Cross-Lingual Speech-to-Text Systems with Low-Latency Neural Networks for Real-Time Applications.

Imprecise Neural Computation in Real-Time Neural System Design

This research optimizes neural network topologies for real-time image and video processing to achieve high-speed, accurate performance in dynamic contexts. The project aims to find efficient optimization methodologies, track neural network model progress, and highlight visual media applications. A secondary data review synthesizes peer-reviewed literature, technical reports, neural network design, and optimization advances. The research found that lightweight neural network architectures like MobileNet and Transformer-based Vision Transformers (ViTs) boost the computing economy without losing accuracy. Real-time applications need model pruning, quantization, knowledge distillation, and hardware-aware design. From real-time object identification in surveillance and autonomous driving to medical imaging and creative media creation, neural networks have transformed many applications. Despite these advances, balancing accuracy and economy, addressing hardware variability, and assuring ethical usage in face recognition remain issues. The report emphasizes the need for privacy-friendly and egalitarian AI rules. These results may help future research improve real-time visual processing systems and legislators control their responsible use in real-world applications.

Harmonic pollution minimisation in voltage-source programmed pulse-width modulation (PPWM) inverters is defined as a time-limited optimisation problem in real-time applications with variable DC sources. In order to obtain minimum total harmonic distortion (THD) as the objective function, shuffled-frog-leaping algorithm (SFLA) is modified and employed to calculate the switching angles and compared with non-linear programming as a traditional optimisation method. In addition, particle swarm optimisation and three of its modified versions as popular evolutionary optimisation algorithms are employed to ensure the capability of the proposed optimisation method. Moreover, modified sinusoidal PWM (MSPWM) THD is compared with PPWM THD. Furthermore, as the DC bus voltage in some applications might have high variations (in amplitude or frequency of fluctuations) in a short time, to acquire adequate response speed to this variation of DC source of inverters in real-time control applications, a neural network (NN) is trained by the off-line calculated results of MSFLA for various desired modulation indexes (various DC voltages). Simulation results demonstrate the accurate and high-speed response of the designed NN. The main contribution of this study is to provide a fast accurate method which can track the variation of DC source of inverters with high-quality solutions in real-time control applications.

Deaf people use sign languages to communicate with other people in the community. Although the sign language is known to hearing-impaired people due to its widespread use among them, it is not known much by other people. In this article, we have developed a real-time sign language recognition system for people who do not know sign language to communicate easily with hearing-impaired people. The sign language used in this paper is American sign language. In this study, the convolutional neural network was trained by using dataset collected in 2011 by Massey University, Institute of Information and Mathematical Sciences, and 100% test accuracy was obtained. After network training is completed, the network model and network weights are recorded for the real-time system. In the real-time system, the skin color is determined for a certain frame for hand use, and the hand gesture is determined using the convex hull algorithm, and the hand gesture is defined in real-time using the registered neural network model and network weights. The accuracy of the real-time system is 98.05%.

