Recurrent Neural Networks

As the volume of text data continues to grow rapidly, efficient and accurate text summarization has become crucial for managing and processing this information. Traditional summarization methods, including extractive and abstractive approaches, face challenges in maintaining relevance, fluency, and factual consistency, particularly in complex domains such as legal documents. In this research, we propose an advanced summarization framework leveraging transformer networks, specifically designed to enhance the quality of English text summaries. The framework begins with the preprocessing of text data, including lowercasing, elimination of punctuation, HTML tags, stopwords, and stemming/lemmatization to standardize the input. Text feature extraction is then performed using the Term Frequency-Inverse Document Frequency (TF-IDF) method, converting textual data into numerical representations that highlight the significance of terms within documents. The core of the framework employs a transformer model that eschews traditional RNNs in favor of multiple attention layers and positional encoding to capture global dependencies and enable parallel processing. The model's encoder-decoder structure uses self-attention mechanisms and multi-head attention layers to effectively encode the sequence and context of words, leading to high-quality abstractive summaries. The proposed architecture's effectiveness is evaluated using the BillSum dataset, which includes over 22,000 U.S. Congressional bills and human-written summaries. The results highlight the superior performance of the proposed Transformer model, which achieved the highest ROUGE-1, ROUGE-2, and ROUGE-L scores, with values of 37.03, 16.89, and 28.34, respectively and the Proposed method is implemented using python software. Our approach demonstrates significant improvements in summary accuracy and relevance compared to existing methods, making it a promising solution for automated summarization tasks across various domains.

The development of machine learning (ML) technologies provide a new development direction for cryptanalysis. Several ML research in the field of cryptanalysis was carried out to identify the cryptographic algorithm used, find out the secret key, and even recover the secret message The first objective of this study is to see how much influence optimization and activation function have on the multi-layer perceptron (MLP) model in performing cryptanalysis. The second research objective, which is to compare the performance of cryptanalysis in recovering keys and the plaintext. Several experiments have been carried out, the observed parameters found that the use of the rectified linear unit (ReLU) activation function and the ADAM optimizer improves the performance of deep learning (DL)-based cryptanalysis as evidenced by a significantly smaller error rate. DL-based cryptanalysis works more effectively in recovering keys than recovering plaintext. DL-based cryptanalysis managed to recover the keys with an average loss of 0.007, an average of 49 epochs, and an average time of 0.178 minutes.

Discourse coherence modeling evaluation remains a challenge task in all Natural Language Processing subfields. Most proposed approaches focus on feature engineering, which accepts the sophisticated features to capture the logic, syntactic or semantic relationships between all sentences within a text. This paper investigates the automatic evaluation of text coherence. We introduce a fully-automatic rich statistical model of local and global coherence that uses word2vec approach to assess the coherence a document. Our modeling approach relies on numerical vectors derived from word2vec algorithm applied on a very large collection of texts. We successfully combined the word2vec vectors and most likely n-grams with cohesive LD-ngrams perplexity to assess the coherence and topic integrity of document. We present experimental results that assess the predictive power that it does not depend on the language and its semantic concepts. So it has the ability to apply on any language. Our model achieves state-of-the-art performance in coherence evaluation and order discrimination task on two datasets widely used in the previous methods.

This paper aims to investigate whether applying the Singular Value Decomposition (SVD) technique can reduce the workload of Convolutional Neural Networks (CNNs) without compromising image classification results. Usual methods for reducing file size are generally unsuitable for deployment because they are costly to train, not straightforward to use, or require additional resources. We performed a low-rank approximation of the CNN weights using the singular value decomposition (SVD) approach. Results for ResNet-50 were obtained from CIFAR-10, and EfficientNet-B0 was assessed using ImageNet. The assessment used metrics including accuracy, precision, recall, and F1-score. Using SVD at a 0.5 compression level, we reduced the network size of ResNet-50 on CIFAR-10 by 41.7% without significantly compromising accuracy (95.4%). With aggressive compression at a ratio of 0.1, the model achieved an 88.9% reduction in parameters, but accuracy decreased to 72.9%. Furthermore, the model increased inference speed slightly (up to 1.0x) and maintained an approximately 89.96 MB size. The uncompressed model was 10.4% accurate before training. Rather than other techniques, SVD offers straightforward yet practical suggestions for enhancing CNN performance on small datasets, without significantly compromising accuracy. These results demonstrate that SVD-based compression is a promising approach for utilizing CNNs in limitedresource systems. It achieves nearly complete accuracy with fewer parameters and has a minimal impact on the speed of CNNs.

