VFAST - Virtual Foundation for Advancement of Science and Technology (Pakistan)
Not a member yet
1255 research outputs found
Sort by
Recent Advances in Machine Learning Models for Antiviral Peptide Prediction
Full text linkViral diseases are widespread, and their impact is expressed in millions of cases of infection and mortality around the world. Chronic viral diseases include COVID-19, HIV, and hepatitis. To prevent and treat these viral infections, novel agents and antiviral peptides (AVPs) have been developed. Thus, identifying AVPs is crucial because these pieces of information are invaluable throughout the entire pharmaceutical industry and other sciences. This has been done experimentally and computationally, but the need for better, more accurate, and efficient predictors persists. This research also reviews current AVP predictors, including the datasets employed, feature representation methods, classification algorithms, and assessment criteria. In our paper, we discuss the weaknesses of the existing techniques, overview the most efficient strategies, and evaluate the benefits and drawbacks of the classifiers. Furthermore, several directions for future work are discussed in detail, including enhanced feature representation techniques, feature selection, and classification approaches. The following advancements aim to improve the effectiveness of algorithms for predicting AVPs, enabling the development of new antiviral drugs more successfully
Analyzing the Impact of Machine Learning Algorithms on Software Requirements Classification
Full text linkAlong with the rapid growth of the world, the demand for efficient and successful software has increased swiftly. Any software has many steps for developing software and the most important step is software requirements engineering. Requirements classification can be applied manually, which requires great effort, time, and cost and the accuracy may vary. Many previous studies utilized machine learning algorithms to automate the classification process but traditional classification algorithms often require a large amount of labeled data, which can be expensive and time-consuming to collect. Few-Shot Learning (FSL) excels in situations with limited data, making it a promising alternative. This paper investigates the potential of applying Few-Shot Learning (FSL) algorithms for classifying software requirements. This study explores three prominent FSL algorithms: Prototypical Networks, Matching Networks, and Model-Agnostic Meta-Learning (MAML). These algorithms are evaluated on their ability to classify software requirements using a publicly available dataset. The results demonstrate that Prototypical Networks outperforms Matching Networks and MAML in this specific application. Matching Networks, designed for visual similarity tasks, struggle with textual data. Prototypical Networks achieve a remarkable accuracy of 82 percent, suggesting their effectiveness in learning class representations from a small number of samples. MAML also shows promising results with an accuracy of 76.9 percent. While acknowledging limitations in data pre-processing, the study concludes that FSL holds significant potential for efficient and cost-effective software requirement classification, particularly when dealing with limited labeled data
Scalable Hybrid Deep Learning-Based Architecture for Glaucomatous and Healthy Eye Classification in Retinal Fundus Images
Full text linkGlaucoma remains one of the leading preventable causes of irreversible blindness worldwide, with early detection being essential for preserving vision. This study presents a hybrid deep learning framework that integrates VGG16 and ResNet-50 architectures to improve the classification of glaucoma severity using retinal fundus images. A balanced dataset of 2,081 images was utilized, with data augmentation and the Synthetic Minority Over-sampling Technique (SMOTE) applied to address class imbalance and enhance model generalization. All images were normalized and resized to 224 × 224 pixels, and training was conducted for 50 epochs with a batch size of 32, resulting in approximately 14.7 million trainable parameters. The proposed hybrid model achieved an average accuracy of 83%, surpassing standalone VGG16 (80%), ResNet-50 (70%), and EfficientNet-B0 (51%), underscoring the benefit of combining hierarchical and residual feature extraction. In addition, it achieved a precision of 82%, recall of 81%, and an F1-score of 82%, with a final loss value of 0.36. Quantization-aware training was employed to optimize computational efficiency, reducing the average prediction time to 95 milliseconds per image and enabling near real-time deployment in low-resource clinical environments. While some misclassifications were observed due to close visual similarities among glaucoma stages, the proposed approach demonstrates strong potential as a scalable and efficient solution for automated glaucoma screening and early detection