Contents: D. Sobajic, Foreword. Part I:Perspectives. Y-H. Pao, G-H. Park, Learning and Generalization Characteristics of the Random Vector Functional-Link Net. C-C. Liu, M. Damborg, Artificial Neural Networks and Expert Systems in the Power System Operation Environment. E. Bradley, A Utility Perspective on Neural Networks, Fuzzy Logic, and Artificial Intelligence. Part II:Neural Network Methodologies. B. Widrow, M.A. Lehr, Backpropagation and Its Applications. F. Beaufays, E.A. Wan, Using Flow Graph Interreciprocity to Relate Recurrent-Backpropagation and Backpropagation-Through-Time. A. Guha, Neural Network Based Inferential Sensing and Instrumentation. S.A. Harp, T. Samad, Optimizing Neural Networks Using Genetic Algorithms. Part III:Nuclear Power Plants. R. Uhrig, Potential Use of Neural Networks in Nuclear Power Plants. M. Khadem, A. Ipakchi, F.J. Alexandro, R.W. Colley, Sensor Validation in Power Plants Using Neural Networks. A. Ikonomopoulos, L. Tsoukalas, R. Uhrig, Measuring Fuzzy Variables in a Nuclear Reactor Using Artificial Neural Networks. Y.D. Lukic, C.R. Stevens, J. Si, Application of a Real Time Artificial Neural Network for Classifying Nuclear Power Plant Transient Events. J.A. Boshers, C.H.M. Saylor, S. Kamadolli, R. Wood, C. Isik, Control Rod Wear Recognition Using Neural Nets. R. Doremus, Severe Accident Management System On-Line Network (SAMSON). Part IV:Power System Operation. H. Ren-mu, A.J. Germond, Comparison of Dynamic Load Models Extrapolation Using Neural Networks and Traditional Methods. B. Avramovic, On Neural Network Voltage Assessment. D. Sobajic, Y-H. Pao, M. Djukanovic, Neural Network Synthesis of Tangent Hypersurfaces for Transient Security Assessment of Electric Power Systems. D. Niebur, A.J. Germond, Power System Static Security Assessment Using the Kohonen Neural Network Classifier. H. Mori, Voltage Stability Monitoring with Artificial Neural Networks. D. Novosel, A.B. Boveri, R.L. King, Intelligent Load Shedding. E. Chan, N. Markushevich, R. Adapa, Considerations in Intelligent Alarm Processing. Part V:Modeling and Prediction. D.J. Sobajic, Y-H. Pao, D.T. Lee, Predictive Security Monitoring with Neural Networks. A.G. Parlos, A.D. Patton, Empirical Modeling in Power Engineering Using the Recurrent Multilayer Perceptron Network. T. Samad, Modeling and Identification with Neural Networks. E. Wan, Autoregressive Neural Network Prediction: Learning Chaotic Time Series and Attractors. Part VI:Control. B. Widrow, F. Beaufays, Neural Control Systems. R.L. King, M.L. Oatts, Potential Uses of Intelligent and Adaptive Controls for Electric Power System Operations in the Year 2000 and Beyond. F. Beaufays, B. Widrow, Load-Frequency Control Using Neural Networks. L.L. Adams, Reinforcement Learning for Adaptive Control. Part VII:Load Forecasting. A.J. Germond, N. Macabrey, T. Baumann, Application of Artificial Neural Networks to Load Forecasting. M. Khadem, A. Lago, E. Dobrowolski, Short-Term Electric Load Forecasting Using Neural Networks. J.Y. Cheung, J. Fagan, D.C. Chance, Load Forecasting by Hierarchical Neural Networks that Incorporate Known Load Characteristics. Part VIII:Scheduling and Optimization. H. Sasaki, Y. Takiuchi, J. Kubokawa, A Solution Method for Maintenance Scheduling of Thermal Units by Artificial Neural Networks. H. Saitoh, Y. Shimotori, J. Toyoda, Generation Dispatch Algorithm Coordinating Economy and Stability by Using Artificial Neural Networks. Part IX:Fault Diagnosis. T. Baumann, A.J. Germond, D. Tschudi, Impulse Test Fault Diagnosis on Power Transformers Using Kohonen's Self-Organizing Neural Network. Y. Du, F. Wang, T.C. Cheng, A Case Study of Neural Network Application: Power Equipment Application Failure. A. Agogino, M-L. Tseng, P. Jain, Integrating Neural Networks with Influence Diagrams for Power Plant Monitoring and Diagnostics. W.L. Biach, Use of Neural Network in Optimizing RPV Bolting Procedures.

Currently, there are some emerging online learning applications handling data streams in real-time. The On-line Sequential Extreme Learning Machine (OS-ELM) has been successfully used in real-time condition prediction applications because of its good generalization performance at an extreme learning speed, but the number of trainings by a second (training frequency) achieved in these continuous learning applications has to be further reduced. This paper proposes a performance-optimized implementation of the OS-ELM training algorithm when it is applied to real-time applications. In this case, the natural way of feeding the training of the neural network is one-by-one, i.e., training the neural network for each new incoming training input vector. Applying this restriction, the computational needs are drastically reduced. An FPGA-based implementation of the tailored OS-ELM algorithm is used to analyze, in a parameterized way, the level of optimization achieved. We observed that the tailored algorithm drastically reduces the number of clock cycles consumed for the training execution up to approximately the 1%. This performance enables high-speed sequential training ratios, such as 14 KHz of sequential training frequency for a 40 hidden neurons SLFN, or 180 Hz of sequential training frequency for a 500 hidden neurons SLFN. In practice, the proposed implementation computes the training almost 100 times faster, or more, than other applications in the bibliography. Besides, clock cycles follows a quadratic complexity O ( N ˜ 2 ) , with N ˜ the number of hidden neurons, and are poorly influenced by the number of input neurons. However, it shows a pronounced sensitivity to data type precision even facing small-size problems, which force to use double floating-point precision data types to avoid finite precision arithmetic effects. In addition, it has been found that distributed memory is the limiting resource and, thus, it can be stated that current FPGA devices can support OS-ELM-based on-chip learning of up to 500 hidden neurons. Concluding, the proposed hardware implementation of the OS-ELM offers great possibilities for on-chip learning in portable systems and real-time applications where frequent and fast training is required.