In recent years, there has been a significant increase in investment in renewable energy sources, leading to the decarbonization of the electricity sector. Accordingly, a key concern is the influence of this process on future electricity market prices, which are expected to decrease with the increasing generation of renewable power. This is important for both current and future investors, as it can affect profitability. To address these concerns, a long-term analysis is proposed here to examine the influence of the future electricity mix on Iberian electricity prices in 2030. In this study, we employed artificial intelligence forecasting models that incorporated the main electricity price-driven components of MIBEL, providing accurate predictions for the real operation of the market. These can be extrapolated into the future to predict electricity prices in a scenario with high renewable power penetration. The results, obtained considering a framework featuring an increase in the penetration of renewables into MIBEL of up to 80% in 2030, showed that electricity prices are expected to decrease by around 50% in 2030 when compared to 2019, and there will be a new pattern of electricity prices throughout the year due to the uneven distribution of renewable electricity. The study's findings are relevant for ongoing research on the unique challenges of energy markets with high levels of renewable generation.

Recurrent neural network (RNN) has been widely used as a tool in the data classification. This network can be educated with gradient descent back propagation. However, traditional training algorithms have some drawbacks such as slow speed of convergence being not definite to find the global minimum of the error function since gradient descent may get stuck in local minima. As a solution, nature inspired metaheuristic algorithms provide derivative-free solution to optimize complex problems. This paper proposes a new metaheuristic search algorithm called Cuckoo Search (CS) based on Cuckoo bird’s behavior to train Elman recurrent network (ERN) and back propagation Elman recurrent network (BPERN) in achieving fast convergence rate and to avoid local minima problem. The proposed CSERN and CSBPERN algorithms are compared with artificial bee colony using BP algorithm and other hybrid variants algorithms. Specifically, some selected benchmark classification problems are used. The simulation...

Identification of nonlinear dynamic systems is a critical task in various fields. Artificial neural networks have been widely used for this purpose due to their ability to approximate complex functions. However, their computational efficiency and stability often pose challenges, especially in real-time applications. Quantum computation has shown potential for enhancing computational performance, but its integration with neural networks is still under investigation. The primary motivation addressed in this paper is the development of an effective strategy for synthesizing and applying recurrent quantum neural networks based on Lyapunov stability criteria (RQNN-LS) for nonlinear system identification. This model enhances the computational efficiency of recurrent neural networks by incorporating quantum computation into the neural network characteristics by using qubit neurons for data processing. Additionally, adaptive learning rates are derived based on Lyapunov stability theory for online tuning of the parameters to guarantee the stability of the proposed technique. The applicability and superiority of the presented RQNN-LS identifier are verified through the simulation and practical results of nonlinear system identification, comparing its performance with other existing identification techniques. The comparative results demonstrated significant improvements in computational efficiency with the proposed technique and highlighted the merits and superiority of the developed model over other methodologies.

Predicting traffic volume in real-time can improve both traffic flow and road safety. A precise traffic volume forecast helps alert drivers to the flow of traffic along their preferred routes, preventing potential deadlock situations. Existing parametric models cannot reliably forecast traffic volume in dynamic and complex traffic conditions. Therefore, in order to evaluate and forecast the traffic volume for every given time step in a real-time manner, we develop non-linear memory-based deep neural network models. Our extensive experiments run on the Metro Interstate Traffic Volume dataset demonstrate the effectiveness of the proposed models in predicting traffic volume in highly dynamic and heterogeneous traffic environments.

As deep learning models continue to grow in complexity, the computational and energy demands associated with their training and deployment are becoming increasingly significant, particularly for convolutional neural networks (CNNs) deployed on CPU-bound and resource-limited devices. Fully connected (FC) layers, while vital, are energy-intensive, accounting for 85.7% of a network's parameters but contributing only 1% of the computations. This research proposes a novel approach to optimising these layers for greater energy efficiency by integrating low-rank factorisation with differential partial differential equations (PDEs). The introduction of the LowRankDense layer, which combines low-rank matrix factorisation with a differential PDE solver, aims to reduce both the parameter count and energy consumption of FC layers. Experiments conducted on the MNIST, Fashion MNIST, and CIFAR-10 datasets demonstrate the effectiveness of this approach, yielding promising results in terms of reduced energy usage and maintaining comparable performance, thereby enhancing the practicality and sustainability of CNNs for widespread use in environments with limited computational resources.

HAL is a multi-disciplinary open access archive for the deposit and dissemination of scientific research documents, whether they are published or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers. L'archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d'enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.

Hyperspectral image unmixing is an inverse problem aiming at recovering the spectral signatures of pure materials of interest (called endmembers) and estimating their proportions (called abundances) in every pixel of the image. However, in spite of a tremendous applicative potential and the avent of new satellite sensors with high temporal resolution, multitemporal hyperspectral unmixing is still a relatively underexplored research avenue in the community, compared to standard image unmixing. In this paper, we propose a new framework for multitemporal unmixing and endmember extraction based on a state-space model, and present a proof of concept on simulated data to show how this representation can be used to inform multitemporal unmixing with external prior knowledge, or on the contrary to learn the dynamics of the quantities involved from data using neural network architectures adapted to the identification of dynamical systems.

