{ "exported_at": "2026-11-15 19:45:26", "count": 68, "records": [ { "id": 53, "title": "Evaluating word-vector generation techniques for use in semantic role labeling with recurrent LSTM neural networks", "title_ar": "", "title_en": "Evaluating word-vector generation techniques for use in semantic role labeling with recurrent LSTM neural networks", "author": "Daniel Toom", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "", "university_source": "", "college": "", "department": "", "country": "", "governorate": "", "publication_year": 2017, "language": "en", "research_domain": "nlp", "research_subdomain": "rnn", "categories": "machine-learning; deep-learning; nlp; self-semi-supervised; rnn", "keywords": "", "abstract": "This is an investigation of how the performance on a natural language task, semantic role labeling (SRL), is affected by how the input text is encoded. Four methods for encoding words as dense vectors of floating-point numbers are evaluated: Noisecontrastive estimation (NCE), Continuous bag-of-words (CBOW), Skip-grams and Global vectors for word representation (GloVe). Vectors are generated using a corpus of Wikipedia articles as the training set, with different values for various parameters, such as vector length, context length and training methods. In order to evaluate the generated vectors, they are then used as the input into a recurrent bi-directional neural network using LSTM neurons. This network is trained using the standard CoNLL-2005 shared task data set, and evaluated with the associated test sets. The results show that there is no large or consistent advantage to using word encodings from any of the tried methods or any of the tried parameter settings over any of the other methods or parameter settings. The SRL system that was used thus seems to be fairly robust in regard to the choice of input vectors. All methods generate vectors that outperform random vectors, however, indicating that pretraining vectors has a positive effect. The labeling accuracy is close to but slightly worse than previously published state-of-the-art performance results from the use of a similar network on the SRL task.", "source_url": "https:\/\/openalex.org\/W2745353936", "pdf_url": "", "doi": "", "handle": "", "repository_name": "", "source_record_id": "W2745353936", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:46", "updated_at": "2026-11-15 16:52:46" }, { "id": 66, "title": "Deep Learning for Medical Image Segmentation", "title_ar": "", "title_en": "Deep Learning for Medical Image Segmentation", "author": "Binyan Hu", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "", "university_source": "", "college": "", "department": "", "country": "Australia", "governorate": "", "publication_year": 2026, "language": "en", "research_domain": "medical-ai", "research_subdomain": "medical-imaging", "categories": "machine-learning; deep-learning; computer-vision; medical-ai; responsible-ai; ai-applications; self-semi-supervised; image-segmentation; medical-imaging; federated-learning; fairness; ai-cybersecurity", "keywords": "", "abstract": "Medical image segmentation (MIS) plays a fundamental role in a wide range of medical image analysis applications. Over the past decade, MIS has been significantly advanced by deep learning (DL). However, the effectiveness of DL-based methods typically depends on large-scale, precisely annotated datasets, which are often unavailable in the medical domain due to the high cost and expertise required for medical data acquisition and annotation. When trained on limited data, deep neural networks (DNNs) are prone to over-fitting, resulting in poor generalisation performance when deployed on previously unseen data. In MIS, this generalisation gap is exacerbated by two key challenges: 1) data scarcity: due to the limited quantity of annotated images available for training, over-fitting the training data causes poor generalisation performance on images unseen during training; 2) expert-specific annotation: variability in annotations across experts arises from the inherent complexity of medical images and differences in expert knowledge. Consequently, a model trained on certain experts' annotations may under-perform when deployed in clinical contexts with new images and new experts. This thesis aims to address these domain-specific limitations, with an overarching goal of generating enhanced supervision signals to improve the generalisation capability of DNNs for MIS across both images and user experts under constrained data settings. Toward achieving this goal, our contributions are threefold.First, we tackle the expert-specific annotation problem by leveraging the supervision from multiple experts via multi-expert annotation, aiming to achieve adaptive MIS that generalises to new experts' annotations. Through extensive empirical analysis, we investigate when and how multi-expert annotation improves a DNN model's generalisation capability to new experts' annotations and what DNN training method can best achieve such generalisation. Our findings provide detailed evidence for the efficacy of multi-expert annotation and practical insights into developing generalisable models that can efficiently adapt to new experts in the presence of expert-specific annotations.Second, we intend to mitigate the impact of data scarcity and improve the model's generalisation to new images by leveraging self-supervision.We develop an adaptive training method, which generates diverse self-supervised tasks from the existing annotated dataset, dynamically selects the most beneficial combination of tasks, and adaptively transfers the knowledge learnt from them to enrich the segmentation model's knowledge.Compared to existing approaches, our method integrates more diverse self-supervised tasks and offers superior effectiveness and efficiency in exploiting their supervision signals to benefit segmentation performance.The third study moves beyond the single-site setting and generates enhanced supervision via multi-site collaboration, where the goal is to improve generalisation performance for participating sites while navigating strict privacy constraints and potential security threats arising from non-transparent local processes. We propose a new federated learning method, tailored for the multi-site MIS scenario, to enable privacy-preserving defence against poisoning attacks by sustaining robust performance and localising the compromised site. Compared to previous approaches, our method provides utility-based defence and diagnosis, does not depend on public datasets, and works for the setting with few participating sites. These features enable more practical and effective federated learning with enhanced privacy preservation and security for multi-site MIS.Extensive experiments across diverse datasets and clinical settings demonstrate the effectiveness of the proposed methods in improving DNN model generalisation for MIS tasks. Collectively, this thesis contributes practical methodologies to address domain-specific challenges in MIS and advances the practical deployment of DNNs in real-world clinical environments.", "source_url": "https:\/\/openalex.org\/W7161281791", "pdf_url": "", "doi": "10.25916\/sut.32150593", "handle": "", "repository_name": "Swinburne figshare (Swinburne University of Technology)", "source_record_id": "W7161281791", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:46", "updated_at": "2026-11-15 16:52:46" }, { "id": 67, "title": "Deep-Learning-Based 3D Medical Image Segmentation and Registration", "title_ar": "", "title_en": "Deep-Learning-Based 3D Medical Image Segmentation and Registration", "author": "Hao Xu", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "The University of Sydney", "university_source": "repository_host", "college": "", "department": "", "country": "Australia", "governorate": "", "publication_year": 2025, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; computer-vision; medical-ai; few-zero-shot; self-semi-supervised; transformers; multimodal; image-segmentation; medical-imaging", "keywords": "", "abstract": "Understanding the importance of image segmentation and registration in clinical diagnosis and treatment is critical for advancing precision medicine. Medical image analysis involves complex data processing tasks, particularly in scenarios with limited annotated data and multi-modality feature inconsistencies. Recent deep learning-based methods have provided breakthroughs in image segmentation and registration, especially in modalities like MRI and CT. However, challenges such as one-shot learning, incremental learning, and efficient deformable registration still persist. This thesis introduces a series of innovative deep learning frameworks, focusing on three one-shot segmentation settings and three Segment Anything Model (SAM) -assisted registration methods, to address the issues of annotation scarcity and feature misalignment in medical imaging. For segmentation, we develop three methods for one-shot, class-incremental, and novel-class tract segmentation. These incorporate uncertainty estimation, knowledge distillation, and voxel-level contrastive learning to enhance performance and generalization in one-shot settings. For registration, we propose three SAM-guided frameworks that leverage segmentation masks and anatomical prompts, as well as attention mechanisms, prototype learning, and contour-aware losses. These approaches achieve state-of-the-art accuracy in unsupervised deformable registration. Overall, our methods yield substantial improvements in segmentation and registration accuracy across diverse datasets and demonstrate strong potential for clinical application.", "source_url": "https:\/\/openalex.org\/W7126877936", "pdf_url": "", "doi": "", "handle": "", "repository_name": "The Sydney eScholarship Repository (The University of Sydney)", "source_record_id": "W7126877936", "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": 88, "title": "Semi-Supervised Deep Learning using Consistency-Based Methods for Segmentation of Medical Images", "title_ar": "", "title_en": "Semi-Supervised Deep Learning using Consistency-Based Methods for Segmentation of Medical Images", "author": "Axel Rönnberg", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "KTH Royal Institute of Technology", "university_source": "repository_host", "college": "", "department": "", "country": "Sweden", "governorate": "", "publication_year": 2020, "language": "en", "research_domain": "medical-ai", "research_subdomain": "medical-imaging", "categories": "machine-learning; deep-learning; computer-vision; medical-ai; ai-applications; self-semi-supervised; predictive-analytics; image-segmentation; medical-imaging; ai-education", "keywords": "", "abstract": "In radiation therapy, a form of cancer treatment, accurately locating the anatomical structures is required in order to limit the impact on healthy cells. The automatic task of delineating these structures and organs is called segmentation, where each pixel in an image is classified and assigned a label. Recently, deep neural networks have proven to be efficient at automatic medical segmentation. However, deep learning requires large amounts of training data. This is a restricting feature, especially in the medical field due to factors such as patient confidentiality. Nonetheless, the main challenge is not the image data itself but the lack of high-quality annotations. It is thus interesting to investigate methods for semi-supervised learning, where only a subset of the images re- quires annotations. This raises the question if these methods can be acceptable for organ segmentation, and if they will result in an increased performance in comparison to supervised models. A category of semi-supervised methods applies the strategy of encouraging consistency between predictions. Consistency Training and Mean Teacher are two methods in which the network weights are updated in order to minimize the impact of input perturbations such as data augmentations. In addition, the Mean Teacher method trains two models, a Teacher and a Student. The Teacher is updated as an average of consecutive Student models, using Temporal Ensembling. To resolve the question whether semi-supervised learning could be beneficial, the two mentioned techniques are investigated. They are used in training deep neural networks with an U-net architecture to segment the bladder and anorectum in 3D CT images. The results showed signs of promise for Consistency Training and Mean Teacher, with nearly all model configurations having improved segmentations. Results also showed that the methods caused a reduction in performance variance, primarily by limiting poor delineations. With these results in hand, the use of semi-supervised learning should definitely be considered. However, since the segmentation improvement was not repeated in all experiment configurations, more research needs to be done.", "source_url": "https:\/\/openalex.org\/W3111652972", "pdf_url": "", "doi": "", "handle": "", "repository_name": "KTH Publication Database DiVA (KTH Royal Institute of Technology)", "source_record_id": "W3111652972", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:46", "updated_at": "2026-11-15 16:52:46" }, { "id": 92, "title": "Explainable Artificial Intelligence in the Image Domain and its Applications to the Medical Field", "title_ar": "", "title_en": "Explainable Artificial Intelligence in the Image Domain and its Applications to the Medical Field", "author": "Mirtha Lucas", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "DePaul University", "university_source": "repository_host", "college": "", "department": "", "country": "United States", "governorate": "", "publication_year": 2026, "language": "en", "research_domain": "medical-ai", "research_subdomain": "medical-imaging", "categories": "machine-learning; deep-learning; computer-vision; generative-ai; medical-ai; responsible-ai; self-semi-supervised; predictive-analytics; cnn; gan; medical-imaging; xai; fairness", "keywords": "", "abstract": "This dissertation investigates the development of Explainable Artificial Intelligence (XAI) methods for deep learning models in the image domain, with a particular focus on medical imaging applications. Although neural networks achieve high predictive performance, their lack of interpretability limits their adoption in critical domains such as healthcare, where transparency and trust are essential. This work addresses this challenge by proposing novel approaches that improve the interpretability and reliability of model predictions. A primary contribution is the introduction of Riemann–Stieltjes Integrated Grad-CAM (RSI Grad-CAM), a gradient-based attribution method that generates more relevant and spatially localized saliency maps. The method is evaluated using quantitative metrics and statistical analysis, demonstrating improved performance over established techniques such as Grad-CAM and Integrated Grad-CAM. A secondary contribution is the analysis of the interpretability of intermediate layers in convolutional neural networks, demonstrating that clinically meaningful features, such as spiculation in lung nodules, can be identified within specific layers of the model. A third contribution is the development of a human-centered framework based on a Variational Autoencoder-Auxiliary Classifier Wasserstein Generative Adversarial Network (VAE-AC-WGAN), which generates contrastive explanations aligned with clinical reasoning. Experimental results demonstrate that the proposed method produces higher-quality reconstructions and more informative anomaly maps than baseline approaches (e.g., medXGAN), as reflected in improved structural similarity (SSIM) and classifier-based evaluation metrics (AUROC). Furthermore, human evaluation indicates that the generated explanations better capture clinically relevant features, such as spiculation, supporting their interpretability in medical imaging contexts. Overall, this work advances XAI by enhancing model transparency and supporting the trustworthy integration of AI systems into medical decision-making.", "source_url": "https:\/\/openalex.org\/W7172068913", "pdf_url": "", "doi": "", "handle": "", "repository_name": "Digital Commons-DePaul (DePaul University)", "source_record_id": "W7172068913", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:46", "updated_at": "2026-11-15 16:52:46" }, { "id": 94, "title": "A Novel Explainable Artificial Intelligence Framework with Improved Learning Mechanisms", "title_ar": "", "title_en": "A Novel Explainable Artificial Intelligence Framework with Improved Learning Mechanisms", "author": "Kabir, Sami", "supervisor": "", "degree_type": "phd", "degree_evidence": "نص السجل: «doctoral thesis»", "university": "Luleå University of Technology", "university_source": "repository_host", "college": "", "department": "", "country": "Sweden", "governorate": "", "publication_year": 2025, "language": "en", "research_domain": "medical-ai", "research_subdomain": "cnn", "categories": "machine-learning; deep-learning; computer-vision; medical-ai; responsible-ai; expert-systems; ai-applications; transfer-learning; self-semi-supervised; predictive-analytics; optimization; cnn; multimodal; xai; fairness; ai-smart-cities", "keywords": "", "abstract": "Artificial Intelligence (AI) models are increasingly being deployed in various critical domains, such as healthcare, finance, law, and autonomous systems, resulting in remarkable benefits for the society. Due to black-box (sub-symbolic) nature, these AI models do not provide explanation of the predictive output. This lack of explanation results in lack of transparency between human and machine. Such opacity of black-box AI models, particularly deep learning architectures, is a significant concern. To address this concern, Explainable Artificial Intelligence (XAI) has emerged as a vital research area. The aim of XAI is to enhance transparency, trust, and accountability of AI models. Various post-hoc XAI tools explain the outputs of black-box AI models based on training datasets. However, the use of training datasets, rather than domain knowledge, makes such explanations proxy. A biased training dataset may lead to a misleading post-hoc explanation as well. In contrast, an explanation based on domain knowledge will be more trustworthy to an end user. Motivated by this, we propose a novel XAI framework, consisting of Belief Rule Based Expert System (BRBES), to predict output and explain it with reference to domain knowledge. BRBES represents domain knowledge with its rule base. It outperforms other knowledge-driven AI models to handle uncertainty due to ignorance. To improve the accuracy of BRBES, we fine-tune its parameters and structure through evolutionary learning. Moreover, to overcome the scarcity of labeled dataset in this learning mechanism, we integrate semi-supervised and self-supervised learning with BRBES. We explain the output of BRBES with respect to the most influential rule of the rule base of BRBES. Thus, the output of our proposed XAI framework becomes not only accurate, but also explainable. This doctoral thesis delves into the challenges and opportunities surrounding XAI in order to provide a comprehensive understanding of the AI output. It presents a novel XAI framework to provide domain knowledge based energy consumption prediction with improved accuracy, and explain this predictive output in non-technical human language. Then it deals with multi-modal air quality data by integrating deep learning model with BRBES. Moreover, to reduce dependence on labeled data for evolutionary learning, this thesis integrates semi-supervised and self-supervised learning with BRBES. This thesis presents six significant contributions. First, we conduct a Systematic Literature Review (SLR) on XAI using the PRISMA guidelines, delving into the numerous challenges and opportunities of XAI. Extensive research is conducted to explore the definition, terminologies, taxonomy, and application domains of XAI. It highlights various challenges of XAI, such as, no universal definition, trade-off between accuracy and explainability, and lack of standardized evaluation metrics. To address this lack of standardized evaluation metrics, we also propose a unified framework to evaluate XAI. Secondly, we present an innovative explainable BRBES (eBRBES) framework, which offers accurate prediction of building energy consumption phenomenon, while providing insightful explanation and counterfactual based on domain knowledge. As part of eBRBES framework, we also present a novel Belief Rule Based adaptive Balance Determination (BRBaBD) algorithm to assess the optimal balance of the proposed framework between explainability and accuracy. Thirdly, we propose a mathematical model to integrate BRBES with the Convolutional Neural Network (CNN). We leverage the domain knowledge of BRBES, and image data patterns discovered by CNN with this integrated approach. We predict air quality with this integrated model using outdoor ground images and numerical sensor data. Fourth, we integrate two-layer BRBES with CNN to monitor air quality using satellite images, and relevant environmental parameters, such as, cloud, relative humidity, temperature, and wind speed. The two-layer BRBES showcases the strength of BRBES in conducting reasoning across multiple layers. Fifth, we enhance the learning