In this paper, the implementation of a new pupil detection system based on artificial intelligence techniques suitable for real-time and real-word applications is presented. The proposed AI-based pupil detection system uses a classifier implemented with slim-type neural networks, with its classes being defined according to the possible positions of the pupil within the eye image. In order to reduce the complexity of the neural network, a new parallel architecture is used in which two independent classifiers deliver the pupil center coordinates. The training, testing, and validation of the proposed system were performed using almost 40,000 eye images with a resolution of 320 × 240 pixels and coming from 20 different databases, with a high degree of generality. The experimental results show a detection rate of 96.29% at five pixels with a standard deviation of 3.38 pixels for all eye images from all databases and a processing speed of 100 frames/s. These results indicate both high accuracy and high processing speed, and they allow us to use the proposed solution for different real-time applications in variable and non-uniform lighting conditions, in fields such as assistive technology to communicate with neuromotor-disabled patients by using eye typing, in computer gaming, and in the automotive industry for increasing traffic safety by monitoring the driver’s cognitive state.

The prevalence of financial fraud poses significant challenges to global financial stability, resulting in billions of dollars in losses annually and undermining consumer trust in financial institutions. With the increasing complexity and volume of financial transactions driven by the rapid growth of digital banking and e-commerce, traditional fraud detection methodologies have proven inadequate in addressing the scale and sophistication of modern fraudulent activities. This paper seeks to investigate and delineate the development of advanced data science and artificial intelligence (AI) methodologies aimed at detecting, mitigating, and preventing financial fraud in real-time systems. By exploring a range of state-of-the-art models, algorithms, and technologies, this research aims to provide comprehensive insights into how these systems can be deployed effectively to safeguard financial operations and maintain systemic integrity. Financial fraud detection is inherently challenging due to the dynamic and evolving nature of fraudulent tactics. The emergence of techniques such as machine learning (ML) and deep learning (DL) has significantly enhanced the ability to identify complex, non-linear patterns within large datasets that were previously undetectable by conventional rule-based systems. This paper focuses on the integration of supervised, unsupervised, and semi-supervised learning methods, as well as hybrid approaches that combine different algorithmic strategies for greater detection accuracy. In the context of financial fraud, algorithms such as decision trees, support vector machines (SVM), random forests, and neural network architectures have been adapted and fine-tuned to operate under stringent latency constraints inherent in real-time processing systems. Moreover, the adaptation of generative adversarial networks (GANs) for synthetic data generation and anomaly detection is examined to bolster the robustness and adaptability of fraud detection models. A critical aspect of this research lies in the exploration of feature engineering and data pre-processing techniques to optimize the input datasets for AI models. Given that the quality of data directly influences the efficacy of predictive algorithms, innovative feature extraction, dimensionality reduction, and data augmentation methods are discussed in detail. The use of time-series analysis and sequence modeling, especially through recurrent neural networks (RNNs) and long short-term memory (LSTM) networks, is emphasized for fraud detection in transactions that require contextual and sequential understanding. Such methodologies enable the capture of temporal dependencies that are essential for detecting anomalous behaviors indicative of fraudulent activities. Additionally, the paper addresses the significance of explainable AI (XAI) in the realm of financial fraud prevention. Trust in AI-driven fraud detection systems can be undermined by their "black-box" nature, where decision-making processes remain opaque to users and regulators. As such, incorporating interpretable models and explainability tools is essential for meeting regulatory requirements and fostering confidence in automated systems. This research evaluates various XAI techniques, such as SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations), and their integration with AI models to ensure that the decision-making process can be audited and understood by human analysts. The paper also explores the real-world applicability of AI and data science-based fraud detection through case studies of financial institutions and tech firms that have implemented such systems. These case studies illustrate the challenges faced, such as the need for real-time processing, false positive management, and system scalability. Furthermore, it provides an analysis of the trade-offs between model accuracy, computational resources, and real-time performance requirements. The dynamic nature of fraud tactics demands adaptive learning mechanisms that can update models in response to new data, which brings attention to the necessity of continuous learning and model retraining protocols. Techniques such as online learning and active learning are discussed as viable solutions to ensure that models remain effective against emerging fraud patterns. The challenges of data privacy and security are also examined, given the sensitive nature of financial data. AI and ML models, particularly those deployed in real-time environments, must comply with stringent data protection laws such as the General Data Protection Regulation (GDPR) and regional financial regulations. The implications of privacy-preserving machine learning, differential privacy, and federated learning as methods to process data without compromising individual user privacy are evaluated. This aspect is critical for building trust between financial institutions and customers, ensuring that fraud detection efforts do not come at the expense of user data confidentiality. Lastly, the research covers future directions and emerging trends that could shape the landscape of financial fraud detection and prevention. The integration of blockchain technology and distributed ledger systems is considered for enhancing transparency and reducing opportunities for fraudulent activities. Advanced threat intelligence platforms that leverage cross-industry data sharing and the collective insights of AI models trained on diverse datasets are also discussed as potential avenues for mitigating fraud in a proactive manner. The role of collaborative networks and the potential for AI-driven fraud detection to be part of a larger cybersecurity framework are posited as next-generation solutions to create a more secure financial ecosystem. The findings of this research underline the significance of continuous advancements in data science and AI to stay ahead of increasingly sophisticated financial fraud tactics. While AI models have shown promising capabilities in detecting fraudulent activities in real-time, challenges such as model interpretability, scalability, and adaptability remain prominent. This paper concludes with a strategic roadmap for financial institutions, policymakers, and technology developers to enhance the efficacy of fraud prevention strategies, which include fostering innovation in AI-driven solutions, promoting the development of robust real-time processing infrastructures, and encouraging collaborative research efforts that leverage cross-sector knowledge and resources.