Recent hand pose estimation methods require large numbers of annotated training data to extract the dynamic information from a hand representation. Nevertheless, precise and dense annotation on the real data is difficult to come by and the amount of information passed to the training algorithm is significantly higher. This paper presents an approach to developing a hand pose estimation system which can accurately regress a 3D pose in an unsupervised manner. The whole process is performed in three stages. Firstly, the hand is modelled by a novel latent tree dependency model (LTDM) which transforms internal joints location to an explicit representation. Secondly, we perform predictive coding of image sequences of hand poses in order to capture latent features underlying a given image without supervision. A mapping is then performed between an image depth and a generated representation. Thirdly, the hand joints are regressed using convolutional neural networks to finally estimate the l...

This paper introduces AliceSkyGardenT3, a novel energy-efficient language model framework based on ternary parameters {-1, 0, 1}. The architecture implements a 1.58-bit quantization scheme for linear layers combined with Straight-Through Estimator (STE) gradients, achieving 85% weight sparsity while maintaining 97% of full-precision model accuracy. Our framework integrates compression and decompression functions that reduce model size by 6.3× compared to FP32 baselines, requiring only 2.1GB storage for a 7B parameter model. The ternary activation architecture enables energyefficient inference with theoretical 3.8× speedup on specialized hardware. Experiments demonstrate that the model preserves linguistic capabilities while reducing energy consumption by 42% during inference. The integrated compression API allows model deployment with just two commands: model.compress_model_weights() for storage and model.load_compressed_model() for inference. This work establishes a new paradigm for green AI that balances computational efficiency with environmental sustainability without compromising model performance.

Human activity recognition (HAR) is a rapidly growing field that utilizes smart devices, sensors, and algorithms to automatically classify and identify the actions of individuals within a given environment. These systems have a wide range of applications, including assisting with caring tasks, increasing security, and improving energy efficiency. However, there are several challenges that must be addressed in order to effectively utilize HAR systems in multi-resident environments. One of the key challenges is accurately associating sensor observations with the identities of the individuals involved, which can be particularly difficult when residents are engaging in complex and collaborative activities. This paper provides a brief overview of the design and implementation of HAR systems, including a summary of the various data collection devices and approaches used for human activity identification. It also reviews previous research on the use of these systems in multi-resident environments and offers conclusions on the current state of the art in the field.

Artificial intelligence (AI) and machine learning (ML) tools play a significant role in the recent evolution of smart systems. AI solutions are pushing towards a significant shift in many fields such as healthcare, autonomous airplanes and vehicles, security, marketing customer profiling and other diverse areas. One of the main challenges hindering the AI potential is the demand for high-performance computation resources. Recently, hardware accelerators are developed in order to provide the needed computational power for the AI and ML tools. In the literature, hardware accelerators are built using FPGAs, GPUs and ASICs to accelerate computationally intensive tasks. These accelerators provide high-performance hardware while preserving the required accuracy. In this work, we present a systematic literature review that focuses on exploring the available hardware accelerators for the AI and ML tools. More than 169 different research papers published between the years 2009 and 2019 are studied and analysed. Hardware accelerators • Artificial intelligence • Machine learning • AI on hardware • Real-time AI * Qassim Nasir

Deep neural networks proved to be a very useful and powerful tool with many practical applications. They especially excel at learning from large data sets with labeled samples. However, in order to achieve good learning results, the network architecture has to be carefully designed. Creating an optimal topology requires a lot of experience and knowledge. Unfortunately there are no practically applicable algorithms which could help in this situation. Using an evolutionary process to develop new network topologies might solve this problem. The limiting factor in this case is the speed of evaluation of a single specimen (a single network architecture), which includes learning based on the whole large dataset. In this paper we propose to overcome this problem by using a fitness prediction technique: use subsets of the original training set to conduct the training process and use its results as an approximation of specimen's fitness. We discuss the feasibility of this approach in context of the desired fitness predictor features and analyze whether subsets obtained in an evolutionary process can be used to estimate the fitness of the network topology. Finally we draw conclusions from our experiments and outline plans for future work.

Identifying "influential spreader" is finding a subset of individuals in the social network, so that when information injected into this subset, it is spread most broadly to the rest of the network individuals. The determination of the information influence degree of individual plays an important role in online social networking. Once there is a list of individuals who have high influence, the marketers can access these individuals and seek them to impress, bribe or somehow make them spread up the good information for their business as well as their product in marketing campaign. In this paper, according to the idea "Information can be spread between two unconnected users in the network as long as they both check-in at the same location", we proposed an algorithm called SMPRank (Social Meta Path Rank) to identify individuals with the largest influence in complex online social networks. The experimental results show that SMPRank performs better than Weighted LeaderRank because of the ability to determinate more influential spreaders.

