{ "exported_at": "2026-11-15 22:54:54", "count": 51, "records": [ { "id": 25, "title": "Deep Learning-Based Breast Cancer Detection with IoMT", "title_ar": "", "title_en": "Deep Learning-Based Breast Cancer Detection with IoMT", "author": "Bo Li", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "University of Auckland", "university_source": "repository_host", "college": "", "department": "", "country": "New Zealand", "governorate": "", "publication_year": 2025, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "deep-learning; computer-vision; generative-ai; medical-ai; cnn; transformers; gan; medical-imaging", "keywords": "", "abstract": "Breast cancer remains a leading cause of cancer-related mortality among women worldwide, which demands innovations in early detection and diagnostic efficiency. Existing approaches often neglect peri-tumoral contextual features, rely on limited or unrealistic synthetic data, and impose latency bottlenecks in distributed healthcare settings. We propose a reserved-space strategy for Region of Interest (ROI) extraction that achieves 97.5% True Positive Rate (TPR) while preserving contextual features, thereby capturing diagnostically critical tumour margins that rigid bounding boxes overlook. To address data scarcity, we develop an enhanced Generative Adversarial Network (GAN) model incorporating Efficient Multi-scale Attention (EMA) modules and Hinge loss, which generates high-fidelity (15.2 Fréchet Inception Distance (FID)) ROI images that maintain anatomical plausibility in lesions and microcalcification diversity, resolving the over smoothing and pathological irrelevance of conventional GAN architectures. This GAN-augmented dataset improves classification AUC by 10% compared to traditional augmentation methods. Through systematic comparison, we demonstrate that the Swin Transformer outperforms the Convolutional Neural Network (CNN) and standard Vision Transformers (ViT) models in mammography classification with 86.8% Area Under the Curve (AUC). Finally, we design a distributed edge-fog-cloud Internet of Medical Things (IoMT) framework that implements our workflow to mitigate computational bottlenecks in clinical deployment.", "source_url": "https:\/\/openalex.org\/W7135345431", "pdf_url": "", "doi": "", "handle": "", "repository_name": "ResearchSpace (University of Auckland)", "source_record_id": "W7135345431", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:46", "updated_at": "2026-11-15 16:52:46" }, { "id": 31, "title": "Transformer-based Transfer Learning for Breast Cancer Detection via Digital Breast Tomosynthesis", "title_ar": "", "title_en": "Transformer-based Transfer Learning for Breast Cancer Detection via Digital Breast Tomosynthesis", "author": "Mikea Dimech", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "", "university_source": "", "college": "", "department": "", "country": "", "governorate": "", "publication_year": 2023, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; computer-vision; generative-ai; medical-ai; transfer-learning; cnn; transformers; medical-imaging", "keywords": "", "abstract": "Breast cancer has emerged as the most prevalent form of cancer worldwide and the leading cause of mortality from any cancer in women, emphasising the critical need for its early detection. Traditional 2D mammography has long been the cornerstone of breast cancer screening, given its proven efficacy in early prognosis and reduction of mortality. However, the limitations of this modality have precipitated the development and adoption of more advanced imaging techniques, such as Digital Breast Tomosynthesis (DBT), which yields a 3D view of the breast. The application of deep learning to classify these scans is paramount to increasing the efficiency and efficacy of screening. However, this task presents arduous challenges such as high computational demands, scarcity of diverse data, and generalisation across varying demographics and imaging equipment. Transfer learning has proven to be effective in overcoming these challenges by enabling models to leverage pre-existing knowledge and adapt to new tasks. However, its application to DBT has been confined to Convolutional Neural Network (CNN) architectures. Meanwhile, the Transformer paradigm, which has revolutionised various deep learning domains and shown promise in traditional mammography, has yet to be explored in the context of DBT via transfer learning, highlighting a notable research gap. This study thus investigates the performance of a pre-trained Transformer model, fine-tuned on a public DBT dataset, comparing it with a fine-tuned state-of-the-art CNN. While the Transformer outperforms the convolutional architecture, the results reveal severe challenges posed by the imbalanced dataset, with models developing biases towards the majority class. Data augmentation, while beneficial, could not fully address the issue, and the need for more diverse datasets became clear. Moreover, the results underscore the importance of preserving the high resolution of DBT images, with fine-grained details, such as subtle changes in breast tissue captured at higher resolutions, proving to be pivotal for accurate diagnosis. Finally, a comparative analysis between CNNs and transformer-based models revealed that Transformers excel in recognising broader patterns across the entire image, which aids to mitigate overfitting. By using pre-trained transformers and fine-tuning them on this dataset, their effectiveness was highlighted, suggesting they can match or even surpass CNNs in breast cancer detection via DBT.", "source_url": "https:\/\/openalex.org\/W4385850049", "pdf_url": "", "doi": "10.5281\/zenodo.8243430", "handle": "", "repository_name": "Zenodo (CERN European Organization for Nuclear Research)", "source_record_id": "W4385850049", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:46", "updated_at": "2026-11-15 16:52:46" }, { "id": 68, "title": "Efficient and Generalised Deep Learning Models for Medical Image Segmentation", "title_ar": "", "title_en": "Efficient and Generalised Deep Learning Models for Medical Image Segmentation", "author": "Zhendi Gong", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "University of Nottingham", "university_source": "repository_host", "college": "", "department": "", "country": "United Kingdom", "governorate": "", "publication_year": 2026, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; computer-vision; medical-ai; ai-applications; few-zero-shot; self-semi-supervised; cnn; transformers; image-segmentation; medical-imaging; ai-industry", "keywords": "", "abstract": "Medical image segmentation is a critical task in computer-assisted diagnosis and therapy planning. Accurate and efficient fully automated segmentation methods are essential for accelerating decision-making processes. Compared with natural image segmentation methods, current deep learning–based segmentation models for medical imaging face challenges related to data scarcity, inconsistent annotations, and high computational cost. Consequently, this thesis aims to develop efficiency and generalised medical image segmentation models, which are fundamental to the clinical adoption and industrial deployment of this technology in healthcare settings. This thesis systematically addresses these challenges through a series of progressively advanced deep learning models. The foundational work, CTranS, forms a baseline medical image segmentation model that combines a convolutional neural network (CNN) based encoder with a Transformer-based decoder. This initial contribution leverages the strengths of CNNs for local feature extraction and Transformers for global contextual understanding, achieving state-of-the-art (SOTA) performance. Subsequently, to enhance the generalisability, MO-CTranS is introduced, which is capable of learning from a mixture of multiple partially labelled datasets. MO-CTranS employs a CNN encoder and Transformer decoder with task-specific tokens to handle label inconsistencies and class imbalances, tackling the challenges of training a single unified model from fragmented and heterogeneously annotated medical data. This innovation moves beyond single-task, fully-annotated model training towards a more scalable and clinically realistic approach. Despite the superior performance of CTranS and MO-CTranS, the computational efficiency of the Transformer blocks hinders their clinical application. Thus, CRFTrans is proposed to enhance the model efficiency. As a novel recursive Vision Transformer layer designed to replace the commonly used cascaded Transformer layers, this approach represents a fundamental shift from simply stacking layers to a recursive refinement process, demonstrating that comparable performance can be achieved with significantly reduced computational complexity, thus facilitating deployment in resource-constrained clinical environments. However, these methods still depend on extensive manual annotation, which remains challenging and time-consuming in clinical practice. Thus, to further enhance both efficiency and generality, a novel semi-supervised medical image segmentation framework, referred to as SSL-MedSAM2, has been developed. SSL-MedSAM2 integrates a training-free few-shot segmentation branch for pseudo labelling and a fully supervised learning branch for label refinement. This design allows the model to leverage a small fraction of labelled images together with a large unlabelled set, dramatically decreasing the annotation effort. Together, these contributions result in deep learning models that are significantly more data-efficient and broadly applicable than conventional approaches. The overall impact of this work is to facilitate the reliable and scalable deployment of segmentation models in real-world clinical settings by bridging the gap among accuracy, efficiency, and generalisability.", "source_url": "https:\/\/openalex.org\/W7140609462", "pdf_url": "https:\/\/repository.nottingham.ac.uk\/bitstreams\/1a58d983-6aaa-4bbb-a40b-688e6d3af90c\/download", "doi": "10.17639\/7969", "handle": "", "repository_name": "Nottingham ePrints (University of Nottingham)", "source_record_id": "W7140609462", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:46", "updated_at": "2026-11-15 16:52:46" }, { "id": 70, "title": "GENERALIZING MEDICAL IMAGE SEGMENTATION TASK WITH EFFICIENT DEEP LEARNING MODELS", "title_ar": "", "title_en": "GENERALIZING MEDICAL IMAGE SEGMENTATION TASK WITH EFFICIENT DEEP LEARNING MODELS", "author": "Abel A. Reyes-Angulo", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Michigan Technological University", "university_source": "repository_host", "college": "", "department": "", "country": "United States", "governorate": "", "publication_year": 2025, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; computer-vision; medical-ai; predictive-analytics; cnn; transformers; image-segmentation; medical-imaging", "keywords": "", "abstract": "Medical Image Segmentation is a critical task in the field of medical imaging, playing a crucial role in diagnostics, treatment planning, and disease monitoring. The emergence of Deep Learning (DL) has ushered in a new era in Artificial Intelligence (AI), propelling remarkable advancements in key domains like language translation, object recognition, and recommendation systems. This evolution has been accompanied by continuous enhancements in computational efficiency and improvements in predictive accuracy. The introduction of sophisticated algorithms, such as convolutional neural networks (CNNs) and transformers, exemplifies these advancements. DL algorithms have demonstrated exceptional efficacy in medical image segmentation tasks, showcasing the potential for AI-driven early diagnostics. However, the deployment of AI systems in clinical environments is often hindered by the substantial computational demands and complexity of cutting-edge DL models. In this research proposal, we explore various methodologies to enrich the visual feature representation for medical images. We focus on integrating global context-oriented techniques, such as attention mechanisms, into the development of parameter-efficient deep learning models. Our goal is to create a generalized, end-to-end medical image segmentation framework that can accurately and efficiently segment medical images across different modalities and conditions. By leveraging advanced deep learning techniques and optimizing model architectures, we aim to enhance the performance and generalization capabilities of medical image segmentation models, ultimately contributing to improved clinical outcomes", "source_url": "https:\/\/openalex.org\/W4410497124", "pdf_url": "", "doi": "10.37099\/mtu.dc.etdr\/1941", "handle": "", "repository_name": "Digital Commons - Michigan Tech (Michigan Technological University)", "source_record_id": "W4410497124", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:46", "updated_at": "2026-11-15 16:52:46" }, { "id": 79, "title": "Semantic segmentation of 3D medical images with deep learning", "title_ar": "", "title_en": "", "author": "Olivier Petit", "supervisor": "", "degree_type": "phd", "degree_evidence": "theses.fr: الفهرس الوطني الفرنسي لأطاريح الدكتوراه", "university": "", "university_source": "", "college": "", "department": "", "country": "France", "governorate": "", "publication_year": 2021, "language": "fr", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "deep-learning; computer-vision; medical-ai; transformers; image-segmentation; medical-imaging", "keywords": "", "abstract": "Segmentation sémantique d'images médicales 3D par deep learning L’apprentissage profond a récemment montré des résultats impressionnants en vision par ordinateur. En particulier avec les réseaux de neurones convolutifs (ConvNets) qui ont redéfini l’état-de-l’art dans de nombreuses applications telles que la segmentation d’images médicales. Dans cette thèse nous abordons des problèmes en segmentation d’organes de l’abdomen en utilisant ces modèles. Premièrement, nous nous sommes intéressés à l’entrainement de ConvNets avec des bases de données partiellement étiquetées. Les professionnels se concentrant, souvent sur des régions anatomiques précises, les bases de données sont de ce fait hétérogènes et partiellement étiquetées. Entraîner un modèle de segmentation directement donne de tr`es mauvais résultats. Nous proposons donc un schéma d’entrainement qui utilise toutes les étiquettes disponibles sans être affect´e par les mauvaises. De plus, un schéma itératif permet de progressivement ré-étiqueter les organes manquants ce qui permet d’améliorer encore notre modèle. La seconde partie étudie l’utilisation d’un a priori spatial sur la position absolue des organes afin d’améliorer la détection des structures et réduire les erreurs aberrantes. Les ConvNets sont par construction incapables de capturer l’information de position spatiale absolue. Cependant, les images médicales sont très structurées et les positions des organes sont connues. Dans ces travaux nous proposons d’utiliser un a priori spatial 3Dqui capture la position des organes et qui va explicitement biaiser le modèle grâce à une fonction d’activation «prior-driven ». Pour finir, nous étudions les Transformers qui permettent de mod´eliser des interactions `a long terme entre les structures anatomiques dans un modèle de segmentation. Les ConvNets ne permettent pas de capturer ces interactions globales principalement `a cause de leur champ réceptif limité. Utiliser le mécanisme d’attention propose dans les Transformers permet de connecter tous les pixels entre eux, ayant pour effet de modéliser des interactions complexes. Nous proposons le modèle U-Transformer et montrons qu’il améliore la qualité de la segmentation sur plusieurs bases de données.", "source_url": "https:\/\/openalex.org\/W4392220284", "pdf_url": "", "doi": "10.70675\/0c16f3d3z8e39z4f4az97f2ze8a6f2043cc3", "handle": "", "repository_name": "theses.fr (ABES)", "source_record_id": "W4392220284", "other_sources": "", "verification_status": "verified", "created_at": "2026-11-15 16:52:46", "updated_at": "2026-11-15 16:52:46" }, { "id": 104, "title": "Applications of artificial intelligence for explainability and uncertainty quantification for performance-based design and damage classification in civil engineering", "title_ar": "", "title_en": "Applications of artificial intelligence for explainability and uncertainty quantification for performance-based design and damage classification in civil engineering", "author": "Shabbir, Khurram", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Aalto University", "university_source": "repository_host", "college": "", "department": "", "country": "Finland", "governorate": "", "publication_year": 2025, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; computer-vision; medical-ai; responsible-ai; ai-applications; predictive-analytics; transformers; xai; fairness; ai-smart-cities", "keywords": "", "abstract": "Structural resilience and safety are critical concerns in modern infrastructure, particularly in the face of increasing seismic risks and aging buildings. While advanced artificial intelligence (AI) techniques offer promising solutions for structural design, damage assessment, and retrofitting, the \"black-box\" nature of many AI models limits their trustworthiness and adoption in safety-critical fields. This research integrates explainable AI (XAI) techniques to bridge the gap between complex machine learning models and engineering practice, developing innovative frameworks for enhanced interpretability in civil engineering applications. A primary contribution is Seismo-XAI, a web-based tool that delivers transparent, interpretable insights into building performance under seismic loads, empowering engineers to optimize designs and retrofits. The study also introduces a novel hybrid framework combining microstructure-informed finite element modelling with computer vision to advance damage classification in concrete, synergizing vision transformers' computational efficiency with physics-based finite element analysis. To address data limitations, a synthetic dataset of concrete microstructures was generated using stochastic Monte Carlo augmentation. For uncertainty quantification in seismic assessment, this research employs artificial neural networks with the quality-driven lower upper bound estimation method to establish prediction intervals for long-term ground motion effects. This distribution-free machine learning approach provides both local and global levels probabilistic damage assessment, enhancing reliability in post-earthquake evaluations and early warning systems. Validation through benchmark structures and real-world case studies demonstrates the frameworks’ effectiveness in promoting reliable design practices for seismic resilience and damage assessment. The findings highlight the potential of XAI to transform structural engineering by providing transparent, interpretable, and actionable insights, fostering trust in AI-driven solutions. This research lays the foundation for future advancements in and development of AI applications for structural health monitoring and disaster management for urban infrastructure.", "source_url": "https:\/\/openalex.org\/W7115275889", "pdf_url": "https:\/\/aaltodoc.aalto.fi\/bitstreams\/22ee9f47-ed26-415f-8ef3-c22a3082d694\/download", "doi": "", "handle": "", "repository_name": "Aaltodoc (Aalto University)", "source_record_id": "W7115275889", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:46", "updated_at": "2026-11-15 16:52:46" }, { "id": 117, "title": "Interpretable Measures for Arabic Readability", "title_ar": "", "title_en": "Interpretable Measures for Arabic Readability", "author": "Rabih, Nour", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "", "university_source": "", "college": "", "department": "", "country": "United Arab Emirates", "governorate": "", "publication_year": 2025, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; nlp; responsible-ai; expert-systems; ai-applications; predictive-analytics; transformers; arabic-nlp; xai; ai-education", "keywords": "", "abstract": "Readability assessment is a crucial task