mechanism of BRBES by utilizing the extensive unlabeled energy data, along with limited labeled data. For this purpose, we synthetically generate unlabeled data through weak and strong augmentation. We then pseudo-label these unlabeled data by integrating semi-supervised self-training model with BRBES. By learning from both labeled and pseudo-labeled data, initial BRBES transitions to semi-supervised BRBES. Sixth, to conduct the learning mechanism of BRBES without relying on any labeled data, we integrate self-supervised learning with it. To this effect, we pseudo-label the synthetically generated unlabeled energy data with BRBES in the pretext tasks of self-supervised learning. Then, the initial BRBES learns deeper representation from these pseudo-labeled data, and transitions to self-supervised BRBES.", "source_url": "https:\/\/openalex.org\/W7112031890", "pdf_url": "https:\/\/ltu.diva-portal.org\/smash\/get\/diva2:2007851\/FULLTEXT01", "doi": "", "handle": "", "repository_name": "DiVA (Luleå University of Technology)", "source_record_id": "W7112031890", "other_sources": "", "verification_status": "verified", "created_at": "2026-11-15 16:52:46", "updated_at": "2026-11-15 16:52:46" }, { "id": 111, "title": "Explainable artificial intelligence (XAI) for healthcare decision-making", "title_ar": "", "title_en": "Explainable artificial intelligence (XAI) for healthcare decision-making", "author": "Zhiwei Zeng", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Nanyang Technological University", "university_source": "repository_host", "college": "", "department": "", "country": "Singapore", "governorate": "", "publication_year": 2022, "language": "en", "research_domain": "medical-ai", "research_subdomain": "xai", "categories": "machine-learning; medical-ai; responsible-ai; ai-applications; self-semi-supervised; xai; fairness; ai-cybersecurity", "keywords": "", "abstract": "With the increasing ubiquity of Artificial Intelligence (AI) in daily life and into more critical applications, explainability has been brought under the spotlight due to ethical, social, and legal concerns. Explainability is particularly important in application domains involving high-stakes decision-making, such as healthcare, medicine, law, and defence, as simply having high performance may not satisfy all the desiderata of stakeholders. Stakeholders in these domains often require greater insights into the decision-making process, better support for tracing decision trails and identifying errors, as well as greater trustworthiness from an AI system. An eXplainable AI (XAI) is one that has the ability to explain the outputs or functioning of a model in a clear and understandable manner to a human. In the explanation process of XAI, explainability approaches provide explanatory information to human stakeholders in order to facilitate their understanding of the outputs and functioning of an AI model. The renewed understanding then affects the extent to which the desiderata are satisfied. Over the past decade, a plethora of explainability approaches have been proposed, ranging from ante-hoc methods that seek to build explainability features into a model itself to post-hoc methods that use techniques to improve the explainability of an already developed model. Despite the rapid growth of the XAI field, three key open problems emerge from a review of the existing XAI methods. Firstly, post-hoc methods have been increasingly used for high-stakes decisions, but most of them provide explanatory information that is not faithful or not insightful to what the original model computes. Secondly, most existing methods can only provide associational explanations and lack the structure and reasoning ability to provide contrastive and counterfactual explanations. Thirdly, most existing methods do not provide explanations to suit different levels of understanding. Research problem 1 - Post-hoc explainability methods have quickly gained popularity as they are mostly model-agnostic and highly extensible. However, without elucidating or understanding the model’s inner workings, explanations generated by post-hoc methods may not faithfully represent the underlying relationships of the model and often do not provide enough information to facilitate the understanding of how the output is obtained. In this thesis, I propose an ante-hoc explainability methodology to support the making and explaining of high-stakes decisions in healthcare. It is an argumentative methodology that can infuse explainability into the model and generate explanations that are faithful and insightful to the inner workings and internal representations of the model. It follows a “debate-like” process that shares many similarities with the way humans deliberate and make decisions, thereby giving it unique advantages in generating human-understandable explanations. I instantiate the argumentative methodology in three explainable decision-making approaches for the modelling and reasoning of contexts, qualitative preferences as well as both attack and support relationships. Research problem 2 - Although contrastive and counterfactual explanations have attracted more research interest recently, existing XAI methods mostly generate associational explana- tions. Most existing methods are good at answering “Why X? ” but poor at answering questions that require higher-level inference, such as “Why not-X? ”. This can be attributed to the fact that most of them lack the necessary structures and mechanisms to perform high-level logical reasoning. Building on the structures and reasoning ability of the three proposed explain- able decision-making approaches, I construct different forms of contrastive and counterfactual explanations that can answer questions such as “Why X instead of Y? ” and “Why not-X? ”. Research problem 3 - Different levels of understanding are often required in most practical scenarios. However, most existing XAI methods do not provide explanations that can cater to the different levels of understanding. In this thesis, I propose and study formal properties of explanatory information in order to generate such explanations. I propose notions and computational constructs for generating explanations that can provide different amounts of justification or have different dialectical strengths, as well as focused explanations that only include selective information in the decision-making process. As a case study, I implement a proposed explainable decision-making approach for the diagnostics and prognostics of Alzheimer’s Disease (AD). Empirical evaluation with real-world AD datasets and human evaluations were conducted to assess the performance and explainability of the approach.", "source_url": "https:\/\/openalex.org\/W4224077403", "pdf_url": "https:\/\/dr.ntu.edu.sg\/bitstream\/10356\/155849\/4\/Thesis_final_Zhiwei%20Zeng_signed%20with%20watermark.pdf", "doi": "10.32657\/10356\/155849", "handle": "", "repository_name": "DR-NTU (Nanyang Technological University)", "source_record_id": "W4224077403", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:47", "updated_at": "2026-11-15 16:52:47" }, { "id": 137, "title": "Machine Learning-Based Network Intrusion Detection for Industrial Control", "title_ar": "", "title_en": "Machine Learning-Based Network Intrusion Detection for Industrial Control", "author": "Dehlaghi Ghadim, Alireza", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Mälardalen University College", "university_source": "repository_host", "college": "", "department": "", "country": "Sweden", "governorate": "", "publication_year": 2025, "language": "en", "research_domain": "responsible-ai", "research_subdomain": "federated-learning", "categories": "machine-learning; responsible-ai; ai-applications; self-semi-supervised; federated-learning; ai-cybersecurity; ai-industry; ai-smart-cities", "keywords": "", "abstract": "The growing connectivity of Industrial Control Systems (ICS) has increased their exposure to cyber threats, posing serious risks to both critical infrastructure (e.g., power plants, water systems) and industrial operations. Although Machine Learning (ML)-based Intrusion Detection Systems (IDS) show potential to detect complex and novel attacks, their practical application in ICS environments remains challenging. This thesis investigates the feasibility, design considerations, and barriers to applying ML-based IDS in ICSs. Key challenges include limited labeled data, high sensitivity to false alarms, distributed architectures, and constrained hardware. Through empirical evaluations and prototype implementations, we show how tailored ML and system design strategies can address these issues. Our first goal is to enable experimentation with cybersecurity and intrusion detection in industrial systems. The thesis introduces ICSSIM, a novel framework for creating flexible, scalable, and cost-effective ICS testbeds. It also presents ICSFlowGenerator, a tool for analyzing network traffic and computing customized network flow parameters. Using this tool, the thesis presents ICS-Flow, a new dataset developed to train and evaluate anomaly detection models, which serves as a realistic benchmark for assessing ML-based intrusion detection in ICS networks. In addition, the thesis tackles several technical deployment barriers. To support distributed ICS architectures and minimize the reliance on central servers, federated learning is explored as a decentralized training strategy. It also investigates semi-supervised learning techniques for detecting anomalies using only normal traffic. To reduce the burden of false alarms, we integrate a decision support system with the IDS to filter alerts and suggest mitigation actions. Furthermore, the thesis emphasizes the importance of temporal traffic patterns in identifying attack types and evaluates the efficiency and resource demands of various ML models on ICS-representative hardware. Collectively, these contributions advance the practical application of ML-based intrusion detection in ICSs.", "source_url": "https:\/\/openalex.org\/W7112960186", "pdf_url": "", "doi": "", "handle": "", "repository_name": "DiVA (Mälardalen University College)", "source_record_id": "W7112960186", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:47", "updated_at": "2026-11-15 16:52:47" }, { "id": 146, "title": "Evaluating Self-Supervised Learning for Network Intrusion Detection", "title_ar": "", "title_en": "Evaluating Self-Supervised Learning for Network Intrusion Detection", "author": "Bryan A. Marin Samano", "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": "transfer-learning", "categories": "machine-learning; deep-learning; ai-applications; transfer-learning; self-semi-supervised; ai-cybersecurity; ai-smart-cities", "keywords": "", "abstract": "This study examines whether self-supervised learning (SSL) improves anomaly detection performance in network intrusion detection systems when labeled data is limited. Although machine learning—based systems have advanced intrusion detection, their effectiveness depends heavily on large, up-to-date labeled datasets, which are costly to obtain and quickly become outdated. In contrast, real-world networks continuously generate vast amounts of unlabeled traffic. The central objective of this project is to evaluate whether representations learned from unlabeled network data can enhance classification performance under constrained labeling conditions. To investigate this question, I implemented a two-stage SSL pipeline using a synthetic dataset constructed to approximate the statistical properties of the CIC-IDS2017 benchmark. A baseline neural network trained exclusively on labeled data was compared to a model pretrained on unlabeled traffic using an autoencoder and subsequently fine-tuned for binary intrusion detection. The experimental design intentionally limited labeled training samples while providing abundant unlabeled data for representation learning. Preliminary results indicate that the SSL-pretrained model achieved higher test accuracy than the supervised baseline under constrained labeling conditions. Initial experiments suggest that, in this configuration, self-supervised pretraining did outperform fully supervised learning. This study highlights important implementation and design considerations for applying SSL in cybersecurity contexts and provides a reproducible framework for further experimentation using real-world datasets and refined pretraining strategies.", "source_url": "https:\/\/openalex.org\/W7171335856", "pdf_url": "", "doi": "", "handle": "", "repository_name": "SUNY Digital Repository Support (State University of New York System)", "source_record_id": "W7171335856", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:47", "updated_at": "2026-11-15 16:52:47" }, { "id": 181, "title": "SPARSITY AND WEAK SUPERVISION IN QUANTUM MACHINE LEARNING", "title_ar": "", "title_en": "SPARSITY AND WEAK SUPERVISION IN QUANTUM MACHINE LEARNING", "author": "Seyran Saeedi", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Virginia Commonwealth University", "university_source": "repository_host", "college": "", "department": "", "country": "United States", "governorate": "", "publication_year": 2020, "language": "en", "research_domain": "computer-vision", "research_subdomain": "self-semi-supervised", "categories": "machine-learning; computer-vision; self-semi-supervised; quantum-ml; predictive-analytics", "keywords": "", "abstract": "Quantum computing is an interdisciplinary field at the intersection of computer science, mathematics, and physics that studies information processing tasks on a quantum computer. A quantum computer is a device whose operations are governed by the laws of quantum mechanics. As building quantum computers is nearing the era of commercialization and quantum supremacy, it is essential to think of potential applications that we might benefit from. Among many applications of quantum computation, one of the emerging fields is quantum machine learning. We focus on predictive models for binary classification and variants of Support Vector Machines that we expect to be especially important when training data becomes so large that a quantum algorithm with a guaranteed speedup becomes useful. We present a quantum machine learning algorithm for training Sparse Support Vector Machine for problems with large datasets that require a sparse solution. We also present the first quantum semi-supervised algorithm, where we still have a large dataset, but only a small fraction is provided with labels. While the availability of data for training machine learning models is steadily increasing, oftentimes it is much easier to collect feature vectors to obtain the corresponding labels. Here, we present a quantum machine learning algorithm for training Semi-Supervised Kernel Support Vector Machines. The algorithm uses recent advances in quantum sample-based Hamiltonian simulation to extend the existing Quantum LS-SVM algorithm to handle the semi-supervised term in the loss while maintaining the same quantum speedup as the Quantum LS-SVM.", "source_url": "https:\/\/openalex.org\/W3112171833", "pdf_url": "", "doi": "10.25772\/037r-wn88", "handle": "", "repository_name": "VCU Scholars Compass (Virginia Commonwealth University)", "source_record_id": "W3112171833", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:47", "updated_at": "2026-11-15 16:52:47" }, { "id": 191, "title": "Learning latent graph structures for graph neural networks", "title_ar": "", "title_en": "Learning latent graph structures for graph neural networks", "author": "Haicang Zhou", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Nanyang Technological University", "university_source": "repository_host", "college": "", "department": "", "country": "Singapore", "governorate": "", "publication_year": 2025, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; self-semi-supervised; data-mining; transformers; gnn", "keywords": "", "abstract": "Graphs are powerful tools for modeling complex relationships between entities, making them essential in various domains, such as social networks, citation networks, molecular graphs, and road networks. In recent years, Graph Neural Networks (GNNs) have gained significant attention for learning on graphs. However, existing GNNs are very limited in effectively learning various latent graph structures, which are not immediately apparent but very important for graph learning, such as graph clusters. Consequently, there is a pressing need to design novel approaches that enable GNNs to capture diverse latent structures in graph learning. To this end, this dissertation aims to design novel methodologies that can integrate latent structures into graph neural networks to bolster their learning performance. In particular, the focus is on three types of latent graph structures: graph clusters, implicit sparsity and latent structure from the Third Law of Geography. With our investigation on these structures and method designs, we can obtain more insights on both graph structures and graph learning, and further improve the performance of graph neural networks and their real-world applications. First, we highlight the importance of graph clusters in graph learning, and identify the benefits of incorporating graph clusters into graph attention which lacks cluster information. To incorporate graph clusters into graph attention, we proposed Differentiable Clustering for Graph Attention (DCAT), which enhances graph attention by learning cluster-aware attention scores. In DCAT, graph clusters are learned through a differentiable objective based on a relaxed form of modularity maximization. Thus the objective can be optimized together with the semi-supervised loss in an end-to-end manner. Our analysis indicates that DCAT can allocate higher attention scores to nodes within the same cluster, allowing them to have a higher influence on node representation learning. Second, we investigate the latent sparsity phenomenon in real-world graphs, where a large portion of edges connect irrelevant nodes, potentially degrading GNN performance. To address this issue, we propose a novel graph selective attention mechanism,which employs various forms of learnable node-node dissimilarity to learn the latent sparsity. With selective attention, each node can effectively filter out irrelevant edges and selectively attend to its most irrelevant neighbors. We further build the graph selective attention network (SAT) based on it, and conduct theoretical analysis on its expressiveness. Third, we investigate a latent structure in road networks according to the Third Law of Geography, which has not been explored in previous literature. This law suggests a latent structure which goes beyond the spatial relationships of geospatial entities according to their geographic configurations. To learn from the latent structure, we propose a novel graph contrastive learning method, with geographic configuration aware graph augmentation and spectral negative sampling. Our proposed approaches have been evaluated against a number of state-of-the-art baselines across various real-world tasks. The results demonstrate our methods can perform robustly, showcasing learning latent graph structures is indeed an effective way to advance graph neural networks. The methodologies proposed in this thesis inaugurate new directions that can further strengthen graph representation learning for various real-world tasks.", "source_url": "https:\/\/openalex.org\/W4410047771", "pdf_url": "", "doi": "10.32657\/10356\/184288", "handle": "", "repository_name": "DR-NTU (Nanyang Technological University)", "source_record_id": "W4410047771", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:47", "updated_at": "2026-11-15 16:52:47" }, { "id": 214, "title": "Engineering Fairness in Machine Learning Systems", "title_ar": "", "title_en": "", "author": "Moses Openja", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "École Polytechnique de Montréal", "university_source": "repository_host", "college": "", "department": "", "country": "Canada", "governorate": "", "publication_year": 2026, "language": "fr", "research_domain": "nlp", "research_subdomain": "llm", "categories": "machine-learning; nlp; medical-ai; responsible-ai; self-semi-supervised; llm; fairness", "keywords": "", "abstract": "s systèmes d’apprentissage automatique (AA), y compris les modèles classiques, les réseaux de neurones profonds (DNN) et les grands modèles de langage (LLMs), sont de plus en plus déployés dans des domaines à enjeux élevés tels que la santé, le recrutement, la fi-nance et les politiques publiques, où des décisions injustes ou biaisées peuvent avoir des conséquences sociétales majeures. Bien que des progrès substantiels aient été réalisés en matière d’apprentissage équitable, la plupart des approches existantes considèrent l’équité comme une contrainte statique lors de l’entraînement et supposent des distributions de don-nées fixes, des attributs sensibles prédéfinis et une possibilité de réentraîner sans restriction. En pratique, toutefois, les systèmes d’apprentissage automatique évoluent en permanence : le biais peut être introduit lors du prétraitement des données et de la sélection des carac-téristiques, émerger au cours de l’entraînement à travers l’apprentissage des représentations, persister dans des modèles préentraînés, ou être amplifié après le déploiement en raison de