This study proposes a novel framework to explain a key driver of technological change: the invasive behavior of general-purpose technologies (GPTs). These technologies disrupt established scientific and industrial systems by rapidly introducing innovations that render existing solutions obsolete. We explore how this invasive dynamic-characterized by accelerated innovation and radical displacement-drives transformative shifts across technological, economic, and social domains. To empirically validate this framework, we analyze the evolution of neural network architectures, focusing on the rise of transformer models compared to earlier technologies such as Long Short-Term Memory (LSTM) and Recurrent Neural Networks (RNNs). Patent data from 2020 to 2024 reveal that transformer technologies experienced an exponential growth rate of 45.91%, more than double that of LSTM (21.17%) and RNN (18.15%). We also introduce a new metric-the invasive rate-to quantify a technology's penetration into its domain. Transformers scored 2.2%, significantly higher than LSTM (1.39%) and RNN (1.22%). These findings underscore the disruptive and dominant nature of invasive technologies, particularly in the field of Artificial Intelligence. The study offers new insights into the mechanisms of technological evolution and the forces shaping global innovation for economic and social change.

This paper presents a User Intent (UI) mining scheme based on an emerging neural machine intelligence technique called the Neuronal Auditory Machine Intelligence (NeuroAMI) and considering a Wi-Fi sessions dataset containing about 8000 data points for intent prediction. The results of simulations considering graded increase in samples from 50samples to 999samples showed that substantial increases in accuracies are achievable. When compared to the Long Short-Term Memory (LSTM) method, the results showed that the NeuroAMI will outperform the LSTM by a factor of about 2 considering percentage accuracies.

Technique for selective localization of four-photon mixing and its use for physical parameter extraction in fiber is described. The measurement method relies on localized countercolliding power transfer and was applied to map dispersion and corresponding transverse geometry of different types of optical fibers. The technique is capable of resolving fiber core fluctuations comparable to a single silica molecular diameter.

Effective and efficient mitigation of malware is a long-time endeavor in the information security community. The development of an anti-malware system that can counteract an unknown malware is a prolific activity that may benefit several sectors. We envision an intelligent anti-malware system that utilizes the power of deep learning (DL) models. Using such models would enable the detection of newly-released malware through mathematical generalization. That is, finding the relationship between a given malware $x$ and its corresponding malware family $y$, $f: x \mapsto y$. To accomplish this feat, we used the Malimg dataset (Nataraj et al., 2011) which consists of malware images that were processed from malware binaries, and then we trained the following DL models 1 to classify each malware family: CNN-SVM (Tang, 2013), GRU-SVM (Agarap, 2017), and MLP-SVM. Empirical evidence has shown that the GRU-SVM stands out among the DL models with a predictive accuracy of ~84.92%. This stands to...

Accurate weather forecasting plays a vital role in sectors such as agriculture, transportation, energy, and disaster management. With the increasing availability of historical meteorological data, time series forecasting has emerged as a powerful approach to predicting weather conditions. This study explores the application of various time series models including statistical, machine learning, and deep learning techniques to forecast key weather parameters such as temperature, humidity, and precipitation. The model is trained using Adam optimizer with 0.001 learning rate and calculate the loss using Mean Squared Error. Historical weather data is pre-processed to address issues like missing values, seasonality, and noise, and models are evaluated based on metrics such as Mean Squared Error. Training Loss and Validation Loss is calculated for various epochs. The prediction is done by using validation set. Results demonstrate that while statistical models perform adequately for short-term forecasts, deep learning methods capture complex temporal patterns more effectively.

Combining convolutional neural networks (CNNs) and recurrent neural networks (RNNs) produces a powerful architecture for video classification problems as spatial–temporal information can be processed simultaneously and effectively. Using transfer learning, this paper presents a comparative study to investigate how temporal information can be utilized to improve the performance of video classification when CNNs and RNNs are combined in various architectures. To enhance the performance of the identified architecture for effective combination of CNN and RNN, a novel action template-based keyframe extraction method is proposed by identifying the informative region of each frame and selecting keyframes based on the similarity between those regions. Extensive experiments on KTH and UCF-101 datasets with ConvLSTM-based video classifiers have been conducted. Experimental results are evaluated using one-way analysis of variance, which reveals the effectiveness of the proposed keyframe extrac...

Background: Cloud computing (CC), artificial intelligence (AI), and the Internet of Things (IoT) are combining to revolutionise healthcare by making it possible to process data in real time and diagnose diseases. For increased accuracy and efficiency, optimised solutions are required because traditional approaches are unable to handle the volume and complexity of data from IoT-enabled devices. Objective: The objective of this research is to develop and apply a hybrid model that combines Radial Basis Function Networks (RBFN), Genetic Algorithms (GA), and Particle Swarm Optimisation (PSO) to improve accuracy, shorten processing times, and improve real-time healthcare disease diagnosis for chronic sickness identification. Methods: Optimisation of Recurrent Neural Networks (RNNs) and RBFNs for processing realtime data gathered from IoT devices was done using a hybrid PSO-GA framework. Preprocessing of the data occurred with cloud computing for extensive analysis, and optimisation were done to improve sensitivity and specificity metrics for various medical applications. Results: The suggested PSO-GA RNN-RBFN model performed exceptionally well, with 94% specificity, 92% sensitivity, and 93% accuracy. It reduced processing time to 54 seconds and outperformed conventional techniques like CNN, DeepDR, and EKF-SVM, making it ideal for real-time healthcare monitoring. Conclusion: A reliable method for real-time illness detection is provided by the combination of PSO, GA, and RBFN in a cloud-based architecture. As compared to conventional methods, the hybrid model dramatically increases accuracy, sensitivity, and specificity, proving its usefulness in managing intricate, urgent healthcare data.

