Evaluating VGG16 and Mobilenetv2 for Automatic Recognition of Art Periods Using Transfer Learning (Published)
Artificial intelligence offers new ways to organize and share cultural heritage collections. This paper evaluates its potential for automatically classifying paintings by art period, as a tool for the digitization and accessibility of art-related information. Two convolutional neural network (CNN) models, VGG16 and MobileNetV2, were adapted through transfer learning from ImageNet-pretrained weights and trained on the WikiArt dataset, which contains 81,444 paintings across 27 art periods. Class imbalance was addressed with data augmentation and class weights, and the data were split into stratified training, validation and test sets (70/15/15). VGG16 achieved the best performance, with an AUC of 0.966, an accuracy of 0.763 and an F1-score of 0.701, compared with 0.946, 0.703 and 0.699 for MobileNetV2. MobileNetV2, however, trained faster and used less memory. The results reveal a trade-off between accuracy and efficiency that can guide the choice of model for galleries, museums and educational platforms.
Keywords: Computer Vision, MobileNetV2, VGG16, art period classification, convolutional neural networks, transfer learning
Leakage-Aware Evaluation of a Dual-Backbone Feature-Fusion Ensemble for Multi-Class Brain Tumor MRI Classification (Published)
Deep-learning studies of brain tumor MRI classification often report accuracies close to or above 99%, yet public benchmarks contain duplicate and near-duplicate images that can leak between training and test sets. This study audits the widely used 7,200-image Brain Tumor MRI Dataset and proposes a leakage-aware protocol that removes exact duplicates and groups near-duplicates, using perceptual hashing with pixel correlation, before a group-stratified split. The audit shows that 36.25% of the official test images have a near-duplicate in the official training set. Under the leakage-free protocol, an EfficientNetV2-B0 + ResNet50V2 feature-fusion network and a three-model ensemble were trained with test-time augmentation. The ensemble, selected on validation data alone, achieved 97.80% accuracy (95% CI 96.86–98.64%), macro-F1 of 0.978 and ROC-AUC of 0.998 on 954 unseen images, missing one of 740 tumors; the fusion model alone reached 98.32%. Rigorous de-duplication thus yields credible, still clinically promising, performance estimates.
Keywords: Brain tumor classification, Data Leakage, Ensemble learning, Feature fusion, MRI, transfer learning
Training AI Models with Minimal Data: Strategies for High Accuracy on a Lean Dataset (Published)
Data scarcity presents a significant challenge in artificial intelligence implementation across industries, constraining organizations from deploying effective machine learning solutions. This article explores strategic approaches that transform limited datasets from barriers into competitive advantages through methodological innovation. By examining transfer learning mechanisms that leverage pre-existing knowledge, data augmentation techniques that artificially expand available examples, few-shot and zero-shot learning paradigms that function with minimal labeled instances, and active learning strategies that optimize annotation resource allocation, a framework emerges for maximizing model performance under severe data constraints. These complementary strategies, when thoughtfully integrated, enable high-accuracy AI models in domains previously considered impractical due to insufficient training data. The economic, regulatory, and practical implications extend beyond technical performance enhancement to fundamentally alter the feasibility landscape of AI adoption, particularly in specialized domains where data collection faces inherent limitations such as healthcare, manufacturing, and low-resource languages.
Keywords: active learning, data augmentation, few-shot learning, minimal data learning, transfer learning
A Comprehensive Guide to Optimizing Machine Learning and Deep Learning Models (Published)
Machine learning and deep learning model optimization remain a pivotal aspect of artificial intelligence development, encompassing crucial elements from data preprocessing to deployment monitoring. The optimization process involves multiple interconnected stages, including data quality management, algorithm selection, feature engineering, hyperparameter tuning, transfer learning, and model deployment strategies. Each stage presents unique challenges and opportunities for enhancing model performance, with modern techniques offering solutions for improved accuracy, efficiency, and reliability. From addressing data quality issues through systematic preprocessing to implementing sophisticated deployment monitoring systems, the various aspects of model optimization work together to create robust and effective machine learning solutions that can be successfully deployed in real-world applications.
Keywords: MLOps deployment, feature engineering, hyperparameter tuning, model optimization, transfer learning
Cracking the Code: How Deep Learning unmasks Complex Fraud Schemes (Published)
In the fast-paced and high-stakes world of finance, the fight against fraud is a continuous and evolving challenge. Deep learning has emerged as a revolutionary tool, capable of processing vast amounts of data and predicting sophisticated fraud patterns with unprecedented accuracy. Unlike traditional rule-based systems, which remain static and predictable, deep learning models dynamically adapt to the ever-changing tactics employed by fraudsters, offering a level of detection that was previously unattainable. Our research delves into the use of advanced transformer models and pre-training techniques, which significantly enhance the precision and flexibility of fraud detection systems. However, implementing deep learning is not without its challenges, including issues related to data quality and the inherent complexity of these models, often referred to as their “black box” nature. Despite these challenges, the benefits are substantial: deep learning not only identifies elusive fraud schemes but also reduces the incidence of false positives, which can be costly and disruptive. Financial institutions are increasingly integrating deep learning with traditional detection methods to create a more robust and comprehensive defense against fraud. Advances in explainable AI are helping to demystify these complex models, making them more transparent and easier to understand. Additionally, transfer learning is enhancing the efficiency of these systems, allowing models trained on one task to be adapted for others with minimal data. This research underscores the critical role of deep learning in strengthening financial systems, providing a formidable barrier against fraud that evolves as quickly as the threats themselves. As financial institutions continue to adopt and refine these technologies, the potential for deep learning to transform fraud detection and prevention is immense. This makes deep learning an indispensable asset in the ongoing battle to protect financial integrity and security.
Keywords: deep learning, explainable AI, financial fraud detection, transfer learning, transformer models