In recent years, the rapid advancement of Large Language Models (LLMs) has significantly transformed natural language processing (NLP), enabling impressive performance across a wide range of tasks. However, these developments have largely benefited high-resource languages, leaving many low-resource and underrepresented languages at risk of further digital marginalization. Addressing this imbalance is crucial to building more inclusive and culturally sustainable AI systems, which is motivating growing research interest in adapting LLMs for linguistically diverse and resource-scarce communities. This systematic review examines recent progress (2020–2025) in the pretraining and adaptation of LLMs for Low-Resource Languages (LRLs). Analysed 812 records obtained in the large databases and using PRISMA criteria, 140 core studies were identified. The innovations in data augmentation and parameter-efficient fine-tuning approaches can be outlined in this selection process. It combines major innovations on data-driven augmentation, parameter-efficient fine-tuning and morphologically rich and underrepresented language script-sensitive tokenization. The results highlight the growing effectiveness of culturally aware standards such as IrokoBench and BLEnD and show that approaches to lightweight adaptation eliminate high computational costs while maintaining language accuracy. The review focuses on the ethics in AI practice, the development of corpora through communities, and interdisciplinary research collaboration among computational linguists, social scientists, and digital humanists. The task of generating a diversified dataset, typology-conscious modelling strategies, and open-source multilingual benchmarks should be prioritized in future research as one possible solution to the existing digital language gap worldwide.

As healthcare systems increasingly embrace digitalization, effective management of electronic health records (EHRs) has emerged as a critical priority, particularly in inpatient settings where data sensitivity and real-time access are paramount. Traditional EHR systems face significant challenges, including unauthorized access, data breaches, and inefficiencies in tracking follow-up appointments, which heighten the risk of misdiagnosis and medication errors. To address these issues, this research proposes a hybrid blockchain-based solution for securely managing EHRs, specifically designed as a framework for tracking inpatient follow-ups. By integrating QR code-enabled data access with a blockchain architecture, this innovative approach enhances privacy protection, data integrity, and auditing capabilities, while facilitating swift and real-time data retrieval. The architecture adheres to Role-Based Access Control (RBAC) principles and utilizes robust encryption techniques, including SHA-256 and AES-256-CBC, to secure sensitive information. A comprehensive threat model outlines trust boundaries and potential adversaries, complemented by a validated data transmission protocol. Experimental results demonstrate that the framework remains reliable in concurrent access scenarios, highlighting its efficiency and responsiveness in real-world applications. This study emphasizes the necessity for hybrid solutions in managing sensitive medical information and advocates for integrating blockchain technology and QR code innovations into contemporary healthcare systems.

Manual inspection of onba earing casting defects is not realistic and unreliable, particularly in the case of some micro-level anomalies which lead to major defects on a large scale. To address these challenges, we propose BearFusionNet, an attention-based deep learning architecture with multi-stream, which merges both DenseNet201 and MobileNetV2 for feature extraction with a classification head inspired by VGG19. This hybrid design, figuratively beaming from one layer to another, extracts the enormity of representations on different scales, backed by a pre-preprocessing pipeline that brings defect saliency to the fore through contrast adjustment, denoising, and edge detection. The use of multi-head self-attention enhances feature fusion, enabling the model to capture both large and small spatial features. BearFusionNet achieves an accuracy of 99.66% and Cohen’s kappa score of 0.9929 in Kaggle’s Real-life Industrial Casting Defects dataset. Both McNemar’s and Wilcoxon signed-rank statistical tests, as well as five-fold cross-validation, are employed to assess the robustness of our proposed model. To interpret the model, we adopt Grad-Cam visualizations, which are the state of the art standard. Furthermore, we deploy BearFusionNet as a web-based system for near real-time inference (5–6 s per prediction), which enables the quickest yet accurate detection with visual explanations. Overall, BearFusionNet is an interpretable, accurate, and deployable solution that can automatically detect casting defects, leading to significant advances in the innovative industrial environment.

This research introduces an innovative lightweight image segmentation framework where models of hybrid architectures work together to predict the output and also have self-adapting ability, along with maintaining data privacy. In this framework, data is distributed and trained in a decentralized way using different deep learning architectures. That is how the advantages of all these models will be integrated into the system. Each trained model makes its own prediction, and the final output is determined through cooperation among these models. Here, the confidence-level and pixel-wise voting majority algorithms will be utilized for the co-operation-based output prediction system. Due to the efficient setup of the operations of these two algorithms, each input will get its accurate output. Additionally, the federated learning-based self-adapting feature facilitated the proposed framework for advancing its performance consistently by interacting with the inputs. Here, UNet, SegNet and FCNN models have been trained and integrated into the prediction framework. Here, the Oxford-IIT pet dataset was used. And all the data of this dataset is distributed among these three models. The framework’s effectiveness was measured using metrics like average pixel accuracy, IoU, F1 score, precision, and recall, which resulted in scores of 89.26%, 71.48%, 81.29%, 83.96% and 81.16%, respectively. Another notable feature of this proposed framework is allocating comparatively fewer computational resources and taking less time. To validate these claims, the proposed system is compared with three other state-of-the-art models, and the proposed system delivered superior performance among all.