This  study  investigates  the  use  of  Recurrent  Neural  Networks  (RNNs)  and  Long  Short-Term  Memory  (LSTM)  networks  to  predict  crime  patterns  in  Bucaramanga,  Colombia.  A temporal approach is presented, which starts by splitting the city into 17 communes. Using a dataset of robbery incidents from 2016 to January 2023, we developed individual time  series  models  for  each  commune.  Then,  we  used  the  Root  Mean  Squared  Error  (RMSE) as the evaluation metric in these regression tasks. The LSTM models consistently outperformed both the RNN and ARIMA models, a classical methodology for time series prediction,  achieving  lower  RMSE  scores.  The  LSTM  model  yielded  an  average  RMSE  of  2.875  (with  a  standard  deviation  of  1.657),  which  is  considerably  lower  than  that  obtained by the RNN model 3.101 (1.82) and the ARIMA model 3.428 (2.57). These results show that LSTM better captures the complex temporal dependencies in the data. Future work should explore hybrid models and the incorporation of additional data sources to enhance predictive accuracy further.

The predictive model is used to calculate artificial neural networks algorithm factor of relationships for dengue hemorrhagic fever outbreaks in Northeast of Thailand can use analysis factors for dengue outbreaks by using data in past from 2007-2016 (total of ten years). Furthermore, the factor is complex and contains fourteen variables ( temperature or weather, rainfall and humidity, mosquito growth periods, areas and residence and population) for dengue hemorrhagic fever. Data analysis objectives were ( 1) to factor analyzed climate change for dengue hemorrhagic fever outbreaks in Northeast of Thailand and (2) to create a model optimization for prediction by neural network statistic which is called Artificial Neural Networks. The evaluation model results is prediction of error values mean square error (MSE). This model is accurate to more than 80%.

Recurrent neural network (RNN) has been widely used as a tool in the data classification. This network can be educated with gradient descent back propagation. However, traditional training algorithms have some drawbacks such as slow speed of convergence being not definite to find the global minimum of the error function since gradient descent may get stuck in local minima. As a solution, nature inspired metaheuristic algorithms provide derivative-free solution to optimize complex problems. This paper proposes a new metaheuristic search algorithm called Cuckoo Search (CS) based on Cuckoo bird’s behavior to train Elman recurrent network (ERN) and back propagation Elman recurrent network (BPERN) in achieving fast convergence rate and to avoid local minima problem. The proposed CSERN and CSBPERN algorithms are compared with artificial bee colony using BP algorithm and other hybrid variants algorithms. Specifically, some selected benchmark classification problems are used. The simulation...

This research focuses on predicting stock prices using Gated Recurrent Units (GRUs), a type of Recurrent Neural Network (RNN) that effectively captures sequential dependencies in time series data. The model leverages historical stock data and presents results using interactive dashboards in Power BI, enhancing decision-making and interpretability for stakeholders. The study evaluates the accuracy and efficiency of GRU-based models against traditional approaches, demonstrating improved forecasting capabilities. Additionally, the research explores the impact of various hyper-parameters on model performance and compares GRU with other deep learning architectures, such as LSTMs, to determine the optimal approach for financial time series forecasting. The system's ability to detect trends, mitigate noise, and provide real-time insights makes it a valuable tool for investors and financial analysts. Furthermore, this study examines real-world applications, industry adoption, and the scalability of GRU-based predictive models in financial markets, ensuring robust performance across varying market conditions.

The Stock market index in India play a crucial role in the country's economic growth and development. The Investors, policymakers, and market participants closely monitor the performance of stock market indices. In general, people tend to be unaware of Indian indices, their weights, and their composition patterns. This research paper aims to analyze the performance of stock market indices and focuses on important indexes from the Bombay Stock Exchange and the National Stock Exchange while taking into account their development, calculation method, and historical value. This Study is mainly based on three Objectives and data selected and synthesizes for the year 2001 to 2022, due to base year limitations of various indices. By utilizing quantitative analysis methods and examining relevant economic indicators, this paper seeks to provide valuable insights into the Indian stock market Indices and its role as a barometer of economic health. This article makes it simple for people to understand the number of companies that make up the important indices, their weight in the indexes, their performance on a sectoral basis, and other methodologies. Finally, this paper attempts to provide a brief overview of various index which is traded in Indian Stock market.