changements de distribution. Il est donc essentiel de garantir que l’équité n’est pas compro-mise tout au long du développement et de l’évolution des systèmes d’AA—non seulement pour produire des systèmes logiciels de qualité, mais aussi pour la conformité réglementaire et l’audit. Cette thèse envisage l’équité comme une propriété de qualité au niveau logi-ciel des systèmes d’AA, qui doit être continuellement testée, diagnostiquée et corrigée tout au long du cycle de vie du système. S’inspirant de l’ingénierie logicielle et de l’assurance qualité, la thèse adopte une démarche proactive en matière d’équité, mettant l’accent sur la détection précoce et la correction à faible coût des composants générant des biais, tout en développant également des mécanismes de correction post-entraînement et post-déploiement spécifiques aux contextes où le réentraînement n’est pas faisable. Aux stades des données et de l’entraînement, la thèse propose une méthode systématique pour identifier les carac-téristiques du modèle susceptibles de produire des biais, même sans connaître à l’avance les attributs sensibles. Grâce à une analyse cause-à-effet, cette approche permet aux experts du domaine de raisonner sur les sources de biais et démontre empiriquement que corriger ou supprimer les caractéristiques moins importantes mais génératrices de biais peut améliorer significativement l’équité sans détériorer la performance prédictive. Pour les DNN entraînés, la thèse introduit FairFLRep, un cadre de localisation et de répa-ration des défauts sensible à l’équité qui considère les comportements injustes comme des défectuosités localisées au sein des représentations apprises. FairFLRep identifie les poids de neurones contribuant de manière disproportionnée aux disparités entre sous-groupes et applique des mises à jour ciblées et minimales des paramètres, inspirées par les techniques de correction logicielle issues de l’ingénierie logicielle traditionnelle. En outre, la thèse étudie l’équité dans les modèles transformateurs préentraînés au travers de deux études complé-mentaires. La première examine la manière dont les biais sociaux sont représentés en interne dans les transformateurs, à partir d’un ensemble de données de relations biaisées — des triplets codant des stéréotypes selon divers types de biais — et de techniques d’attribution neuronale. L’étude montre que la connaissance biaisée est souvent concentrée dans de petits sous-ensembles de neurones tout en étant aussi redondamment encodée à travers le réseau. Ces résultats établissent le biais comme un phénomène à la fois architectural et représen-tatif, plutôt qu’un simple artéfact lié aux données, et en évaluent l’impact sur des tâches d’ingénierie logicielle en aval. Sur la base de ces observations, la seconde étude propose un cadre de réparation rigoureux pour les transformateurs préentraînés. Elle introduit une lo-calisation neuronale contrastive et optimale selon Pareto afin d’identifier les neurones de biais de manière concentrée et fiable, puis applique une optimisation axée sur la justesse à l’aide de la méthode d’optimisation par essaim particulaire (PSO) pour les corriger. Les résultats démontrent que la réparation ciblée au niveau des neurones est efficace pour atténuer les biais et constitue une alternative légère et pratique au réentraînement global. Enfin, la thèse aborde la réparation de l’équité sous changement de distribution, un phénomène fréquent dans les déploiements réels, même lorsque les facteurs sources du décalage de données ou de biais sont inconnus ou impossibles à mesurer. Pour dépasser les limites des approches de réparation statiques comme FairFLRep, la thèse propose DShift-FRep, un cadre adaptatif intégrant un alignement des données, un finetuning léger à point de départ chaud, et une réparation ciblée de neurones via PSO.", "source_url": "https:\/\/openalex.org\/W7220748911", "pdf_url": "https:\/\/publications.polymtl.ca\/76114\/1\/2026_MosesOpenja.pdf", "doi": "", "handle": "", "repository_name": "PolyPublie (École Polytechnique de Montréal)", "source_record_id": "W7220748911", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:47", "updated_at": "2026-11-15 16:52:47" }, { "id": 292, "title": "Resources and Applications for Dialectal Arabic: the Case of Levantine", "title_ar": "", "title_en": "Resources and Applications for Dialectal Arabic: the Case of Levantine", "author": "Chatrine Qwaider ; Kathrein Abu Kwaik", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Gothenburg University", "university_source": "repository_host", "college": "", "department": "", "country": "Sweden", "governorate": "", "publication_year": 2022, "language": "en", "research_domain": "nlp", "research_subdomain": "llm", "categories": "machine-learning; deep-learning; computer-vision; nlp; ai-applications; transfer-learning; self-semi-supervised; transformers; llm; sentiment-analysis; arabic-nlp; ai-industry", "keywords": "", "abstract": "This is a thesis about the computational study of Dialectal Arabic (DA). In particular, the thesis studies DA, with a special emphasis on Levantine Arabic, and develops tools and resources for the computational study of Dialectal Arabic Natural Language Processing (DANLP). It investigates the creation of fine-grained resources that can be used for a variety of computational tasks, and a number of effective models that can deal with the complexity of fine-grained dialectal data. Dialect Identification (DI), as well as Sentiment Analysis (SA) are the Natural Language Processing (NLP) tasks investigated in this thesis. In the first part (Study 1 and Study 2), I study the DI task on both coarse-grained and fine-grained levels. For this reason, I build the first annotated Levantine (SHAMI) Dialect Corpus (SDC). Furthermore, I explore the ability of n-gram language models, Machine Learning (ML) algorithms and ensemble learning techniques to classify and detect 26 Arabic varieties. In the second part, I conduct a linguistic study to measure the lexical distance between MSA and DA, and between the dialects themselves. This is done to check whether transferring knowledge from one variety to another is possible. In the third part, studies 4,5 and 6, I explore Arabic Sentiment Analysis (SA). I investigate the idea of knowledge transfer between MSA and the dialects using SA as a case study. Furthermore, I implement various models such as the pre-trained language model BERT, Deep Learning (DL), ML and feature engineering approaches to detect the sentimental polarity of DA data. I introduce two valuable resources for this task, one focusing on Levantine sentiment (Shami-Senti), and the other for DA in general (ATSAD). I exploit different ways of annotation, e.g. human, lexicon-based and automatic distant supervision annotation. The last study is about choosing the best model for DI and SA. I exploit well-known models and approaches using various kinds of DA resources. The thesis contributes to the field of DANLP in a number of ways. The introduced valuable resources can be seen as a stepping stone for a deeper investigation and understanding of issues in DANLP. They are also reliable and can be used by researchers to address different NLP tasks. The cross-dialectal linguistic studies will open up prospects for researchers to fine-tune models and transfer knowledge among Arabic varieties. A big part of the contribution lies in designing DI and SA models. I implement several ML models that use feature engineering approaches and N-gram language models to identify the dialect or detect the sentiment. For DI, I design and implement an ensemble learning model that is able to handle fine-grained dialects. Additionally, I exploit the usage of DL models on different SA dialectal datasets and achieve competitive results. For both tasks, I exploit the recent pre-trained language models and perform a comparison to choose the best model. I also implement a semi-supervised approach for automatic labelling and annotating data with the help of self-training techniques to improve the performance of the dataset. These models will help researchers dive deeper into DANLP and create practical and industrial systems.", "source_url": "https:\/\/openalex.org\/W7216829323", "pdf_url": "", "doi": "", "handle": "", "repository_name": "Gothenburg University Publications Electronic Archive (Gothenburg University)", "source_record_id": "W7216829323", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:48", "updated_at": "2026-11-15 16:52:48" }, { "id": 381, "title": "CONTRASTIVE ATTENTION TRANSFORMER FOR CLASSIFICATION", "title_ar": "", "title_en": "CONTRASTIVE ATTENTION TRANSFORMER FOR CLASSIFICATION", "author": "Yunjiao Bai", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Rutgers University", "university_source": "repository_host", "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; self-semi-supervised; predictive-analytics; transformers", "keywords": "", "abstract": "Most traditional classification methods process each data point in isolation, ignoring the structural information embedded in the relationships among data points. This dissertation develops the contrastive attention transformer, a novel classification architecture that attends a given point separately to each candidate class, producing one representation per class, and predicts the label by contrasting these representations through a shared classifier. We develop the framework for both binary and multiclass classification, proposing three loss functions under two complementary approaches. The comparative approach enforces the relative ordering of confidences across representations through the probability difference and proportional losses. The aggregation approach maximizes joint correctness through the conditional probability loss, which we show to coincide with a forecast aggregation model from decision analysis and to be equivalent to classification on the sum of the representations, motivating a flexible concatenation architecture. A key-value caching scheme further reduces the computational cost of attending to each class separately by 50 to 60 percent for large numbers of classes. We evaluate the framework on five binary classification tasks, four multiclass classification tasks, and a controlled CIFAR-100 experiment, spanning image, text, and tabular data. The best contrastive variant outperforms a feedforward network baseline and a standard transformer baseline with unrestricted attention on every binary and multiclass task. The CIFAR-100 experiment identifies the distinguishability of the classes, rather than the number of classes, as the primary driver of the contrastive method's advantage over unrestricted attention.", "source_url": "https:\/\/openalex.org\/W7220982118", "pdf_url": "", "doi": "10.7282\/t3-ger4-v617", "handle": "", "repository_name": "Rutgers University Community Repository (Rutgers University)", "source_record_id": "W7220982118", "other_sources": "", "verification_status": "needs_review", "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": 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": 418, "title": "Low-shot Learning based on Pre-trained Model", "title_ar": "", "title_en": "Low-shot Learning based on Pre-trained Model", "author": "Ziqin Zhou", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "University of Adelaide", "university_source": "repository_host", "college": "", "department": "", "country": "Australia", "governorate": "", "publication_year": 2024, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; computer-vision; responsible-ai; transfer-learning; few-zero-shot; self-semi-supervised; cnn; transformers; multimodal; image-classification; image-segmentation; fairness", "keywords": "", "abstract": "In the field of computer vision, deep learning has demonstrated remarkable accomplishments. However, the dependence on extensive annotated datasets poses significant challenges, particularly when such data is scarce, costly, or ethically sensitive. This thesis addresses the crucial issue of low-shot learning by aiming to equip learning models with the capability to generalize effectively from limited examples. It explores various low-supervision methodologies, including self-supervised, unsupervised, semi-supervised, and active learning, with the goal of reducing the reliance on labeled data. One potential solution to this challenge is leveraging pre-trained models trained on large-scale datasets and transferring them to downstream tasks and unknown classes. Nevertheless, directly applying pre-trained models in a frozen state often fails to meet the requirements of specific tasks. Fine-tuning these models on base classes with limited training samples and then transferring them to novel domains presents obstacles, as it may disrupt the general knowledge encoded in the model, leading to overfitting on the base classes and suboptimal performance on unseen categories. The central focus of this research is to fully exploit the knowledge encoded in large pre-trained models and enhance their capacity to transfer to novel classes. This thesis introduces a systematic approach to balance generalization and task-specific adaptation, ensuring that fine-tuning preserves pre-trained knowledge while effectively adapting to new data. These models encapsulate extensive knowledge from diverse data sources, enabling swift adaptation to new tasks and classes with minimal additional training. The downstream tasks investigated in this doctoral work range from the low-shot problem for few-shot image classification to more intricate endeavors such as zero-shot and few-shot pixel-level semantic segmentation. The exploration encompasses various pre-trained model architectures, including convolutional and transformer-based models, spanning both pure vision and cross-modal vision-language paradigms. By harnessing the representational capabilities of state-of-the-art pre-trained models, this research aims to push the boundaries of what is achievable with limited data. The novel methodologies proposed strive to maximize performance, maintain generalization across domains, and improve inference efficiency, paving the way for broader adoption in low-data scenarios.", "source_url": "https:\/\/openalex.org\/W7216837125", "pdf_url": "https:\/\/digital.library.adelaide.edu.au\/server\/api\/core\/bitstreams\/c8d7f0de-908a-4f17-b259-cc6fb0be75e4\/content", "doi": "", "handle": "", "repository_name": "Adelaide Research & Scholarship (AR&S) (University of Adelaide)", "source_record_id": "W7216837125", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:49", "updated_at": "2026-11-15 16:52:49" }, { "id": 439, "title": "Natural Language Processing for Slang", "title_ar": "", "title_en": "Natural Language Processing for Slang", "author": "Zhewei Sun", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "University of Toronto", "university_source": "repository_host", "college": "", "department": "", "country": "Canada", "governorate": "", "publication_year": 2024, "language": "en", "research_domain": "nlp", "research_subdomain": "llm", "categories": "machine-learning; nlp; generative-ai; self-semi-supervised; predictive-analytics; llm", "keywords": "", "abstract": "Slang is a common type of language that makes creative and highly flexible use of words. A basic problem that language users tackle is how to develop and interpret novel slang terms for communication in a community. This problem is relevant for natural language processing since new slang expressions often emerge in daily conversations and online social media. However, principled computational approaches to modeling slang are lacking, which presents key challenges to the effective natural language processing of slang. In this dissertation, I develop a computational framework that offers new methodologies for the automated generation, interpretation, and translation of English slang word usages, as well as for characterizing the principles in slang variation across language communities. My dissertation is organized into three main parts. The first part addresses the under-explored problem of slang semantic extension, namely how existing words in the lexicon take on new meanings in informal context. I develop a generative framework that combines contrastive learning with probabilistic models of semantic chaining to capture slang semantic extension. By leveraging dictionary-based resources of slang, I show how the learned semantic representations more accurately predict slang word choices compared to existing approaches that rely more exclusively on corpus data. The second part of my dissertation tackles the inverse problem of slang interpretation by applying these semantic representations to interpret and translate novel slang usages in natural text. I show how this approach provides better accuracy and sample efficiency in both slang interpretation and translation, in comparison to baseline contextualized language models. Finally, the third part of my dissertation investigates semantic variation of slang across different language communities focusing on a comparative study of US and UK. I show that models incorporating either communicative need or semantic chaining can predict the regional identity of slang usages. In summary, my dissertation contributes a principled framework for modeling the lexical semantics and usages of English slang, and it opens up future opportunities for the computational investigation and automated processing of informal language across a diverse set of communities and languages.", "source_url": "https:\/\/openalex.org\/W7133071557", "pdf_url": "https:\/\/utoronto.scholaris.ca\/bitstreams\/2f5619f5-5d4c-40a8-bf85-3e19e6c0bfa5\/download", "doi": "", "handle": "", "repository_name": "TSpace (University of Toronto)", "source_record_id": "W7133071557", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:49", "updated_at": "2026-11-15 16:52:49" }, { "id": 452, "title": "Computer Vision for Aircraft Fuselage Defect Detection: A YOLO11-Based Approach", "title_ar": "", "title_en": "Computer Vision for Aircraft Fuselage Defect Detection: A YOLO11-Based Approach", "author": "Pedro Afonso Bruno Arroteia", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "", "university_source": "", "college": "", "department": "", "country": "Portugal", "governorate": "", "publication_year": 2025, "language": "en", "research_domain": "computer-vision", "research_subdomain": "object-detection", "categories": "machine-learning; deep-learning; computer-vision; ai-applications; transfer-learning; self-semi-supervised; predictive-analytics; object-detection; image-segmentation; ai-industry", "keywords": "", "abstract": "Aircraft fuselages are exposed to operational and environmental stresses that generate defects such as scratches, paint peeling, rust, and rivet damage. Conventional inspection techniques, though widely used, are labor-intensive, time-consuming, and dependent on inspector expertise, which can lead to variability and overlooked anomalies. These chal- lenges highlight the need for automated solutions that ensure consistency and efficiency in quality inspection processes. This thesis presents the development of a computer vision-based system for auto- mated aircraft fuselage defect detection, combining deep learning with a user-oriented interface. The system employs the YOLO11 object detection architecture, fine-tuned on a specialized dataset to detect and classify four defect types. To mitigate data scarcity and class imbalance, data augmentation and transfer learning were applied. The model was evaluated on training, validation, and testing sets, achieving 0.814 precision and 0.735 recall on the training set, and on the validation set 0.722 precision, 0.787 recall, 0.659 accuracy, mAP@0.5 of 0.82, and mAP@0.5:0.95 of 0.524. Beyond model training, a RESTful API and web-based Human–Machine Interface were developed, enabling inspection images to be uploaded and processed, with results presented as bounding boxes, defect types, and confidence scores. This integration reflects Industry 5.0 principles of human–AI collaboration by allowing technicians to interpret, adjust, and validate predictions. The results demonstrate the feasibility of an intelligent inspection system that enhances accuracy, efficiency, and transparency in aircraft maintenance workflows. While challenges such as dataset scarcity and background variability remain, the framework establishes a solid foundation for future work in semi-supervised learning, segmentation-based approaches, and real-time deployment on edge devices. By contributing both a validated detection model and an operational platform, this study advances the application of AI in aviation quality inspection and supports safer, more efficient maintenance practices.", "source_url": "https:\/\/openalex.org\/W7224044969", "pdf_url": "", "doi": "", "handle": "", "repository_name": "Portuguese National Funding Agency for Science, Research and Technology (RCAAP Project by FCT)", "source_record_id": "W7224044969", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:49", "updated_at": "2026-11-15 16:52:49" }, { "id": 511, "title": "Graph Neural Networks for Natural Language Processing", "title_ar": "", "title_en": "Graph Neural Networks for Natural Language Processing", "author": "Kunze Wang", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "The University of