Journal of Ambient Intelligence and Humanized Computing最新文献

Machine translation has revolutionized the field of language translation in the last decade. Initially dominated by statistical models, the rise of deep learning techniques has led to neural networks, particularly Transformer models, taking the lead. These models have demonstrated exceptional performance in natural language processing tasks, surpassing traditional sequence-to-sequence models like RNN, GRU, and LSTM. With advantages like better handling of long-range dependencies and requiring less training time, the NLP community has shifted towards using Transformers for sequence-to-sequence tasks. In this work, we leverage the sequence-to-sequence transformer model to translate Urdu (a low resourced language) to English. Our model is based on a variant of transformer with some changes as activation dropout, attention dropout and final layer normalization. We have used four different datasets (UMC005, Tanzil, The Wire, and PIB) from two categories (religious and news) to train our model. The achieved results demonstrated that the model’s performance and quality of translation varied depending on the dataset used for fine-tuning. Our designed model has out performed the baseline models with 23.9 BLEU, 0.46 chrf, 0.44 METEOR and 60.75 TER scores. The enhanced performance attributes to meticulous parameter tuning, encompassing modifications in architecture and optimization techniques. Comprehensive parametric details regarding model configurations and optimizations are provided to elucidate the distinctiveness of our approach and how it surpasses prior works. We provide source code via GitHub for future studies.

The keyboard, a major mean of interaction between human and internet devices, should beset right for good performance during authentication task. To guarantee that one legitimate user can interleave or simultaneously interact with two or more devices with protecting user privacy, it is essential to build device adaptation free-text keystroke dynamics authentication (free-KDA) model based on multi-teacher knowledge distillation methods. Some multi-teacher knowledge distillation methods have shown effective in C-way classification task. However, it is unreasonable for free-KDA model, since free-KDA model is one-class classification task. Instead of using soft-label to capture useful knowledge of source for target device, we propose a device adaptation free-KDA model. When one user builds the authentication model for target device with limited training samples, we propose a novel optimization objective by decreasing the distance discrepancy in Euclidean distance and cosine similarity between source and target device. And then, we adopt an adaptive confidence gate strategy to solve different correlation for each user between different source devices and target device. It is verified on two keystroke datasets with different types of keyboards, and compared its performance with the existing dominant multi-teacher knowledge distillation methods. Extensive experimental results demonstrate that AUC of target device reaches up to 95.17%, which is 15.28% superior to state-of-the-art multi-teacher knowledge distillation methods.

Covid-19 infection influenced the screening test rate of breast cancer worldwide due to the quarantine measures, routine procedures reduction, and delay of early diagnosis, causing high mortality risk and severity of the disease. X-ray mammography is the gold standard for diagnosing early signs of breast cancer, and Artificial Intelligence enables the detection of suspicious lesions and classifying them in terms of malignancy. This paper aimed to investigate mass detection and classification in a large-scale OPTIMAM dataset with 6000 cases and extracted 3524 images with masses in the mammograms of the Hologic manufacturer. The methodology of the detection step is to train the RetinaNet architecture of ResNet50, ResNet101, and ResNet152 backbones with three types of initializations by ImageNet and COCO weights and from scratch. The dataset was pre-processed to generate two types of input with entire mammograms and patches, which are stated as the first and the second approaches. The results show that in the first approach, RetinaNet of ResNet50 backbone with ImageNet and COCO weights and ResNet152 with the same weights performed 0.91 True Positive Rate at 0.78 False Positive Per Image, respectively. In contrast, in the second approach, ResNet152 with ImageNet weights reached 0.88 TPR at 0.78 FPPI. In the classification step, the Transfer Learning approach was applied with fine-tuning by adding L2-regularization and class weights to balance class distribution in the datasets.

Driver’s mental stress is known as a prime factor in road crashes. The devastation of these crashes often results in losses of humans, vehicles, and infrastructure. Likewise, persistent mental stress could develop mental, cardiovascular, and abdominal disorders. Preceding research in this domain mostly focuses on feature engineering and conventional machine learning (ML) approaches. These approaches recognize different stress levels based on handcrafted features extracted from various modalities including physiological, physical, and contextual data. Acquiring the good quality features from these modalities using feature engineering is often a difficult job. The recent developments in the form of deep learning (DL) algorithms have relieved feature engineering by automatically extracting and learning resilient features. Conventional DL models, however, frequently over-fit due to large number of parameters. Thus, large networks face gradient vanishing issues causing an increase in learning failure and generalization errors. Furthermore, it is often hard to acquire a large dataset for training a deep learning model from scratch. To overcome these problems for driver’s stress recognition domain, this paper proposes fast and computationally efficient deep transfer learning models based on Xception pre-trained neural networks. These models classify the driver’s Low, Medium, and High stress levels through electrocardiogram (ECG), heart rate (HR), galvanic skin response (GSR), electromyogram (EMG), and respiration (RESP) signals. Continuous Wavelet Transform (CWT) acquires the scalograms for ECG, HR, GSR, EMG, and RESP signals separately. Then unimodal Xception models are trained based on these scalograms to classify the three stress levels. The proposed Xception models have achieved 97.2%, 86.4%, 82.7%, 71.9%, and 68.9% average validation accuracies based on ECG, RESP, HR, GSR, and EMG signals, respectively. The fuzzy EDAS (evaluation based on distance from average solution) approach also evaluates the performance of proposed models based on accuracy, recall, precision, F-score, and specificity. For the driver’s three stress levels, fuzzy EDAS performance estimation shows that the proposed ECG, RESP, and HR based Xception models achieved 1st, 2nd, and 3rd positions, respectively.