E-commerce and financial transaction platforms are increasingly vulnerable to cyber threats and fraudulent activities due to the rapid digitization of global markets. Anomaly detection plays a vital role in identifying unusual behavior indicative of fraud, security breaches, or financial manipulation. Traditional methods such as rule-based systems and statistical models often fall short in adapting to evolving patterns of fraud. Deep learning, with its ability to extract complex features and learn non-linear relationships from massive datasets, offers a transformative approach to anomaly detection. This paper explores the use of deep learning techniques—such as Autoencoders, Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Generative Adversarial Networks (GANs)—to enhance fraud prevention and cybersecurity in the e-commerce and financial sectors. The paper highlights the comparative effectiveness of different models, challenges such as data imbalance and explainability, and future prospects for integrated intelligent systems.

Nous proposons dans cet article un système de reconnaissance de mots manuscrit arabes basé sur les modèles perceptifs d'activation interactive et de vérification, définis par des psychologues. Le système proposé est basé sur un Réseau de Neurones faisant partie des systèmes Transparent (RNT). Il procède par une vision globale des caractéristiques structurelles apparentes durant la première phase de propagation et une vision locale par des Descripteurs de Fourier normalisés en phase de rétro-propagation (DF). L'avantage du système appelé RNT-DF, est qu'il ne nécessite pas d'étape d'apprentissage. Il se réfère à des descriptions standard de l'écriture imprimée aussi bien au niveau global (mot) qu'au niveau de la normalisation locale (lettres). Une évaluation expérimentale des deux visions est effectuée sur une base de données des montants littéraux de chèques et de noms de villes tunisiennes.

Automatic Text Summarization (ATS) compacts source content into a concise format while preserving core information. While extensively studied for resource-rich languages, ATS remains challenging for low-resource languages like Kannada due to limited corpora and NLP tools. This work introduces an extractive, topic-driven method for summarizing Kannada news articles from multiple documents. We developed a custom dataset of 100 Kannada news story sets (3 articles per set) to address the lack of standardized benchmarks. The proposed approach leverages Latent Dirichlet Allocation (LDA) to identify latent themes across documents, followed by sentence selection using vector-space modeling. Sentences are scored based on their relevance to identified topics (via cosine similarity) and prioritized to maximize informational value while minimizing redundancy through Maximum Marginal Relevance (MMR). Evaluations using ROUGE metrics demonstrate that the LDA-based method outperforms existing summarization algorithms, producing summaries closer to human-generated references. The system achieves higher F-scores (e.g., 0.68 at 40% compression) compared to baseline models like TextRank and approaches for other Indian languages, validating its efficacy for low-resource linguistic contexts.

We investigate how different forms of plasticity shape the dynamics and computational properties of simple recurrent spiking neural networks. In particular, we study the effect of combining two forms of neuronal plasticity: spike timing dependent plasticity (STDP), which changes the synaptic strength, and intrinsic plasticity (IP), which changes the excitability of individual neurons to maintain homeostasis of their activity. We find that the interaction of these forms of plasticity gives rise to interesting network dynamics characterized by a comparatively large number of stable limit cycles. We study the response of such networks to external input and find that they exhibit a fading memory of recent inputs. We then demonstrate that the combination of STDP and IP shapes the network structure and dynamics in ways that allow the discovery of patterns in input time series and lead to good performance in time series prediction. Our results underscore the importance of studying the interaction of different forms of plasticity on network behavior.

Recognizing human actions through video analysis has gained significant attention in applications like surveillance, sports analytics, and human-computer interaction. While deep learning models such as 3D convolutional neural networks (CNNs) and recurrent neural networks (RNNs) deliver promising results, they often struggle with computational inefficiencies and inadequate spatial-temporal feature extraction, hindering scalability to larger datasets or high-resolution videos. To address these limitations, we propose a novel model combining a two-dimensional convolutional restricted Boltzmann machine (2D Conv-RBM) with a long short-term memory (LSTM) network. The 2D Conv-RBM efficiently extracts spatial features such as edges, textures, and motion patterns while preserving spatial relationships and reducing parameters via weight sharing. These features are subsequently processed by the LSTM to capture temporal dependencies across frames, enabling effective recognition of both short-and long-term action patterns. Additionally, a smart frame selection mechanism minimizes frame redundancy, significantly lowering computational costs without compromising accuracy. Evaluation on the KTH, UCF Sports, and HMDB51 datasets demonstrated superior performance, achieving accuracies of 97.3%, 94.8%, and 81.5%, respectively. Compared to traditional approaches like 2D RBM and 3D CNN, our method offers notable improvements in both accuracy and computational efficiency, presenting a scalable solution for real-time applications in surveillance, video security, and sports analytics.

This paper studies the use of discrete-time recurrent neural networks for predicting the next symbol in a sequence. The focus is on online prediction, a task much harder than the classical offline grammatical inference with neural networks. The results obtained show that the performance of recurrent networks working online is acceptable when sequences come from finite-state machines or even from some chaotic sources. When predicting texts in human language, however, dynamics seem to be too complex to be correctly learned in real-time by the net. Two algorithms are considered for network training: real-time recurrent learning and the decoupled extended Kalman filter.