Sydney", "university_source": "repository_host", "college": "", "department": "", "country": "Australia", "governorate": "", "publication_year": 2023, "language": "en", "research_domain": "nlp", "research_subdomain": "gnn", "categories": "machine-learning; deep-learning; nlp; self-semi-supervised; predictive-analytics; gnn", "keywords": "", "abstract": "By constructing graph-structured data from the input data, Graph Neural Network (GNN) enhances the performance of numerous Natural Language Processing (NLP) tasks. In this thesis, we mainly focus on two aspects of NLP: text classification and knowledge graph completion. TextGCN shows excellent performance in text classification by leveraging the graph structure of the entire corpus without using any external resources, especially under a limited labelled data setting. Two questions are explored: (1) Under the transductive semi-supervised setting, how to utilize the documents better and learn the complex relationship between nodes. (2) How to transform TextGCN into an inductive model and also reduce the time and space complexity? In detail, firstly, a comprehensive analysis was conducted on TextGCN and its variants. Secondly, we propose ME-GCN, a novel method for text classification that utilizes multi-dimensional edge features in a graph neural network (GNN) for the first time. It uses the corpus-trained word and document-based edge features for semi-supervised classification and has been shown to be effective through experiments on benchmark datasets under the limited labelled data setting. Thirdly, InducT-GCN, an inductive framework for GCN-based text classification that does not require additional resources is introduced. The framework introduces a novel approach to make transductive GCN-based text classification models inductive, improving performance and reducing time and space complexity. Most existing work for Temporal Knowledge Graph Completion (TKGC) overlooks the significance of explicit temporal information and fails to skip irrelevant snapshots based on the entity-related relation in the query. To address this, we introduced Re-Temp (Relation-Aware Temporal Representation Learning), a model that leverages explicit temporal embedding and a skip information flow after each timestamp to eliminate unnecessary information for prediction.", "source_url": "https:\/\/openalex.org\/W7126651815", "pdf_url": "", "doi": "", "handle": "", "repository_name": "The Sydney eScholarship Repository (The University of Sydney)", "source_record_id": "W7126651815", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:50", "updated_at": "2026-11-15 16:52:50" }, { "id": 525, "title": "Robot Learning Through World Models", "title_ar": "", "title_en": "Robot Learning Through World Models", "author": "Baheyeldin Tharwat Moustafa Abdelrazek", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "", "university_source": "", "college": "", "department": "", "country": "", "governorate": "", "publication_year": 2026, "language": "en", "research_domain": "computer-vision", "research_subdomain": "video-analysis", "categories": "machine-learning; computer-vision; robotics; responsible-ai; self-semi-supervised; predictive-analytics; video-analysis; fairness", "keywords": "", "abstract": "Embodied AI introduces a new generation of intelligent agents capable of closed-loop perception and control in real-world environments. However, despite recent progress in robot learning, Most current approaches to robot manipulation are heavily based on supervised learning frameworks. Methods like imitation learning require a large number of examples to succeed in prediction and control. These examples are usually collected by teleoperation, which is expensive to collect, difficult to scale, and prone to bias, limiting the scalability and generalizability of these systems across tasks, environments, and embodiments. On the other hand, humans accumulate knowledge and experience by building internal models of how the world around them works. These world models allow them to predict the future outcomes of their potential actions, reducing the trial and error in the real world needed to learn successful behaviors. This thesis focuses on the problem of equipping robots with the ability to understand the world, learn action representations, imagine the future, and make decisions based on their predictions. We introduce methods for both pretraining and post-training of imitation learning models using world models, where we improve these models without any physical robot interactions. Pretraining in a self-supervised way, by learning latent action representations from large-scale unlabeled human video data, sourced from the Internet, performing actions with everyday objects. Then, training on downstream tasks of a target embodiment using a paired image action dataset using supervised learning. Finally, post-training in the sense of test-time decision-making, given the trained imitation learning model, a learned world model predicts different future states, and a reward model evaluates these imagined outcomes to select the highest-value sequence for action execution. Our experiments show that the proposed methods outperform previous approaches without world models on LIBERO and CALVIN benchmarks and also on real-world experiments, not only by requiring fewer examples to learn, but also by achieving higher performance. The proposed methods open up the potential for building foundation models for robotics by pretraining large models on much larger web-scale video data, and that are able to think and reason in their own imagination to adapt to environments.", "source_url": "https:\/\/openalex.org\/W7164150232", "pdf_url": "", "doi": "10.82419\/356", "handle": "", "repository_name": "", "source_record_id": "W7164150232", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:50", "updated_at": "2026-11-15 16:52:50" }, { "id": 561, "title": "Evolutionary game theoretic models for natural language processing", "title_ar": "", "title_en": "Evolutionary game theoretic models for natural language processing", "author": "Rocco Tripodi", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Università Ca' Foscari Venezia", "university_source": "repository_host", "college": "", "department": "", "country": "Italy", "governorate": "", "publication_year": 2016, "language": "en", "research_domain": "nlp", "research_subdomain": "self-semi-supervised", "categories": "machine-learning; nlp; ai-applications; self-semi-supervised; data-mining; optimization; ai-education", "keywords": "", "abstract": "This thesis is aimed at discovering new learning algorithms inspired by principles of biological evolution, which are able to exploit relational and contextual information, viewing clustering and classification problems in a dynamical system perspective. In particular, we have investigated how game theoretic models can be used to solve different Natural Language Processing tasks. Traditional studies of language have used a game-theoretic perspective to study how language evolves over time and how it emerges in a community but to the best of our knowledge, this is the first attempt to use game-theory to solve specific problems in this area. These models are based on the concept of equilibrium, a state of a system, which emerges after a series of interactions among the elements, which are part of it. Starting from a situation in which there is uncertainty about a particular phenomenon, they describe how a disequilibrium state resolves in equilibrium. The games are situations in which a group of objects has to be classified or clustered and each of them has to choose its collocation in a predefined set of classes. The choice of each one is influenced by the choices of the other and the satisfaction that a player has, about the outcome of a game, is determined by a payoff function, which the players try to maximize. After a series of interactions the players learn to play their best strategies, leading to an equilibrium state and to the resolution of the problem. From a machine-learning perspective this approach is appealing, because it can be employed as an unsupervised, semi-supervised or supervised learning model. We have used it to resolve the word sense disambiguation problem. We casted this task as a constraint satisfaction problem, where each word to be disambiguated is con- strained to choose the most coherent sense among the available, according to the sense that the words around it are choosing. This formulation ensures the mainte- nance of textual coherence and has been tested against state-of-the-art algorithms with higher and more stable results. We have also used a game theoretic formulation, to improve the clustering results of dominant set clustering and non-negative matrix factorization technique. We evaluated our system on different document datasets through different approaches, achieving results, which outperform state-of-the-art algorithms. This work opened new perspectives in game theoretic models, demonstrating that these approaches are promising and that they can be employed also for the resolution of new problems.", "source_url": "https:\/\/openalex.org\/W2509969451", "pdf_url": "", "doi": "", "handle": "", "repository_name": "ARCA (Università Ca' Foscari Venezia)", "source_record_id": "W2509969451", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:50", "updated_at": "2026-11-15 16:52:50" }, { "id": 564, "title": "Self- and semi-supervised machine learning techniques for discovering novel ocular and neuroimaging biomarkers of Alzheimer's disease", "title_ar": "", "title_en": "Self- and semi-supervised machine learning techniques for discovering novel ocular and neuroimaging biomarkers of Alzheimer's disease", "author": "Charlene Z. L. Ong", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Nanyang Technological University", "university_source": "repository_host", "college": "", "department": "", "country": "Singapore", "governorate": "", "publication_year": 2025, "language": "en", "research_domain": "medical-ai", "research_subdomain": "medical-imaging", "categories": "machine-learning; deep-learning; computer-vision; generative-ai; medical-ai; responsible-ai; self-semi-supervised; predictive-analytics; cnn; gnn; multimodal; image-segmentation; gan; medical-imaging; xai", "keywords": "", "abstract": "There is increasing prevalence of Alzheimer’s disease worldwide. To manage the progression of Alzheimer’s disease, biomarkers can be utilized to identify patients with the disease or predict subsequent progression. While neuroimaging biomarkers have been long established to visualize structural changes associated with Alzheimer’s disease, there are other promising biomarkers such as retinal imaging biomarkers which offer the possibility of detecting early pathophysiological changes and are non-invasive and widely available. Nevertheless, the link between retinal imaging biomarkers to Alzheimer’s disease remains unclear and there is much difficulty in procuring datasets of patients for investigation, in particular longitudinal datasets. At the same time, advances in deep learning also offer much promise for furthering our understanding of the link between ophthalmic biomarkers and dementia. Self-supervised learning techniques offer potential for learning new biomarkers which might not have been defined previously. However, there is often a lack of sufficient ground truth annotations, hence hampering the application of deep learning models. In light of this, the work in this thesis aims to address these issues by exploring the use of self or semi-supervised learning approaches and leveraging on both neuroimaging and ophthalmic imaging. By utilizing unlabeled datasets, self or semi-supervised learning could potentially improve the model performance on a range of tasks, such as segmentation, disease classification, and prognosis prediction. More importantly, self-supervised learning techniques could uncover novel biomarkers in a data-driven manner. Some of these methods include generative adversarial networks, graph neural networks, and masked autoencoders, which are explored in this study for biomarker discovery. Firstly, the challenge of lack of sufficient ground truth annotation to develop models for retinal layer segmentation is investigated. Retinal layer segmentation is critical for quantifying the retinal layer thickness, which has been shown to be a useful diagnostic biomarker for age-related macular degeneration and a potential biomarker for Alzheimer’s disease. While there have been existing approaches developed which reduce annotation need, we propose to develop a deep learning approach using generative adversarial learning that utilizes a very limited number of labels for segmentation. We demonstrated that our approach showed comparable or superior performance to other state-of-the-art approaches, particularly with limited labels. In addition, we also evaluated our approach under various noisy conditions and showed it was robust in most noisy conditions. Secondly, current dementia diagnosis is based primarily on clinical assessment. Neuroimaging in the form of magnetic resonance imaging is used to provide additional granularity. It will hence be practical to explore neuroimaging biomarkers as a value add to clinical assessment, especially in a pre-dementia cohort. Our proposed approach uses a graph convolutional neural network to combine both clinical and non-clinical features such as neuroimaging for the classification of cognitive syndromes. It integrates a Cross Attention Module to highlight the salient features in the non-clinical features, attaining the highest accuracy and Matthew’s correlation coefficient compared to existing state-of-the-art methods. In addition, by investigating the feature interpretability via Integrated Gradients, several nonclinical features, such as cerebrospinal fluid volume, right hippocampal fissure volume, right frontal pole thickness, and left pallidum volume were identified as key features for the classification. Thirdly, we explore the utility of ophthalmic biomarkers as prognostic biomarkers for cognitive decline in Alzheimer’s disease patients. In parallel, we explore the possibility of combining both ophthalmic and neuroimaging features to form multimodal features. The application of these multimodal features is investigated in a clinical trial dataset consisting of participants with mild-to-moderate Alzheimer’s disease. To learn key feature representations associated with the ophthalmic images, we utilize self-supervised learning via masked autoencoders to extract feature representations. For neuroimaging, due to the availability of established neuroimaging software, we utilize standard preprocessing pipelines to obtain these neuroimaging features. We investigated the combination of these features using a graph convolutional network model for the prediction of cognitive decline in Alzheimer’s disease and showed initial promising results with the addition of optical coherence tomography data.", "source_url": "https:\/\/openalex.org\/W7127051058", "pdf_url": "", "doi": "10.32657\/10356\/205534", "handle": "", "repository_name": "DR-NTU (Nanyang Technological University)", "source_record_id": "W7127051058", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:50", "updated_at": "2026-11-15 16:52:50" }, { "id": 565, "title": "Semi-Self-Supervised Learning for Deep Models: Reducing Annotation Costs in Dense-Pattern Agricultural Image Analysis", "title_ar": "", "title_en": "Semi-Self-Supervised Learning for Deep Models: Reducing Annotation Costs in Dense-Pattern Agricultural Image Analysis", "author": "Najafian, Keyhan", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "University of Saskatchewan", "university_source": "repository_host", "college": "", "department": "", "country": "Canada", "governorate": "", "publication_year": 2025, "language": "en", "research_domain": "computer-vision", "research_subdomain": "object-detection", "categories": "machine-learning; deep-learning; computer-vision; generative-ai; ai-applications; transfer-learning; self-semi-supervised; object-detection; image-segmentation; video-analysis; ai-agriculture", "keywords": "", "abstract": "Most of the success of deep learning in computer vision depends on supervised learning (SL) and large-scale, manually annotated datasets. Creating such datasets is a major bottleneck due to the substantial cost and labor required, particularly for pixel-level tasks on images with dense and irregular object distributions, as commonly encountered in precision agriculture. While existing semi- and self-supervised methods attempt to reduce this dependency by requiring minimal to no annotation, they still require massive, difficult-to-acquire unlabeled datasets. Self-supervised learning typically involves pretraining on extensive unlabeled data to learn generalizable representations, followed by fine-tuning on task-specific labeled samples, whereas semi-supervised learning combines a small set of labeled data with larger unlabeled collections. To mitigate these challenges, this dissertation introduces a semi-self-supervised learning paradigm that integrates data synthesis, domain adaptation, and spatiotemporal modeling to develop high-performing models across various tasks. Specifically, our paradigm leverages high-frame-rate unannotated video, supplemented with minimal human-annotated static images and video frames, to create large-scale annotated datasets more efficiently than conventional data collection. The produced data includes synthetic training samples, which are rendered through data simulation (synthetic data), and automatically labeled real samples (generated data), where all labels originate from the initial minimal human annotations, thus eliminating the need for further manual labeling and directly addressing the data scarcity problem. We demonstrate the efficacy of our proposed semi-self-supervised paradigm through four key studies on progressively more challenging computer vision tasks within the demanding context of precision agriculture. The framework is first applied to object detection, where a computational synthesis method creates large-scale bounding-box annotated datasets from minimal manually annotated video frames, achieving robust performance of over 80%. Extending this to the pixel-level, the method generates high-quality semantic masks for scenes characterized by dense and overlapping objects, reaching a 91% Dice score with merely 19 annotated video frames and static images. For the more granular task of instance segmentation, our approach enables scalable instance-level object analysis, obtaining over 90% accuracy. Finally, by integrating optical flow and spatiotemporal modeling, our method effectively segments objects in complex video sequences with minimal supervision, outperforming conventional approaches. Together, these studies validate the framework’s ability to deliver high-performing models across diverse and complex tasks while drastically reducing manual labeling effort. These studies establish that our proposed semi-self-supervised paradigm offers a transformative approach for developing DL models. By effectively leveraging video data and supervision derived from minimal human annotations, coupled with domain adaptation, this methodology demonstrates significant scalability and generalizability. It successfully enables instance-level and pixel-level reasoning even with limited annotation, confirming its potential to overcome annotation bottlenecks and advance DL applications in complex, data-constrained real-world scenarios, particularly within the field of precision agriculture.", "source_url": "https:\/\/openalex.org\/W7113662752", "pdf_url": "", "doi": "", "handle": "", "repository_name": "HARVEST (University of Saskatchewan)", "source_record_id": "W7113662752", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:50", "updated_at": "2026-11-15 16:52:50" }, { "id": 566, "title": "From unsupervised to semi-supervised : Leveraging self-supervised embeddings with partial anomaly supervision for tabular anomaly detection", "title_ar": "", "title_en": "From unsupervised to semi-supervised : Leveraging self-supervised embeddings with partial anomaly supervision for tabular anomaly detection", "author": "Jesper Lampinen", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Dalarna University Library", "university_source": "repository_host", "college": "", "department": "", "country": "Sweden", "governorate": "", "publication_year": 2026, "language": "en", "research_domain": "computer-vision", "research_subdomain": "object-detection", "categories": "machine-learning; computer-vision; transfer-learning; self-semi-supervised; predictive-analytics; object-detection", "keywords": "", "abstract": "Self-supervised learning (SSL) has shown promise for extracting useful representations from unlabelled tabular data, but recent work has found that SSL embeddings do not improve unsupervised anomaly detection. This thesis investigates whether these limitations can be overcome by pairing SSL representations with semi-supervised detector heads, a direction identified as future work by Mai et al. (2024). Using SCARF as the SSL method, embeddings from a frozen, contrastively pre-trained encoder are passed to four downstream models: Isolation Forest and One-Class SVM in the unsupervised setting, and self-training with XGBoost