Machine and Deep Learning Approaches for Alzheimer\u27s Disease Classification with EEG Signals and MRI Images
Full text linkAlzheimer’s disease (AD) is a progressive neurodegenerative disorder and the main cause of dementia globally, making early and accurate diagnosis crucial for early and effective treatment. Traditional diagnostic approaches, such as MRI and EEG-based diagnosis, have limitations due to their high cost, lengthy execution times, and the need for specialized clinical expertise. Recent advances in machine learning (ML) and deep learning (DL) offer new prospects for automating and improving AD detection by extracting discriminative structures from multimodal biomedical data. This study proposes a smart system for Alzheimer’s disease classification based on the MRI images. Various deep learning models such as VGG16, InceptionV3, and ResNet50 were used for MRI-based classification. Various preprocessing, enhancement, and feature selection techniques were applied to enhance data quality and model performance. Experimental results show that VGG16 achieves the best accuracy of 98% on MRI images followed by the Inception 3 model. The RestNet50 model performed worst compared to these two models with respect to accuracy, F1 score and Recall. The proposed work shows the potential of using advanced learning algorithms to achieve robust and scalable early diagnosis of Alzheimer’s disease, thereby assisting clinicians in reducing misdiagnosis and improving patient treatment outcomes
Enhancing Traffic Sign Recognition in Rough Weather to Reduce Traffic Accidents
Full text linkBackground: Automated traffic sign recognition is a technology that helps in detecting and understanding the traffic signs on the road. These signs may include stop signs, speed limits and some other warning signs. These signs are important in maintaining road safety and helping drivers obeying the traffic rules. Machine learning and computer vision together have advanced in this field and many methods have been proposed for automated traffic sign detection and recognition. Despite these advancements, there are still many challenging situations including occlusion, varying lighting conditions, difficult environmental conditions and sign variations which still need attention. Methods: We use Arabic traffic sign dataset to train EfficientNetB3 architecture with attention mechanism to classify the traffic sign under the unique visual and linguistic complexities as well as diverse environmental conditions. We also improved the dataset by augmenting and adding extra images to cover actual scenarios like fog, heavy rain, low light etc. making it stronger for testing and future research. Results: Our trained model achieved the accuracy of 99.61\% in testing better than the compared to the baseline CNN-Resnet model. Conclusion: In this research we addressed the existing limitations and sets a benchmark for the effective and efficient classification of Arabic traffic signs, particularly in challenging weather conditions
Predictive Modeling of Parameter Variations in Conveyor Belt Dryers using Machine Learning for Improved Drying Performance
Full text linkThis study takes an in-depth look at the mathematical model governing the operation of a tangential flow conveyor dryer operating in a co-current configuration. The Conveyor Belt Dryer (CBD) is represented as an Ordinary Differential Equation (ODE), and our research focuses on studying the influence of parameter variation and symmetry on the Rate of Exchange of Moisture Content (RMC). To achieve this, we employ a system studying framework based on artificial Neural Networks (NN) and the Levenberg-Marquardt Training (LMT) algorithm, enabling the examination of symmetry within surrogate solutions. The RK-four method is applied to generate target data points for supervised learning in the NN-LMT structure. Furthermore, we investigate various scenarios of the mathematical model related to the rate of change of moisture content. Detailed graphical representations, including histograms, absolute error plots, curve fitting graphs, and regression graphs, are employed to facilitate comprehensive explanations. Additionally, a comparative analysis between the numerical solutions obtained through the machine learning technique is provided, followed by graphical and statistical representations of the determined errors
A Review of Developments in Generative AI, Machine Learning, and Neuroimaging for the Diagnosis of Autism.