Abstract. Measurements of the surface soil moisture (SM) content are important for a wide range of applications. Among them, operational hydrology and numerical weather prediction, for instance, need SM information in near-real-time (NRT), typically not later than 3 h after sensing. The European Space Agency (ESA) Soil Moisture and Ocean Salinity (SMOS) satellite is the first mission specifically designed to measure SM from space. The ESA Level 2 SM retrieval algorithm is based on a detailed geophysical modelling and cannot provide SM in NRT. This paper presents the new ESA SMOS NRT SM product. It uses a neural network (NN) to provide SM in NRT. The NN inputs are SMOS brightness temperatures for horizontal and vertical polarizations and incidence angles from 30 to 45°. In addition, the NN uses surface soil temperature from the European Centre for Medium-Range Weather Forecasts (ECMWF) Integrated Forecast System (IFS). The NN was trained on SMOS Level 2 (L2) SM. The swath of the NRT SM retrieval is somewhat narrower (∼ 915 km) than that of the L2 SM dataset (∼ 1150 km), which implies a slightly lower revisit time. The new SMOS NRT SM product was compared to the SMOS Level 2 SM product. The NRT SM data show a standard deviation of the difference with respect to the L2 data of < 0.05 m3 m−3 in most of the Earth and a Pearson correlation coefficient higher than 0.7 in large regions of the globe. The NRT SM dataset does not show a global bias with respect to the L2 dataset but can show local biases of up to 0.05 m3 m−3 in absolute value. The two SMOS SM products were evaluated against in situ measurements of SM from more than 120 sites of the SCAN (Soil Climate Analysis Network) and the USCRN (US Climate Reference Network) networks in North America. The NRT dataset obtains similar but slightly better results than the L2 data. In summary, the NN SMOS NRT SM product exhibits performances similar to those of the Level 2 SM product but it has the advantage of being available in less than 3.5 h after sensing, complying with NRT requirements. The new product is processed at ECMWF and it is distributed by ESA and via the European Organisation for the Exploitation of Meteorological Satellites (EUMETSAT) multicast service (EUMETCast).