and logistic regression at label budgets of 1%, 5%, 10%, and 20%. Raw and PCA-transformed features serve as baselines. Experiments are conducted on three tabular benchmark datasets from ADBench (Cardiotocography, Shuttle, SpamBase), across five random seeds. The results show that SCARF embeddings consistently underperform raw and PCA features across all datasets, detectors, and supervision levels. Semi-supervision improves detection for all feature conditions, but the improvement is proportionally similar across conditions, and the performance gap between SCARF and baselines does not close as more labels are introduced. These findings extend the negative results of Mai et al. into the semi-supervised regime, suggesting that the contrastive pre-training objective of SCARF, applied without supervised fine-tuning, does not produce representations sufficiently aligned with the anomaly detection task to benefit from semi-supervised detector heads.", "source_url": "https:\/\/openalex.org\/W7165072400", "pdf_url": "", "doi": "", "handle": "", "repository_name": "Diva portal (Dalarna University Library)", "source_record_id": "W7165072400", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:50", "updated_at": "2026-11-15 16:52:50" }, { "id": 567, "title": "SELF-SUPERVISED LEARNING FOR REFRIGERATION SYSTEM MONITORING IN IIOT", "title_ar": "", "title_en": "SELF-SUPERVISED LEARNING FOR REFRIGERATION SYSTEM MONITORING IN IIOT", "author": "Xavier Alexander Lopez", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "", "university_source": "", "college": "", "department": "", "country": "United States", "governorate": "", "publication_year": 2026, "language": "en", "research_domain": "machine-learning", "research_subdomain": "self-semi-supervised", "categories": "machine-learning; expert-systems; self-semi-supervised", "keywords": "", "abstract": "This thesis evaluates self-supervised, semi-supervised, supervised, and rule-based machine learning approaches for anomaly detection in low-frequency LoRaWAN refrigeration monitoring across 84 sensors.", "source_url": "https:\/\/openalex.org\/W7160421932", "pdf_url": "", "doi": "10.25394\/pgs.32129659", "handle": "", "repository_name": "Purdue", "source_record_id": "W7160421932", "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": 569, "title": "Efficient Rare Disease Detection with Self-Supervised and Transfer Learning", "title_ar": "", "title_en": "Efficient Rare Disease Detection with Self-Supervised and Transfer Learning", "author": "Miftahul Jannat", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "University of Eastern Finland", "university_source": "repository_host", "college": "", "department": "", "country": "Finland", "governorate": "", "publication_year": 2025, "language": "en", "research_domain": "medical-ai", "research_subdomain": "medical-imaging", "categories": "machine-learning; deep-learning; computer-vision; medical-ai; transfer-learning; self-semi-supervised; predictive-analytics; cnn; medical-imaging", "keywords": "", "abstract": "Accurate diagnosis of rare diseases by analyzing medical images is often faced with the problem of limited availability of annotated datasets, as the manual annotation process, being time-consuming, requires expert knowledge in addition to a significant time investment.The research explores the use of Self-Supervised Learning (SSL) combined with Transfer Learning as a method to reduce the dependency on labeled data without compromising diagnostic performance.The approach designed involves pretraining convolutional neural networks using large unlabeled datasets through self-supervised learning (SSL) methods.For this purpose, the CheXpert dataset of over 200000 X-ray images is employed.To enable the model to imbibe generalizable features from the unlabeled pictures, several SSL pretext tasks are applied, namely rotation prediction, jigsaw puzzles, as well as contrastive learning (SimCLR).Following the pretraining process, the models are fine-tuned using the ISIC 2018 dataset, which contains annotated dermoscopic images for the task of skin lesion classification.Feature extraction and classification performance are evaluated using different architectures, such as ResNet-50.Finetuning is performed using different percentages of labeled data (10%, 25%, 50%, and 100%) to evaluate the efficiency of annotations in semi-supervised learning (SSL).The performance and models have been assessed with respect to Accuracy, Precision, Recall, and F1 Score.From the results, it appears that SSL-pretrained models using SimCLR outperformed fully supervised models even when they were trained on significantly fewer labeled samples.It is astonishing that models which were trained with only 50% of the labeled data actually achieved over 75% accuracy in comparison to those trained on the complete dataset.This work demonstrates that the semi-supervised learning (SSL) approach made possible through transfer learning is a highly effective and efficient in terms of labeling resources for detecting rare diseases in medical images.", "source_url": "https:\/\/openalex.org\/W7215667449", "pdf_url": "https:\/\/erepo.uef.fi\/bitstreams\/9b495e9a-88c8-4914-8f52-550745585eb9\/download", "doi": "", "handle": "", "repository_name": "UEF eRepo (University of Eastern Finland)", "source_record_id": "W7215667449", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:50", "updated_at": "2026-11-15 16:52:50" }, { "id": 570, "title": "Explainable Supervised and Semi-Supervised Learning for Breast Cancer Risk Prediction from Questionnaires: A Study on BCSC and UAE Datasets", "title_ar": "", "title_en": "Explainable Supervised and Semi-Supervised Learning for Breast Cancer Risk Prediction from Questionnaires: A Study on BCSC and UAE Datasets", "author": "Omar Ahmad Alsarookh", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "American University of Sharjah", "university_source": "repository_host", "college": "", "department": "", "country": "United Arab Emirates", "governorate": "", "publication_year": 2025, "language": "en", "research_domain": "medical-ai", "research_subdomain": "medical-imaging", "categories": "machine-learning; medical-ai; responsible-ai; self-semi-supervised; predictive-analytics; medical-imaging; xai", "keywords": "", "abstract": "Breast cancer is one of the most prevalent cancers globally and remains a leading cause of death among women. While mammography helps detect existing abnormalities, it offers limited insight into the future risk of developing cancer. This highlights the need for proactive risk assessment which enables early intervention before symptoms appear. This thesis explores supervised and semi-supervised machine learning methods to classify women into low and high-risk groups using environmental and lifestyle features. Three datasets were utilized: the labeled Breast Cancer Surveillance Consortium (BCSC) Risk Estimation dataset, the unlabeled BCSC Risk Factors dataset, and a labeled private region-specific dataset from University Hospital Sharjah (UHS) in the UAE. In the supervised approach, multiple models were evaluated on the BCSC Risk Estimation dataset including XGBoost, TabNet, Random Forest, and Logistic Regression. XGBoost achieved the best performance with an F1-score of 0.93. A separate region-specific supervised model was also developed using the UHS dataset, with XGBoost again performing best (F1-score: 0.85). Since labeled medical data is often scarce and expensive to obtain, semi-supervised learning allows us to leverage large volumes of unlabeled data to improve model performance and generalization. In the semi-supervised approach, two techniques were used on the unlabeled BCSC Risk Factors dataset: Label Spreading, a graph-based method, and Self-Training, where the classifier iteratively labels data based on confidence thresholds. Pseudo-labeled samples from both methods were combined with labeled data to retrain classifiers. The best semi-supervised model (Self-Training with XGBoost) achieved an F1-score of 0.91. Generalizability was evaluated by testing the BCSC-trained models on the UHS dataset. The supervised BCSC model achieved an F1-score of 0.92, while the semi-supervised model’s F1-score dropped to 0.81 when tested on UHS data, which highlights domain shift and pseudo-label noise. SHAP and LIME were used for explainability, confirming the influence of key risk factors such as breast density, family history, and BMI. These results validate the feasibility of integrating supervised and semi-supervised approaches for proactive, population-specific breast cancer risk assessment.", "source_url": "https:\/\/openalex.org\/W7215125748", "pdf_url": "", "doi": "10.82366\/568", "handle": "", "repository_name": "DSpace at AUS (American University of Sharjah)", "source_record_id": "W7215125748", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:50", "updated_at": "2026-11-15 16:52:50" }, { "id": 571, "title": "Semi-supervised learning using pseudo-labels: A case study in Northern Sámi ASR", "title_ar": "", "title_en": "Semi-supervised learning using pseudo-labels: A case study in Northern Sámi ASR", "author": "Pal, Priyanshi", "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": "speech-recognition", "categories": "machine-learning; nlp; ai-applications; transfer-learning; self-semi-supervised; speech-recognition; ai-education", "keywords": "", "abstract": "While Self-Supervised Learning has advanced Automatic Speech Recognition (ASR) for low-resource languages, limited labeled data remains a bottleneck, often preventing reasonable downstream performance. In line with this requirement, Semi-Supervised Learning approaches have gained prominence as providing a means to data-efficient methods that can be helpful, even with limited supervised data. This work investigates the usefulness of semi-supervised learning, specifically Pseudo-Labeling with self-supervised speech models for Northern Sámi, an under-resourced agglutinative language. Within a teacher–student framework, pseudo-labels are incorporated using two strategies: (a) filtering them through WER-based agreement between the models, and (b) using all pseudo-labels generated by a teacher model. Results show that pseudo-labels from the larger teacher foundation model can improve the downstream performance of a smaller one without additional labeled data, yielding 33.30–3.84% relative character error rate reduction (CERR) across three out-of-distribution test sets. The smaller student model is evaluated in a supervised setting (20 hours of labeled data) and semi-supervised settings with 20, 26 and 75 hours of pseudo-labeled data. Notably, the student model can benefit as much, or even more, from pseudo-labels as from labeled data. Further gains can be achieved when both annotated and pseudo-labels are combined in a two-stage fine-tuning process, rather than mixing all data in a single training phase.", "source_url": "https:\/\/openalex.org\/W7115821482", "pdf_url": "", "doi": "", "handle": "", "repository_name": "Aaltodoc (Aalto University)", "source_record_id": "W7115821482", "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": 573, "title": "Segmentation Label Generation Based on Self-Supervised and Semi-Supervised Learning", "title_ar": "", "title_en": "Segmentation Label Generation Based on Self-Supervised and Semi-Supervised Learning", "author": "Berna de Araújo Nogueira", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "", "university_source": "", "college": "", "department": "", "country": "Portugal", "governorate": "", "publication_year": 2022, "language": "en", "research_domain": "computer-vision", "research_subdomain": "image-segmentation", "categories": "machine-learning; computer-vision; self-semi-supervised; image-segmentation", "keywords": "", "abstract": "", "source_url": "https:\/\/openalex.org\/W7111104835", "pdf_url": "", "doi": "", "handle": "", "repository_name": "Portuguese National Funding Agency for Science, Research and Technology (RCAAP Project by FCT)", "source_record_id": "W7111104835", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:50", "updated_at": "2026-11-15 16:52:50" }, { "id": 574, "title": "Self and semi-supervised Learning for vehicle make and model recognition in images", "title_ar": "", "title_en": "Self and semi-supervised Learning for vehicle make and model recognition in images", "author": "Ricardo Nuno Alves Teixeira", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "", "university_source": "", "college": "", "department": "", "country": "Portugal", "governorate": "", "publication_year": 2024, "language": "en", "research_domain": "computer-vision", "research_subdomain": "transfer-learning", "categories": "machine-learning; computer-vision; ai-applications; transfer-learning; few-zero-shot; self-semi-supervised; ai-education", "keywords": "", "abstract": "Recent advances in computer vision have demonstrated remarkable success in general object recognition tasks. However, fine-grained visual classification remains challenging due to subtle inter-class variations and the requirement for extensive labeled datasets. While self-supervised learning has shown promise in learning rich representations from unlabeled data, its application to specialized domains like fine-grained recognition has been less explored. This study explores various self-supervised learning architectures, such as a contrastive learning method and a teacher-student framework , namely SimCLR and DINO respectively, to evaluate their effectiveness in fine-grained classification tasks. Specifically, it focuses on the application of self and semisupervised learning techniques to vehicle make and model recognition in image datasets with few labels, providing a comprehensive analysis of their performance for this specialized classification challenge. We propose a multi-stage framework that leverages a backbone pretrained on the ImageNet dataset, followed by self-supervised fine-tuning on the Stanford Cars dataset, finalizing with few-shot classification with a linear classifier, also on Stanford Cars, to evaluate representation quality. Our experiments reveal significant challenges in domain transfer from general to specialized visual domains. We observe catastrophic forgetting when adapting models from broad object recognition to fine-grained automotive features, particularly pronounced with limited data availability. This work highlights critical research directions for improving domain adaptation in self-supervised learning, particularly for fine-grained applications with data constraints.", "source_url": "https:\/\/openalex.org\/W7124516569", "pdf_url": "", "doi": "", "handle": "", "repository_name": "Portuguese National Funding Agency for Science, Research and Technology (RCAAP Project by FCT)", "source_record_id": "W7124516569", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:50", "updated_at": "2026-11-15 16:52:50" }, { "id": 575, "title": "Self-Supervised Learning for Geometry", "title_ar": "", "title_en": "Self-Supervised Learning for Geometry", "author": "Huangying Zhan", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "University of Adelaide", "university_source": "repository_host", "college": "", "department": "", "country": "Australia", "governorate": "", "publication_year": 2020, "language": "en", "research_domain": "computer-vision", "research_subdomain": "video-analysis", "categories": "machine-learning; deep-learning; computer-vision; robotics; ai-applications; self-semi-supervised; predictive-analytics; video-analysis; ai-smart-cities", "keywords": "", "abstract": "This thesis focuses on two fundamental problems in robotic vision, scene geometry understanding and camera tracking. While both tasks have been the subject of research in robotic vision, numerous geometric solutions have been proposed in the past decades. In this thesis, we cast the geometric problems as machine learning problems, specifically, deep learning problems. Differ from conventional supervised learning methods that using expensive annotations as the supervisory signal, we advocate for the use of geometry as a supervisory signal to improve the perceptual capabilities in robots, namely Geometry Self-supervision. With the geometry self-supervision, we allow robots to learn and infer the 3D structure of the scene and ego-motion by watching videos, instead of expensive ground-truth annotation in traditional supervised learning problems. Followed by showing the use of geometry for deep learning, we show the possibilities of integrating self-supervised models with traditional geometry-based methods as a hybrid solution for solving the mapping and tracking problem. We focus on an end-to-end mapping problem from stereo data in the first part of this thesis, namely Deep Stereo Matching. Stereo matching is one of the oldest problems in computer vision. Classical approaches to stereo matching typically rely on handcrafted features and a multiple-step solution. Recent deep learning methods utilize deep neural networks to achieve end-to-end trained approaches while significantly outperforming classic methods. We propose a novel data acquisition pipeline using an untethered device (Microsoft HoloLens) with a Time-of-Flight (ToF) depth camera and stereo cameras to collect real-world data. A novel semi-supervised method is proposed to train networks with ground-truth supervision and self-supervision. The large scale real-world stereo dataset with semi-dense annotation and dense self-supervision allow our deep stereo matching network to generalize better when compared to prior arts. Mapping and tracking using a single camera (Monocular) is a harder problem when compared to that using a stereo camera due to varies well-known challenges. In the second part of this thesis, We decouple the problem into single view depth estimation (mapping) and two view visual odometry (tracking) and propose a self-supervised framework, namely SelfTAM, which jointly learns the depth estimator and the odometry estimator. The self-supervised problem is usually formulated as an energy minimization problem consist of an energy of data consistency in multi-view (e.g. photometric) and an energy of prior regularization (e.g. depth smoothness prior). We strengthen the supervision signal with a deep feature consistency energy term and a surface normal regularization term. Though our method trains models with stereo sequence such that a real-world scaling factor is naturally incorporated, only monocular data is required in the inference stage. In the last part of this thesis, we revisit the basics of visual odometry and explore the best practice to integrate deep learning models with geometry-based visual odometry methods. A robust visual odometry system, DF-VO, is proposed. We use deep networks to establish 2D-2D\/3D-2D correspondences and pick the best correspondences from the dense predictions. Feeding the high-quality correspondences into traditional VO methods, e.g. Epipolar Geometry and Prospective-n-Points, we can solve visual odometry problem within a more robust framework. With the proposed self-supervised training, we can even allow the models to perform online adaptation in the run-time and take a step toward a lifelong learning visual odometry system.", "source_url": "https:\/\/openalex.org\/W3132493307", "pdf_url": "", "doi": "", "handle": "", "repository_name": "Adelaide Research & Scholarship (AR&S) (University of Adelaide)", "source_record_id": "W3132493307", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:50", "updated_at": "2026-11-15 16:52:50" }, { "id": 576, "title": "Optimization Methods for Semi-Supervised Learning", "title_ar": "", "title_en": "Optimization Methods for Semi-Supervised Learning", "author": "Edward Cheung", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "University of Waterloo", "university_source": "repository_host", "college": "", "department": "", "country": "Canada", "governorate": "", "publication_year": 2018, "language": "en", "research_domain": "medical-ai", "research_subdomain": "self-semi-supervised", "categories": "machine-learning; medical-ai; self-semi-supervised; predictive-analytics; optimization", "keywords": "", "abstract": "The goal of this thesis is to provide efficient optimization algorithms for some semi-supervised learning (SSL) tasks in machine learning. For many machine learning tasks, training a classifier requires a large amount of labeled data; however, providing labels typically requires costly manual annotation. Fortunately, there is typically an abundance of unlabeled data that can be easily collected for many domains. In this thesis, we focus on problems where an underlying structure allows us to leverage the large amounts of unlabeled data, while only requiring small amounts of labeled data. In particular, we consider low-rank matrix completion problems with applications to recommender systems, and semi-supervised support vector machines (S3VM) to solve binary classification problems, such as digit recognition or disease classification. For the first class of problems, we study convex approximations to the low-rank matrix completion problem. Instead of restricting the solution space to low-rank matrices, we use the trace norm as a convex surrogate. Unfortunately, many trace norm minimization algorithms scale very poorly in practice since they require a full singular value decomposition (SVD) at each