Full text link Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder with be- havioral and neurological variations. Diagnosing ASD is complicated because behavioral assessments vary from person to person, and it is challenging as it relies on subjective behavioral assessments rather than objective medical tests. Following the PRISMA guide- lines for a systematic review, 53 papers from ScienceDirect, IEEE Xplore, and PubMed databases published between 2015 and 2024 have been selected, to determine how ML (Machine Learning) and GANs (Generative Adversarial Network) enhance the use of neuroimaging techniques for the early detection of ASD. The study addresses significant dataset issues by combining discriminative machine learning models, such as SVMs (Sup- port Vector Machine) and CNNs (Convolutional Neural Network), for pattern recognition in sMRI (Structural Magnetic Resonance Imaging), fMRI (Functional Magnetic Resonance Imaging), and EEG (Electroencephalography) with GANs for data augmentation and en- hanced feature learning. The review highlights how AI algorithms can be helpful in the diagnosis of ASD, particularly in areas with limited medical facilities. For example, due to limited resources, just 1 in 628 children in Pakistan are diagnosed with autism. Future re- search will focus on methods that combine explainable AI, generative and discriminative techniques, and structured datasets, while improving model transparency, generalizabil- ity, and avoiding biased data. By integrating AI with clinical information, this research helps to design treatment protocols that may be customized to each patient’s specific requirements and enables earlier identification of autism
Mitigating Cyber Threats: Machine Learning and Explainable AI for Phishing Detection
Full text linkThe exponential growth of organizations and users has accelerated the adoption of new technologies, increasing the complexity of online security. Phishing attacks have surged significantly in 2024, with over 932,923 incidents reported in Q3 alone, driven by advanced AI-enabled social engineering tactics. From simple scams to sophisticated schemes exploiting emails, URLs, text messages, and social media platforms, phishing attacks deceive victims into disclosing sensitive information or inadvertently installing malware, often compromising devices as part of more extensive botnet networks. Despite advancements in Cyber-security measures, phishing remains a critical threat, causing substantial financial and reputational damage to businesses. Recently, Machine Learning (ML) algorithms have demonstrated remarkable efficacy in phishing detection; however, many high-performing models operate as black boxes, raising concerns about transparency, interpretability, and trustworthiness—factors essential in high-stakes applications for ensuring reliability, accountability, and regulatory compliance. This research integrates ML techniques with Explainable Artificial Intelligence (XAI) methodologies to address this issue and enhance model interpretability and transparency in phishing detection. The proposed approach employs Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), Random Forest, k-Nearest Neighbors (KNN), Twin Support Vector Machine (Twin SVM), and Convolutional Neural Networks (CNN), evaluated across four publicly available datasets to assess performance and interpretability. The research findings reveal that XGBoost achieved the highest accuracy at 99.65%. The Local Interpretable Model-agnostic Explanations (LIME) method was applied to elucidate the importance of feature and model decision-making processes. This comprehensive approach aims to strengthen Cyber-security resilience against phishing threats while promoting model transparency and regulatory compliance
How Effective are AI Tools for Diverse Learners? A Case Study on Python and Data Science Education
Full text linkArtificial Intelligence (AI) based tools (such as ChatGPT) has revolutionized the learning methodologies for complex subjects by introducing the innovative methods in education. This study focuses on how AI tools can be helpful for the undergraduate students at Shah Abdul Latif University in learning the Python programming and Data science concepts. The study was conducted on beginners and intermediate groups of students through an organized survey by comparing their experiences gained through interactive tasks. The survey results reflect that intermediate users get more benefits from AI tools due to their familiarity with the technologies, whereas beginners face challenges in comprehension and ease of use. The study concludes recommending some practical suggestions to enable AI tools more effective, comprehensive and user-friendly
Harnessing Machine Learning for Accurate Smog Level Prediction: A Study of Air Quality in India
Full text linkAccurate prediction of smog concentrations is needed to mitigate the harm of AP on public health and the environment. This research proposes a new method to combine machine learning (ML) models with live data from Central Pollution Control Board (CPCB) to fill in the smog prediction accuracy gaps. The data consist of hourly AQI readings from different towns in India which were preprocessed to adjust for missing values and normalize data before ML models. The algorithms were tested with 8 ML algorithms, and hyper-parameter settings were tuned using the GridSearchCV method. The results show that XG Boost Regressor (XGBR) and Extra Tree Regressor (ETR) models significantly surpass other ML algorithms and traditional techniques with better accuracy on predicting smog. These results are useful for policymakers and environmental agencies to implement sustainable air quality management