Shah Md Tanvir Siddiquee

The advancement of Intelligent Information Processing, such as machine learning or deep learning models, increasingly depends on the ability to execute complex deep learning models on resource-constrained edge devices, because converting a trained model into a mobile or embedded app requires domain expertise, which is time-consuming and expensive. The deployment bottleneck exists across many applications, including medical diagnosis, industrial quality control, and agricultural disease detection. To solve this problem, we propose a domain-agnostic, configuration-driven framework that enables researchers to deploy their models to the field without deep programming or mobile app development expertise. The proposed framework provides three features: dynamic metadata mapping for runtime model initialization, automatic architecture adaptation for handling different input size and tensor types, and built-in performance benchmarking for comparing models across different heterogeneous hardware, which, unlike traditional workflows, where recompilation of the app is required for every model update, allows researchers to "drop-in" replace compatible models through file swaps, thus greatly accelerating the prototyping loop. We evaluate the proposed framework in multiple case studies. Field tests on different Android devices show stable performance, with inference times ranging from 45 to 120 ms, and reliable behavior under real-world conditions. By lowering the technical barrier to deployment and accelerating the feedback loop between lab and field, this framework democratizes AI deployment, enabling researchers across disciplines to deploy their models in real-world settings, learn from real-world use, and rapidly iterate to develop practical applications.

Text normalization methods such as Stemming and Lemmatization are key components of NLP pipelines. Due to the advancement of Artificial Intelligence, normalization methods for new Languages are being introduced, but current evaluation methodologies are limited and often rely on Compression Ratio (CR), downstream model accuracy, or Seq2Seq prediction scores, and typically use only one of these metrics in isolation, which fails to distinguish between helpful vocabulary reduction and harmful semantic distortion. This study proposes a data-driven evaluation framework comprising five components: Compression Ratio (CR), Model Performance Delta (MPD), Information Retention Score (IRS), Algorithm Effectiveness Score (AES), and Average Normalized Levenshtein Distance (ANLD). AES measures the limit to which normalization reduces lexical variability, MPD assesses its impact on the downstream model's performance, and ANLD serves as a safety indicator for potential meaning distortion. Applying this framework in a multilingual environment, evaluating multiple algorithms in both English and Bangla, shows two important findings. First, improvements in model performance often do not correlate with the quality of normalization algorithms, suggesting that downstream accuracy alone is unreliable for assessing normalization success. Second, aggressive algorithms are highly efficient at compression but significantly alter word meaning and formation. This framework offers NLP practitioners a standardized methodology for distinguishing between useful reduction and semantic distortion.

Acute Lymphoblastic Leukemia (ALL), a cancer affecting the blood and bone marrow, requires precise classification for accurate diagnosis, personalized treatment plans, and improved predictive assessments to enhance patient survival and quality of life. This study presents LEU3, a novel classification model designed to improve the accuracy of leukemia detection from peripheral blood smear (PBS) images. LEU3 leverages an attention-based convolutional neural network (CNN) architecture, incorporating pooling layers, a global average pooling layer, and dense layers with dropout for regularization. The model is trained with an Adam optimizer comprising with four classes: Benign, early malignant pre-B, malignant pre-B, and malignant pro-B. Data augmentation techniques were employed to increase training set diversity. Additionally, Shapley Additive Explanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) are used to enhance interpretability and transparency in the model’s decision-making process. LEU3 achieved a test accuracy of 99% and a validation accuracy of 99% on 484 PBS images, demonstrating a 3% improvement over the baseline model. These results underline the potential of LEU3 in supporting medical professionals by reducing diagnostic workload and improving the accuracy of leukemia classification.

Potato crops and their salability are influenced by potato pests in that both crop yield and quality are reduced. This in turn reduces the income for potato farmers due to lower prices for the crop, lower crop yield, trade restriction and reduced market access. Agricultural viability over the long run therefore depends on sustainable pest management. In order to efficiently detect potato pests, a dataset was constructed which contains eight prevalent potato species that were taken from several sources. Image pre-processing techniques were employed enhance image quality for compatibility with deep learning models. Among InceptionV3, VGG-16, and MobileNetV2 models, VGG-16 attained the highest accuracy of 94.44%, outperforming others. Inception-V3 achieved 58% accuracy, while MobileNetV2 reached 75%. Pre-processing has a major influence on improving result accuracy, which emphasizes its significance in enhancing model performance, according to an evaluation of its effects. These findings might lead to the development of pest management strategies for potato farming that are more effective. The efficient use of VGG-16 in potato pest identification systems is demonstrated by its excellent performance. Using deep learning models can therefore reduce financial losses and promote sustainable potato production. This study provides an approach for further investigation into the best ways to control pests in potato production, allowing farmers to overcome the obstacles and take advantage of valuable market prospects even in the face of pest threats.

The detection of coronavirus disease (COVID-19) using standard laboratory tests, such as reverse transcription polymerase chain reaction (RT-PCR), is time-consuming. Complex medical imaging problems are currently being solved using machine learning and deep learning techniques. Our proposed solution utilizes chest radiography imaging techniques, which have shown to be a faster alternative for detecting COVID-19. We present an efficient and lightweight deep learning architecture for identifying COVID-19 using chest X-ray images which achieve 99.81% accuracy in intra-database testing and 100% accuracy in cross-validation testing on a separate data set. The results demonstrate the potential of our proposed model as a reliable tool for COVID-19 diagnosis using chest X-ray images, which can have a significant impact on improving the efficiency of COVID-19 diagnosis and treatment.