Simulations are often used to model real physical systems prior to electrical, mechanical, and computer hardware development. This allows engineers and scientists to experiment with various concepts before committing time and effort into hardware. Simulations can also be run concurrently with real-time systems to build knowledge of the environment that the real-time system is operating in, then provide feedback to the system to optimize its performance. For both of these types of simulations, the simulation must accurately model the real physical system. A comparison of simulations to real-time controlled physical systems is illustrated in this paper using several simple robotic and artificial Neural Network examples. The robotics examples show how real-time control of mobile robots and robotic-arms, and the resultant governing equations and software algorithms can provide several interesting simulation problems to overcome if the simulation is to accurately model the physical system. The Neural Network example demonstrates how computational speed and numerical precision can become an issue when comparing simulations to real-time Neural Network hardware. In general, comparison of simulations to real physical systems often enhances understanding of the underlying governing principles and equations, and results in simulations that accurately model the real world.

Transient stability analysis of electric energy systems via a fuzzy ART-ARTMAP neural network

Solution of the inverse kinematics (IK) problem with the aid of neural networks is considered. The control system for a multi-link redundant manipulator is synthesized. The configuration of a control system on the basis of servo drives is presented. The proposed control system is based on an algorithm involving a new hybrid method for solving the IK problem. This method combines the ANFIS adaptive neural fuzzy inference network and an iterative refinement algorithm (according to the Newton-Raphson method). Thus, the proposed algorithm combines the advantages of the neural network and iterative approaches, namely, high precision and high response speed. The required coordinates (link rotation angles) for the inverse problem are calculated in the neural network (ANFIS), after which they are refined by the iteration method. Hence, a much fewer number of iterations have to be carried out in the numerical method, and, accordingly, much shorter time is taken to execute the algorithm. The developed algorithm ensures a controlled accuracy of calculations with due regard of its application in real-time control systems. The developed control system is universal in nature and can be used for manipulators having different design parameters. For investigating the control system synthesized proceeding from the developed method for solving the IK problem, a three-link manipulator design was considered. The mathematical description used for constructing the manipulator’s operating space and for training the control system neural network is presented. The results from experimental investigations of applying the hybrid algorithm for calculating the link and actuator coordinates are given. The method was investigated in the Matlab environment. The results of the performed experiments allowed a conclusion to be drawn about the possibility of using the developed method in real-time control systems. To solve the problem of shaping the operating spaces of certain types of manipulators, a graphical interface was developed using the Matlab GUI. The application includes features for adjusting the design parameters (the mechanical structure and accuracy). As an example, the workspaces of both planar and spatial manipulator designs with the specified parameters are taken.

Optimized co-scheduling of mixed-precision neural network accelerator for real-time multitasking applications

Artificial neural networks (ANNs, or simply NNs) are inspired by biological nervous systems and consist of simple processing units (artificial neurons) that are interconnected by weighted connections. Neural networks can be trained to solve problems that are difficult to solve by conventional computer algorithms. The usage of the FPGA (Field Programmable Gate Array) for neural Network implementation provides flexibility in programmable systems. For the neural network based instrument prototype in real time application, conventional specific VLSI neural chip design suffers the limitation in time and cost. With low precision artificial neural network design, FPGAs have higher speed and smaller size for real time application than the VLSI design. Several work show that FPGA are real opportunity for flexible hardware implementation of neural network and yet representation of standard neural network face some problem. The difficulty such as limit of size and architecture of neural network that can mapped on to FPGA. This paper discuss the usage of neural network implementations. Both assets and obstacles are described and various solution are outlined.

Objective: This study developed a real-time system to detect driver"s cognitive load using a multi-layer artificial neural network (MANN) based on electrocardiography (ECG) signals. The real-time system was aimed at classifying driver"s status into either normal or overload.Background: Driving with cognitive load is considered as one of significant factors for traffic accidents. Thus, an early detection of this risky status while driving is needed to prevent vehicle accidents.Method: The ECG signals of this study were measured from 22 participants who performed simulator-based driving experiment under two different conditions (1: normal driving, 2: overload driving (driving while doing a two-back task or an arithmetic task)). A real-time detection system was developed using MANN on the ECG signals and its effectiveness was evaluated for two new participants who drove under the two driving conditions.Results: The MANN model used for the real-time detection system showed perfect accuracy (100%), sensitivity (100%), and specificity (100%) for both of the training and testing data sets. In addition, the proposed real-time detection system successfully detected the change of participant"s status with a reasonable time delay (mean = 4.5 seconds).Conclusion: This study demonstrated that the ECG signals can be used as a biometric measure for the detection of the driver"s cognitive status in real-time.Application: The proposed detection system would be useful for the development of an intelligent vehicle that can provide timely interventions and/or warnings at the early onset of cognitive overload.