Predictions arising from deep neural networks may be very accurate but not very robust, leading to uncertainty in their outcome. This critical problem is receiving growing attention from the Machine Learning (ML) community. A practical solution that is increasingly applied is calculating confidence bounds for the ML predictions. Most confidence bounds available in the literature are theoretically sound but unfeasible from a practical viewpoint. In this paper, we contribute to the literature with probabilistic confidence bounds based on conditional probabilities, and we demonstrate their operational validity by means of a real-world application that concerns the prediction of the sleeping states of car drivers.

Background: Active Learning with AI-tutoring in Higher Education tackles dropout rates. Objectives: To investigate teaching-learning methodologies preferred by students. ChatGPT-based gamified learning methodology is compared to another active learning methodology and a traditional methodology. Study with Learning Analytics to evaluate alternatives, their implementation, and help students elect the best strategies according to their preferences. Methods: Comparative study of three learning methodologies in a Single-Group counterbalanced with 45 university students. It follows a pretest/post-test approach using AHP and SAM. HRV and GSR used for emotional state estimation. Findings: Criteria related to in-class experiences valued higher than test-related criteria. Chat-GPT integration was well regarded compared to well-established methodologies. Student emotion self-assessment correlated with physiological measures, validating used Learning Analitycs. Conclusions: Proposed model AI-Tutoring classroom integration functions effectively at increasing engagement and avoiding false information. AHP with the physiological measuring allows students to determine preferred learning methodologies, avoiding biases, and acknowledging minority groups.

In the environmental security monitoring of smart cities, the infrared and visible image fusion method deployed on intelligent systems based on cloud and fog computing plays an vital role in providing enhanced images for target detection systems. However, the fusion quality can be significantly influenced by the illumination of the monitoring scenario in visible images. Therefore, conventional methods typically suffer a severe performance drop under the condition of insufficient illumination. To tackle this issue, we propose an illumination enhancement and target-aware fusion method (IETAFusion) based on artificial intelligence, which breaks the boundaries between the task of illumination enhancement and image fusion and provide a fusion result with better visual perception in nighttime scene. Specifically, we use a light-weight contrast enhancement module (CEM) restore the brightness of the visble image. Moreover, a Swin Transformer-based backbone network (STBNet) is utilized to facilitate information exchangement between the source images and enhance the capabilities of target awareness. Finally, the fused images are reconstructed by the contrast-texture retention module (CTRM) and reconstructor. The extensive experiments indicates that the proposed approach achieves improved performance both in human perception and quantitative analysis compared with the state-of-the-art (SOTA) methods.

Voice Liveness Detection (VLD) has become one of the hot research topics in the Internet of Things era. Conventional VLD methods are centralized solutions that are trained on abundant data collected from local clients and stored on the server, however, they have the risk of causing data islands and privacy leakage. To address this problem, we propose a novel word-level VLD framework based on asynchronous federated learning (FL) with pop noise, named FL-VLD. Structurally, FL-VLD takes the preprocessed voice for local model training and constructs a global model by only transmitting the learned weights after differential privacy with FL’s central server in an asynchronous manner. In addition, the local network of the framework incorporates the residual network and the spatial grouping enhancement module to optimize the complexity and accuracy of the global model. With the advantage of FL’s distributed structure, FL-VLD solves the data island problem in VLD scenario without threatening users’ privacy. Experimental results on the popular POCO dataset show that our proposal is clearly superior to the traditional centralized methods as well as overperforming other federated schemes in terms of fairness, stability, accuracy, and lightweightness. Further, for attacks involving far-field replay, synthesis, and conversion, FL-VLD has high generalisation capabilities. Finally, the ablation study attests to its efficacy.