in natural language processing (NLP), with wide reaching implications in education, accessibility, and language learning. While prior research has primarily focused on predicting readability levels, this work shifts the focus toward understanding why a text is assigned a particular level. By identifying key linguistic features that drive readability, the goal is to enhance interpretability in Arabic text complexity assessment. Arabic poses unique challenges for automatic readability evaluation due to its morphological richness, variation across dialects, and limited availability of annotated corpora. The complex structure of Arabic words, and extensive derivational morphology add layers of difficulty in determining textual complexity. These challenges necessitate a more comprehensive approach to readability evaluation beyond traditional classification. To address these limitations, this work leverages the Balanced Arabic Readability Evaluation Corpus (BAREC)—a large-scale, manually annotated dataset of 68,182 sentences across 19 readability levels, from kindergarten to postgraduate comprehension. BAREC is designed to ensure genre diversity, topical variety, and balanced target audiences, making it a strong foundation for Arabic readability modeling. This study systematically engineers and evaluates a comprehensive set of linguistic features from BAREC’s annotation guidelines, including morphology, syntax, vocabulary, syllables, and content and vocabulary-based measures. Through extensive experimentation, it investigates the effectiveness of these features in explaining and predicting readability levels across three task configurations: foundational levels (1–11), advanced levels (12–19), and the full 19-level scale. In this study, we explore a variety of modeling approaches, including traditional machine learning classifiers, a fully interpretable rule-based model, and an AraBERT model based on transformers. Furthermore, a hybrid model is proposed that integrates linguistic features with contextual embeddings to improve both performance and transparency. The results demonstrate that linguistic features significantly enhance interpretability, especially in low-resource settings, and that feature-driven models can provide human readable justifications aligned with expert annotation practices. This research not only advances Arabic readability prediction but also contributes interpretable insights into text complexity, paving the way for more accessible and pedagogically meaningful NLP tools in Arabic education.", "source_url": "https:\/\/openalex.org\/W7106299741", "pdf_url": "", "doi": "10.82419\/276", "handle": "", "repository_name": "MBZUAI iRep", "source_record_id": "W7106299741", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:47", "updated_at": "2026-11-15 16:52:47" }, { "id": 201, "title": "Scaling graph neural networks to larger graphs", "title_ar": "", "title_en": "Scaling graph neural networks to larger graphs", "author": "Zhang, William", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Harvard University", "university_source": "repository_host", "college": "", "department": "", "country": "United States", "governorate": "", "publication_year": 2022, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; data-mining; predictive-analytics; transformers; gnn", "keywords": "", "abstract": "Graph-structured data appears abundantly in both the social and natural sciences. How- ever, the common supervised-learning techniques in machine learning are not readily ap- plicable to graph data because they require their inputs to be structured as feature vectors. The current paradigm of learning on graph data is using graph neural networks (GNN). Computation on GNN’s follow the pattern of message-passing of vectors between neigh- boring nodes and message-aggregation. This thesis first motivates and then provides an overview of GNNs. It then discusses the connections between GNNs and Transformers, another recent machine learning model that has taken over the state-of-the-art in many ap- plications. We exploit this connection to develop a procedure that trades off model perfor- mance for training time and execution speed. By clustering nodes together, we can obtain approximations of the representations of each node, and we can leverage this approxima- tion to train and predict with the models much faster. We establish that our method. Addi- tionally, we open source our code for reproducibility of the experiment.", "source_url": "https:\/\/openalex.org\/W7112491723", "pdf_url": "", "doi": "", "handle": "", "repository_name": "Digital Access to Scholarship at Harvard (DASH) (Harvard University)", "source_record_id": "W7112491723", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:47", "updated_at": "2026-11-15 16:52:47" }, { "id": 231, "title": "Machine Learning for Cooperative Logistics: Multi-Agent Reinforcement Learning to Negotiate Fair Outcomes", "title_ar": "", "title_en": "Machine Learning for Cooperative Logistics: Multi-Agent Reinforcement Learning to Negotiate Fair Outcomes", "author": "Stephen Mak", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "University of Cambridge", "university_source": "repository_host", "college": "", "department": "", "country": "United Kingdom", "governorate": "", "publication_year": 2024, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; nlp; reinforcement-learning; transformers", "keywords": "", "abstract": "Effectively managing businesses is perhaps the ultimate testbed for AI. It requires creativity. It requires reasoning. It requires intelligence. Logistics forms the backbone of our economy. From the food we eat, to the computers we work on, to the planes we fly on – logistics has played a central role. Yet, profit margins for logistics providers are thin, and customers have ever-increasing expectations. Worse, humanity is not on track to decarbonise road freight transportation. One effective method to alleviate this pressure is Cooperative Logistics. Here, companies cooperate by pooling their resources together (such as trucks, drivers, and\/or warehouses) to improve their individual performance (such as profitability and carbon emissions). But if companies cooperate, how much should everyone be fairly paid? This is known as fair gain sharing. Prior methods do not scale because they require evaluating O(2n) NP-hard vehicle routing problems (VRPs) for n companies – limiting practical cooperation to ~20 companies – forgoing significant economies of scale. Scaling existing methods to 1,000 companies would therefore require solving 1 × 10^301 NP-hard VRPs. We propose MAGS (Multi-Agent Gain Sharing) to enable fair gain sharing at scale. MAGS effectively handles over 1,000 agents using a principled, game-theoretic approach. This is achieved by training deep multi-agent reinforcement learning agents to negotiate fair gain shares. This is challenging due to the n-player, mixed-motive environment with agents learning concurrently. Crucially, we eliminate reliance on the characteristic function. We additionally leverage transformers to design a reward model for coalition structure generation. We show that in a real-world setting with 1,000 carriers, we can potentially reduce carbon emissions by 22.1%1, whilst improving average profitability from £66.71 per day to £259.26 per day. This is particularly important as global freight activity is expected to double by 2050. Furthermore, this thesis lays the principled foundations for Agentic AI and Human-AI natural language negotiation in future work. This thesis therefore contributes to both the artificial intelligence (AI) and supply chain management (SCM) literature.", "source_url": "https:\/\/openalex.org\/W7147016666", "pdf_url": "https:\/\/www.repository.cam.ac.uk\/bitstreams\/898c5761-d491-476e-90be-19e4e014764a\/download", "doi": "10.17863\/cam.128509", "handle": "", "repository_name": "Apollo (University of Cambridge)", "source_record_id": "W7147016666", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:47", "updated_at": "2026-11-15 16:52:47" }, { "id": 256, "title": "Comparative Performance Analysis of Deep Learning Models for Brain Tumor Classification from MRI Images", "title_ar": "", "title_en": "Comparative Performance Analysis of Deep Learning Models for Brain Tumor Classification from MRI Images", "author": "Hassan Khan", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "University of Eastern Finland", "university_source": "repository_host", "college": "", "department": "", "country": "Finland", "governorate": "", "publication_year": 2026, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; computer-vision; medical-ai; transfer-learning; cnn; transformers; medical-imaging", "keywords": "", "abstract": "Brain tumors are among the most serious categories of neurological disease, with accurate classification from magnetic resonance imaging (MRI) playing a critical role in treatment planning and patient outcomes.Manual interpretation of MRI scans by radiologists is timeconsuming, subject to inter-observer variability, and constrained by the availability of specialist expertise, particularly in resource-limited settings.These factors have driven growing interest in automated deep learning solutions for brain tumor MRI classification.While convolutional neural networks (CNNs) have demonstrated strong performance on this task through transfer learning, the emergence of Vision Transformers and hybrid CNN-Transformer architectures has created a diverse model landscape whose relative merits remain poorly understood.Existing comparative studies benchmark either multiple CNN variants in isolation or transformer-based models separately, without unified head-tohead evaluation across all three architectural families under identical experimental conditions.Furthermore, deployment-relevant efficiency metrics -inference latency, model size, and GPU memory -are rarely reported alongside accuracy, limiting practical model selection guidance for clinical settings with varying hardware constraints.This thesis addresses these gaps by conducting a systematic benchmark of nine deep learning architectures -four CNN-based (VGG16, ResNet50, DenseNet121, EfficientNet-B0), three Transformer-based (ViT-B\/16, Swin Transformer, DeiT-B), and two Hybrid\/Lightweight (MobileNetV2, MaxViT-Tiny) -evaluated on the Brain Tumor MRI Dataset comprising 7,200 T1-weighted MRI images across four classes (glioma, meningioma, pituitary tumor, no tumor).All models were trained under identical conditions using ImageNet transfer learning and assessed on both classification performance (accuracy, precision, recall, F1-score, AUC-ROC) and computational efficiency (parameter count, model size, inference latency, peak GPU memory).The Swin Transformer achieves the highest test accuracy (98.52%,F1 = 0.9852), demonstrating that hierarchical transformer architectures can outperform CNNs when appropriately designed for the task.MobileNetV2 leads all models on efficiency metrics with 2.2M parameters and 7.55 ms inference latency.Pareto analysis identifies three deployment-optimal models: Swin-T for high-resource GPU environments, DenseNet121 (98.15%, 7.0M parameters) for standard workstations, and MobileNetV2 (97.04%) for edge devices.Efficiency varies by a factor of 61× in parameter count and 25× in GPU memory across the nine models, while accuracy spans only ~2 percentage points, underscoring that architectural design matters far more than raw model size.This study provides the first unified benchmark comparing CNN, Transformer, and Hybrid architectures for brain tumor MRI classification under identical conditions with both accuracy and efficiency metrics reported, yielding evidence-based model selection guidelines for clinical deployment across diverse computational environments.", "source_url": "https:\/\/openalex.org\/W7215551487", "pdf_url": "https:\/\/erepo.uef.fi\/bitstreams\/57ebe372-c179-491f-bae8-4ad920908485\/download", "doi": "", "handle": "", "repository_name": "UEF eRepo (University of Eastern Finland)", "source_record_id": "W7215551487", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:48", "updated_at": "2026-11-15 16:52:48" }, { "id": 306, "title": "Optimization of video surveillance networks in smart cities using video compressive sensing and deep learning techniques", "title_ar": "", "title_en": "", "author": "Wael Saideni", "supervisor": "", "degree_type": "phd", "degree_evidence": "theses.fr: الفهرس الوطني الفرنسي لأطاريح الدكتوراه", "university": "", "university_source": "", "college": "", "department": "", "country": "France", "governorate": "", "publication_year": 2022, "language": "fr", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "deep-learning; computer-vision; ai-applications; transformers; video-analysis; ai-smart-cities", "keywords": "", "abstract": "Optimisation d'un système de video surveillance dans les villes intelligentes en utilisant le Deep Learning et la video compressive sensing La technique de compressive sensing joue un rôle important dans le traitement des données vu que l’acquisition et la compression se font simultanément grâce à un processus de prise de mesures. Cette technique optimise les capacités de stockage des systèmes ainsi que la vitesse et le coût d’acquisition. Récemment, cette technique est devenue de plus en plus utilisée grâce à l’optimisation des algorithmes de reconstruction en utilisant les architectures du Deep Learning. L’objectif principal de cette thèse est de tirer profit des architectures de Deep Learning pour optimiser la technique de compressive sensing en l’appliquant sur des signaux vidéo et par la suite optimiser l’acquisition, la transmission et la reconstruction des vidéos dans les systèmes numériques modernes. Ainsi, la stratégie adoptée au cours de ces travaux de recherche consiste à commencer par établir une étude comparative sur les approches de vidéo compressive sensing (VCS) basées sur le Deep Learning en évaluant la qualité et la vitesse de reconstruction ainsi que les différentes architectures. Puis, deux environnements de VCS ont été conçus : le premier se base sur la prédiction des frames vidéo en implémentant une approche basée sur un nouveau réseau récurrent et le deuxième exploite les dernières performances réalisées avec les Transformers et le mécanisme d’attention. Alors, la démarche adoptée dans ces deux approches repose sur une analyse de l’état de l’art suivie d’une explication de chaque architecture et une validation expérimentale. Les différentes contraintes rencontrées au cours de ces travaux sont discutées et des solutions appropriées sont proposées.", "source_url": "https:\/\/openalex.org\/W7151367475", "pdf_url": "https:\/\/theses.hal.science\/tel-03850837\/document", "doi": "10.70675\/ad16c2f9zf249z4f36z80ddz17f162cfe1a5", "handle": "", "repository_name": "theses.fr (ABES)", "source_record_id": "W7151367475", "other_sources": "", "verification_status": "verified", "created_at": "2026-11-15 16:52:48", "updated_at": "2026-11-15 16:52:48" }, { "id": 375, "title": "Leveraging Self-Attention in Transformer Models for Improved Retail Demand Forecasting", "title_ar": "", "title_en": "Leveraging Self-Attention in Transformer Models for Improved Retail Demand Forecasting", "author": "Rohith Sardena", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "National College of Ireland", "university_source": "repository_host", "college": "", "department": "", "country": "Ireland", "governorate": "", "publication_year": 2025, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; predictive-analytics; rnn; transformers", "keywords": "", "abstract": "Inventory management and operation planning in retail sector involve accurate demand forecasting as a key element of the process. Linear models and tree-based models are typical examples of traditional forecasting models that may not work well in capturing complex, non-linear, and time-evolving demands that retail demand exhibits especially when it is affected by promotions, seasonal trends, and other external market factors. The study examines the possibility of deep learning models based on Transformers, which will use self-attention mechanisms, to improve the accuracy and scalability of retail demand forecasting systems. Model development was done using a detailed retail dataset with historical sales, inventory, pricing, discounts, promotions, weather conditions, and other contextual characteristics. A number of models were applied and tested in the study: Linear Regression, Decision Tree, Random Forest, Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Transformer. Mean Absolute Error (MAE), Root Mean Squared Error (RMSE) and Median Absolute Error (MedianAE) were used to evaluate performance. The classical models were moderate in their performance, and Linear Regression and Random Forest injected MAEs of 7.47 and 7.55 respectively, and RMSEs of more than 8.6. Deep learning models, on the contrary, showed a dramatic increase. Transformer model performed the best and its MAE was 0.81, RMSE was 0.99, and MedianAE was 0.74, which are better than LSTM (MAE: 0.83) and GRU (MAE: 0.83). The results verify that self-attention mechanism of Transformer is capable of learning long-range dependencies and complex temporal interactions, giving significant improvement on the baseline models. This study shows the ability of Transformers to be an efficient and scalable method of real-time retail demand forecasting. Further development will consider multi-step forecasting, probabilistic additions and deployment in other retail areas.", "source_url": "https:\/\/openalex.org\/W7204618203", "pdf_url": "https:\/\/norma.ncirl.ie\/9664\/2\/rohithsardenaconfigurationmanual.pdf", "doi": "", "handle": "", "repository_name": "TRAP@NCI (National College of Ireland)", "source_record_id": "W7204618203", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:49", "updated_at": "2026-11-15 16:52:49" }, { "id": 376, "title": "Understanding Human-Model Alignment in Visual Social Attention with Video Transformers", "title_ar": "", "title_en": "Understanding Human-Model Alignment in Visual Social Attention with Video Transformers", "author": "Sabrina M. Lopez", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "", "university_source": "", "college": "", "department": "", "country": "United States", "governorate": "", "publication_year": 2026, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; computer-vision; responsible-ai; predictive-analytics; cnn; transformers; video-analysis; fairness", "keywords": "", "abstract": "Artificial Neural Networks (ANNs) are among the leading computational frameworks for approximating human visual processing across a range of visual tasks. Despite this, a notable gap remains between human and ANN performance in tasks that require social perception and cognition such as intent prediction and social interaction recognition. Among ANN architectures, vision transformers have recently emerged as particularly promising approximations of human visual processing as they exhibit a stronger shape bias and produce errors more consistent with those of human observers, and develop more globally uniform internal representations. This paper examines how transformers align with human perception when identifying social interactions in videos, in which two agents help, hinder, or have neutral interactions with each other in a 2D simulated environment. With the use of a dataset containing videos paired with human gaze recordings captured during the classification of helping and hindering social interactions, we assess the degree of human-model alignment through both classification accuracy and visual attention metrics across different models (i.e., TimeSformer, ViViT, V-JEPA 2), different pre-training datasets (e.g., Something Something V2, Kinetics 400), and different visualization methods (i.e., different GradCam methods and integrated gradients). Our findings suggest TimeSformer and ViViT models have strong potential as cognitive models of social interaction, with attention patterns — both static and dynamic — that align meaningfully with those of human participants. Meanwhile, despite their higher classification