The Mixed Radial Basis Function Neural Network (MRBFNN) is an artificial neural network that employs Radial Basis Functions (RBFs) as activation functions in its hidden layer. The number of neurons in the hidden layer and the choice of RBF functions used in these neurons significantly affect the convergence of MRBFNN learning algorithms and impact the overall performance of neural networks. This article presents a nonlinear optimization model and an algorithm to select an appropriate architecture and learning strategy for MRBFNN. To approximate the solution of our model, we utilized an algorithm based on Particle Swarm Optimization (PSO) techniques. We will apply our approach in Medical Diseases Diagnosis (MDD). The numerical results obtained demonstrate the effectiveness of the proposed theoretical approach and underscore the advantages of the new modeling methodology.

At present, there are some problems in strength quality training methods, such as the sports training effect can not reach the expected goal, the feasibility is not good, and the lack of pertinence leads to the unsatisfactory training effect of some athletes. In order to improve the quality training effect of competitive Taekwondo special strength, this paper studies the quality training method of competitive Taekwondo special strength based on multi-scale Retinex algorithm under the background of “Internet+”. In this method, the strength quality training video image acquisition module is used to collect the strength quality training video image and transmit it to the strength quality training video image processing module, and the multi-scale Retinex algorithm is used to enhance the strength quality training video image and correct the measurement error of special strength quality training; Analyze the data related to the special strength quality training of competitive Taekwondo athletes in the intelligent evaluation module of strength quality training results, construct an evaluation index system, and evaluate the strength quality training results of competitive Taekwondo athletes; According to the evaluation results of the evaluation module, provide targeted special strength quality training programs for competitive taekwondo athletes; All sports training data are stored in the database management module. The experimental results show that this method can significantly improve the special strength quality training effect of competitive Taekwondo athletes, and can better grasp the training content and movement control accuracy.

Aptamers are very interesting peptide molecules or oligonucleic acid. They are used to bind particular target molecules. Aptamers play vital roles in various practical applications and physiological functions. Consequently, several diseases can be treated using therapies based on aptamer proteins and designing the binding of aptamers to specific proteins is essential to advance understanding into processes of interaction between aptamer-protein. Despite the wide applications of aptamers, identification of interaction between aptamer protein is always inadequate and challenging. Therefore, it is necessary to develop a computational approach for achieving good predictions of interaction between aptamer-protein. In the present study, a novel method for enhancing the prediction of interacting aptamer-target pairs based on sequence features obtained from both aptamers and their target proteins by employing a novel k-mean based intuitionistic fuzzy rough feature selection method is proposed. Firstly, an intuitionistic fuzzy rough set model based on k nearest neighbour concept is proposed. Then, a novel feature selection technique is introduced by using this model. Furthermore, non-redundant and relevant features are selected from training as well as testing datasets by using proposed feature selection technique. Secondly, SMOTE (Synthetic Minority Oversampling Technique) is applied to obtain the optimal balanced training and testing datasets. Thirdly, we apply various machine learning algorithms on optimally balanced reduced training and testing datasets to evaluate their performances. Experimental results shows that the best prediction performance is obtained by boosted random forest learning algorithm. Using a 10 fold cross-validation test, the proposed method is a good performer, with sensitivity of 91.3, 86.4, specificity of 91.9, 84.8, overall accuracy of 91.60%, 85.60%, Mathews correlation coefficient of 0.832, 0.713, AUC (area under curve) of 0.969, 0.908, and g-means of 91.5, 85.5 on optimal balanced reduced training and testing datasets consisting of aptamer-protein interacting pairs. Finally, a comparative study of the best obtained results with the existing best results is presented, which clearly indicates that our proposed approach is the best performing approach till date.

Since the advent of mass access to the Internet, aggressive behaviors such as cyberbullying have become widespread on social networking sites. An aggressive online environment can lead to negative attitudes that negatively impact the victim, bystanders, and the bullies themselves. One of the main reasons for the increase in this type of behavior is contagion from bystanders—a phenomenon that needs to be stopped. In recent years, many studies have looked at cyberbullying detection, considering various factors to improve detection, such as extracting different types of features, comparing the performance of different classifiers, and processing datasets in myriad ways. It is evident from our findings that previous works in the literature fell short of detecting cyberbullying by ignoring the characteristics of bystanders and their roles. Thus, this paper aims to present a systematic literature review of research conducted over the past 10 years to determine which methods encompassed features related to bystanders and their role and analyzed the contagion and causal factors of the spread of cyberbullying. There are different studies confirmed the existence of bystander contagion, which researchers rarely consider to detect cyberbullying. This gap could be exploited in future studies and used to improve the detection of cyberbullying. Therefore, in this paper, the summary and comparison of findings from the selected studies that examined the role of bystanders in cyberbullying are presented, concluding how bystander-related features could contribute to the detection of cyberbullying.

Financial and credit institutions need to evaluate and rank their subsidiaries to control and improve their performances. There are several methods to evaluate the performance of such branches. In order to take advantage of the strengths of each of these methods and cover some of the limitations that exist in each of these methods alone, in this study, an algorithm which is a combination of multi-criteria decision-making methods, statistical analysis, and data envelopment analysis is proposed. The location of each of the methods mentioned in the steps of the algorithm, and its simulation to a standard linear programming model in MATLAB software, is the main research problem that is designed and presented for fuzzy type uncertain data. The proposed algorithm was used for 1736 branches of a certain bank in banking sector of Iran with uncertain data. Analysis of the results for different alpha-cuts and testing them with SPSS software show that with increasing the range of fuzzy numbers, the number of efficient branches increases and also affect the ranking. Nevertheless, there is still a significant correlation even in the alpha-cut changes in the ranking results.