The expansion of machine learning applications across dif- ferent domains has given rise to a growing interest in tap- ping into the vast reserves of data that are being generated by edge devices. To preserve data privacy, federated learn- ing was developed as a collaboratively decentralized privacy- preserving technology to overcome the challenges of data silos and data sensibility. This technology faces certain lim- itations due to the limited network connectivity of mobile devices and malicious attackers. In addition, data samples across all devices are typically not independent and iden- tically distributed (non-IID), which presents additional chal- lenges to achieving convergence in fewer communication rounds. In this paper, we have simulated attacks, namely Byzantine, label flipping, and noisy data attacks, besides non-IID data. We proposed Robust federated learning against poisoning attacks (RFCaps) to increase safety and accelerate conver- gence. RFCaps incorporates a prediction-based clustering and a gradient quality evaluation method to prevent attack- ers from the aggregation phase by applying multiple filters and also accelerate convergence by using the highest quality gradients. In comparison to MKRUM, COMED, TMean, and FedAvg algorithms, RFCap has high robustness in the pres- ence of attackers and has achieved a higher accuracy of up to 80% on both MNIST and Fashion-MNIST datasets.

Cardiovascular disease (CVD) believes to be a major cause of transience and indisposition worldwide. Early diagnosis and timely intervention are critical in preventing the progression of CVD and improving patient outcomes. Machine learning (ML) algorithms have emerged as powerful tools in CVD recognition, with the potential to assist physicians in making accurate and efficient diagnoses. This research paper explores the combination of multiple ML algorithms for CVD recognition, utilizing diverse datasets such as the Cleveland, Hungarian, Switzerland, statlog, and VA Long Beach datasets. Additionally, a CVD dataset comprising 12 attributes and 70,000 records is employed, demonstrating improved results through the proposed and trained model compared to previous prediction techniques for CVD. The performance of various ML techniques, including support vector machines (SVM), Naive Bayes (NB), K-Nearest Neighbor (KNN), Random Forest (RF), and Logistic Regression (LR), is evaluated and compared. The impact of feature selection and feature scaling on the models’ performance is also examined. An ensemble bagging techniques is applied which is being embedded with other classifiers. LR classifier embedded with bagging techniques proved to be our proposed model. The findings reveal that the proposed Hybrid Linear Regression Bagging Model (HLRBM) outperforms other models. Furthermore, the study highlights the significance of data preprocessing techniques, such as data normalization and class balancing, which significantly enhance the performance of all models. To this end, standard scalar and Synthetic Minority Over-sampling Technique (SMOTE) are employed. The study emphasizes the importance of selecting an appropriate ensemble technique in conjunction with various ML algorithms and preprocessing methods for CVD prediction. Overall, the research provides valuable insights into the potential of ML in improving CVD risk assessment.

This paper presents a new Multi Agent Collaborative Search Algorithm with Adaptive Weights (named MACSAW). MACS is a memetic scheme for multi-objective optimization which contains two kind of actions, the local actions and social actions. The former explore the neighborhood of some virtual agents and the latter push the individual towards the Pareto front. On the base of the latest version of MACS, MACS2.1, we improve the old algorithm from three direction. First, a new kind of utility function is introduced to enhance the convergence. Next, a new social action process which contains more operators and adaptive parameters is embedded in MACSAW. Finally, MACS2.1 lacks the weight vectors adjustment process which leads to diversity losing in some real problems and MACSAW adds it. Further, MACSAW is compared with some state-of-art algorithms and MACS2.1 on some standard benchmarks. It gets competitive results. Two real optimization problems is tackled and the results are analyzed in details.