accuracy, V-JEPA 2 models achieve saliency alignment at or below that of convolutional neural network baselines. Notably, classification performance is not a reliable predictor of human-model attention alignment. Regarding pretraining datasets, larger and more diverse datasets tend to improve saliency alignment for classification-supervised models, though the optimal visualization method for extracting alignment varies by architecture.", "source_url": "https:\/\/openalex.org\/W7155351216", "pdf_url": "", "doi": "10.18130\/08r1-yn85", "handle": "", "repository_name": "Libra", "source_record_id": "W7155351216", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:49", "updated_at": "2026-11-15 16:52:49" }, { "id": 377, "title": "Hybrid Deep Learning Model: LLM-Enhanced Linear Self-Attention Transformer-CNN for Stock Price Prediction", "title_ar": "", "title_en": "Hybrid Deep Learning Model: LLM-Enhanced Linear Self-Attention Transformer-CNN for Stock Price Prediction", "author": "Yilong Chen", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Auckland University of Technology", "university_source": "repository_host", "college": "", "department": "", "country": "New Zealand", "governorate": "", "publication_year": 2025, "language": "en", "research_domain": "nlp", "research_subdomain": "llm", "categories": "machine-learning; deep-learning; computer-vision; nlp; predictive-analytics; cnn; transformers; multimodal; llm", "keywords": "", "abstract": "The unpredictability and inherent complexities of financial markets pose persistent challenges in accurately forecasting stock prices. Recent innovations in deep learning methodologies, specifically the integration of technical analysis with linguistic processing through large language models (LLMs), have demonstrated significant potential in financial forecasting applications. Despite substantial advances in deep learning for stock price prediction, current approaches face critical limitations in effectively integrating multi-modal data while lacking sophisticated technical analysis capabilities, particularly when simultaneously modeling both temporal dependencies in financial time series and spatial features from stock chart visualizations across diverse market regimes and volatility conditions. This study proposes an innovative hybrid deep learning framework aimed at addressing these multi-modal integration challenges, technical analysis limitations, and cross-market adaptability issues through a structured three-phase methodology. In the initial phase, we leverage an LLM to conduct thorough technical analysis of historical financial data, transforming raw market information into rich textual representations. These representations encapsulate intricate financial patterns using well-established technical indicators such as Bollinger Bands (BBs). By interpreting these financial metrics, the LLM generates structured textual summaries that enhance the downstream predictive modeling process. The second phase involves dual processing pipelines. One pipeline utilizes an optimized attention-based Transformer to process the LLM-generated embeddings, capturing long-range market dependencies. Concurrently, the second pipeline deploys a Convolutional Neural Network (CNN) to analyze visual stock chart patterns, extracting spatially relevant features that complement the temporal insights derived from the Transformer model. In the concluding phase, the outputs from both pipelines—textual, sequential, and visual data—are aggregated to generate holistic stock price predictions. The architecture employs Transformers for extracting extended temporal relationships within market sequential information, whereas Convolutional Networks detect regional spatial configurations from graphical stock representations. Large language models enhance our dataset by generating sophisticated technical metrics, subsequently embedded into numerical vectors through FinBERT encoding techniques. The fusion of these multimodal representations enables a more robust and accurate forecasting model, capitalizing on the complementary strengths of each component to enhance predictive performance.", "source_url": "https:\/\/openalex.org\/W7131750035", "pdf_url": "", "doi": "", "handle": "", "repository_name": "Tuwhera (Auckland University of Technology)", "source_record_id": "W7131750035", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:49", "updated_at": "2026-11-15 16:52:49" }, { "id": 382, "title": "Attention Mechanisms in Transformers: A Comparative Survey and Structural Enhancements to Linear Attention", "title_ar": "", "title_en": "Attention Mechanisms in Transformers: A Comparative Survey and Structural Enhancements to Linear Attention", "author": "Nizar Bennani", "supervisor": "", "degree_type": "master", "degree_evidence": "نص السجل: «master»", "university": "", "university_source": "", "college": "", "department": "", "country": "", "governorate": "", "publication_year": 2025, "language": "en", "research_domain": "nlp", "research_subdomain": "llm", "categories": "deep-learning; nlp; transformers; llm", "keywords": "", "abstract": "This master’s thesis explores the trade-off between computational efficiency and modeling performance in Transformer architectures by studying and improving attention mechanisms. The work consists of two parts: (1) a comparative empirical study evaluating five attention variants—standard Multi-Head Attention, FlashAttention, Sparse Attention, Sliding-Window Attention, and Linear Attention—across multiple Transformer model families, including encoder-only (BERT-style), decoder-only (GPT-style), and full encoder–decoder architectures. This broader evaluation highlights how each mechanism behaves under different structural constraints and reveals the consistent underperformance of linear attention in decoder-only setups due to weaker long-range modeling. (2) A technical contribution proposing two new hybrid mechanisms, Linear SparseAttention and Linear Sliding-Window Attention, which enhance the expressiveness of linear attention while preserving its linear-time complexity. Experiments show that both hybrids significantly outperform standard linear attention and narrow the performance gap with full attention, offering a promising path toward efficient and scalable Transformer models deployable in resource-constrained settings.", "source_url": "https:\/\/openalex.org\/W7114772223", "pdf_url": "", "doi": "10.5281\/zenodo.17879303", "handle": "", "repository_name": "Zenodo (CERN European Organization for Nuclear Research)", "source_record_id": "W7114772223", "other_sources": "", "verification_status": "verified", "created_at": "2026-11-15 16:52:49", "updated_at": "2026-11-15 16:52:49" }, { "id": 383, "title": "Transformers and other Attention-Based Algorithms for Biomedicine", "title_ar": "", "title_en": "Transformers and other Attention-Based Algorithms for Biomedicine", "author": "Marco Hernández Pérez", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "University of Salamanca", "university_source": "repository_host", "college": "", "department": "", "country": "Spain", "governorate": "", "publication_year": 2025, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; computer-vision; medical-ai; self-semi-supervised; predictive-analytics; rnn; transformers; image-segmentation; medical-imaging", "keywords": "", "abstract": "[EN] This research explores the application of Transformer-based models and Hybrid Attention Mechanisms across three critical biomedical tasks: meningioma segmentation, epileptic seizure detection, and pathogenicity prediction of genomic variants. For meningioma segmentation, this research investigates refining skip connections in a U-Net architecture by incorporating Swin Transformers for Magnetic Resonance Imaging. The proposed SwinC U-Net model, evaluated on the Brain Tumor Segmentation Meningioma 2023 challenge dataset, achieved a Dice score of 0.8933±0.0016, precision of 0.9020±0.0041, and recall of 0.8939±0.0044, surpassing U-Net and Attention U-Net in precision and accuracy, while slightly underperforming in recall compared to U-Net. In epileptic seizure detection, this work addresses the challenge of real-time monitoring using microelectrode array data from rat brain slices with 4-aminopyridine-induced epileptic activity. Lightweight neural models were trained to detect seizures. The best performing model, GRU+Attention, achieved an event-level F1 score of 0.839, Jaccard index of 0.722, recall of 0.813, and precision of 0.867 on a held-out test set. The model was quantized to INT8, maintaining high event-level detection performance (F1 = 0.812, Jaccard = 0.709) and demonstrating real-time feasibility on embedded platforms like the Raspberry Pi 5 (latency: 2.61 ms) and Coral Dev Board (latency: 23.7 ms), with memory usage under 614 MB. For pathogenicity prediction, a Feature Tokenizer Transformer model was developed to classify genetic variants. The model used a set of inputs derived from next-generation sequencing data from the CLINVAR database, after being processed through the regular pipeline consisting of quality control, alignment, variant calling, filtering, indexing, and annotation. The semi-supervised approach allowed the model to extract insight from uncertain-labeled data. The model achieved a precision of 0.98, sensitivity of 0.92, specificity of 0.99, and F1 score of 0.95, outperforming ClinPred and REVEL in precision and specificity, but with slightly lower sensitivity than ClinPred. These studies demonstrate the versatility and efficacy of Attention-based models in handling diverse biomedical signal modalities. The results underscore the potential of Transformer architectures to advance biomedical data analysis, enhancing diagnostics and personalized medicine.", "source_url": "https:\/\/openalex.org\/W4414483146", "pdf_url": "", "doi": "10.14201\/gredos.167164", "handle": "", "repository_name": "Gredos (University of Salamanca)", "source_record_id": "W4414483146", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:49", "updated_at": "2026-11-15 16:52:49" }, { "id": 384, "title": "Higher Order Transformers With Kronecker-Structured Attention", "title_ar": "", "title_en": "Higher Order Transformers With Kronecker-Structured Attention", "author": "Soroush Omranpour", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "", "university_source": "", "college": "", "department": "", "country": "", "governorate": "", "publication_year": 2025, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; computer-vision; medical-ai; predictive-analytics; transformers; image-classification; medical-imaging", "keywords": "", "abstract": "Modern datasets are increasingly high-dimensional and multiway, naturally represented as tensors with multiple indexed variables. Such data arise in numerous applications, including multivariate time series forecasting, 3D medical imaging, and climate modeling. While Transformers have demonstrated exceptional performance in sequence modeling and high-dimensional learning tasks, their direct application to multiway data remains computationally prohibitive due to the quadratic complexity of dot-product attention. Moreover, standard self-attention mechanisms do not inherently preserve the tensor structure, leading to inefficiencies in capturing cross-dimensional dependencies.In this work, we introduce the Higher-Order Transformer (HOT), a novel attention mechanism designed to process multiway data efficiently. HOT leverages Kronecker factorization to express multiway attention as either a sum of Kronecker products or mode-wise attention matrices, effectively capturing structured dependencies while significantly reducing computational and memory overhead. Unlike conventional self-attention, HOT maintains the original tensor structure, ensuring efficient representation of both dense and sparse interactions across dimensions. Through theoretical analysis, we show that HOT retains the expressiveness of full high-order attention while providing a mechanism to control complexity via the factorization rank.We evaluate HOT on challenging real-world datasets, including multivariate time series forecasting, 3D medical image classification, and stock market prediction. Experimental results demonstrate that HOT achieves competitive or superior performance compared to state-of-the-art Transformer-based baselines while requiring significantly fewer computational resources. Additionally, ablation studies highlight the impact of factorization rank, positional encoding strategies, and attention patterns, providing further insights into HOT’s efficiency and adaptability. Our findings underscore HOT's potential as a scalable and effective solution for high-dimensional multiway learning, paving the way for future extensions in autoregressive modeling, large-scale vision tasks, and cross-modal learning", "source_url": "https:\/\/openalex.org\/W7115035808", "pdf_url": "", "doi": "10.82308\/30876", "handle": "", "repository_name": "", "source_record_id": "W7115035808", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:49", "updated_at": "2026-11-15 16:52:49" }, { "id": 387, "title": "Explainable Transformers in Natural Language Processing", "title_ar": "", "title_en": "Explainable Transformers in Natural Language Processing", "author": "George Mihaila", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "", "university_source": "", "college": "", "department": "", "country": "United States", "governorate": "", "publication_year": 2026, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "deep-learning; nlp; medical-ai; responsible-ai; transformers; sentiment-analysis; speech-recognition; xai; fairness", "keywords": "", "abstract": "Transformer-based models have revolutionized natural language processing while becoming increasingly opaque in their decision-making processes. As these systems deploy in high-stakes domains such as healthcare, legal decision-making, and content moderation, their lack of transparency poses critical challenges, particularly where explainability is legally mandated or ethically required. This dissertation introduces ExpNet, a novel framework that learns to generate human-interpretable explanations from transformer attention mechanisms through supervised learning. Unlike existing approaches that apply fixed mathematical heuristics or rely on computationally expensive perturbation methods, ExpNet treats explanation generation as a learning problem. The framework extracts attention features and processes them through a lightweight neural network trained on human-annotated rationales, discovering optimal mappings from attention patterns to token importance scores. ExpNet demonstrates strong cross-task generalization capability through rigorous evaluation on diverse natural language processing tasks including sentiment analysis, grammatical acceptability judgment, and hate speech detection. When trained on tasks from certain domains and evaluated on entirely different tasks, ExpNet consistently outperforms thirteen baseline methods spanning model-agnostic, gradient-based, and attention-based approaches. This generalization achieved without task-specific retraining suggests that transformer attention encodes universal linguistic patterns, enabling the framework to explain models across a broad range of applications beyond its training domains. Quantitative results show ExpNet achieves substantially higher alignment with human rationales than all baselines across varied linguistic phenomena. The framework generates explanations orders of magnitude faster than perturbation-based methods, making it suitable for real-time deployment. This combination of superior explanation quality and computational efficiency addresses the critical gap between interpretability research and practical production requirements. The dissertation develops extensive analysis of explanation evaluation challenges in natural language processing. By learning from human reasoning, generalizing across diverse tasks, and maintaining computational efficiency, ExpNet provides a pathway toward transparent transformer-based systems suitable for high-stakes applications where explainability is imperative.", "source_url": "https:\/\/openalex.org\/W7166919207", "pdf_url": "", "doi": "", "handle": "", "repository_name": "University of North Texas Digital Library (University of North Texas)", "source_record_id": "W7166919207", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:49", "updated_at": "2026-11-15 16:52:49" }, { "id": 388, "title": "Transformer Block Coupling and Self-Attention Flows for Token Evolution", "title_ar": "", "title_en": "Transformer Block Coupling and Self-Attention Flows for Token Evolution", "author": "Anton Pavlovych Sugolov", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "University of Toronto", "university_source": "repository_host", "college": "", "department": "", "country": "Canada", "governorate": "", "publication_year": 2025, "language": "en", "research_domain": "nlp", "research_subdomain": "llm", "categories": "deep-learning; nlp; transformers; llm", "keywords": "", "abstract": "Transformers and Large Language Models have achieved incredible performance in deep learning, especially in natural language processing, and a precise understanding of the internal mechanisms driving their success is essential. We first provide an empirical analysis of structural characteristics of transformer blocks, by analyzing the trajectories of LLM token embeddings as they traverse the model depth. We linearize the system along the trajectories through their Jacobian matrices and uncover transformer block coupling in a multitude of LLMs, characterized by the coupling of their top singular vectors across tokens and depth. Our findings reveal that coupling positively with model performance, and that this relationship is stronger than with other hyperparameters such as parameter count, model depth, and embedding dimension. Further, we investigate recent theoretical perspectives that cast transformers as iterative dynamical systems which modify embeddings. We provide an analysis of self-attention as a vector field that acts on Gaussian mixtures, building on recent works that construct interaction functionals on measures with similar properties to self-attention. Collectively, these insights offer a novel perspective on token interactions in transformers, opening new directions for studying their mechanisms as well as improving training and generalization.", "source_url": "https:\/\/openalex.org\/W7132879793", "pdf_url": "", "doi": "", "handle": "", "repository_name": "TSpace (University of Toronto)", "source_record_id": "W7132879793", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:49", "updated_at": "2026-11-15 16:52:49" }, { "id": 389, "title": "Exploring Resource-Efficient Methods on Inhibitor Attention-Based Transformers", "title_ar": "", "title_en": "Exploring Resource-Efficient Methods on Inhibitor Attention-Based Transformers", "author": "Simon Engström", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Luleå University of Technology", "university_source": "repository_host", "college": "", "department": "", "country": "Sweden", "governorate": "", "publication_year": 2026, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "deep-learning; transformers", "keywords": "", "abstract": "In this thesis, quantization and knowledge distillation are applied to a transformer model with the inhibitor attention mechanism and the results are compared to that of a scaled dot-product attention (SDPA) model. The two main questions that are explored in this thesis are (1) whether an inhibitor model can be trained with knowledge distillation to match the performance of a SDPA model and (2) if inhibitor models are more robust than SDPA at low bit-width quantization. For the knowledge distillation, previous work trained an inhibitor model with DistilBERT as the teacher, yielding a GLUE score of 74.3. This work used a similar distillation method but replaced the teacher model with BERT, which resulted in a score of 75.0 on GLUE. This is not a significant increase, so using a bigger teacher model does not yield better results in this case. Quantization on the inhibitor model was also investigated and results of equivalent (number of parameters) inhibitor and dot-product based models were compared. 