iteration. Recently, there has been renewed interest in the trace norm constrained problem utilizing the Frank-Wolfe algorithm, which only requires calculating the leading singular vector pair, providing an order of magnitude improvement on the iteration complexity. However, the Frank-Wolfe algorithm empirically has very slow convergence and in practice yields high-rank solutions, which greatly increases computational costs. To address this issue, we investigate a rank-drop step for Frank-Wolfe, which solves a subproblem specifically designed to decrease the rank of the iterate, ensuring that the Frank-Wolfe algorithm converges along a low-rank path. We show that this rank-drop subproblem can be decomposed into two cases, where each subproblem can be solved efficiently and we guarantee that the iterates remain feasible, preserving the projection-free property of Frank-Wolfe. Next we show that these ideas can be used to provide scalable algorithms for simultaneously sparse and low-rank matrix completion problems. We extend the Frank-Wolfe analysis to accommodate nonsmooth objectives, which can be used to solve the simultaneously sparse and low-rank problem. We replace the traditional linear approximation used in Frank-Wolfe by a uniform affine approximation to better address poor local approximations given by the first-order Taylor approximation. We show that this naturally leads to a sequence of smooth functions that uniformly converges to the original nonsmooth objective, allowing for a careful balance between approximation quality and convergence that is closely related to the step sizes of the Frank-Wolfe algorithm. We apply this algorithm to solve sparse covariance estimation problems, graph link prediction, and robust matrix completion problems. Finally, we propose a variant of self-training for the semi-supervised binary classification problem by leveraging ideas from S3VM. To address common issues associated with self-training, such as error propagation and label imbalances, we proposed an adaptive scheme using the functional margin of S3VM to construct a confidence measure. The confidence score is used to create rules to adapt the optimization problems to incorporate label uncertainty and class imbalances. Moreover, we show that the incremental training approach leverages warm-starts very well, leading to much faster training than standard S3VM methods alone, with much stronger empirical performance on imbalanced datasets.", "source_url": "https:\/\/openalex.org\/W2803599927", "pdf_url": "", "doi": "", "handle": "", "repository_name": "UWSpace (University of Waterloo)", "source_record_id": "W2803599927", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:50", "updated_at": "2026-11-15 16:52:50" }, { "id": 577, "title": "Semi-supervised learning for classification of Nordic news articles", "title_ar": "", "title_en": "Semi-supervised learning for classification of Nordic news articles", "author": "Nils Magne Fossåen", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "", "university_source": "", "college": "", "department": "", "country": "Norway", "governorate": "", "publication_year": 2020, "language": "en", "research_domain": "machine-learning", "research_subdomain": "self-semi-supervised", "categories": "machine-learning; self-semi-supervised", "keywords": "", "abstract": "Semi-supervised learning defines the techniques that fall in between supervised and unsupervised learning. It is commonly used in classification settings where one has a lesser amount of labeled data compared to unlabeled. The goal is to extract extra learning from the unlabeled data to improve on the supervised classification. We will explore some of the approaches to semi-supervised learning to improve on the classification of Nordic news articles in the corpus provided. We will be exploring the methods of self-training in several different configurations and methods of feature extraction and engineering. We will also provide some background and baseline using common supervised methods for improving results as well as different document representations like word-embedding so that we will be able to compare and put our semi-supervised results in relation to these methods. We will see that while some of the methods explored did not succeed, others did and in relation to some of the supervised methods their performance is comparable. We will also see some promising approaches for countering the imbalance problem when considering confident pseudo-labels.", "source_url": "https:\/\/openalex.org\/W3089233220", "pdf_url": "", "doi": "", "handle": "", "repository_name": "NORA - Norwegian Open Research Archives", "source_record_id": "W3089233220", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:50", "updated_at": "2026-11-15 16:52:50" }, { "id": 578, "title": "Safe Semi-Supervised Learning", "title_ar": "", "title_en": "Safe Semi-Supervised Learning", "author": "Andrea Bertazzi", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Delft University of Technology", "university_source": "repository_host", "college": "", "department": "", "country": "Netherlands", "governorate": "", "publication_year": 2018, "language": "en", "research_domain": "generative-ai", "research_subdomain": "self-semi-supervised", "categories": "machine-learning; generative-ai; self-semi-supervised", "keywords": "", "abstract": "Semi-supervised algorithms have been shown to possibly have a worse performance than the corresponding supervised model. This may be due to a violation of the assumptions on the data that are introduced in most classification systems. We study an approach that was previously shown to have guarantees of improvement for the LDA classifier in terms of log-likelihood on the full data-set of labeled and unlabeled observations. This method is based on two key concepts: contrast and pessimism. We extend this approach to a broader class of probabilistic generative models, in which the class conditional distributions can be modeled with any parametric class that belongs to the exponential family. In this case, we prove that the classifier is never worse and, under mild assumptions, strictly improves the log-likelihood on the complete data-set. The case of Gaussian densities is analyzed in detail, both for LDA and QDA. Moreover, we study this method in the case of least squares classification. In terms of square loss we prove the contrastive pessimistic classifier is guaranteed not to degrade the performance and, with a further requirement on the data, it strictly outperforms the supervised model. Finally, we apply contrast and pessimism to the task of parameter estimation of a multivariate Gaussian density in a missing data framework. We fully characterize the case of a missing block of data and we show that a strictly increased likelihood on the complete data-set is obtained for any monotone sample. In other terms, the contrastive pessimistic estimates are guaranteed to fit better the complete data-set composed of both observed and hidden components.", "source_url": "https:\/\/openalex.org\/W2996176234", "pdf_url": "https:\/\/repository.tudelft.nl\/file\/File_bd7b1eec-f0aa-4e2d-9825-87b65336c01f", "doi": "", "handle": "", "repository_name": "Research Repository (Delft University of Technology)", "source_record_id": "W2996176234", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:50", "updated_at": "2026-11-15 16:52:50" }, { "id": 579, "title": "Semi-Supervised Hybrid Windowing Ensembles for Learning from Evolving Streams", "title_ar": "", "title_en": "Semi-Supervised Hybrid Windowing Ensembles for Learning from Evolving Streams", "author": "Sean Louis Alan Floyd", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "University of Ottawa", "university_source": "repository_host", "college": "", "department": "", "country": "Canada", "governorate": "", "publication_year": 2019, "language": "en", "research_domain": "computer-vision", "research_subdomain": "object-detection", "categories": "machine-learning; computer-vision; self-semi-supervised; predictive-analytics; object-detection", "keywords": "", "abstract": "In this thesis, learning refers to the intelligent computational extraction of knowledge from data. Supervised learning tasks require data to be annotated with labels, whereas for unsupervised learning, data is not labelled. Semi-supervised learning deals with data sets that are partially labelled. A major issue with supervised and semi-supervised learning of data streams is late-arriving or missing class labels. Assuming that correctly labelled data will always be available and timely is often unfeasible, and, as such, supervised methods are not directly applicable in the real world. Therefore, real-world problems usually require the use of semi-supervised or unsupervised learning techniques. For instance, when considering a spam detection task, it is not reasonable to assume that all spam will be identified (correctly labelled) prior to learning. Additionally, in semi-supervised learning, \"the instances having the highest [predictive] confidence are not necessarily the most useful ones\" [41]. We investigate how self-training performs without its selective heuristic in a streaming setting. This leads us to our contributions. We extend an existing concept drift detector to operate without any labelled data, by using a sliding window of our ensemble's prediction confidence, instead of a boolean indicating whether the ensemble's predictions are correct. We also extend selective self-training, a semi-supervised learning method, by using all predictions, and not only those with high predictive confidence. Finally, we introduce a novel windowing type for ensembles, as sliding windows are very time consuming and regular tumbling windows are not a suitable replacement. Our windowing technique can be considered a hybrid of the two: we train each sub-classifier in the ensemble with tumbling windows, but delay training in such a way that only one sub-classifier can update its model per iteration. We found, through statistical significance tests, that our framework is (roughly 160 times) faster than current state of the art techniques, and achieves comparable predictive accuracy. That being said, more research is needed to further reduce the quantity of labelled data used for training, while also increasing its predictive accuracy.", "source_url": "https:\/\/openalex.org\/W2954664475", "pdf_url": "", "doi": "10.20381\/ruor-23520", "handle": "", "repository_name": "uO Research (University of Ottawa)", "source_record_id": "W2954664475", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:50", "updated_at": "2026-11-15 16:52:50" }, { "id": 580, "title": "Semi-supervised dictionary learning and Semi-supervised deep neural network", "title_ar": "", "title_en": "", "author": "Khanh-Hung Tran", "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": "computer-vision", "research_subdomain": "self-semi-supervised", "categories": "machine-learning; deep-learning; computer-vision; robotics; ai-applications; self-semi-supervised; ai-cybersecurity", "keywords": "", "abstract": "Apprentissage semi-supervisé de dictionnaire et de réseaux de neurones profonds Depuis les années 2010, l’apprentissage automatique (ML) est l’un des sujets qui retient beaucoup l'attention des chercheurs scientifiques. De nombreux modèles de ML ont démontré leur capacité produire d’excellent résultats dans des divers domaines comme Vision par ordinateur, Traitement automatique des langues, Robotique… Toutefois, la plupart de ces modèles emploient l’apprentissage supervisé, qui requiert d’un massive annotation. Par conséquent, l’objectif de cette thèse est d’étudier et de proposer des approches semi-supervisées qui ont plusieurs avantages par rapport à l’apprentissage supervisé. Au lieu d’appliquer directement un classificateur semi-supervisé sur la représentation originale des données, nous utilisons plutôt des types de modèle qui intègrent une phase de l’apprentissage de représentation avant de la phase de classification, pour mieux s'adapter à la non linéarité des données. Dans le premier temps, nous revisitons des outils qui permettent de construire notre modèles semi-supervisés. Tout d’abord, nous présentons deux types de modèle qui possèdent l’apprentissage de représentation dans leur architecture : l’apprentissage de dictionnaire et le réseau de neurones, ainsi que les méthodes d’optimisation pour chaque type de model, en plus, dans le cas de réseau de neurones, nous précisons le problème avec les exemples contradictoires. Ensuite, nous présentons les techniques qui accompagnent souvent avec l’apprentissage semi-supervisé comme l’apprentissage de variétés et le pseudo-étiquetage. Dans le deuxième temps, nous travaillons sur l’apprentissage de dictionnaire. Nous synthétisons en général trois étapes pour construire un modèle semi-supervisée à partir d’un modèle supervisé. Ensuite, nous proposons notre modèle semi-supervisée pour traiter le problème de classification typiquement dans le cas d’un faible nombre d’échantillons d’entrainement (y compris tous labellisés et non labellisés échantillons). D'une part, nous appliquons la préservation de la structure de données de l’espace original à l’espace de code parcimonieux (l’apprentissage de variétés), ce qui est considéré comme la régularisation pour les codes parcimonieux. D'autre part, nous intégrons un classificateur semi-supervisé dans l’espace de code parcimonieux. En outre, nous effectuons le codage parcimonieux pour les échantillons de test en prenant en compte aussi la préservation de la structure de données. Cette méthode apporte une amélioration sur le taux de précision par rapport à des méthodes existantes. Dans le troisième temps, nous travaillons sur le réseau de neurones. Nous proposons une approche qui s’appelle \"manifold attack\" qui permets de renforcer l’apprentissage de variétés. Cette approche est inspirée par l’apprentissage antagoniste : trouver des points virtuels qui perturbent la fonction de coût sur l’apprentissage de variétés (en la maximisant) en fixant les paramètres du modèle; ensuite, les paramètres du modèle sont mis à jour, en minimisant cette fonction de coût et en fixant les points virtuels. Nous fournissons aussi des critères pour limiter l’espace auquel les points virtuels appartiennent et la méthode pour les initialiser. Cette approche apporte non seulement une amélioration sur le taux de précision mais aussi une grande robustesse contre les exemples contradictoires. Enfin, nous analysons des similarités et des différences, ainsi que des avantages et inconvénients entre l’apprentissage de dictionnaire et le réseau de neurones. Nous proposons quelques perspectives sur ces deux types de modèle. Dans le cas de l’apprentissage de dictionnaire semi-supervisé, nous proposons quelques techniques en inspirant par le réseau de neurones. Quant au réseau de neurones, nous proposons d’intégrer \"manifold attack\" sur les modèles génératifs.", "source_url": "https:\/\/openalex.org\/W7149497300", "pdf_url": "https:\/\/theses.hal.science\/tel-03206023v1\/document", "doi": "10.70675\/42000ac0z65ebz4a00z9dd3z6af0c02d8611", "handle": "", "repository_name": "theses.fr (ABES)", "source_record_id": "W7149497300", "other_sources": "", "verification_status": "verified", "created_at": "2026-11-15 16:52:50", "updated_at": "2026-11-15 16:52:50" }, { "id": 581, "title": "Semi-Supervised Self-Organizing Maps with Time-Varying Structures for Clustering and Classification", "title_ar": "", "title_en": "Semi-Supervised Self-Organizing Maps with Time-Varying Structures for Clustering and Classification", "author": "Pedro H. M. Braga", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "", "university_source": "", "college": "", "department": "", "country": "", "governorate": "", "publication_year": 2019, "language": "en", "research_domain": "deep-learning", "research_subdomain": "self-semi-supervised", "categories": "machine-learning; deep-learning; self-semi-supervised; data-mining", "keywords": "", "abstract": "In recent years, the advances in technology have produced datasets of increasing size, not only regarding the number of samples but also the number of features. Unfortunately, despite these advances, creating a sufficiently large amount of properly labeled data with enough examples for each class is not an easy task. Organizing and labeling such data is challenging, expensive, and time-consuming. Also, it is usually done manually, and people can label with different formats and styles, incorporating noise and errors to the dataset. Hence, there is a growing interest in semi-supervised learning, since, in many learning tasks, there is a plentiful supply of unlabeled data, but insufficient labeled ones. Therefore, at the current stage of research, it is of great importance to put forward semi-supervised learning models aiming to combine both types of data, in order to benefit from the distinct information they can provide, to obtain better performances of both clustering and classification tasks, that would expand the range of machine learning applications. Moreover, it is also important to develop methods that are easy to parameterize in a way that become robust to the different characteristics of the data at hand. In this sense, the Self-Organizing Maps (SOM) can be considered as good options to address such objectives. It is a biologically inspired neural model that uses unsupervised and incremental learning to produce prototypes of the input data. However, such an unsupervised characteristic makes it unfeasible for SOM to execute Semi-Supervised Learning. In that way, this Dissertation presents some new proposals based on SOM to perform Semi-Supervised learning tasks for both clustering and classification. It is done by introducing to SOM the standard concepts of Learning Vector Quantization (LVQ), which can be seen as its supervised counterpart, to build hybrid approaches. Such proposals can dynamically switch between the two types of learning at training time, according to the availability of labels and automatically adjust themselves to the local variance observed in each data cluster. In the course of this work, the experimental results show that the proposed models can surpass the performance of other traditional methods not only in terms of classification but also regarding clustering quality. It also enhances the range of possible applications of a SOM and LVQ-based models by combining them with recent and promising techniques from Deep Learning to solve more complex problems commonly found in such field.", "source_url": "https:\/\/openalex.org\/W3119985271", "pdf_url": "", "doi": "", "handle": "", "repository_name": "Americanae (AECID Library)", "source_record_id": "W3119985271", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:50", "updated_at": "2026-11-15 16:52:50" }, { "id": 582, "title": "Study of Semi-supervised Deep Learning Methods on Human Activity Recognition Tasks", "title_ar": "", "title_en": "Study of Semi-supervised Deep Learning Methods on Human Activity Recognition Tasks", "author": "Shiping Song", "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": "rnn", "categories": "machine-learning; deep-learning; nlp; self-semi-supervised; predictive-analytics; rnn", "keywords": "", "abstract": "This project focuses on semi-supervised human activity recognition (HAR) tasks, in which the inputs are partly labeled time series data acquired from sensors such as accelerometer data, and the outputs are predefined human activities. Most state-of-the-art existing work in HAR area is supervised now, which relies on fully labeled datasets. Since the cost to label the collective instances increases fast with the increasing scale of data, semi-supervised methods are now widely required. This report proposed two semi-supervised methods and then investigated how well they perform on a partly labeled dataset, comparing to the state-of-the-art supervised method. One of these methods is designed based on the state-of-the-art supervised method, Deep-ConvLSTM, together with the semi-supervised learning concepts, self-training. Another one is modified based on a semi-supervised deep learning method, LSTM initialized by seq2seq autoencoder, which is firstly introduced for natural language processing. According to the experiments on a published dataset (Opportunity Activity Recognition dataset), both of these semi-supervised methods have better performance than the state-of-the-art supervised methods.", "source_url": "https:\/\/openalex.org\/W3036519956", "pdf_url": "", "doi": "", "handle": "", "repository_name": "KTH Publication Database DiVA (KTH Royal Institute of Technology)", "source_record_id": "W3036519956", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:50", "updated_at": "2026-11-15 16:52:50" }, { "id": 583, "title": "Semi-supervised learning methods for unsupervised domain adaptation in medical imaging segmentation", "title_ar": "", "title_en": "Semi-supervised learning methods for unsupervised domain adaptation in medical imaging segmentation", "author": "Pedro L. Ballester", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "", "university_source": "", "college": "", "department": "", "country": "", "governorate": "", "publication_year": 2019, "language": "en", "research_domain": "medical-ai", "research_subdomain": "medical-imaging", "categories": "machine-learning; computer-vision; medical-ai; transfer-learning; self-semi-supervised; predictive-analytics; image-segmentation; medical-imaging", "keywords": "", "abstract": "Machine