Self-organization theory is an interdisciplinary scientific direction to improve the quality of acquired knowledge and skills. The article aims to study the development of interdisciplinary research activities among students based on self-organization theory. The study was conducted using an integrated approach, methods of analysis and synthesis, modeling, pedagogical experiments, and mathematical statistics. An elective course, "synergetic-interdisciplinary scientific theory", and a model for preparing students for interdisciplinary research activities were developed and implemented. The knowledge and skills of 272 physics students in the Republic of Kazakhstan universities in control and experimental groups improved after the training. The knowledge highest level is observed among the experimental group students while studying the natural experiments' theoretical aspects necessary for researching the natural objects' sustainability. Experimental group students demonstrated better results in completing the task based on the acquired knowledge and skills. They rated the education quality significantly better on a short educational assessment scale. The results-developed course and the students' preparing models for interdisciplinary research can be used in advanced courses for different specialties educators. Research perspectives could potentially contribute to the development of integrated subjects and interdisciplinary programs in natural sciences and humanities.

This project is dedicated to my family. Without their encouragement, understanding, support, and unconditional love, completion of this study could not have been possible. appreciation also goes to my moderator and chairman for their invaluable efforts to perfect this work. Further, I am grateful to my colleagues in The University of Nairobi, whose assistance to this research project cannot be overlooked, for their inspirations, encouragements, guidance and helpful recommendations concerning the procedures through academic discussions and always raising the benchmark.

In the era of modern technologies, distributed web systems play an important role in global IT infrastructures, providing scalability and resilience. However, verification, validation, and testing of these increasingly complex systems remain a major challenge for both specialists working in the field and researchers who want to explore this field. Due to their heterogeneous nature and dynamic interdependencies between existing system components, this important stage of system testing and validation has high complexity. In this context, the paper explores automated testing based on machine learning techniques for anomaly detection in distributed web systems. The methodology used uses advanced machine learning algorithms such as Convolutional Neural Networks, Recurrent Neural Networks and Long-Short Term Memory, K-means Clustering, and Random Forest to analyze data flows and interactions between microservices in real time. It aims to develop a robust framework that includes modules for data collection and preprocessing, model training and validation, as well as an automatic alert system for detected anomalies. Experimental results demonstrate that the proposed approach outperforms traditional testing methods by increasing accuracy and reducing fault identification time.

This research paper presents an innovative methodology for the identification and detection of objects in autonomous driving systems that employ field-programmable gate arrays (FPGAs). Through the integration of deep learning methodologies with FPGA hardware acceleration, the approach successfully attains the minimal latency and optimal precision necessary for secure navigation. By conducting data acquisition, preprocessing, and model training, this can refine the system's performance. By employing parallel computing and hardware optimisation techniques, the FPGA implementation achieves these objectives. Based on experimental data, the FPGA-based approach outperforms conventional CPU and GPU implementations in terms of power efficiency, inference latency, and detection precision. The widespread adoption of field-programmable gate arrays (FPGAs) for enhanced object recognition and identification in autonomous vehicles is imminent due to their exceptional compatibility with autonomous driving systems.

Recurrent neural networks functioning as associative memories are often studied and optimized for recall quality and capacity, with the focus primarily on the network's stability, i.e., convergence to stored attractors. However, the ability of networks to switch between attractors in a controlled way is also potentially a useful phenomenon. Networks that are stable under most conditions, but can be switched by specific stimuli may be used to model cognitive control and other timevarying cognitive phenomena. Such networks, which we term stable-yet-switchable (SyS) networks, are also of interest from the networks perspective, and the SyS properties of scale-free networks have been noted by researchers. In this paper, we consider networks with bimodal connectivity -a core of densely connected neurons and a larger periphery with sparser connectivity -and compare their SyS performance with random and scale-free recurrent neural networks. The results show that core-periphery networks have much better SyS performance than scale-free networks.

Current approaches for (multi-horizon) time-series forecasting using recurrent neural networks (RNNs) focus on issuing point estimates, which are insufficient for informing decision-making in critical application domains wherein uncertainty estimates are also required. Existing methods for uncertainty quantification in RNNbased time-series forecasts are limited as they may require significant alterations to the underlying architecture, may be computationally complex, may be difficult to calibrate, may incur high sample complexity, and may not provide theoretical validity guarantees for the issued uncertainty intervals. In this work, we extend the inductive conformal prediction framework to the time-series forecasting setup, and propose a lightweight uncertainty estimation procedure to address the above limitations. With minimal exchangeability assumptions, our approach provides uncertainty intervals with theoretical guarantees on frequentist coverage for any multi-horizon forecast p...