8-bit integer quantization showed little performance degradation (compared to floating-point models) for both quantization-aware training as well as post-training quantization for both model types. In the case of 4-bit quantization-aware training, the performance degradation was greater for the dot-product based model compared to the inhibitor model, where the inhibitor model scored 64.2 on GLUE while the dot-product model scored 59.3, down from 74.3 and 77.5 respectively. At this bit-width, quantization-aware training outperformed post-training quantization with the latter scoring 36.6 and 45.5 (GLUE score) for dot-product and inhibitor models respectively. With these results it can be concluded that inhibitor attention outperforms scaled dot-product attention at low bit-width quantization (specifically 4-bit in this case). It was found that inhibitor attention brought some challenges for integer-only calculations but one potential solution for computing inhibitor attention with integer-only arithmetic was found. It is not entirely efficient however, as many rescaling operations are required. It is therefore unclear how an efficient integer-only inhibitor model can be implemented using standard quantization schemes.", "source_url": "https:\/\/openalex.org\/W7133038095", "pdf_url": "https:\/\/ltu.diva-portal.org\/smash\/get\/diva2:2041996\/FULLTEXT01", "doi": "", "handle": "", "repository_name": "DiVA (Luleå University of Technology)", "source_record_id": "W7133038095", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:49", "updated_at": "2026-11-15 16:52:49" }, { "id": 393, "title": "MAWT: Multi-Attention-Weight Transformers", "title_ar": "", "title_en": "MAWT: Multi-Attention-Weight Transformers", "author": "Enver Askin", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "", "university_source": "", "college": "", "department": "", "country": "Canada", "governorate": "", "publication_year": 2023, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; nlp; predictive-analytics; transformers; machine-translation", "keywords": "", "abstract": "Transformers are machine learning models designed to learn and predict sequential and structured data, which are crucial to tasks such as neural machine translation and semantic parsing.They have become state-of-the-art engines for both of these tasks and much research in natural language processing is devoted to increasing their performance by introducing modifications to their architectures.In light of this trend, this thesis introduces a new Transformer architecture called MAWT: Multi-Attention-Weight Transformers in an attempt to increase the accuracy and variety of the acceptable predictions of a Transformer.It attempts to achieve this by training multiple weights per each Transformer attention head, which then are used to test the accuracy of the engine.This creates a new architecture under which the system produces a candidate set of outputs (instead of a singly output), along with a method for selecting from the candidate set.My proposal rests on the assumption --motivated by statistical considerations --that having a candidate set increases the probability of finding an exact match within the set.Upon testing, I observed that my system outperforms the regular transformer on 5\/6 benchmark neural machine translation and semantic parsing datasets, where engine performance is measured by exact match accuracy.Exact match accuracy demands syntactic identity between the output and the target.In order to investigate how well my new architecture generalizes to measures of semantic equivalence that don't also demand syntactic identity, I also recorded the BLEU scores on these datasets.The BLEU score is a measure of performance based on n-grams rather than exact symbolic match (i.e., how many contiguous sequence of n-many strings from the predicted output match the desired output).The results I report on the BLEU scores are more iii mixed, raising important questions that I highlight about the role of syntax in measures of semantic equivalence.", "source_url": "https:\/\/openalex.org\/W4362577036", "pdf_url": "http:\/\/repository.library.carleton.ca\/downloads\/g732d990n", "doi": "10.22215\/etd\/2023-15442", "handle": "", "repository_name": "Carleton University's Institutional Repository (MacOdrum Library, Carleton University)", "source_record_id": "W4362577036", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:49", "updated_at": "2026-11-15 16:52:49" }, { "id": 394, "title": "Sparse Attention Mechanisms in Vision Transformers: from Spatiotemporal Transformers to Instance-Dependent Attention Patterns", "title_ar": "", "title_en": "Sparse Attention Mechanisms in Vision Transformers: from Spatiotemporal Transformers to Instance-Dependent Attention Patterns", "author": "Brendan William Duke", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "University of Toronto", "university_source": "repository_host", "college": "", "department": "", "country": "Canada", "governorate": "", "publication_year": 2023, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; computer-vision; nlp; predictive-analytics; cnn; transformers; image-segmentation; video-analysis; speech-recognition", "keywords": "", "abstract": "Deep learning (DL) algorithms have produced superior results for a variety of vision, language, and speech processing tasks, but have proven to be power and compute-hungry. In the past, hardware advancements have met the growing need for compute. Yet in recent years the end of Dennard scaling and Moore's law makes hardware advancements unlikely to continue at the same fast pace. Hence innovative algorithmic solutions are needed to meet DL algorithms' ever-growing demand for compute. In this dissertation we investigated methods using sparsity and conditional computation to meet this compute demand. The two main components of this contribution involve applying structure and dynamic prediction of sparsity patterns by machine learning algorithms. The hypothesis was that structured and dynamic sparse computation improves latency and memory footprint compared to existing approaches with equal algorithm effectiveness. To show this, we built on the state of the art in sparse computations. This dissertation fills in previously unexplored areas of applying structured sparsity to videos, and of predicting per-instance sparsity that is unstructured. The purpose was to investigate using structured and per-example sparsity to design performant and effective deep learning algorithms. Existing literature had investigated sparsity in computer vision, but only in the context of either convolutional neural networks (CNNs) or Transformers applied to image processing. Given the increased computational complexity, this raised the question of how sparsity fares when applied to video processing. We investigated this question in work presented at the CVPR 2021 conference called Sparse Spatiotemporal Transformers for Video Object Segmentation (SSTVOS). SSTVOS looks at the highly computationally demanding problem of applying Transformers to video by using sparse attention blocks. The sparse attention blocks in SSTVOS use a fixed sparsity pattern that takes advantage of a priori knowledge of which visual features to attend to. However, using fixed sparsity patterns limits us to problems where we have this a priori knowledge. Therefore we further executed a research direction to learn sparsity patterns that are performant (low latency and memory usage) and effective (superior predictive power) in a method called Sparsifiner. Sparsifiner predicts per-example sparsity patterns as a more flexible generalization of the fixed sparsity patterns in SSTVOS. With the expectation that a machine learning model learns to predict sparsity patterns superior to any fixed pattern designed by hand, the idea for Sparsifiner was born. Sparsifiner improves over existing sparsity methods in Vision Transformers~(ViTs) by predicting per-instance sparsity patterns with a computationally-efficient connectivity predictor module. The method serves as empirical evidence of the effectiveness of per-instance sparsity patterns compared to fixed sparsity patterns. Sparsifiner shows a Pareto-optimal trade-off between FLOPs and top-1 accuracy on ImageNet when compared to contemporary structured-sparse methods. Together, both SSTVOS and Sparsifiner constitute a basis for future work in the direction of structured and dynamic sparsity in deep learning.", "source_url": "https:\/\/openalex.org\/W7132914290", "pdf_url": "", "doi": "", "handle": "", "repository_name": "TSpace (University of Toronto)", "source_record_id": "W7132914290", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:49", "updated_at": "2026-11-15 16:52:49" }, { "id": 396, "title": "Arbitrary-order Sampling and Hand Motion Modeling with Transformers", "title_ar": "", "title_en": "Arbitrary-order Sampling and Hand Motion Modeling with Transformers", "author": "Maxwell Clarke", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "", "university_source": "", "college": "", "department": "", "country": "New Zealand", "governorate": "", "publication_year": 2023, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; responsible-ai; self-semi-supervised; transformers; fairness", "keywords": "", "abstract": "Transformer models have been quickly taking over many tasks in the field of deep learning, due to their versatility while maintaining high performance. This thesis provides an introduction to transformer models, presents experiments with new ways of sampling data with them, and then applies them to the domain of hand motion modeling. Firstly, a comprehensive introduction to transformer models is given, including the attention operation, masking, architecture variants, and different pre-training tasks. Secondly, an experiment is presented using a probabilistic transformer model on the MNIST dataset, which was trained so that it is capable of arbitrary-order sampling. The experiment compares different sampling orders, including some dynamic sampling order heuristics based on the entropy. The experiments find that such sampling orders introduce a statistical bias into the samples. Lastly, the problem domain of hand motion modeling is introduced, and transformer models are trained to generate hand-motion sequences, via self-supervised learning on a motion-capture dataset. Both deterministic and probabilistic models are trained. The deterministic models can generate realistic-looking hand motions, but cannot be directed to generate specific motions. The probabilistic model performs poorly.", "source_url": "https:\/\/openalex.org\/W4389570359", "pdf_url": "https:\/\/openaccess.wgtn.ac.nz\/articles\/thesis\/Arbitrary-order_Sampling_and_Hand_Motion_Modeling_with_Transformers\/24785205\/1\/files\/43587534.pdf", "doi": "10.26686\/wgtn.24785205", "handle": "", "repository_name": "Open Access Victoria University of Wellington | Te Herenga Waka (Figshare)", "source_record_id": "W4389570359", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:49", "updated_at": "2026-11-15 16:52:49" }, { "id": 398, "title": "Vision transformers with Inductive Bias introduced through self-attention regularization", "title_ar": "", "title_en": "Vision transformers with Inductive Bias introduced through self-attention regularization", "author": "MORMILLE LUIZ HENRIQUE BARBOSA", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "", "university_source": "", "college": "", "department": "", "country": "Japan", "governorate": "", "publication_year": 2023, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "deep-learning; computer-vision; responsible-ai; cnn; transformers; fairness", "keywords": "", "abstract": "In recent years, the Transformer achieved remarkable results in computer vision related tasks, matching, or even surpassing those of convolutional neural networks (CNN). However, unlike CNNs, those vision transformers lack strong inductive biases and, to achieve state-of-the-art results, rely on large architectures and extensive pre-training on tens of millions of images. Introducing the appropriate inductive biases to vision transformers can lead to better convergence and generalization on settings with fewer training data. This work presents a novel way to introduce inductive biases to vision transformers: self-attention regularization. Two different methods of self-attention regularization were devised. Furthermore, this work proposes ARViT, a novel vision transformer architecture, where both self-attention regularization methods are deployed. The experimental results demonstrated that self-attention regularization leads to better convergence and generalization, especially on models pre-trained on mid-size datasets.", "source_url": "https:\/\/openalex.org\/W7145074315", "pdf_url": "", "doi": "", "handle": "", "repository_name": "Institutional Repositories DataBase (IRDB)", "source_record_id": "W7145074315", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:49", "updated_at": "2026-11-15 16:52:49" }, { "id": 486, "title": "Deep learning approaches for yield prediction and crop disease recognition", "title_ar": "", "title_en": "Deep learning approaches for yield prediction and crop disease recognition", "author": "Luning Bi", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Iowa State University", "university_source": "repository_host", "college": "", "department": "", "country": "United States", "governorate": "", "publication_year": 2022, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; computer-vision; generative-ai; medical-ai; ai-applications; predictive-analytics; optimization; cnn; rnn; transformers; gan; ai-agriculture", "keywords": "", "abstract": "The increase of the world population has brought significant challenges to the agriculture production system. Although mechanization has been realized in agriculture, many tasks (e.g., breeding, field inspection) are still labor-intensive and time-consuming. Therefore an automatic and intelligent solution is needed for the advancement of agricultural production. During this process, the biggest challenge is how to teach computers to understand the concepts in the real world. For example, an experienced expert can easily determine whether a plant is diseased or healthy. However, this may be challenging for the computer. Thus, the motivation of this dissertation study is to tackle these challenges in precision agriculture. This dissertation consists of four papers that propose different deep learning methods for the most challenging problems in agriculture. In the first paper, a genetic algorithm (GA)-assisted deep neural network was built for yield prediction using genetic information and environmental factors. In the global search phase, the GA was introduced to help determine the best initial weights of the neural network. In the local phase, random perturbation was used to avoid the local optimum. By using the proposed method, the root mean square error can be reduced by up to 10%. In the second paper, we proposed a generative adversarial network (GAN)-based approach to generate additional images for the classification of plant species and diseases using limited data. CNN was used as the basic network to classify species and diseases. GAN and label smoothing regularization (LSR) were combined to generate additional training images. Regular data augmentation techniques were also used to expand the dataset. The results showed that compared with using the real dataset only, the proposed method can improve the prediction accuracy by 6%. In the third paper, the potential of using satellite imagery for plant disease detection was explored. A gated recurrent units (GRU)-based model was presented for early detection of soybean sudden death syndrome (SDS) through time-series satellite imagery. The results showed that, compared to XGBoost and fully connected deep neural network (FCDNN), the GRU-based can improve the overall prediction accuracy by 7%. In addition, the proposed method can also be adapted to predict the future development of SDS. In the fourth paper, a transformer-based approach was proposed for soybean yield prediction using time-series camera images and seed treatments information. First, a vision transformer (ViT) base model was designed to extract features from the images. Then another transformer-based model was established to predict the yield using the time-series features. A case study was been conducted using a data set that was collected during the 2020 soybean-growing seasons in Canada. The experiment results show that compared to non-time series prediction and other baseline models, the proposed approach can reduce the mean squared error by 25%-40%. In conclusion, this dissertation aims to apply different state-of-art deep learning methods in agriculture. The study covers different topics, which range from yield prediction, species classification, to plant disease classification and prediction. At the model level, the application of linear models, tree-based methods, fully connected neural networks, convolutional neural networks, time-series models and transformers to different tasks have been investigated. In terms of the learning type, both unsupervised learning and supervised learning have been utilized. The experimental results have shown that appropriate deep learning methods can achieve better performance than traditional methods on specific tasks. Based on our work, more applications of deep learning techniques can be developed in the future.", "source_url": "https:\/\/openalex.org\/W7135318077", "pdf_url": "", "doi": "", "handle": "", "repository_name": "Iowa State University Digital Repository (Iowa State University)", "source_record_id": "W7135318077", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:50", "updated_at": "2026-11-15 16:52:50" }, { "id": 542, "title": "Natural language processing for business documents", "title_ar": "", "title_en": "", "author": "Ελευθέριος-Παναγιώτης Λούκας", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "", "university_source": "", "college": "", "department": "", "country": "", "governorate": "", "publication_year": 2025, "language": "el", "research_domain": "nlp", "research_subdomain": "llm", "categories": "machine-learning; deep-learning; nlp; few-zero-shot; rnn; transformers; llm", "keywords": "", "abstract": "Η Επεξεργασία Φυσικής Γλώσσας (NLP) για επιχειρηματικά και χρηματοοικονομικά κείμενα (Hahn et al., 2018; Chen et al., 2022) αποτελεί ένα ταχέως αναπτυσσόμενο ερευνητικό πεδίο που εφαρμόζει υπολογιστικές τεχνικές σε κείμενα όπως εταιρικά έγγραφα, αναφορές αναλυτών και οικονομικές ειδήσεις. Τα έγγραφα αυτά παρουσιάζουν μοναδικές προκλήσεις λόγω του εξειδικευμένου λεξιλογίου (El-Haj et al., 2019), του κρίσιμου ρόλου των αριθμητικών δεδομένων, των ιδιαίτερων συντακτικών δομών και της σημασιολογίας του εκάστοτε πεδίου. Οι δυσκολίες αυτές εντείνονται από ευρύτερα προβλήματα, όπως η επεξεργασία μεγάλου όγκου μη δομημένων δημόσιων δεδομένων και η οικονομικά αποδοτική ανάπτυξη γλωσσικών μοντέλων, ιδίως για οργανισμούς με περιορισμένους πόρους όπως οι small-to-medium enterprises (SMEs). Η αντιμετώπιση αυτών των προκλήσεων είναι κρίσιμη για εφαρμογές που εκτείνονται από την ανίχνευση απάτης (Goel and Gangolly, 2012), τη long-form summarization (Cao et al., 2024) και την εξαγωγή πληροφορίας έως τη χρηματοοικονομική question answering (Maia et al., 2018). Η παρούσα διατριβή στοχεύει στην προώθηση της εφαρμοσμένης NLP και της αξιοποίησης επιχειρηματικών εγγράφων σε πραγματικές συνθήκες, αντιμετωπίζοντας σύγχρονες βιομηχανικές προκλήσεις σε διαφορετικά επίπεδα της τεχνητής νοημοσύνης, και πιο συγκεκριμένα στο επίπεδο των δεδομένων, των εφαρμογών και της ανάπτυξης (deployment), με σταθερή έμφαση σε περιβάλλοντα περιορισμένων πόρων. Διατυπώνουμε τρία βασικά ερευνητικά ερωτήματα: (1) Πώς μπορούν τα ανοικτής πρόσβασης, μη δομημένα επιχειρηματικά έγγραφα να αξιοποιηθούν αποτελεσματικά για NLP; (2) Πώς μπορούν οι σύγχρονες μέθοδοι NLP που βασίζονται σε deep learning (DL) να