Human activity recognition is a significant area of research in artificial intelligence for surveillance, healthcare, sports, and human-computer interaction applications. The article benchmarks the performance of You Only Look Once version 11-based (YOLOv11-based) architecture for multi-class human activity recognition. The article benchmarks the performance of You Only Look Once version 11-based (YOLOv11-based) architecture for multi-class human activity recognition. The dataset consists of 14,186 images across 19 activity classes, from dynamic activities such as running and swimming to static activities such as sitting and sleeping. Preprocessing included resizing all images to 512 $\times$ 512 pixels, annotating them in YOLO’s bounding box format, and applying data augmentation methods such as flipping, rotation, and cropping to enhance model generalization. The proposed model was trained for 100 epochs with adaptive learning rate methods and hyperparameter optimization for performance improvement, with a mAP@0.5 of 74.93% and a mAP@0.5-0.95 of 64.11%, outperforming previous versions of YOLO (v10, v9, and v8) and general-purpose architectures like ResNet50 and EfficientNet. It exhibited improved precision and recall for all activity classes with high precision values of 0.76 for running, 0.79 for swimming, 0.80 for sitting, and 0.81 for sleeping, and was tested for real-time deployment with an inference time of 8.9 ms per image, being computationally light. Proposed YOLOv11’s improvements are attributed to architectural advancements like a more complex feature extraction process, better attention modules, and an anchor-free detection mechanism. While YOLOv10 was extremely stable in static activity recognition, YOLOv9 performed well in dynamic environments but suffered from overfitting, and YOLOv8, while being a decent baseline, failed to differentiate between overlapping static activities. The experimental results determine proposed YOLOv11 to be the most appropriate model, providing an ideal balance between accuracy, computational efficiency, and robustness for real-world deployment. Nevertheless, there exist certain issues to be addressed, particularly in discriminating against visually similar activities and the use of publicly available datasets. Future research will entail the inclusion of 3D data and multimodal sensor inputs, such as depth and motion information, for enhancing recognition accuracy and generalizability to challenging real-world environments.

This study investigates the application of Learnable Memory Vision Transformers (LMViT) for detecting metal surface flaws, comparing their performance with traditional CNNs, specifically ResNet18 and ResNet50, as well as other transformer-based models including Token to Token ViT, ViT without memory, and Parallel ViT. Leveraging a widely-used steel surface defect dataset, the research applies data augmentation and t-distributed stochastic neighbor embedding (t-SNE) to enhance feature extraction and understanding. These techniques mitigated overfitting, stabilized training, and improved generalization capabilities. The LMViT model achieved a test accuracy of 97.22%, significantly outperforming ResNet18 (88.89%) and ResNet50 (88.90%), as well as the Token to Token ViT (88.46%), ViT without memory (87.18), and Parallel ViT (91.03%). Furthermore, LMViT exhibited superior training and validation performance, attaining a validation accuracy of 98.2% compared to 91.0% for ResNet18, 96.0% for ResNet50, and 89.12%, 87.51%, and 91.21% for Token to Token ViT, ViT without memory, and Parallel ViT, respectively. The findings highlight the LMViT’s ability to capture long-range dependencies in images, an area where CNNs struggle due to their reliance on local receptive fields and hierarchical feature extraction. The additional transformer-based models also demonstrate improved performance in capturing complex features over CNNs, with LMViT excelling particularly at detecting subtle and complex defects, which is critical for maintaining product quality and operational efficiency in industrial applications. For instance, the LMViT model successfully identified fine scratches and minor surface irregularities that CNNs often misclassify. This study not only demonstrates LMViT’s potential for real-world defect detection but also underscores the promise of other transformer-based architectures like Token to Token ViT, ViT without memory, and Parallel ViT in industrial scenarios where complex spatial relationships are key. Future research may focus on enhancing LMViT’s computational efficiency for deployment in real-time quality control systems.

This research introduces a unique approach to segmenting breast cancer images using a U-Net-based architecture. However, the computational demand for image processing is very high. Therefore, we have conducted this research to build a system that enables image segmentation training with low-power machines. To accomplish this, all data are divided into several segments, each being trained separately. In the case of prediction, the initial output is predicted from each trained model for an input, where the ultimate output is selected based on the pixel-wise majority voting of the expected outputs, which also ensures data privacy. In addition, this kind of distributed training system allows different computers to be used simultaneously. That is how the training process takes comparatively less time than typical training approaches. Even after completing the training, the proposed prediction system allows a newly trained model to be included in the system. Thus, the prediction is consistently more accurate. We evaluated the effectiveness of the ultimate output based on four performance matrices: average pixel accuracy, mean absolute error, average specificity, and average balanced accuracy. The experimental results show that the scores of average pixel accuracy, mean absolute error, average specificity, and average balanced accuracy are 0.9216, 0.0687, 0.9477, and 0.8674, respectively. In addition, the proposed method was compared with four other state-of-the-art models in terms of total training time and usage of computational resources. And it outperformed all of them in these aspects.

Machine fault diagnostics are essential for industrial operations, and advancements in machine learning have significantly advanced these systems by providing accurate predictions and expedited solutions. Machine learning models, especially those utilizing complex algorithms like deep learning, have demonstrated major potential in extracting important information from large operational datasets. Despite their efficiency, machine learning models face challenges, making Explainable AI (XAI) crucial for improving their understandability and fine-tuning. The importance of feature contribution and selection using XAI in the diagnosis of machine faults is examined in this study. The technique is applied to evaluate different machine-learning algorithms. Extreme Gradient Boosting, Support Vector Machine, Gaussian Naive Bayes, and Random Forest classifiers are used alongside Logistic Regression (LR) as a baseline model because their efficacy and simplicity are evaluated thoroughly with empirical analysis. The XAI is used as a targeted feature selection technique to select among 29 features of the time and frequency domain. The XAI approach is lightweight, trained with only targeted features, and achieved similar results as the traditional approach. The accuracy without XAI on baseline LR is 79.57%, whereas the approach with XAI on LR is 80.28%.