A nonlinear dynamic handling model for a tractor-semitrailer combination vehicle is presented in this report. The equations of motion are derived from the fundamental equations of dynamics in Euler's formulation without approximations. The model is modular in the sense that it is easy to change axle configuration, tyre model, suspension model, or to add new features. The primary aim of the model is simulations of handling scenarios with active yaw control, using unilateral braking and possibly tractor rear wheel steering. Other applications of the model may include real-time hardware-in-the-loop simulations of tractor-semitrailer handling scenarios. The model is formulated as a state-space model that may be implemented in standard simulation environments. A Simulink implementation is presented. Simulation results are compared with experiments to validate the model.

Cloud services are among the technologies that are developing the fastest. Additionally, it is acknowledged that load balancing poses a major obstacle to reaching energy efficiency. Distributing the load among several resources in order to provide the best possible services is the main purpose of load balancing. The network's accessibility and dependability are increased through the usage of fault tolerance. An approach for hybrid deep learning (DL)-based load balancing is proposed in this paper. Tasks are first distributed in a round-robin fashion to every virtual machine (VM). When assessing whether a VM is overloaded or underloaded, the deep embedding cluster (DEC) also considers the central processing unit (CPU), bandwidth, memory, processing elements, and frequency scaling factors. For cloud load balancing, the tasks completed on the overloaded VM are assigned to the underloaded VM based on their value. To balance the load depending on many aspects like supply, demand, capacity, load, resource utilization, and fault tolerance, the deep Q recurrent neural network (DQRNN) is also suggested. Additionally, load, capacity, resource consumption, and success rate are used to evaluate the efficacy of this approach; optimum values of 0.147, 0.726, 0.527, and 0.895 are attained.

Research on image validation models is an interesting topic. The application of deep learning (DL) for object detection has been demonstrated to effectively and efficiently address the challenges in this field. Deep neural networks (DNN) are deep learning algorithms capable of handling large datasets and effectively solving complex problems due to their robust learning capacity. Despite their ability to address complex problems, DNN encounter challenges related to the necessity for intricate architectures and a large number of hidden layers. The objective of this research is to identify the most effective model for achieving optimal performance in image validation. This study investigates target image validation using DNN algorithms, examining architectures with 3, 4, 5, and 6 hidden layers. This study also evaluates the performance of image validation across various activation functions, batch sizes, and numbers of neurons. The results of the study show that the best performance for image validation is achieved using the Leaky-ReLU and Sigmoid activation functions, with a batch size of 64, and an architecture consisting of 3 hidden layers with neuron sizes of 256, 128, and 64. This model is capable of providing real-time target image validation with an accuracy of up to 94.31%.

This paper contributes the first published evaluation of the quality of automatic translation between Khmer (the official language of Cambodia) and twenty other languages, in both directions. The experiments were carried out using three different statistical machine translation approaches: phrase-based, hierarchical phrase-based, and the operation sequence model (OSM). In addition two different segmentation schemes for Khmer were studied, these were syllable segmentation and supervised word segmentation. The results show that the highest quality machine translation was attained with word segmentation in all of the experiments. Furthermore, with the exception of very distant language pairs the OSM approach gave the highest quality translations when measured in terms of both the BLEU and RIBES scores. For distant languages, our results showed a hierarchical phrase-based approach to be the most effective. An analysis of the experimental results indicated that Kendall’s tau may be direc...

Falls have become a relevant public health issue due to their high prevalence and negative effects in elderly people. Wearable fall detector devices allow the implementation of continuous and ubiquitous monitoring systems. The effectiveness for analyzing temporal signals with low energy consumption is one of the most relevant characteristics of these devices. Recurrent neural networks (RNNs) have demonstrated a great accuracy in some problems that require analyzing sequential inputs. However, getting appropriate response times in low power microcontrollers remains a difficult task due to their limited hardware resources. This work shows a feasibility study about using RNN-based deep learning models to detect both falls and falls’ risks in real time using accelerometer signals. The effectiveness of four different architectures was analyzed using the SisFall dataset at different frequencies. The resulting models were integrated into two different embedded systems to analyze the execut...

Communicable diseases pose significant threats at local, regional, and global levels, often leading to epidemics or pandemics. An epidemic refers to a sudden increase in the number of cases of an infectious disease above what is normally expected in a given population. Examples include cholera, measles, malaria, and dengue fever. Pandemics, however, can result in widespread illness, significant loss of life, and severe social and economic consequences. Concerns about potential pandemic diseases, such as new strains of influenza and severe acute respiratory syndrome (SARS) remain critical. This study presents a deep learning framework based on Geographic Information Systems (GIS) to predict the spatial distribution of epidemic infectious diseases. The framework combines the strengths of deep learning and GIS techniques, both of which offer exceptional capabilities in the field of epidemiology. The study outlines the key steps involved in developing the proposed framework and explains its operational functionality. The proposed framework aims to enhance decision-making efficiency, assist governmental authorities in generating sustainable strategies, and establish appropriate protocols to control epidemics, particularly in highrisk areas. By predicting vulnerable areas, the framework helps mitigate the risks associated with outbreaks and protects social and economic stability. Choosing an appropriate framework requires consideration of several key factors, including those relevant to the spread of diseases or epidemics, accuracy, flexibility, and validation.