang. Click on the document to view its contents.

Artificial Intelligence (AI) is a technology that has significant impact on the economy and the people. AI entails development of machines and computer system that emulate human intelligence and develops algorithms that help with decision-making and predictions. The following paper entails a literature review analysis on the impact of AI on the economy and the people. The review of various peer-reviewed sources highlighted various themes on the impact of AI on the economy, which included productivity, labor market, and economic inequality. However, this review is specifically on the analysis of these reviewed sources that overall helps us understand the consequences and benefits of AI on the economy. The impact of AI on the people includes ethical concerns, language barrier and social interactions.