προσαρμοστούν και να επεκταθούν ώστε να δημιουργούν επιχειρηματική αξία σε εργασίες όπως το automatic document tagging, λαμβάνοντας υπόψη τις ιδιαιτερότητες της χρηματοοικονομικής γλώσσας και την έντονη εξάρτησή της από αριθμητικά δεδομένα; (3) Για τις συνήθεις βιομηχανικές εργασίες text classification, ποιες προσεγγίσεις είναι οι πιο ακριβείς και οικονομικά αποδοτικές σε συνθήκες περιορισμένων πόρων; Το τρίτο ερώτημα μελετάται μέσω μιας πραγματικής εφαρμογής οικονομικά αποδοτικής αναγνώρισης πρόθεσης σε διαλόγους πελατών, με σύγκριση μεταξύ BERT-based models και Large Language Models (LLMs) και με βελτιστοποίηση του κόστους ανάπτυξης των LLMs. Για την αντιμετώπιση των παραπάνω, αρχικά εστιάζουμε στον «εκδημοκρατισμό» της πρόσβασης στα επιχειρηματικά έγγραφα μέσω της ανάπτυξης του EDGAR-CORPUS, του μεγαλύτερου δημόσια διαθέσιμου financial NLP corpus στα Αγγλικά, των domain-specific word embeddings EDGAR-W2V, τα οποία υπερτερούν των εναλλακτικών προσεγγίσεων, καθώς και του EDGAR-CRAWLER, ενός open-source εργαλείου εξαγωγής χρηματοοικονομικών δεδομένων με εκατοντάδες χρήστες από τον ακαδημαϊκό χώρο, τη FinTech κοινότητα και την ανάπτυξη web εφαρμογών. Στη συνέχεια, εισάγουμε το XBRL tagging ως μια πραγματική εργασία NLP, όπου δημιουργούμε το FiNER-139, το πρώτο dataset για το συγκεκριμένο πρόβλημα, και δείχνουμε ότι τα LSTM μοντέλα μπορούν να υπερέχουν έναντι του BERT λόγω περιορισμών της αρχιτεκτονικής των transformers και της τυπικής τεχνικής tokenization σε σχέση με τον κατακερματισμό αριθμητικών tokens. Μετά από εκτενή αξιολόγηση διαφορετικών μεθόδων, προτείνουμε μια νέα τεχνική tokenization με χρήση pseudo-tokens για transformer μοντέλα, η οποία βελτιώνει σημαντικά την απόδοση σε numeric-first tasks, οδηγώντας στην κυκλοφορία νέων state-of-the-art BERT μοντέλων (SEC-BERT) ειδικά σχεδιασμένων για το χρηματοοικονομικό πεδίο. Τέλος, για την οικονομικά αποδοτική αναγνώριση πρόθεσης, πραγματοποιούμε μια εκτενή συγκριτική μελέτη στο Banking77 dataset (Casanueva et al., 2020), δείχνοντας ότι μικρότερα BERT-based models μπορούν να είναι πιο αποδοτικά και οικονομικά από τα LLMs, απαιτώντας μόνο ελαφρώς περισσότερα επισημασμένα δεδομένα εκπαίδευσης. Παράλληλα, παρουσιάζουμε το “Dynamic Few-Shot Prompting”, μια προσέγγιση βασισμένη σε Retrieval-Augmented Generation (RAG) που μειώνει δραστικά το κόστος inference των LLMs διατηρώντας υψηλή ακρίβεια, και διερευνούμε τη χρησιμότητα της synthetic data generation. Η διατριβή καθιστά δημόσια διαθέσιμους βασικούς πόρους: το EDGAR-CORPUS, τα EDGAR-W2V embeddings, το λογισμικό EDGAR-CRAWLER, το dataset FiNER-139, την οικογένεια μοντέλων SEC-BERT, καθώς και ένα επιμελημένο υποσύνολο του Banking77 dataset, το οποίο περιλαμβάνει expert-selected παραδείγματα για κάθε intent class και το οποίο αποδεικνύεται κρίσιμο για την επίτευξη υψηλής απόδοσης σε few-shot learning σενάρια.", "source_url": "https:\/\/openalex.org\/W4417302042", "pdf_url": "", "doi": "10.12681\/eadd\/60507", "handle": "", "repository_name": "", "source_record_id": "W4417302042", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:50", "updated_at": "2026-11-15 16:52:50" }, { "id": 546, "title": "Natural Language Processing for Discussion Platforms", "title_ar": "", "title_en": "Natural Language Processing for Discussion Platforms", "author": "Maike Katrin Behrendt", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "", "university_source": "", "college": "", "department": "", "country": "Germany", "governorate": "", "publication_year": 2026, "language": "en", "research_domain": "nlp", "research_subdomain": "llm", "categories": "machine-learning; deep-learning; nlp; medical-ai; transfer-learning; transformers; llm", "keywords": "", "abstract": "Online participation platforms and digital discussion spaces have become central arenas for public discourse, particularly in times of crisis such as the COVID-19 pandemic. While these environments expand opportunities for civic engagement, they are frequently characterized by incivility, misinformation, and unstructured exchanges. Drawing on the concept of deliberation, defined by the core dimensions of rationality, civility, interactivity and equality, this thesis investigates how Natural Language Processing (NLP) methods can be leveraged to foster more respectful, reasoned, and inclusive online discussions. The dissertation addresses five central aspects: (i) the current approaches in applying NLP to online discussion platforms, (ii) the optimization of smaller language models for efficient real-time deployment, (iii) the integration of NLP-based interventions into discussion platforms, (iv) user perceptions of Artificial Intelligence (AI)-supported moderation and assistance, and (v) the measurable effects of such interventions on discussion quality. First, a systematic review provides a comprehensive overview of existing NLP applications in political and civic online discussions, identifying established approaches and tasks that could potentially be used to support discussions (see Chapter 4). Second, the thesis develops and evaluates methods to improve the efficiency and effectiveness of small Transformer-based models, particularly Bidirectional Encoder Representations from Transformers (BERT), for tasks relevant to online deliberation. This includes adapting pre-training strategies for argument similarity detection (ArgueBERT, see Chapter 5) and optimizing fine-tuning procedures for classification performance (MaxPoolBERT, see Chapter 6). In addition, the Additive deliberative Quality score with Adapters (AQuA) is introduced to assess the deliberative quality of user contributions (see Chapter 7). This model, along with a comment recommendation module based on stance detection, has been integrated into an online discussion platform and tested in a large-scale, real-time experiment (see Chapter 8). The findings provide insights into the technical integration of AI-supported tools into online discussion platforms, as well as into users’ perceptions of such systems and their observed effects in a specific participatory context. By combining perspectives from computer science, political science, and communication science, this interdisciplinary work contributes to the development of AI-supported tools that aim to support deliberative processes in digital public spheres.", "source_url": "https:\/\/openalex.org\/W7213477619", "pdf_url": "", "doi": "", "handle": "", "repository_name": "Univ. Duesseldorf: Duesseldorfer Dokumenten- und Publikationsserver", "source_record_id": "W7213477619", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:50", "updated_at": "2026-11-15 16:52:50" }, { "id": 568, "title": "Improving Human Behavior Understanding with Semi-supervised Learning", "title_ar": "", "title_en": "Improving Human Behavior Understanding with Semi-supervised Learning", "author": "Bruno Manuel Degardin", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "", "university_source": "", "college": "", "department": "", "country": "Portugal", "governorate": "", "publication_year": 2026, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; computer-vision; nlp; generative-ai; responsible-ai; self-semi-supervised; cnn; transformers; multimodal; video-analysis; gan; diffusion; fairness", "keywords": "", "abstract": "Recent advancements in Computer Vision (CV) and Machine Learning (ML) have fundamentally revolutionized the field of human behaviour understanding, enabling intelligent systems to move beyond simple motion detection to the interpretation of complex, semantic actions from vast streams of visual data. This paradigm shift has been driven by the exponential growth of Deep Learning algorithms, particularly the evolution from 3D Convolutional Neural Networks (3D CNNs) to modern Vision Transformers (ViTs) , which have achieved remarkable success in modeling long-range spatiotemporal dependencies in RGB videos and multimodal signals. However, despite their impressive performance on controlled benchmarks, the deployment of these pixel-based models in real-world scenarios remains heavily constrained. Their reliance on high-dimensional visual data often incurs prohibitive computational costs that hinder real-time application on edge devices. Furthermore, these methods exhibit significant sensitivity to background clutter and viewpoint variations , and the processing of raw video raises critical privacy concerns by capturing unnecessary appearance details such as identity and clothing. Maximizing the potential of this modality presents a significant set of challenges that hinder widespread deployment. The heavy reliance on massive, high-quality annotated datasets creates a prohibitive bottleneck for supervised learning, while existing generative models used for data augmentation often lack physical plausibility or controllability. Furthermore, self-supervised methods borrowed from video domains frequently fail to capture the fine-grained dynamics of skeletal motion, and current evaluation metrics for synthetic data struggle to distinguish between realistic kinematics and artifacts. Addressing these limitations requires a fundamental shift toward more label-efficient, physically aware, and generative approaches to behaviour analysis. In this thesis, we proposed a comprehensive suite of novel solutions to mitigate these problems by exploiting semi-supervised and self-supervised learning paradigms, as well as advanced generative modeling. First, we addressed the limitations of temporal reasoning in Graph Convolutional Networks (GCNs) by proposing REGINA, a reasoning- augmented architecture. This method injects explicit temporal self-similarity cues into the network through a learnable temporal adjacency matrix, allowing the model to adaptively optimize its receptive fields to capture long-range dependencies that standard convolutions often miss. Addressing data scarcity, we proposed Kinetic-GAN, a generative adversarial network that scales action conditioning by an order of magnitude over prior works. By generating actions directly from a disentangled latent space using a mapping network and stochastic noise injection, it synthesizes diverse, high-fidelity global body movements across more than 100 action classes. We further advanced representation learning without labels through ATOM, a self-supervised strategy that leverages atomic motion representation learning. By decomposing actions into short, fine-grained primitives and training the backbone to recover their temporal order, ATOM learns robust features from the intrinsic structure of motion, transferring effectively to supervised tasks without the computational burden of heavy reconstruction heads. To bridge the gap between generation and recognition, we developed an intrinsic, kinematics-aware quality assessment framework. Recognizing that conventional distribution metrics fail to capture physical validity, this framework utilizes per-sample kinematic indicators within a weaklysupervised discriminator to iteratively filter synthetic samples, ensuring only physically plausible data is used to boost downstream performance. Finally, pushing the boundaries of controllable synthesis, we proposed GLDM, a Spatiotemporal Graph Latent Diffusion Model. GLDM projects skeleton sequences into a compact, graph-aligned latent space and performs diffusion with graph-aware updates. Crucially, by integrating cross-attention mechanisms, GLDM enables precise multi-modal conditioning, allowing generation to be guided not just by class labels but also by natural language prompts, achieving generalization beyond training data, unconstrained temporal flexibility and solving the mode collapse issues inherent to adversarial approaches. Results obtained by the proposed methods show that it is possible to improve HBU systems regarding label efficiency, generative controllability, and physical plausibility, while still obtaining state-of-the-art results. All proposed methods were evaluated on several large-scale benchmarks, demonstrating that robust behaviour understanding can be achieved with significantly reduced reliance on manual annotation.", "source_url": "https:\/\/openalex.org\/W7225277943", "pdf_url": "", "doi": "", "handle": "", "repository_name": "Portuguese National Funding Agency for Science, Research and Technology (RCAAP Project by FCT)", "source_record_id": "W7225277943", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:50", "updated_at": "2026-11-15 16:52:50" }, { "id": 572, "title": "Transformers and contrastive semi-supervised learning for medical image segmentation", "title_ar": "", "title_en": "Transformers and contrastive semi-supervised learning for medical image segmentation", "author": "Qianying Liu", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "The University of Glasgow", "university_source": "repository_host", "college": "", "department": "", "country": "United Kingdom", "governorate": "", "publication_year": 2026, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; computer-vision; medical-ai; responsible-ai; ai-applications; self-semi-supervised; cnn; transformers; image-segmentation; medical-imaging; fairness; ai-education", "keywords": "", "abstract": "Medical Image Semantic Segmentation (MISS), the process of assigning a semantic label to each pixel in an image, is a foundational task in computational medicine, critical for quantitative diagnostics and treatment planning. However, developing robust MISS models faces two intertwined challenges. First, there is an architectural dilemma: Convolutional Neural Networks (CNNs), like U-Net, excel at learning local features but are limited by their receptive fields, failing to capture global context essential for segmenting organs with large deformations. Conversely, Vision Transformers (ViTs) effectively model long-range dependencies but lack the inductive biases of CNNs, leading to poor generalization on the small datasets typical in medicine without extensive pre-training. Second, the prohibitive cost and expertise required for creating pixel-level annotations create a severe data scarcity bottleneck. While Semi-Supervised Learning (SSL) aims to mitigate this by leveraging unlabeled data, existing methods often fail to learn high-level semantic relations and are susceptible to confirmation bias from noisy pseudo-labels, class imbalance, and suboptimal contrastive sample selection. This thesis presents a comprehensive investigation to systematically address these challenges, delivering a cohesive suite of novel deep learning frameworks. The contributions are four-fold: First, to resolve the architectural trade-off, this work introduces CS-Unet, a pure Transformer network built upon a U-Net-like architecture. Its core innovation is the Convolutional Swin Transformer (CST) block, which integrates convolutions directly within the Multi-Head Self-Attention and Feed-Forward Network modules. This design imbues the Transformer with inherent localized spatial context and strong inductive biases, enabling it to efficiently learn both local and global features. Without pre-training, CS-Unet outperforms existing Transformer and CNN-based models on multi-organ and cardiac datasets, achieving state-of-the-art performance with fewer parameters. Second, to address data scarcity, a novel Multi-Scale Cross Supervised Contrastive Learning (MCSC) framework for SSL is developed. MCSC jointly trains CNN and Transformer models, using a cross-teaching paradigm where each network provides pseudo-labels for the other. Crucially, it moves beyond simple output consistency by applying a contrastive loss to feature maps at multiple scales, enforcing hierarchical semantic consistency. To handle the class imbalance endemic to medical imaging, a class-prevalence-aware loss is used to ensure features for infrequent classes are learned robustly. Third, to fortify SSL against noisy pseudo-labels, a certainty-guided contrastive learning strategy is proposed. This approach mitigates the impact of inaccurate pseudo-labels by using a certainty metric to guide the selection of samples for contrastive learning. The framework’s computational efficiency is enhanced through novel sampling strategies that select a few representative samples for contrasting, and a negative memory bank is used to increase sample diversity and eliminate dependence on batch size. Fourth, this thesis introduces a new paradigm for SSL by leveraging external anatomical priors through the Contrastive Cross-Teaching with Registration (CCT-R) framework. CCT-R is the first method to integrate spatial registration transforms into the learning process. It features two novel modules: a Registration Supervision Loss (RSL), which uses transforms between labeled and unlabeled volumes to generate an additional, highly reliable source of pseudo-labels, and Registration-Enhanced Positive Sampling (REPS), which uses registration to identify anatomically-corresponding positive pairs across volumes for contrastive learning. Overall, these contributions provide a powerful toolkit that significantly alleviates the annotation bottleneck in medical AI. The proposed methods demonstrate state-of-the-art performance on challenging segmentation benchmarks, delivering a pathway to develop accurate, data-efficient models for real-world clinical applications and opening new avenues for research into fusing geometric priors with semantic segmentation.", "source_url": "https:\/\/openalex.org\/W7127166515", "pdf_url": "", "doi": "10.5525\/gla.thesis.85734", "handle": "", "repository_name": "Enlighten: Theses (The University of Glasgow)", "source_record_id": "W7127166515", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:50", "updated_at": "2026-11-15 16:52:50" }, { "id": 593, "title": "Generative Models from the Inside: a journey through properties and applications", "title_ar": "", "title_en": "Generative Models from the Inside: a journey through properties and applications", "author": "Barbara Toniella Corradini", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "University of Florence", "university_source": "repository_host", "college": "", "department": "", "country": "Italy", "governorate": "", "publication_year": 2025, "language": "en", "research_domain": "nlp", "research_subdomain": "llm", "categories": "deep-learning; computer-vision; nlp; generative-ai; transformers; image-segmentation; llm; gan; diffusion", "keywords": "", "abstract": "Generative Artificial Intelligence (Gen-AI) is a rapidly evolving field focused on developing models that can generate realistic and novel content, from text to images. Recent advancements such as ChatGPT and DALL-E have revolutionized domains like Natural Language Processing and Computer Vision by creating models capable of producing highly realistic and coherent outputs. This increase in generative capabilities has been driven by foundational models like Transformers, Generative Adversarial Networks (GANs), and Diffusion Models, which have expanded the boundaries of what machines can create. This thesis addresses open challenges in the field, including optimizing computational efficiency, analyzing the latent spaces of generative models, and extracting valuable insights from pre-trained models. The contributions of this work include the development of novel recursive architectures, the application of topological data analysis to better understand GAN training dynamics, and the use of diffusion models for training-free segmentation, highlighting the increasing versatility of generative models across diverse applications. The future of Gen-AI is promising, with numerous opportunities yet to be explored. While progress has been significant, much remains to be learned, and further exploration will be essential to fully realize the potential of these models. We hope that this thesis will provide a meaningful contribution to inspire a deeper understanding and development of generative modelling.", "source_url": "https:\/\/openalex.org\/W7213486992", "pdf_url": "", "doi": "", "handle": "", "repository_name": "Florence Research (University of Florence)", "source_record_id": "W7213486992", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:51", "updated_at": "2026-11-15 16:52:51" }, { "id": 