learning applications make several assumptions regarding the scenario where they are employed. One common assumption is that data distribution in the test environment follows the same distribution of the training set. This assumption is systematically broken in most real-world scenarios; the difference between these distributions is commonly known as domain shift. Unsupervised domain adaptation aims at suppressing this problem by leveraging knowledge with unlabeled data from the test environment. One of the most sensitive fields for domain shift is medical imaging. Due to the heterogeneity in data distributions from scanners, models tend to vary in predictive performance when dealing with images from scanners with no examples in the training set. We propose two contributions in this work. First, we introduce the use of self-ensembling domain adaptation in the field of medical imaging segmentation in a spinal cord grey matter segmentation task. Next, based on the success of self-ensembling, we adapt two other recent work from the semi-supervised learning literature to the same task, namely, unsupervised data augmentation and MixMatch. We conduct ablation studies and other experiments in order to understand the behavior of each method and compare their best results. The results show improvements over training models in a supervised learning fashion and demonstrate that recent semi-supervised learning methods are promising for domain adaptation in medical imaging segmentation.", "source_url": "https:\/\/openalex.org\/W2992030155", "pdf_url": "", "doi": "", "handle": "", "repository_name": "LA Referencia (Red Federada de Repositorios Institucionales de Publicaciones Científicas)", "source_record_id": "W2992030155", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:50", "updated_at": "2026-11-15 16:52:50" }, { "id": 584, "title": "Semi-Supervised Learning with Unlabeled data: from Centralized to Distributed Systems", "title_ar": "", "title_en": "Semi-Supervised Learning with Unlabeled data: from Centralized to Distributed Systems", "author": "Nan Yang", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "The University of Sydney", "university_source": "repository_host", "college": "", "department": "", "country": "Australia", "governorate": "", "publication_year": 2024, "language": "en", "research_domain": "computer-vision", "research_subdomain": "image-classification", "categories": "machine-learning; deep-learning; computer-vision; responsible-ai; self-semi-supervised; cnn; image-classification; federated-learning; fairness", "keywords": "", "abstract": "The rapid increase in data generated by edge devices and IoT technologies demands efficient management solutions, especially in terms of cost and infrastructure. Key challenges include expensive data labeling, requiring significant human resources, and ensuring data privacy and security, with the risk of information leakage during transmission. The limited availability of labeled data versus the exponential growth of new data presents challenges for maintaining accuracy and efficiency in data-driven models. High costs of data annotation, especially involving subject-matter experts, limit model training effectiveness. Privacy concerns are heightened due to edge devices' interaction with sensitive user data, making the affordability of data labeling and data privacy protection at the edge crucial issues. To tackle these challenges, we proposed Data Augmentation Random Padding to increase effective data for model training in CNNs, enhancing image classification accuracy. Despite limitations, we developed Semi-Supervised Learning (SSL) method \"AdaptMatch\" to utilize a large amount of unlabeled data and a small number of labeled data in centralized learning, improving learning speed and reducing label bias. To address the inadequacies of centralized data learning, we introduced a decentralized SSL method, Federated Incremental Learning (FedIL), for learning on edge devices while protecting privacy. However, FedIL's lower training efficiency and struggles with data imbalance led to the development of Federated Masked Autoencoder (FedMAE), which is also a decentralized semi-supervised learning method based on self-supervised learning. FedMAE enables asynchronous training of large-scale unlabeled images in federated learning, outperforming existing methods in handling highly imbalanced data.", "source_url": "https:\/\/openalex.org\/W7126832066", "pdf_url": "", "doi": "", "handle": "", "repository_name": "The Sydney eScholarship Repository (The University of Sydney)", "source_record_id": "W7126832066", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:50", "updated_at": "2026-11-15 16:52:50" }, { "id": 585, "title": "Semi-supervised learning in insurance : fairness and active learning", "title_ar": "", "title_en": "", "author": "François Hu", "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": "responsible-ai", "research_subdomain": "fairness", "categories": "machine-learning; responsible-ai; self-semi-supervised; fairness", "keywords": "", "abstract": "Apprentissage semi-supervisé en assurance : équité et apprentissage actif Les organismes d'assurance stockent quotidiennement des sources de données textuelles volumineuses (zones de texte libre utilisées par les téléconseillers, courriers électroniques, avis des clients, etc.). Cependant, cette masse de données textuelles comporte des enjeux spécifiques en termes de réglementations comme par exemple le respect des contraintes de protection de la vie privée, imposées en Europe par le récent Règlement général sur la protection des données (RGPD) : ces données textuelles peuvent contenir des informations non-conformes aux normes RGPD, soulevant ainsi des questions éthiques et ne peuvent pas être conservées par l'assureur. Aujourd'hui, ces données textuelles sont étiquetées par des experts (oracles) et ce processus n'est pas adapté à la gestion de grands volumes ni à une gestion de l'information en temps quasi réel. Par conséquent, la mise en place d'un système d'apprentissage précis (en termes de prédiction), peu coûteux (en termes d'étiquetage) et éthique (en termes d'équité) est nécessaire en assurance et cette thèse aborde et résout certains de ces défis. Le premier défi est de réduire l'effort d'étiquetage (se concentrant ainsi sur la qualité des données) avec l'aide de l'apprentissage actif, une boucle de rétroaction entre l'inférence du modèle et un oracle : puisqu'en assurance les données non étiquetées sont généralement abondantes, l'apprentissage actif peut devenir un atout important pour réduire le coût de l'étiquetage. Un autre défi majeur est la question de l'équité dans les inférences de modèles ML. Puisque des inégalités et des discriminations peuvent être trouvées dans les données, les modèles d'apprentissage sont susceptibles de reproduire certaines injustices, ce qui les rend inutilisables en production. Cette thèse explore ces problèmes et propose des solutions, notamment pour les tâches de classification multi-classes. En particulier, nous proposons une méthode d'équité algorithmique qui garantit soit une équité exacte au détriment de la précision du modèle, soit un compromis entre équité et précision appelé epsilon-fairness. En outre, nous proposons une méthode d'apprentissage actif équitable qui requête les instances informatives tout en rendant le modèle équitable. Les méthodologies proposées ont l'avantage d'être agnostiques par rapport au modèle d'apprentissage statistique. Ces résultats sont étudiés et appliqués sur des jeux de données réels et synthétiques.", "source_url": "https:\/\/openalex.org\/W7151310309", "pdf_url": "https:\/\/theses.hal.science\/tel-03752063\/document", "doi": "10.70675\/f9cec72az9beaz4c9bz9436z152694ee9162", "handle": "", "repository_name": "theses.fr (ABES)", "source_record_id": "W7151310309", "other_sources": "", "verification_status": "verified", "created_at": "2026-11-15 16:52:50", "updated_at": "2026-11-15 16:52:50" }, { "id": 586, "title": "Semi-Supervised Image Segmentation of Medical Data", "title_ar": "", "title_en": "Semi-Supervised Image Segmentation of Medical Data", "author": "Ove Nicolai Dalheim", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "", "university_source": "", "college": "", "department": "", "country": "Norway", "governorate": "", "publication_year": 2020, "language": "en", "research_domain": "medical-ai", "research_subdomain": "image-segmentation", "categories": "machine-learning; deep-learning; computer-vision; medical-ai; transfer-learning; self-semi-supervised; data-mining; cnn; image-segmentation", "keywords": "", "abstract": "Bladder cancer is the fourth most common cancer type in Norway, and tenth most common on a global scale. More and more tissue samples are sent to pathologists labs, increasing the workload and affecting the waiting time for patients. The corresponding increase is not seen in number of pathologists. Digitization and scanning of the tissue samples unveil the world of computational pathology, along with the many possibilities within it. Supervised approaches have been proposed within deep learning earlier, however, many express the lack of labeled data as a source for performance degradation. Convolutional neural networks have proven effective on image processing within the field of medicine. Some work based on semi-supervised approaches within computational pathology has been published in the recent years, however, most researchers are exploring supervised methods. This thesis proposes several methods to tile-wise segment histological images of bladder cancer into six different classes background, blood, damaged tissue, muscle tissue, stroma tissue and urothelium. A multiscale model is fed the tile at three different levels of magnification, and inherits capabilities from the VGG16 network through transfer learning. The proposed methods are either based on semi-supervised learning utilizing probability or clustering within a self-training process, or based on a combination of both expert and non-expert annotations. The dataset used in this thesis consists of about 360 whole-slide images, of which only 37 contain regions annotated by a pathologist, allowing for only about 10 % of the dataset to be utilized for supervised methods. Through the course of this thesis, a total of 145 new whole-slide images were introduced during training through semi-supervised learning, utilizing about 50 % of the dataset through various methods. The non-expert approach increased the F1-score for the muscle tissue class with 9.18 % from an initial 79.42 %, and the cluster-baser approach saw an increase of 1.38 % in accuracy from the initial 94.61 %. The method involving non-expert annotations outperformed both semi-supervised techniques with regards to segmentation, as a semi-supervised method will introduce new uncertain features to a deep learning model. This will have a strong impact on sensitive classes that are poorly represented in the training dataset. That said, a significant improvement is seen by using semi-supervised techniques as well, and scored better than non-expert with regards to classification.", "source_url": "https:\/\/openalex.org\/W3087951694", "pdf_url": "", "doi": "", "handle": "", "repository_name": "NORA - Norwegian Open Research Archives", "source_record_id": "W3087951694", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:50", "updated_at": "2026-11-15 16:52:50" }, { "id": 587, "title": "Semi-supervised learning for plankton image classification", "title_ar": "", "title_en": "Semi-supervised learning for plankton image classification", "author": "Maram Hamid", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "LUT University", "university_source": "repository_host", "college": "", "department": "", "country": "Finland", "governorate": "", "publication_year": 2020, "language": "en", "research_domain": "computer-vision", "research_subdomain": "image-classification", "categories": "machine-learning; deep-learning; computer-vision; self-semi-supervised; cnn; image-classification", "keywords": "", "abstract": "The production of big data from plankton populations has become feasible with the use of imaging devices. This opens up the possibility of testing key characteristics in planktonic systems. The manual labeling of images is complicated as well as time-consuming. On the other hand, semi-supervised learning makes it possible to automate the process by performing the classification of large unlabeled data with the available labeled data. The classification of plankton images using semi-supervised methods including label propagation, label spreading, and Pseudo labeling utilizing CNN based image features has been implemented in this study. The comparison of the accuracy of the semi-supervised methods and the CNN method using different datasets shows that the semi-supervised learning methods were more accurate than the baseline method where the amount of data available to train the model was sufficient and fully supervised CNN was able to extract the correct features. However, sometimes the semi-supervised learning methods have not been more accurate than CNN, when CNN has struggled to find the right image features.", "source_url": "https:\/\/openalex.org\/W3037808886", "pdf_url": "", "doi": "", "handle": "", "repository_name": "LUTPub (LUT University)", "source_record_id": "W3037808886", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:50", "updated_at": "2026-11-15 16:52:50" }, { "id": 588, "title": "Utilizing unlabeled data in cell type identification : A semi-supervised learning approach to classification", "title_ar": "", "title_en": "Utilizing unlabeled data in cell type identification : A semi-supervised learning approach to classification", "author": "Thijs Quast", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Linkoping University", "university_source": "repository_host", "college": "", "department": "", "country": "Sweden", "governorate": "", "publication_year": 2020, "language": "en", "research_domain": "machine-learning", "research_subdomain": "self-semi-supervised", "categories": "machine-learning; self-semi-supervised", "keywords": "", "abstract": "Recent research in bioinformatics has presented multiple cell type identification meth- dologies using single cell RNA sequence data (scRNA-seq). However, a consensus on which cell typing methodology consistently demonstrates superior performance remains absent. Additionally, very few studies approach cell type identification through a semi- supervised learning study, whereby the information in unlabeled data is leveraged to train an enhanced classifier. This paper presents cell annotation methodologies through self- learning and graph-based semi-supervised learning, in both raw count scRNA-seq data as well as in a latent embedding. I find that a self-learning framework enhances perfor- mance compared to a solely supervised learning classifier. Additionally, modelling on the latent data representations consistently outperforms modelling on the original data. The results show an overall accuracy of 96.12%, whereas additional models achieve an average precision rate of 95.12% and an average recall rate of 94.40%. The semi-supervised learn- ing approaches in this thesis compare favourable to scANVI in terms of accuracy, average precision rate, average recall rate and average f1-score. Moreover, results for alternative scenarios, in which cell types among training and test data do not perfectly overlap, are reported in this thesis.", "source_url": "https:\/\/openalex.org\/W3046627445", "pdf_url": "", "doi": "", "handle": "", "repository_name": "DiVA (Linkoping University)", "source_record_id": "W3046627445", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:50", "updated_at": "2026-11-15 16:52:50" }, { "id": 589, "title": "Adversarial signals against discriminative and generative models for data protections", "title_ar": "", "title_en": "Adversarial signals against discriminative and generative models for data protections", "author": "Mingzhi Lyu", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Nanyang Technological University", "university_source": "repository_host", "college": "", "department": "", "country": "Singapore", "governorate": "", "publication_year": 2025, "language": "en", "research_domain": "computer-vision", "research_subdomain": "diffusion", "categories": "machine-learning; computer-vision; generative-ai; responsible-ai; ai-applications; transfer-learning; self-semi-supervised; diffusion; fairness; ai-cybersecurity", "keywords": "", "abstract": "In recent years, discriminative and generative models have been widely adopted across a broad range of applications—including classification, detection, image synthesis, and editing—significantly enhancing productivity and creativity in various domains. However, their proliferation also introduces critical societal challenges such as copyright infringement, misinformation, and privacy violations. Protecting visual data from unauthorized manipulation and model training has thus become an urgent and essential task. Although adversarial attacks have been explored as protection strategies, achieving transferable and generalizable attacks remains a major open challenge due to the diversity of model architectures and task-specific training procedures. In response, this thesis systematically analyzes the mechanisms underlying discriminative and generative models and proposes a series of attack and defense strategies aimed at either protecting visual data or exposing the vulnerabilities of existing methods to inform future research. For discriminative models, data availability attacks (DAAs) have been widely adopted to protect copyrighted data against unauthorized training, by poisoning training data to degrade the performance of downstream models. However, existing DAAs often suffer from limited robustness or poor image quality. To address this, we introduce Imperfect Restoration Poisoning (IRP), a novel poisoning technique designed for supervised learning (SL), which achieves strong attack performance while preserving high image fidelity. Through extensive comparisons of IRP with eight baselines, coupled with evaluations alongside five representative defense methods, we showcase the superiority of IRP. In addition to proposing improved DAAs, we investigate the defense side, especially how to mitigate the impact of data poisoning. While Self-Supervised Learning (SSL) has been considered a promising paradigm to mitigate poisoning due to its objective mismatch with supervised attacks, our experiments across multiple poison types and datasets (CIFAR-10 and ImageNet-100) reveal that SSL often performs poorly when trained on poisoned data. To address this and further expose the limit of existing DAAs, we introduce VESPR (Vulnerability Exploitation of Supervised Poisoning for Robust SSL), which introduce adversarial training on SL to obfuscate poison features and guide robust feature learning for SSL. VESPR outperforms six existing defenses across seven poisoning strategies, revealing a research gap in understanding and improving poisoning robustness in discriminative models. Beyond classification models, this thesis also investigates visual data protection against unauthorized training in generative models, particularly with diffusion models (DMs) fine-tuned via techniques such as Textual Inversion (TI). We identify critical vulnerabilities in existing generative DAAs through a series of analyses. First, we propose Semantic Sensitivity Maps (SSM) to visualize the influence of poisoning on token-level text embeddings during TI training. Second, we observe that DMs exhibit non-uniform learning behavior across timesteps, with DAAs primarily injecting poison signals at early (low-noise) stages—leaving high-timestep stages insufficiently defended. Third, we show that poisoned signals shift attention away from semantically relevant regions, impairing embedding quality. Based on these insights, we introduce Safe-Zone Training (SZT), a novel training strategy that mitigates poisoning in DMs. SZT comprises 3 key components: (1) JPEG compression to weaken high-frequency poison signals, (2) restriction to higher timesteps during textual inversion training to avoid adversarial signal at lower timesteps, and (3) loss masking to constrain learning to target regions. Our results demonstrate that SZT effectively neutralizes several state-of-the-art DAAs, revealing a significant research gap for poisoning on generative models in the future. Apart from training-time protection, we extend the study to inference-time adversarial attacks to prevent AI-driven visual manipulations. In some tasks, such as visual watermarking removal, some prior adversarial attack methods are typically constrained to specific types of target models and operate in white-box settings, limiting their applicability. Within the context of visual watermarking removal, multiple studies have revealed that watermarks are vulnerable to both inpainting-based removers and blind watermark removers. In the thesis, a novel black-box method is proposed, named Adversarial Watermark Defender with Attribution-Guided Perturbation (AWD-AGP), which is the first attack defending against both types