In the present era, healthcare systems grapple with substan- tial volumes of medical data. However, a significant portion of this data is marked by incompleteness, inconsistency, er- rors, and unsuitability for training Machine Learning (ML) or Deep Learning (DL) algorithms. This necessitates preprocess- ing the data to render it amenable to utilization by ML/DL al- gorithms. Medical datasets predominantly feature two types of attributes: numerical and categorical values. The conver- sion of categorical features into numerical vectors is a crucial step in preparing the data for ML/DL algorithms, known as Feature Engineering (FE) based categorical encoding. Con- ventional and straightforward encoding of categorical fea- tures, termed one-hot encoding, generates multiple columns, thereby transforming data from a lower-dimensional to a higher- dimensional space. This approach poses challenges, includ- ing increased memory requirements due to the proliferation of columns. Considering these issues, this research proposes an encoding technique named “Weight of Evidence with En- tity Embedding” (WoEEE). The WoEEE approach bolsters the predictive capabilities of ML/DL algorithms by calculating the weight of evidence and concurrently mitigates dimension- ality issues. To empirically validate the proposed method, it is tested on six diverse datasets: Breast Cancer, Hospital Readmission, Vadu, Covid-19, Stroke, and Heartstatlog. Four distinct ML/DL algorithms—Decision Tree (DT), Random For- est (RF), Logistic Regression (LR), and a simple Feed-forward Neural Network (NN)— are employed for testing. The re- sults obtained demonstrate that the WoEEE approach yields an average improvement of 11.18%, 10.37%, 5.83%, 7.58%, 7.83%, and 6.83% across all combinations of datasets, classi- fiers, and encoding methods. Furthermore, an Anova test is performed to confirm the effectiveness of WoEEE in encod- ing categorical data, especially for tasks involving binary clas- sification. This enhances the treatment of categorical data in ML and data analytics scenarios. Overall, WoEEE shows po- tential as a valuable approach for categorical data encoding, making a positive contribution to the creation of effective techniques for handling this type of data in real-world appli- cations.

Anemia occurs when the hemoglobin (Hgb) value falls below a certain reference range. It requires many blood tests, radiological images, and tests for diagnosis and treatment. By processing medical data from patients with artificial intelligence and machine learning methods, disease predictions can be made for newly ill individuals and decision-support mechanisms can be created for physicians with these predictions. Thanks to these methods, which are very important in reducing the margin of error in the diagnoses made by doctors, the evaluation of data records in health institutions is also important for patients and hospitals. In this study, three hybrid models are proposed to classify non-anemia records, Hgb-anemia, folate deficiency anemia (FDA), iron deficiency anemia (IDA), and B12 deficiency anemia by combining artificial intelligence and machine learning methods TreeBagger with Crow Search Algorithm (CSO), Chicken Swarm Optimization Algorithm (CSO) and JAYA methods. The proposed hybrid models aim to achieve high performance by better emphasizing the importance of parameters. To solve the multiclass anemia classification problem, fuzzy logic-based parameter optimization is applied to improve the class-based accuracy as well as the overall accuracy in the dataset. The classification performances of the proposed methods are evaluated using accuracy, F-score, precision, and recall criteria to build a prediction model to identify the anemia type of anemic patients. A result of the study on the dataset taken from the Kaggle database found that the three proposed hybrid methods outperformed other studies using the same dataset and similar studies in the literature.

Binary combinatorial optimization plays a crucial role in various scientific and engineering fields. While deterministic approaches have traditionally been used to solve these problems, stochastic methods, particularly metaheuristics, have gained popularity in recent years for efficiently handling large problem instances. Ant Colony Optimization (ACO) is among the most successful metaheuristics and is frequently employed in non-binary combinatorial problems due to its adaptability. Although for binary combinatorial problems ACO can suffer from issues such as rapid convergence to local minima, its eminently parallel structure means that it can be exploited to solve large and complex problems also in this field. In order to provide a versatile ACO implementation that achieves competitive results across a wide array of binary combinatorial optimization problems, we introduce a parallel multicolony strategy with an improved cooperation scheme to maintain search diversity. We evaluate our proposal (Binary Parallel Cooperative ACO, BiPCACO) using a comprehensive benchmark framework, showcasing its performance and, most importantly, its flexibility as a successful all-purpose solver for binary combinatorial problems.

A document by Said Yasin. Click on the document to view its contents.