697, "title": "On the limits of privacy in federated learning : personalization and adversarial analysis", "title_ar": "", "title_en": "On the limits of privacy in federated learning : personalization and adversarial analysis", "author": "Sargam Gupta", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "KTH Royal Institute of Technology", "university_source": "repository_host", "college": "", "department": "", "country": "Sweden", "governorate": "", "publication_year": 2026, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; computer-vision; responsible-ai; ai-applications; predictive-analytics; transformers; face-recognition; federated-learning; fairness; ai-cybersecurity; ai-industry; ai-smart-cities", "keywords": "", "abstract": "With the rapid adoption of Artificial Intelligence across diverse applications, concerns around data privacy have become increasingly prominent. Traditional centralized machine learning approaches require aggregating large volumes of user data at a central server, posing risks related to privacy, scalability, and regulatory compliance. Federated Learning has emerged as a decentralized alternative, enabling clients to train models locally while only sharing model updates with a central server. It addresses some of the concerns by enabling multiple participants to collaboratively train a shared model while keeping their raw data locally stored. However, data locality alone does not guarantee privacy. Information about local training records may still be revealed through model parameters, gradients, updates, or predictions.This makes FL vulnerable to privacy attacks such as model poisoning, data reconstruction and membership inference attacks. This thesis investigates these limitations in FL through two complementary perspectives: adversarial evaluation of the framework and personalization of privacy protection. In this thesis, we first examine the validity of various data anonymization techniques for facial images through the re-identification attacks. A weight-optimized record linkage method is developed to determine whether anonymized image records can be linked to their corresponding original records. The evaluation considers different adversarial settings, including cases in which an attacker possesses only partial information. The findings demonstrate that visually modifying an image or separating identity-related information does not necessarily prevent re-identification. We also experiment with Shapley values to check whether they could be useful in detecting model poisoning attacks.Then, we develop two privacy-enhancing mechanisms for federated time-series forecasting. The first framework is centered around using differential privacy with transformers specifically for traffic flow data whereas the other framework is a microaggregation-based model agnostic k-anonymous federated forecasting framework. The results demonstrate that useful forecasting performance can be maintained while raw data remain decentralized and additional protection is applied at the client level. One of the major challenges that exist in FL is heterogeneity. There have been several studies in the literature on data and device heterogeneity in FL. However, the heterogeneity that exists in the privacy requirements of different clients is widely unexplored. Building on these findings, this thesis introduces a personalized privacy-preserving federated learning framework that allows individual clients to select their preferred privacy mechanism and protection level. The proposed FedPUP framework supports the simultaneous use of differential privacy and k-anonymity within a federated round. We also used Membership Inference attacks to empirically evaluate the information leakage associated with different privacy choices and to identify combinations of ε and k that provide comparable levels of protection. We also propose a privacy-aware server-side aggregation algorithm that incorporates client-level privacy characteristics during the aggregation process in each training round. Our empirical contributions are also complemented by an exploratory study investigating the practical requirements for deploying privacy-preserving federated learning systems in the industry. Our study highlights the importance of transparency, accountability, trust, costs, and privacy–utility trade-offs. Overall, this thesis demonstrates that federated learning should not be considered inherently private merely because data remain decentralized. By combining adversarial evaluation with personalized privacy-preserving mechanisms, it provides a more comprehensive understanding of privacy risks and safeguards in FL. The findings offer practical guidance for developing federated learning systems that are more privacy-aware, adaptable, and suitable for real-world deployment.", "source_url": "https:\/\/openalex.org\/W7214512574", "pdf_url": "", "doi": "", "handle": "", "repository_name": "KTH Publication Database DiVA (KTH Royal Institute of Technology)", "source_record_id": "W7214512574", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:52", "updated_at": "2026-11-15 16:52:52" }, { "id": 717, "title": "Automated Generation of X-Ray Reports by Machine Learning Approach", "title_ar": "", "title_en": "Automated Generation of X-Ray Reports by Machine Learning Approach", "author": "Aigerim Bauyrzhanova", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Nazarbayev University", "university_source": "repository_host", "college": "", "department": "", "country": "Kazakhstan", "governorate": "", "publication_year": 2025, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; computer-vision; medical-ai; transformers; medical-imaging", "keywords": "", "abstract": "Medical reports are essential for conveying conclusions and diagnoses from medical scans, such as X-ray scans, in order to proceed with the patient’s treatment. Especially lung X-rays, which are used in diagnosing and treating diseases such as pneumonia, pulmonary diseases, COVID-19, lung cancer, etc. The traditional way of writing medical reports for such X-ray images is to write them manually, by specialists. This is a time-consuming, hard-to-keep-up with the increasing demand, and error-prone process since it is done by the specialists themselves after taking the scans. This prompts to make the process more automated so as to save time and lessen the probability of a mistake in diagnosis. Therefore, automating the process of writing the medical reports from chest X-ray images has become a critical need in the field of radiology as nowadays there are a lot of detrimental and even fatal diseases of the lungs that need to be treated as soon as possible. For this purpose, the machine learning approach using encoder-decoder transformers and vision transformers is a potential candidate for getting numeric results of an effective model to generate X-ray reports on two datasets: IU-Xray and MIMIC-CXR", "source_url": "https:\/\/openalex.org\/W7198807364", "pdf_url": "https:\/\/nur.nu.edu.kz\/bitstreams\/6cc809dc-79d0-4ea3-a8fe-b30bff270ac5\/download", "doi": "", "handle": "", "repository_name": "Nazarbayev University Repository (Nazarbayev University)", "source_record_id": "W7198807364", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:52", "updated_at": "2026-11-15 16:52:52" }, { "id": 735, "title": "Adaptive Deep Learning for Vision and Multimodal Understanding with Heterogeneous Data", "title_ar": "", "title_en": "Adaptive Deep Learning for Vision and Multimodal Understanding with Heterogeneous Data", "author": "Youwei Liang", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "", "university_source": "", "college": "", "department": "", "country": "United States", "governorate": "", "publication_year": 2026, "language": "en", "research_domain": "nlp", "research_subdomain": "llm", "categories": "machine-learning; deep-learning; computer-vision; nlp; medical-ai; transfer-learning; few-zero-shot; predictive-analytics; transformers; multimodal; image-segmentation; llm", "keywords": "", "abstract": "Large-scale pretraining has produced powerful deep learning representations, but real-world data rarely match the clean, uniform assumptions of standard transfer learning. This dissertation uses heterogeneous data to describe settings in which information differs in modality, representation, relevance, quality, or frequency across inputs, outputs, tasks, or supervision. Examples include molecules represented as graphs, images, strings, and text; images containing both informative and redundant regions; noisy or incomplete labels; and rare but important classes. Applying pretrained models in these settings requires deciding which information to process, how to connect different forms of information, and how much to trust each supervision signal. This dissertation develops adaptive learning as a unifying approach through three paradigms: adaptive selection focuses computation on useful input components; adaptive alignment maps heterogeneous modalities, tasks, and outputs into a shared predictive space; and adaptive supervision adjusts imperfect or imbalanced training signals. The first contribution, EViT, demonstrates adaptive selection for efficient Vision Transformers. Standard Vision Transformers devote equal computation to all image patches despite substantial redundancy. EViT uses class-token attention to identify important tokens, preserves them in subsequent layers, and fuses inattentive tokens into a compact representation rather than discarding their context. This progressive token reorganization reduces computation while maintaining accurate visual recognition. The second contribution, DrugChat, demonstrates adaptive alignment for multimodal molecular prediction. Molecules can be represented as chemical graphs, rendered images, SMILES strings, and natural-language prompts, while predictions may require labels, probabilities, categorical decisions, or free-form descriptions. DrugChat encodes molecular graphs and images with pretrained encoders, projects them into the latent space of an instruction-following language model, and conditions generation on molecular structure and task prompts. It thereby aligns heterogeneous inputs in a shared representation space and expresses diverse prediction targets through a common language-generation interface. The third contribution, RobPicker, demonstrates adaptive supervision for cryo-electron tomography particle picking, where annotations are expensive, incomplete, noisy, and highly imbalanced across macromolecular species. Its meta-learning framework jointly trains a segmentation network, a label correction network, and a data reweighting network. Guided by cleaner validation data, these components refine imperfect labels and adapt each subtomogram's contribution to the loss, giving greater influence to rare or small particle classes. Together, EViT, DrugChat, and RobPicker apply adaptation at complementary levels of the learning pipeline: selecting what data to process, aligning heterogeneous representations and outputs, and learning from imperfect supervision.", "source_url": "https:\/\/openalex.org\/W7222391945", "pdf_url": "https:\/\/escholarship.org\/content\/qt3h17286t\/qt3h17286t.pdf", "doi": "", "handle": "", "repository_name": "eScholarship (California Digital Library)", "source_record_id": "W7222391945", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:52", "updated_at": "2026-11-15 16:52:52" }, { "id": 740, "title": "Deep learning and natural language processing for computer aided diagnosis", "title_ar": "", "title_en": "Deep learning and natural language processing for computer aided diagnosis", "author": "Sadam Hussain", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Tecnológico de Monterrey", "university_source": "repository_host", "college": "", "department": "", "country": "Mexico", "governorate": "", "publication_year": 2025, "language": "en", "research_domain": "nlp", "research_subdomain": "llm", "categories": "machine-learning; deep-learning; computer-vision; nlp; medical-ai; responsible-ai; predictive-analytics; transformers; multimodal; llm; medical-imaging; xai", "keywords": "", "abstract": "Multimodal artificial intelligence (AI) is a cutting-edge technique that integrates diverse modalities, such as imaging and textual data, to enhance classification and regression tasks. This dissertation focuses on the integration, comparison, and evaluation of multimodal AI for breast cancer diagnosis and prognosis. To achieve these objectives, we curated a comprehensive multimodal dataset comprising digital mammograms and corresponding radiological reports. Leveraging this dataset, we introduced and assessed various state-of-the-art (SOTA) multimodal techniques for three key tasks: breast cancer classification, reduction of false-positive biopsies with explainable AI (XAI), and short- term (5-year) risk prediction of breast cancer.In this work, we also introduced a benchmark dataset of radiological reports from breast cancer patients and provided baseline performance evaluations using SOTA machine learning (ML), deep learning (DL), and large language models (LLMs) for BI-RADS category classification. Our approach evaluated the performance of diverse SOTA multimodal architectures, including ResNet, VGG, E!cientNet, MobileNet, and Vision Transformers (ViT). For textual data processing, we employed both general-purpose and domain-specific pretrained LLMs such as BERT, bioGPT, ClinicalBERT, and DeBERTa, which were also integrated into multimodal architectures for enhanced lassification.Notably, our proposed multiview multimodal feature fusion (MMFF) architecture, combining SE-ResNet50 with an artificial neural network (ANN), achieved an AUC of 0.965 for breast cancer classification, significantly outperforming both single-modal and multimodal SOTA architectures. For reducing unnecessary breast biopsies, our multimodal approach achieved an AUC of 0.72, showcasing its clinical utility in minimizing patient burden. Moreover, our ViT and bioGPT-based multimodal architecture achieved an AUC of 0.77 for short-term risk prediction, outperforming the SOTA MIRAI model, which achieved an AUC of 0.59 on our in-house dataset. This work highlights the potential of multimodal AI in advancing breast cancer diagnosis and prognosis, demonstrating its superiority over traditional and unimodal approaches across multiple critical tasks.", "source_url": "https:\/\/openalex.org\/W7225556007", "pdf_url": "", "doi": "", "handle": "", "repository_name": "Repositorio del Tecnológico de Monterrey (Tecnológico de Monterrey)", "source_record_id": "W7225556007", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:52", "updated_at": "2026-11-15 16:52:52" }, { "id": 744, "title": "Multimodal Deep Learning for Driver Monitoring: Integrating EEG and Vision for Robust Drowsiness Detection and Safety Enhancement", "title_ar": "", "title_en": "Multimodal Deep Learning for Driver Monitoring: Integrating EEG and Vision for Robust Drowsiness Detection and Safety Enhancement", "author": "Ur Rahman, Shams", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Dublin City University", "university_source": "repository_host", "college": "", "department": "", "country": "Ireland", "governorate": "", "publication_year": 2025, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; computer-vision; medical-ai; ai-applications; predictive-analytics; cnn; transformers; multimodal; face-recognition; video-analysis; ai-education", "keywords": "", "abstract": "Road accidents remain a major global concern, with driver drowsiness and delayed reaction times recognized as key contributing factors. This thesis advances driver-monitoring research by developing multimodal approaches that integrate electroencephalography (EEG) and vision data to predict reaction times and drowsiness. The investigation first demonstrates that pre-stimulus EEG signals—specifically spectral power in the alpha and theta bands—contain rich information for estimating reaction times to critical road events. Using subject-independent machine-learning pipelines, short EEG windows recorded before event onset effectively differentiate between fast and slow responses. The work then explores the benefits of incorporating vision data as a second modality by fusing EEG signals with camera-based observations of the driver. One branch converts EEG power-spectral-density features into image-like representations for analysis with deep convolutional neural networks and transformer models. Another branch directly integrates raw EEG signals with synchronised video frames through end-toend multimodal transformer architectures. Results indicate that transformers equipped with cross-modal attention capture complex interdependencies between neural and visual cues, yielding significant improvements in driver-drowsiness detection over unimodal approaches. Real-time deployment is addressed by designing and optimising a lightweight pipeline for edge-based processing. This resource-efficient model enables rapid analysis of facial cues under diverse driving conditions, ensuring operation on embedded platforms such as smartphones and automotive edge devices. Extensive evaluations on large-scale simulated datasets confirm the generalisability of the proposed approaches across varied driving scenarios. Experiments reveal that transformer-based fusion significantly enhances predictive performance by effectively combining complementary neural and visual cues. Moreover, the lightweight pipeline maintains high accuracy under stringent computational constraints, enabling real-time, ondevice deployment.", "source_url": "https:\/\/openalex.org\/W7112463353", "pdf_url": "https:\/\/doras.dcu.ie\/31498\/1\/shams_thesis_final_version.pdf", "doi": "", "handle": "", "repository_name": "Dublin City University Open Access Institutional Repository (Dublin City University)", "source_record_id": "W7112463353", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:52", "updated_at": "2026-11-15 16:52:52" }, { "id": 829, "title": "A Novel Hybrid Deep Learning Model using U-net and Vision Transformer for Polyp Segmentation in medical imaging", "title_ar": "", "title_en": "A Novel Hybrid Deep Learning Model using U-net and Vision Transformer for Polyp Segmentation in medical imaging", "author": "Prity Rani Das", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "", "university_source": "", "college": "", "department": "", "country": "", "governorate": "", "publication_year": 2026, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "deep-learning; computer-vision; medical-ai; cnn; transformers; image-segmentation; medical-imaging", "keywords": "", "abstract": "This research represents a hybrid model for polyp segmentation in medical images, which combines convolutional neural networks (CNN) and vision transformers (ViT). We propose a hybrid architecture that integrates UNet with a Vision Transformer, where UNet is used as an encoder to capture local features and ViT is used as a bottleneck module to extract global contextual information. To improve the robustness and generalization of the model, various data augmentation techniques are applied to the training data set, including horizontal flip, vertical flip, and rotation. The model is trained and evaluated on the multicenter PolypDB dataset, with all images preprocessed to a consistent resolution, and the corresponding masks are appropriately managed. A loss function, combining Dice Loss and Binary Cross-Entropy (BCE) Loss, is applied to optimize segmentation performance.", "source_url": "https:\/\/openalex.org\/W7128515784", "pdf_url": "", "doi": "10.5281\/zenodo.18591130", "handle": "", "repository_name": "Zenodo (CERN European Organization for Nuclear Research)", "source_record_id": "W7128515784", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:53", "updated_at": "2026-11-15 16:52:53" }, { "id": 873, "title": "Variational methods and deep learning for semantic image segmentation", "title_ar": "", "title_en": "", "author": "Miguel Chicchón", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Pontifical Catholic University of