This article presents the architecture, design and validation of a microservice orchestration approach, that improves the flexibility of heterogeneous microservice-based platforms. Improving user experience and interaction, for time-critical applications are aspects that were primary objectives for the design of the architecture. Each microservice can provide its own embedded user interface component, also decentralizing it and, in consequence, improving the loosely coupled approach to the architecture. Obtained results are promising, with high throughput and low response times. Also, a key finding was the introduction of benchmarking as a new step in the development lifecycle of performance-critical software components, with an example of how it can be applied within an Agile methodology. Further research is proposed to improve the results and raise the final technology readiness level of the system. Obtained results already make the approach a candidate and viable alternative to classical service composers.

The Siamese tracker consists of two components: a classification and a regression networks. Despite their different roles, most Siamese trackers have similar feature fusion modules in the two networks, leading to the neglect of their unique characteristics. In this work, we experimentally discover that the two networks place different levels of emphasis on different types of information. Specifically, regression tends to rely on semantic information, while classification places more emphasis on global information. Therefore, we propose a new tracking structure named SGTrack, which includes a semantic augmentation fusion (SAF) for regression and a global relevance fusion (GRF) for classification. It allows us to unlock the full potential of both networks. The experimental results of our method on five benchmarks provide evidence of a notable improvement in tracking performance, while preserving real-time speed.

:30.0 The development of machine learning technologies are broadly changing how humans interact with their environments across all sectors. In industrial settings, this is referred to as the fourth industrial revolution, Industry 4.0, and encompasses several technologies that are pushing the boundaries of industrial automation. In this study, a general industrial process optimization (GIPO) methodology is formulated in the context of Industry 4.0 and tested on an industrial Injection Molding Machine (IMM). GIPO aims to encourage the practical inclusion of industrial artificial intelligence at all levels of the manufacturing process while enabling industrial equipment to adapt to a changing processing environment. Special attention is given to the generality of the methodology so that it can be extended to other applications. In the example case study presented here, GIPO combines nearest neighbors classification and nearest neighbors optimization methods to effectively optimize an Injection molding process. Practical implementation conducted on the IMM demonstrates a novel methodology to leverage data mining and machine learning methods in a real-world setting to improve the overall performance regarding production time, energy cost, and production quality.

The paper deals with the development of 2-link lower limb extremity manipulator for rehabilitation purposes in bio-medical engineering. The musculoskeletal modeling was performed on the collected human gait data to obtain the joint angles and driving force values for hip and knee movement. The results obtained from the proposed controller were verified using Matlab-Simulink. The robust adaptive RBF neural network and fuzzy sliding mode controller were developed to track the desired trajectory. The OpenSim software was utilized to perform inverse kinematics and inverse dynamics on the musculoskeletal model which is scaled according to the physical dimensions of the experimental subject. The simulation results stated that SMC-RBF network has better tracking rate, less chattering problems and handle unknown uncertainties well. Finally, the results obtained were compared with the existing method with the help of performance index parameters to prove their effectiveness.

Psychological stress has evolved as an important health concern across the globe. The vulnerability to stress and the ramifications of it have only worsened during the time of the COVID-19 pandemic. This necessitates a timely diagnosis of stress before the condition progresses to chronicity. In this context, the popularity of social media like Twitter, where large numbers of users share opinions without any social stigma, has emerged as a major resource of human opinions. This has led to an increased research interest in social media-based stress detection techniques. However, tweet-level stress detection techniques in the literature have left a void in leveraging the text information in tweets, especially the presence of sarcastic expressions in the tweet’s text content. To this end, a novel method called “Sarcasm-based Tweet-Level Stress Detection” (STSD) is proposed in this work with the modification of the logistic loss function to detect tweet-level stress by availing the information of sarcasm that exists in the tweet-content. The principle of the STSD model is to minimise the loss for non-sarcastic tweets while maximising the loss for sarcastic tweets. Furthermore, an extensive preprocessing and dimensionality reduction is performed using kernel principal component analysis (kernel PCA) to improve the performance by reducing the dimensions. The experimental results show that the proposed STSD model, when applied along with kernel PCA, records a significant improvement in accuracy by a minimum of 5.25% and a maximum of 9.19% over baseline models. Also, there is an increment in F1-score by at least 0.085 points and a maximum of 0.164 points when compared to the baseline models.