Peru", "university_source": "repository_host", "college": "", "department": "", "country": "Peru", "governorate": "", "publication_year": 2025, "language": "es", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; computer-vision; transfer-learning; transformers; image-segmentation", "keywords": "", "abstract": "Actualmente, las aplicaciones de visión por computadora, que incluyen la conducción autónoma, la detección de fallas, la imagenología médica y la teledetección, son cada vez más frecuentes y necesitan mecanismos de segmentación o clasificación de píxeles más precisos. Recientemente, las técnicas de aprendizaje profundo que utilizan redes neuronales se han convertido en la estrategia dominante para resolver diversos problemas en visión por computadora; sin embargo, existen limitaciones y problemas aún por resolver. Es por ello que una integración adecuada con técnicas tradicionales, como los métodos variacionales, con sus propias ventajas y limitaciones, resulta de interés para mejorar la precisión y robustez de la segmentación semántica. Esta tesis se centra en la implementación y validación de algoritmos de segmentación de imágenes que combinan métodos variacionales y aprendizaje profundo en cuatro escenarios diferentes. En el primer escenario, que implica la segmentación de peces en entornos subacuáticos, se compararon métodos variacionales integrados utilizando enfoques de post-procesamiento, extremo a extremo (end-to-end) y una función de pérdida combinada. En el segundo escenario, que aborda la segmentación de baches y grietas en pavimentos de carreteras, se entrenaron tres arquitecturas de redes neuronales SOTA (del estado del arte) utilizando una función de pérdida de entropía cruzada combinada y contornos activos. En el tercer escenario, enfocado en la segmentación de techos de edificios en imágenes satelitales, se entrenó una red neuronal U-Net con diferentes redes SOTA en la ruta del codificador (encoder path) mediante aprendizaje por transferencia (transfer learning), y el mejor modelo fue seleccionado para ser entrenado utilizando una función combinada con pérdida de entropía cruzada y pérdida de contornos activos. Finalmente, en el cuarto escenario, que se refiere a la segmentación de cobertura terrestre en imágenes aéreas, se emplearon aprendizaje por transferencia y la función de contornos activos en el entrenamiento de redes neuronales convolucionales y transformadores (transformers). Los resultados cuantitativos consolidados demuestran mejoras significativas: la función de pérdida combinada logró aumentos promedio de4-6% en el mIoU (mean Intersection over Union) en comparación con los enfoques que solo utilizaban entropía cruzada o Dice. Específicamente, para la segmentación de peces, el modelo U-Net-scSE con un codificador ResNeSt-269e alcanzó 73.1% de mIoU; para la detección de baches, 87.45% de mIoU con una reducción de HD95 a 23.40 mm; para la segmentación de techos, se logró una mejora del 14% en IoU utilizando aprendizaje por transferencia; y para la cobertura terrestre, el modelo U-Net-scSE con un codificador MiT- B4 obtuvo 80.4% de mIoU. Estas validaciones a través de cuatro dominios heterogéneos demuestran que la integración sistemática de métodos variacionales en redes profundas proporciona mejoras cuantificables y transferibles en la exactitud de la segmentación.", "source_url": "https:\/\/openalex.org\/W7138825433", "pdf_url": "", "doi": "", "handle": "", "repository_name": "Communities in DSpace (Pontifical Catholic University of Peru)", "source_record_id": "W7138825433", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:54", "updated_at": "2026-11-15 16:52:54" }, { "id": 902, "title": "Enhancing Domain-Specific Automated Audio Captioning: A Study on Adaptation Techniques and Transfer Learning", "title_ar": "", "title_en": "Enhancing Domain-Specific Automated Audio Captioning: A Study on Adaptation Techniques and Transfer Learning", "author": "Charith Raghavaraju", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Tampere University", "university_source": "repository_host", "college": "", "department": "", "country": "Finland", "governorate": "", "publication_year": 2023, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; nlp; transfer-learning; transformers; speech-recognition", "keywords": "", "abstract": "Automated audio captioning is a challenging cross-modal task that takes an audio sample as input to analyze it and generate its caption in natural language as output.The existing datasets for audio captioning such as AudioCaps and Clotho encompass a diverse range of domains, with current proposed systems primarily focusing on generic audio captioning.This thesis delves into the adaptation of generic audio captioning systems to domain-specific contexts, simultaneously aiming to enhance generic audio captioning performance.The adaptation of the generic models to specific domains has been explored using two different techniques: complete fine-tuning of neural model layers and layer-wise fine-tuning within transformers.The process involves initial training with a generic captioning setup, followed by adaptation using domain-specific training data.In generic captioning, the process for training starts with training the model on the Audio-Caps dataset followed by fine-tuning it using the Clotho dataset.This is accomplished through the utilization of a transformer-based architecture, which integrates a patchout fast spectrogram transformer (PaSST) for audio embeddings and a BART transformer.Word embeddings are generated using a byte-pair encoding (BPE) tokenizer tailored to the training datasets' unique words, aligning the vocabulary with the generic captioning task.Experimental adaptation mainly focuses on audio clips related to animals and vehicles.The results demonstrate notable improvements in the performance of the generic and domain adaptation systems.Generic captioning has demonstrated an improvement in SPIDEr scores, increasing from 0.291 during fine-tuning to 0.301 with layer-wise fine-tuning.Specifically, we observed a notable increase in SPIDEr scores, from 0.315 to 0.323 for animal-related audio clips and from 0.298 to 0.308 for vehicle-related audio clips.", "source_url": "https:\/\/openalex.org\/W7133068187", "pdf_url": "https:\/\/trepo.tuni.fi\/bitstream\/10024\/153190\/2\/RaghavarajuCharitha.pdf", "doi": "", "handle": "", "repository_name": "Tampere University Institutional Repository (Tampere University)", "source_record_id": "W7133068187", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:54", "updated_at": "2026-11-15 16:52:54" }, { "id": 950, "title": "Exploring large language models for clinical code assignment", "title_ar": "", "title_en": "Exploring large language models for clinical code assignment", "author": "Kukkula, Lasse", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Aalto University", "university_source": "repository_host", "college": "", "department": "", "country": "Finland", "governorate": "", "publication_year": 2025, "language": "en", "research_domain": "nlp", "research_subdomain": "llm", "categories": "deep-learning; nlp; medical-ai; cnn; transformers; llm", "keywords": "", "abstract": "Clinical coding, the act of translating clinical transcripts into standardised diagnostic and procedural codes, is an essential statistics and billing tool in hospitals around the world. Despite it carrying a lot of weight in the capitalist healthcare system for billing purposes and for statistical and quality assurance purposes in socialised healthcare systems. The process of clinical coding is labour intensive and error prone. Recent advances in large language models pave a way toward fully or partially automated reliable clinical coding. The research on large language models is constantly evolving and developing into new avenues that could and should be explored in the context of clinical coding. This thesis provides a survey of the advancments in large language model architecture from the almost ancient multilayer perceptron and convolutional neural networks all the way to the current cutting edge advancements such as transformers and state space models. These findings are then taken into the context of clinical coding and some suggestions on future directions in the research into the use of large language models in clinical coding will be presented.", "source_url": "https:\/\/openalex.org\/W7112331338", "pdf_url": "https:\/\/aaltodoc.aalto.fi\/bitstreams\/42c2c5c2-03df-4772-ab73-0042e63ba09e\/download", "doi": "", "handle": "", "repository_name": "Aaltodoc (Aalto University)", "source_record_id": "W7112331338", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:55", "updated_at": "2026-11-15 16:52:55" }, { "id": 1103, "title": "Towards Few-Annotation Learning in Computer Vision : Application to Image Classification and Object Detection tasks", "title_ar": "", "title_en": "", "author": "Quentin Bouniot", "supervisor": "", "degree_type": "phd", "degree_evidence": "theses.fr: الفهرس الوطني الفرنسي لأطاريح الدكتوراه", "university": "", "university_source": "", "college": "", "department": "", "country": "France", "governorate": "", "publication_year": 2023, "language": "fr", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; computer-vision; transformers; image-classification; object-detection", "keywords": "", "abstract": "Vers un apprentissage avec peu d’annotations en vision par ordinateur : Application à la classification et à la détection d’objets dans les images Nous développons dans cette thèse des contributions théoriques, algorithmiques et expérimentales pour l’apprentissage automatique avec peu d'annotations, et plus spécifiquement pour les tâches de classification d’images et de détection d’objets en vision par ordinateur. Dans une première contribution, nous nous intéressons à combler le fossé entre la théorie et la pratique concernant les algorithmes de méta-apprentissage les plus répandus, utilisés pour de la classification avec peu d’exemples. Nous établissons des liens avec l’apprentissage de représentation multi-tâches, qui bénéficie de bases théoriques solides, afin de vérifier les meilleures conditions amenant à un méta-apprentissage plus efficace. Ensuite, afin d’exploiter les données non étiquetées pour entraîner des détecteurs d’objets basés sur l’architecture Transformer, nous proposons à la fois un pré-entraînement non supervisé et une méthode d’apprentissage semi-supervisée dans deux autres contributions distinctes. Pour le pré-entraînement, nous améliorons l’apprentissage contrastif des détecteurs d’objets en introduisant les informations de localisation. Enfin, notre méthode semi-supervisée est la première adaptée aux détecteurs d’objets utilisant des modèles type Transformers.", "source_url": "https:\/\/openalex.org\/W7151562911", "pdf_url": "", "doi": "10.70675\/6796455fzbdb0z4853z9023z932dbb24367b", "handle": "", "repository_name": "theses.fr (ABES)", "source_record_id": "W7151562911", "other_sources": "", "verification_status": "verified", "created_at": "2026-11-15 16:52:58", "updated_at": "2026-11-15 16:52:58" }, { "id": 1252, "title": "Automating news classification with large language models: Exploring fine-tuning, dataset size, and architecture", "title_ar": "", "title_en": "Automating news classification with large language models: Exploring fine-tuning, dataset size, and architecture", "author": "Yesilyurt, Burcu", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "", "university_source": "", "college": "", "department": "", "country": "Portugal", "governorate": "", "publication_year": 2025, "language": "en", "research_domain": "nlp", "research_subdomain": "llm", "categories": "machine-learning; deep-learning; nlp; responsible-ai; transfer-learning; predictive-analytics; transformers; llm; xai", "keywords": "", "abstract": "A comprehensive evaluation benchmarks Large Language Models with traditional machine learning algorithms for automatic news classification is done on three standard news classification datasets: BBC News, 20 Newsgroups, and AG News. We implement traditional models, including Naive Bayes, Logistic Regression, Support Vector Machine, and Random Forest, to provide clear and interpretable baselines using manual term-frequency and syntactic features. Then, fine‐tuned transformer architectures, including BERT, RoBERTa, T5, GPT, and their distilled variants, were used to quantify improvements in predictive accuracy, resource efficiency, and explainability. Performance is measured via 5-fold cross-validation using F1 and accuracy metrics, and statistical significance is assessed with a Friedman test followed by Holm’s correction. Results show that transformer models consistently outperform classical approaches, with BERT achieving the highest scores under both balanced and imbalanced conditions. Distilled models rival or surpass full-size transformers on larger datasets while reducing memory requirements and maintaining comparable inference latency. Attention‐based attribution methods provide semantic explanations on par with feature‐importance metrics, confirming that LLMs deliver superior accuracy, adaptability, and transparency in news classification. Future work should investigate multilingual pretraining, multilabel classification, and ensemble techniques to further strengthen real‐time, explainable news‐analysis pipelines.", "source_url": "https:\/\/openalex.org\/W7113115638", "pdf_url": "", "doi": "", "handle": "", "repository_name": "Portuguese National Funding Agency for Science, Research and Technology (RCAAP Project by FCT)", "source_record_id": "W7113115638", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:53:00", "updated_at": "2026-11-15 16:53:00" }, { "id": 1255, "title": "Federated Fine-Tuning of Mixture of Experts Vision Transformers", "title_ar": "", "title_en": "Federated Fine-Tuning of Mixture of Experts Vision Transformers", "author": "Filippo Serafini", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "University of Bologna", "university_source": "repository_host", "college": "", "department": "", "country": "Italy", "governorate": "", "publication_year": 2026, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; computer-vision; responsible-ai; ai-applications; transfer-learning; transformers; federated-learning; fairness; ai-smart-cities", "keywords": "", "abstract": "Mixture-of-Experts (MoE) architectures decouple representational capacity from per-token computational cost: each input activates only a few specialised experts, yielding far greater capacity than dense models at equal activation cost. Federated Learning (FL) trains models on edge devices—aggregating local updates, never centralising raw data—preserving privacy and unlocking edge-resident data. Their integration is hard: cross-device FL faces tight bandwidth and memory budgets and local datasets are heterogeneous (non-IID); MoE sparsity, however, can overcome these barriers. Adapting the partial-replica paradigm of the FoMoE framework—where each client hosts only a subset of the experts—this thesis investigates the federated fine-tuning of a Vision MoE (V-MoE-S\/32) on edge devices holding severely non-IID data, combined with Low-Rank Adaptation (LoRA) and Sparsity-aware Model Aggregation. Under a static expert assignment, restricting each client to half or fewer of the expert pool costs only 1–2 percentage points of Top-1 accuracy versus the full-replica baseline, while cutting the per-round communication payload by up to 28% and static on-device memory by up to 36%. A static assignment, however, mismatches the experts a client needs most with those initially allocated. To close this gap, this work designs and evaluates a novel server-orchestrated dynamic expert reassignment protocol that periodically migrates experts towards the clients demanding them most. Rather than raising the accuracy ceiling, it mainly stabilises the federation against sub-optimal initial placements, at a communication overhead of 1–3% of total traffic. Partial expert replication, parameter-efficient adaptation, and sparsity-aware aggregation thus compose into a practical recipe for fitting sparse vision models within edge memory, compute, and bandwidth budgets; by decoupling local hardware limits from model size, this design lets edge networks train far larger architectures.", "source_url": "https:\/\/openalex.org\/W7219542350", "pdf_url": "https:\/\/amslaurea.unibo.it\/id\/eprint\/39760\/1\/Master_s_thesis_Filippo_Serafini.pdf", "doi": "", "handle": "", "repository_name": "AMS Degree Thesis (University of Bologna)", "source_record_id": "W7219542350", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:53:00", "updated_at": "2026-11-15 16:53:00" }, { "id": 1256, "title": "SISTEM PENILAIAN FORUM DISKUSI MAHASISWA PADA E�LEARNING BERBASIS MODEL BERT", "title_ar": "", "title_en": "", "author": "Ahmad Habibi ; Endina Putri Purwandari ; Widhia oktoeberza", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "University of Bengkulu", "university_source": "repository_host", "college": "", "department": "", "country": "Indonesia", "governorate": "", "publication_year": 2025, "language": "id", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; nlp; ai-applications; transfer-learning; transformers; ai-education", "keywords": "", "abstract": "Penelitian ini bertujuan untuk merancang dan mengembangkan sistem penilaian otomatis forum diskusi mahasiswa pada platform E-Learning dengan memanfaatkan model BERT (Bidirectional Encoder Representations from Transformers). Penelitian ini dilatarbelakangi oleh tantangan dalam menilai partisipasi mahasiswa dalam forum diskusi E-Learning, nyatanya dengan metode tradisional sering kali bersifat subjektif dan memakan waktu. Sedangkan model BERT dikenal dengan kemampuannya dalam memahami konteks teks secara mendalam, sehingga diharapkan dapat meningkatkan objektivitas dan efisiensi dalam proses penilaian. Penelitian ini mencakup beberapa tahapan, mulai dari analisis kebutuhan, perancangan sistem, pengembangan fine-tuning model BERT, hingga evaluasi dengan membandingkan hasil penilaian otomatis dan manual. Sistem ini dirancang untuk menilai kontribusi mahasiswa berdasarkan kompleksitas dan relevansi kontennya, dengan tujuan meningkatkan interaksi serta kualitas diskusi dalam E-Learning. Hasil yang diharapkan adalah sistem penilaian yang lebih akurat, efisien, dan mampu memberikan umpan balik yang bermanfaat bagi mahasiswa maupun dosen. Selain itu, penelitian ini diharapkan dapat memberikan kontribusi signifikan dalam pengembangan teknologi penilaian otomatis berbasis Natural Language Processing (NLP) di bidang pendidikan. Kata kunci: E-Learning, BERT, Penilaian Otomatis, Forum Diskusi, Teknologi Pendidikan.", "source_url": "https:\/\/openalex.org\/W7126128896", "pdf_url": "", "doi": "", "handle": "", "repository_name": "University of Bengkulu Scholar Repository (University of Bengkulu)", "source_record_id": "W7126128896", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:53:00", "updated_at": "2026-11-15 16:53:00" }, { "id": 1262, "title": "Exploration of fine-tuning and inference time of large pre-trained language models in NLP", "title_ar": "", "title_en": "Exploration of fine-tuning and inference time of large pre-trained language models in NLP", "author": "Анна Юрьевна Короткова", "supervisor": "", "degree_type": "master", "degree_evidence": "نص السجل: «master's»", "university": "Ludwid Maxmilian's Universitat Munchen", "university_source": "repository_host", "college": "", "department": "", "country": "Germany", "governorate": "", "publication_year": 2020, "language": "en", "research_domain": "nlp", "research_subdomain": "llm", "categories": "machine-learning; deep-learning; nlp; transfer-learning; transformers; llm", "keywords": "", "abstract": "In this master's thesis, we perform an analysis based on the language models BERT (Devlin et al., 2019), RoBERTa (Liu et al., 2019), DistilBERT (Sanh et al., 2020), ALBERT (Lan et al., 2019) and XLNET (Yang et al., 2019) whose architectures are based on the Transformer encoder which relies solely on an attention mechanism.We incorporate a measuring of the fine-tuning and inference time in the Python module transformers which was implemented by huggingface (Wolf et al., 2019).In the scope of this master's thesis, the models are fine-tuned on the General Language Understanding Evaluation tasks (Wang et al., 2019), a collection of nine sentenceor sentence-pair language understanding tasks.To monitor our analyses, we used the tool \"Weights & Biases\" (wandb) (Biewald, 2020) where a logging of hyperparameters and output metrics is possible.The goal of this master's thesis is to develop a diverse benchmark setting and evaluate the respective fine-tuning as well as inference times across different hyperparameters, language models and tasks.More precisely, a hyperparameter sweep with the maximum sequence length and the batch size are regarded, taking on the values {128, 256, 512} and {8, 16, 32}, respectively.We show that the maximum sequence length, the task size as well as the model have an effect on the fine-tuning time, whereas the batch size seems to be of secondary importance.Moreover, the maximum sequence length, the batch size and the choice of the model affect the inference time, whereas the task itself does not show such results.", "source_url": "https:\/\/openalex.org\/W3134720031", "pdf_url": "https:\/\/epub.ub.uni-muenchen.de\/75187\/", "doi": "10.5282\/ubm\/epub.75187", "handle": "", "repository_name": "Open access LMU (Ludwid Maxmilian's Universitat Munchen)", "source_record_id": "W3134720031", "other_sources": "", "verification_status": "verified", "created_at": "2026-11-15 16:53:00", "updated_at": "2026-11-15 16:53:00" }, { "id": 1264, "title": "Microbial named entity recognition using BERT models", "title_ar": "", "title_en": "Microbial named entity recognition using BERT models", "author": "Brian K Rao", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "University of Illinois at Urbana-Champaign", "university_source": "repository_host", "college": "", "department": "", "country": "United States", "governorate": "", "publication_year": 2022, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; nlp; medical-ai; ai-applications; transfer-learning; data-mining; transformers; ai-industry", "keywords": "", "abstract": "Bacteria are critical subjects of microbiological research that span many rapidly-growing fields of study.The development and widespread application of high-throughput sequencing has led to more microbial data being collected in recent years than ever before.This study investigates the capabilities of the popular Natural Language Processing (NLP) model Bidirectional Encoder Representations from Transformers (BERT) on the relatively understudied text mining domain of microbiology.This is done by fine-tuning a variety of BERT models (BERT, DistilBERT, SciBERT, BioBERT, PubMedBERT) on the Bacteria Biotope 2019 Open Shared Task (BB2019-OST) corpus of annotated microbial research text and evaluating the best performing models on the Named Entity Recognition (NER) task.Following this, an in-depth error analysis was conducted to gain insights into BERT's entity recognition capabilities.Finally, to investigate performance capabilities further, learning rate and batch size hyperparameters were tuned to increase F1-score.The best BERT model in the comparison was BioBERT, earning an F1-score of 73.82 (±1.04) with default hyperparameters, and 75.35 (±0.62) with tuned hyperparameters.BioBERT had better F1-scores and entity-level statistics, despite PubMedBERT ranking high in biomedical NLP benchmarks.This suggests that the generality of the pretraining corpora of BERT models is particularly important for text mining in the microbial domain.", "source_url": "https:\/\/openalex.org\/W7220763410", "pdf_url": "https:\/\/www.ideals.illinois.edu\/items\/125125\/bitstreams\/410938\/data.pdf", "doi": "", "handle": "", "repository_name": "Illinois Digital Environment for Access to Learning and Scholarship (University of Illinois at Urbana-Champaign)", "source_record_id": "W7220763410", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:53:00", "updated_at": "2026-11-15 16:53:00" }, { "id": 1265, "title": "Sentiment analysis for portuguese with BERT models", "title_ar": "", "title_en": "Sentiment analysis for portuguese with BERT models", "author": "Vinícius Pinheiro Winter", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Federal University of Technology – Paraná", "university_source": "repository_host", "college": "", "department": "", "country": "Brazil", "governorate": "", "publication_year": 2024, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; computer-vision; nlp; medical-ai; transfer-learning; transformers; sentiment-analysis", "keywords": "", "abstract": "This paper presents the development and application of the BERT model (Bidirectional Encoder Representations from Transformers) for sentiment analysis in Portuguese texts, focusing on data from social networks. The study is justified by the need to understand the sentimental nuances expressed in a language with limited NLP resources, such as Portuguese, and the relevance of analyzing sentiments on widely used platforms. The objectives center on training and validating a Portuguese-adapted BERT model, using a manually classified dataset, as well as another dataset collected automatically without classification labels. The methodology involved fine-tuning BERT to adjust it to the linguistic peculiarities of Portuguese, followed by rigorous validation that included comparisons with classic NLP algorithms and extrinsic analyses, such as the generation of word clouds to visualize the frequencies of words associated with different sentiments. The results demonstrate an accuracy of 86.64% in the BERTimbau model, which was chosen due to its pre-training in Portuguese and its superior performance in tests, thus evidencing its effectiveness in classification and its relevance for practical applications. It was also observed through word cloud analysis that a large part of the classifications of the Reddit dataset was effective due to the presence of words commonly associated with the sentiments to which they were classified. This work contributes to the field of NLP by providing a Portuguese-adapted model for sentiment analysis, as well as valuable perspectives for future research and applications in various areas, including behavioral analysis and mental health monitoring.", "source_url": "https:\/\/openalex.org\/W7155474077", "pdf_url": "http:\/\/repositorio.utfpr.edu.br:8080\/jspui\/bitstream\/1\/36753\/1\/analisesentimentosmodelobert.pdf", "doi": "", "handle": "", "repository_name": "Institutional Repository of the Federal Technological University of Paraná (RIUT) (Federal University of Technology – Paraná)", "source_record_id": "W7155474077", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:53:00", "updated_at": "2026-11-15 16:53:00" }, { "id": 1268, "title": "Using Bidirectional Encoder Representations from Transformers for Conversational Machine Comprehension", "title_ar": "", "title_en": "Using Bidirectional Encoder Representations from Transformers for Conversational Machine Comprehension", "author": "Evangelina Gogoulou", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "KTH Royal Institute of Technology", "university_source": "repository_host", "college": "", "department": "", "country": "Sweden", "governorate": "", "publication_year": 2019, "language": "en", "research_domain": "nlp", "research_subdomain": "llm", "categories": "machine-learning; deep-learning; nlp; responsible-ai; predictive-analytics; transformers; llm; xai", "keywords": "", "abstract": "Bidirectional Encoder Representations from Transformers (BERT) is a recently proposed language representation model, designed to pre-train deep bidirectional representations, with the goal of extracting context-sensitive features from an input text [1]. One of the challenging problems in the field of Natural Language Processing is Conversational Machine Comprehension (CMC). Given a context passage, a conversational question and the conversational history, the system should predict the answer span of the question in the context passage. The main challenge in this task is how to effectively encode the conversational history into the prediction of the next answer. In this thesis work, we investigate the use of the BERT language model for the CMC task. We propose a new architecture, named BERT-CMC, using the BERT model as a base. This architecture includes a new module for encoding the conversational history, inspired by the Transformer-XL model [2]. This module serves the role of memory throughout the conversation. The proposed model is trained and evaluated on the Conversational Question Answering dataset (CoQA) [3]. Our hypothesis is that the BERT-CMC model will effectively learn the underlying context of the conversation, leading to better performance than the baseline model proposed for CoQA. Our results of evaluating the BERT-CMC on the CoQA dataset show that the model performs poorly (44.7% F1 score), comparing to the CoQA baseline model (66.2% F1 score). In the light of model explainability, we also perform a qualitative analysis of the model behavior in questions with various linguistic phenomena eg coreference, pragmatic reasoning. Additionally, we motivate the critical design choices made, by performing an ablation study of the effect of these choices on the model performance. The results suggest that fine tuning the BERT layers boost the model performance. Moreover, it is shown that increasing the number of extra layers on top of BERT leads to bigger capacity of the conversational memory.", "source_url": "https:\/\/openalex.org\/W3046466516", "pdf_url": "", "doi": "", "handle": "", "repository_name": "KTH Publication Database DiVA (KTH Royal Institute of Technology)", "source_record_id": "W3046466516", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:53:00", "updated_at": "2026-11-15 16:53:00" }, { "id": 1270, "title": "Analysis of fine-tuning techniques in text classification", "title_ar": "", "title_en": "Analysis of fine-tuning techniques in text classification", "author": "Tobias Gonçalves Pires", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "", "university_source": "", "college": "", "department": "", "country": "", "governorate": "", "publication_year": 2024, "language": "en", "research_domain": "nlp", "research_subdomain": "llm", "categories": "machine-learning; deep-learning; nlp; transfer-learning; transformers; llm", "keywords": "", "abstract": "Natural Language Processing (NLP) aims to develop models that enable computers to understand, interpret, process and generate text in a way similar to human communication. The last decade has seen significant advances in the field, with the introduction of deep neural network models, and the subsequent evolution of the architecture of these models such as the attention mechanism and the Transformers architecture, culminating in language models such as ELMo, BERT and GPT. And later models called Large Language Models (LLMs) improved the ability to understand and generate texts in a sophisticated way. Pre-trained models offer the advantage of reusing knowledge accumulated from vast datasets, although specific fine-tuning is required for individual tasks. However, training and tuning these models consumes a lot of processing resources, making it unfeasible for many organizations due to high costs. In resource-constrained environments, efficient fine-tuning techniques such as LoRA (Low-Rank Adaptation) were developed to optimize the model adaptation process, minimizing the number of adjustable parameters and avoiding overfitting. These techniques allow for faster and more economical training, while maintaining the robustness and generalization of the models. This work evaluates three efficient fine-tuning techniques LoRA, AdaLoRA and IA3 (in addition to full fine-tuning) in terms of memory consumption, training time and accuracy, using the DistilBERT, Roberta-base and TinyLlama models on different datasets (AG News, IMDb and SNLI).", "source_url": "https:\/\/openalex.org\/W7120873934", "pdf_url": "", "doi": "", "handle": "", "repository_name": "LA Referencia (Red Federada de Repositorios Institucionales de Publicaciones Científicas)", "source_record_id": "W7120873934", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:53:00", "updated_at": "2026-11-15 16:53:00" }, { "id": 1279, "title": "ENHANCED DEEP LEARNING NEURAL NETWORKS FOR CLASSIFICATION AND FORECASTING PROBLEMS IN EMBEDDED SYSTEMS", "title_ar": "", "title_en": "ENHANCED DEEP LEARNING NEURAL NETWORKS FOR CLASSIFICATION AND FORECASTING PROBLEMS IN EMBEDDED SYSTEMS", "author": "Pentsos, Vasileios", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Southern Illinois University Carbondale", "university_source": "repository_host", "college": "", "department": "", "country": "United States", "governorate": "", "publication_year": 2025, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; computer-vision; ai-applications; predictive-analytics; cnn; rnn; transformers; image-classification; ai-smart-cities", "keywords": "", "abstract": "The widespread adoption of deep learning has led to a surge in demand for the efficient deployment of Deep Neural Networks (DNNs) on embedded and resource-constrained systems. These platforms are increasingly tasked with performing complex inference workloads in real time, yet they must do so under strict limitations on energy consumption, computation, and memory resources. As DNN models grow in complexity to support advanced applications such as time-series forecasting and image classification, ensuring their efficient execution without compromising performancebecomes a key design challenge. This calls for tailored algorithmic and architectural solutions that can meet application-specific constraints while maximizing the benefits of modern hardware accelerators.This dissertation proposes a set of algorithmic-based approaches aimed at enhancing the efficiency and adaptability of deep learning models for embedded systems. The research centers on three primary contributions: (i) a hybrid Long Short-Term Memory (LSTM)-Transformer architecture optimized for multi-step residential power load forecasting, integrating sequence modeling and attention mechanisms for improved accuracy under training time constraints, (ii) a lightweight hierarchical deep neural network enhancement that augments baseline classifiers through cascading binaryConvolutional Neural Networks (CNNs) and Vision Transformers, achieving higher classification accuracy and reduced inference time, and (iii) an energy-aware scheduling framework for DNN inference, featuring dynamic batch size selection, GPU frequency adjustment, and concurrent task mapping across multi-GPU systems to minimize energy use without violating real-time deadlines.In summary, this dissertation presents a comprehensive approach to designing and managing deep learning workflows on embedded platforms, emphasizing performance-efficiency trade-offs. The proposed methods demonstrated significantimprovements in both predictive accuracy and resource utilization, compared to existing solutions, offering practical pathways for deploying DNNs in sustainable, computing environments.", "source_url": "https:\/\/openalex.org\/W7111621194", "pdf_url": "", "doi": "", "handle": "", "repository_name": "OpenSIUC (Southern Illinois University Carbondale)", "source_record_id": "W7111621194", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:53:00", "updated_at": "2026-11-15 16:53:00" }, { "id": 1285, "title": "Image Denoising via Deep Learning Neural Networks and Transformers", "title_ar": "", "title_en": "Image Denoising via Deep Learning Neural Networks and Transformers", "author": "Jundong Wang", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Tampere University", "university_source": "repository_host", "college": "", "department": "", "country": "Finland", "governorate": "", "publication_year": 2024, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "deep-learning; computer-vision; transformers", "keywords": "image denoising, deep learning, residual learning, Transformers.", "abstract": "This thesis investigates the implementation of image denoising using Neural Networks (NN) in conjunction with wavelet transform techniques. For the purpose of data simulation, images are intentionally corrupted with additive white Gaussian noise (AWGN) at various noise levels. Additionally, the study includes the examination of both random noise and real-world noise scenarios. This research has an emphasis on implementing several novel neural network architectures and strategies. By integrating the wavelet transform, the research aims to leverage its multi-resolution analysis capability to further enhance denoising performance. The proposed methods underscore the ability of residual learning and Transformers to integrate both local and global features, which efficiently capture fine-grained local details and long-range global dependencies. Keywords: image denoising, deep learning, residual learning, Transformers.", "source_url": "https:\/\/openalex.org\/W7133094216", "pdf_url": "https:\/\/trepo.tuni.fi\/bitstream\/10024\/160429\/2\/WangJundong.pdf", "doi": "", "handle": "", "repository_name": "Tampere University Institutional Repository (Tampere University)", "source_record_id": "W7133094216", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:53:00", "updated_at": "2026-11-15 16:53:00" }, { "id": 1326, "title": "Essays on Time Series Forecasting and Machine Learning", "title_ar": "", "title_en": "Essays on Time Series Forecasting and Machine Learning", "author": "Frank, Johannes", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "", "university_source": "", "college": "", "department": "", "country": "", "governorate": "", "publication_year": 2025, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; ai-applications; predictive-analytics; rnn; transformers; ai-industry", "keywords": "", "abstract": "This dissertation investigates how modern machine learning (ML) methods, with a particular focus on multi-task learning (MTL) frameworks, can enhance forecasts in financial and macroeconomic contexts. Chapter 1 outlines the limitations of traditional econometric approaches in environments characterized by nonlinearity, structural breaks, and mixed-frequency data, and motivates the use of ML models capable of exploiting shared information across related prediction tasks. Chapter 2 evaluates the performance of temporal fusion transformers (TFT) for forecasting realized volatility (RV) of US stocks. It shows that models with sector-level pooling significantly outperform asset-specific or fully pooled models, and that TFT surpasses traditional volatility models and common ML benchmarks, especially in volatile market phases. Chapter 3 extends this approach by introducing a two-dimensional MTL framework to jointly forecast RV across industry portfolios and forecast horizons. Using long short-term memory (LSTM) models with structured parameter sharing, the study finds that sharing information across both dimensions improves forecast accuracy, particularly at short horizons. Chapter 4 applies MTL to macroeconomic nowcasting using disaggregated US labor market and inflation data. Results show that MTL enhances nowcast accuracy compared to single-task and traditional models, with the optimal degree of information sharing depending on task-specific heterogeneity. Chapter 5 embeds MTL into a mixed frequency data sampling (MIDAS) setting by combining it with different MIDAS techniques to nowcast sectoral employment using high-frequency indicators. Nonlinear MTL models, particularly those based on extreme gradient boosting (XGBoost), outperform benchmarks and highlight the value of preserving intra-month dynamics. Overall, the dissertation demonstrates that combining MTL with flexible ML architectures offers a powerful framework for improving forecast quality and stability in complex economic settings.", "source_url": "https:\/\/openalex.org\/W7109568429", "pdf_url": "", "doi": "10.25593\/open-fau-2528", "handle": "", "repository_name": "", "source_record_id": "W7109568429", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:53:01", "updated_at": "2026-11-15 16:53:01" } ] }