of watermark removers. The key innovation for superior transferability lies in leveraging the mechanistic differences between inpainting-based and blind removal models.", "source_url": "https:\/\/openalex.org\/W7127049003", "pdf_url": "", "doi": "10.32657\/10356\/205523", "handle": "", "repository_name": "DR-NTU (Nanyang Technological University)", "source_record_id": "W7127049003", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:51", "updated_at": "2026-11-15 16:52:51" }, { "id": 646, "title": "Hybrid Optimization and Learning-Based Techniques for Autonomous Driving", "title_ar": "", "title_en": "Hybrid Optimization and Learning-Based Techniques for Autonomous Driving", "author": "Kürsat Petek", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Universitätsbibliothek Freiburg", "university_source": "repository_host", "college": "", "department": "", "country": "Germany", "governorate": "", "publication_year": 2026, "language": "en", "research_domain": "computer-vision", "research_subdomain": "image-segmentation", "categories": "machine-learning; deep-learning; computer-vision; robotics; ai-applications; self-semi-supervised; predictive-analytics; multimodal; image-segmentation; video-analysis; ai-smart-cities", "keywords": "", "abstract": "Autonomous driving has progressed rapidly in recent years, driven largely by advances in machine learning, computer vision, and probabilistic optimization methods. Despite this progress, the deployment of autonomous systems in complex, safety-critical real- world settings such as urban driving, dynamic traffic scenes, or large-scale outdoor environments remains a major open challenge. Real-world autonomy requires systems that can calibrate and localize themselves reliably, reason about their surroundings with minimal human supervision, and continuously adapt to changing environments. This thesis develops a set of methods that jointly address these challenges by combining geometric optimization with learning-based perception, and by designing label-efficient and self-supervised systems that scale to real-world deployment. In the first part, we study hybrid optimization methods for calibration and localiza- tion, focusing on the tight coupling of classical geometry with learned perception cues. We introduce a fully automatic camera–LiDAR calibration framework that removes the need for targets, initial guesses, or manual intervention by jointly exploiting motion cues and deep 2D–3D correspondences. Building on this idea, we propose a sparse-map localization system that integrates uncertainty-aware semantic predictions into a pose graph, enabling robust operation even in highly occluded urban environments with minimal map information. The second part focuses on online semantic Bird’s-Eye-View (BEV) mapping, where the goal is to infer a semantic map of the surrounding scene directly from frontal-view images in real time. We first investigate multi-modal fusion through BEVCar, which combines cameras and radar to obtain long-range, all-weather BEV maps using low- cost sensor setups. We then reduce the reliance on BEV annotations with SkyEye, a self-supervised method that learns BEV semantics from monocular images while using only frontal-view semantic labels as supervision. Finally, LetsMap removes this remaining supervision by jointly learning geometry and scene representation from raw RGB videos alone, achieving state-of-the-art label-efficient BEV mapping with only a tiny fraction of annotated BEV data. The third part addresses the challenge that modern perception systems rely on thousands of densely annotated images which is a major bottleneck for deploying autonomy at scale. We study how to drastically reduce this dependence by developing label-efficient and continually adapting panoptic segmentation methods that can be trained from data obtained within hours rather than months. SPINO shows that combining foundation-model features with only a handful of labeled images is sufficient to produce high-quality panoptic pseudo-labels at scale. PASTEL extends this idea with normalized-cut–based instance separation and targeted self-training, achieving even stronger performance under strict annotation budgets. Finally, CoDEPS enables continual test-time learning for both depth and panoptic segmentation through a geometry-aware mixing strategy and replay-based stabilization, allowing models to adapt to new environments during deployment without forgetting previous ones. Across these contributions, the thesis shows that combining geometric structure with modern learning techniques enables perception systems that are more robust, more data-efficient, and more adaptable than traditional supervised approaches. We demonstrate these methods extensively on diverse real-world driving datasets and robotic platforms, highlighting their potential to bring scalable, reliable autonomy closer to widespread deployment.", "source_url": "https:\/\/openalex.org\/W7196982350", "pdf_url": "", "doi": "10.6094\/unifr\/285382", "handle": "", "repository_name": "FreiDok plus (Universitätsbibliothek Freiburg)", "source_record_id": "W7196982350", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:51", "updated_at": "2026-11-15 16:52:51" }, { "id": 654, "title": "Application of deep learning for enhancing simultaneous localization and mapping in autonomous driving", "title_ar": "", "title_en": "Application of deep learning for enhancing simultaneous localization and mapping in autonomous driving", "author": "Jintian Ge", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Nanyang Technological University", "university_source": "repository_host", "college": "", "department": "", "country": "Singapore", "governorate": "", "publication_year": 2024, "language": "en", "research_domain": "computer-vision", "research_subdomain": "image-segmentation", "categories": "machine-learning; deep-learning; computer-vision; robotics; transfer-learning; self-semi-supervised; multimodal; image-segmentation", "keywords": "", "abstract": "Simultaneous Localization and Mapping, commonly referred to as SLAM, represents a class of algorithms that involves constructing or updating a map of an unknown environment while simultaneously keeping track of an agent's position within it. This technique is foundational in robotics and autonomous systems, enabling machines and robots to explore, understand, and navigate their surroundings. Deep learning, characterized by its prowess in handling vast data and intricate patterns, has shown promise in several Computer Vision (CV) tasks, i.e., semantic segmentation, optical flow estimation, and disparity estimation. Deep learning excels at these tasks, delivering dense and accurate results. However, as a downstream work of perception, SLAM pays little attention to the application of these results. Therefore, this paper aims to explore possible applications of deep learning (especially CV tasks) in SLAM tasks. In this thesis, we delve into the confluence of Deep Learning and SLAM tasks, emphasizing the potential for performance enhancement and methodological innovation. Our exploration is structured around four SLAM related tasks, each distinct yet synergistically linked around a center topic \"application of deep learning to enhance performance\": an automatic calibration approach for radar-camera system as a pre-process of multi-modal SLAM, which leverages deep learning models to eliminate the need for special markers; the application of deep learning to stereo 3D reconstruction through disparity estimation, presenting its potential for dense environment local mapping; the utility of semantics in initial pose estimation and feature matching within visual odometry for localization, refining the precision of positioning techniques; and introduction of supervised contrastive learning for image representation, targeting the challenges associated with fine-tuning deep learning models in real-world deployments. Collectively, these tasks underline the potential of deep learning in advancing robotic methodologies and offer a vision for future research directions.", "source_url": "https:\/\/openalex.org\/W4396622348", "pdf_url": "", "doi": "10.32657\/10356\/174794", "handle": "", "repository_name": "DR-NTU (Nanyang Technological University)", "source_record_id": "W4396622348", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:51", "updated_at": "2026-11-15 16:52:51" }, { "id": 680, "title": "Leveraging deep learning for visual understanding of videos", "title_ar": "", "title_en": "Leveraging deep learning for visual understanding of videos", "author": "Clement Xian Ren Tan", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Nanyang Technological University", "university_source": "repository_host", "college": "", "department": "", "country": "Singapore", "governorate": "", "publication_year": 2024, "language": "en", "research_domain": "nlp", "research_subdomain": "llm", "categories": "machine-learning; deep-learning; computer-vision; nlp; robotics; self-semi-supervised; predictive-analytics; transformers; multimodal; image-classification; object-detection; image-segmentation; video-analysis; llm", "keywords": "", "abstract": "In the field of Computer Vision (CV), the pursuit of human-level recognition and reasoning of visual scenes has been a long-standing aspiration. Over the past decade, significant contributions to the progress of CV have been made by deep learning, facilitated by the availability of big data and increased computational power. In CV, visual understanding encompasses not only the fundamental aspect of recognition, which involves identifying and categorizing objects and patterns within visual data, but also reasoning, which involves higher-level cognitive processes such as inferring relationships, predicting outcomes, and drawing meaningful conclusions based on the observed visual information. Human-like recognition capabilities across a variety of visual recognition tasks such as image classification, object detection, semantic segmentation, and instance segmentation, have been achieved by machines. The huge success of deep learning for visual recognition tasks has prompted researchers to tackle visual reasoning tasks, which are more challenging. The recognition and reasoning of objects within a scene is of paramount importance to applications such as human-robot collaborations and autonomous vehicles. This thesis aims at addressing the issues associated with current deep learning models for visual understanding and will focus on three tasks: (1) video object segmentation, (2) abductive action inference, and (3) action-conditioned scene graph prediction. Firstly, video object segmentation is an object tracking task that traditionally relies on supervised learning and necessitates extensive annotated datasets for deep learning model training. It is confronted with issues of repetitiveness and impracticality due to the manual labeling of these vast datasets. In addition, tracking objects within videos presents a unique challenge caused by appearance changes and occlusion across video frames. The work proposed in this thesis explores the potential of self-supervised learning for video object segmentation, aiming to harness freely available internet data such as YouTube videos for model training. Different from existing self-supervised learning approaches for video object segmentation that model pixel-to-pixel correspondence, this research shifts the focus towards modeling superpixel-to-superpixel correspondence. To achieve this, a novel approach which involves the tracking of superpixels between video frames through an attention mechanism trained within an end-to-end self-supervised framework is proposed. This approach aims to enhance the performance of existing self-supervised video object segmentation models by leveraging deep learning techniques and mitigating the limitations associated with supervised learning using large-scale annotated datasets. It also aims to enhance existing self-supervised models with the benefits of superpixels. Secondly, this thesis introduces the complex task of abductive action inference to assess the abductive reasoning abilities of deep learning models in comprehending visual scenes. A set of innovative object-relational models designed specifically for this task is presented and compared with existing state-of-the-art image, video, and vision-language models. These models are tasked with abducting the human-performed action that led to the scene depicted in an image or snapshot. The experiments demonstrate the potential of deep learning models in performing abductive action inference. This research is vital in advancing our understanding of the reasoning capacities of deep learning models, representing a significant stride toward attaining human-level reasoning capabilities in visual comprehension. Lastly, in action-conditioned scene graph prediction, the task involves predicting scene graph relations for a future state based on an initial state and an action. Instead of predicting visual representations for future states, each state is organized into a scene graph comprising human-object relationship triplets, thereby succinctly encapsulating the dynamics within the scene. Notably, this task remains underdeveloped in visual understanding research, with prior video scene graph generation approaches largely overlooking actions as a crucial signal. Actions applied to a preconditioned state generate an effect state embodying their anticipated outcomes. Human cognitive reasoning allows the inference of action consequences based on the initial scene context. To address this challenge, this thesis introduces the Action-conditioned Scene Graph dataset. Furthermore, this thesis proposes the AERT (Action-conditioned Effect Relational Transformer) model, designed to capture scene relations and action context to effectively predict future scene graph relations.", "source_url": "https:\/\/openalex.org\/W4402201451", "pdf_url": "", "doi": "10.32657\/10356\/179521", "handle": "", "repository_name": "DR-NTU (Nanyang Technological University)", "source_record_id": "W4402201451", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:52", "updated_at": "2026-11-15 16:52:52" }, { "id": 684, "title": "A study on Human Emotion Recognition in Video Images using Deep Learning", "title_ar": "", "title_en": "A study on Human Emotion Recognition in Video Images using Deep Learning", "author": "Jargalsaikhan Orgil", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "", "university_source": "", "college": "", "department": "", "country": "Japan", "governorate": "", "publication_year": 2022, "language": "en", "research_domain": "nlp", "research_subdomain": "sentiment-analysis", "categories": "machine-learning; deep-learning; computer-vision; nlp; reinforcement-learning; self-semi-supervised; image-classification; image-segmentation; video-analysis; sentiment-analysis; speech-recognition", "keywords": "", "abstract": "From the beginning of this century, Artificial Intelligence (AI) has evolved to handle problems in image recognition, classification, segmentation, etc. AI learning is categorized by supervised, semi-supervised, unsupervised or reinforcement learning. Some researchers have said that the future of AI is self­awareness, which is based on reinforcement learning by rewards based on task success. Moreover, it is said that the reward would be harvested from human reactions, specially emotion recognition. On the other hand, emotion recognition is a new inspiring field, but the lack of enough amount of data for training an AI system is the major problem. Fortunately, in the near future, it will be necessary to correctly recognize human emotions because image and video dataset availability is rapidly increasing. Emotions are mental reactions (such as anger, fear, etc.) marked by relatively strong feelings and usually causing physical reactions to previous actions in a short time duration focused on specific objects. In this Work, we are focusing on emotion recognition using face, body part, and intonation. As stated earlier, automatic understanding of human emotion in a wild setting using audiovisual signals is extremely challenging. Latent continuous dimensions can be used to accomplish the analysis of human emotional states, behaviors, and reactions displayed in real-world settings. Moreover, Valence and Arousal combinations constitute well-known and effective representations of emotions. In this thesis, a new Non-inertial loss function is proposed to train emotion recognition deep learning models. It is evaluated in wild settings using four types of candidate networks with different pipelines and sequence lengths. It is then compared to the Concordance Correlation Coefficient (CCC) and Mean Squared Error (MSE) losses commonly used for training. To prove its effectiveness on efficiency and stability in continuous or non-continuous input data, experiments were performed using the Aff-Wild dataset. Encouraging results were obtained. The contributions of the proposed method Non-Inertial loss function are as follows: 1.The new loss function allows for Valence and Arousal to be viewed together. 2.Ability to train on less data. 3.Better results. 4.Faster training times. The rest of this thesis explains our motivation, the proposed methods and finally presents our results.", "source_url": "https:\/\/openalex.org\/W7146545736", "pdf_url": "", "doi": "", "handle": "", "repository_name": "Institutional Repositories DataBase (IRDB)", "source_record_id": "W7146545736", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:52", "updated_at": "2026-11-15 16:52:52" }, { "id": 736, "title": "Data efficient deep multimodal learning", "title_ar": "", "title_en": "Data efficient deep multimodal learning", "author": "Meng Shen", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Nanyang Technological University", "university_source": "repository_host", "college": "", "department": "", "country": "Singapore", "governorate": "", "publication_year": 2025, "language": "en", "research_domain": "nlp", "research_subdomain": "llm", "categories": "machine-learning; deep-learning; computer-vision; nlp; medical-ai; self-semi-supervised; multimodal; llm", "keywords": "", "abstract": "Multimodal learning, which enables neural networks to process and integrate information from various sensory modalities such as vision, language, and sound, has become increasingly important in applications ranging from affective computing and healthcare to advanced multimodal chatbots. Despite its potential, multimodal learning faces significant challenges, particularly in the area of data efficiency. The requirement for large, high-quality datasets from multiple modalities presents a substantial barrier, limiting the scalability and accessibility of large multimodal models. This dissertation addresses several key issues in data-efficient deep multimodal learning, focusing on the imbalanced multimodal data selection, the cold-start problem in multimodal active learning, and the mitigation of hallucinations in large vision-language models. Firstly, we analyze the limitations of conventional active learning strategies, which tend to favor dominant modalities, leading to unbalanced multimodal models that neglect weaker modalities. To overcome this, we propose a gradient embedding modulation method that ensures a more equitable data selection process across modalities, resulting in models that fairly uilize both weak and strong modalities. Building on our work in warm-start active learning, we tackle the cold-start problem in multimodal active learning, where no initial labels are available for warm-start data selection. We develop a two-stage approach that first reduces the modality representation gap through multimodal self-supervised learning, utilizing unimodal prototypes to harmonize representations across modalities. In the subsequent data selection stage, we introduce a regularization term to maximize modality alignment, leading to improved model performance using the same amount of data compared to existing methods. Extending our focus from data selection to the usage of training data, we address the challenge of hallucinations in large vision-language models, where the models generate content that is incorrect in the context of input images. We investigate the relationship between hallucinations and visual dependence of tokens, revealing that certain tokens contribute disproportionately to these hallucinatory. Based on this insight, we propose an approach that adjusts training weights according to the visual dependence of tokens, thereby reducing the hallucination rate without requiring additional training data or inference costs. The contributions of this thesis offer significant advancements in the field of dataefficient multimodal learning. By developing novel methods for balancing multimodal data selection, addressing cold-start problem in multimodal active learning, and mitigating hallucinations in large vision-language models, this work paves the way for more practical and scalable multimodal learning systems that require less data and computational effort while achieving superior performance.", "source_url": "https:\/\/openalex.org\/W4407130261", "pdf_url": "", "doi": "10.32657\/10356\/182346", "handle": "", "repository_name": "DR-NTU (Nanyang Technological University)", "source_record_id": "W4407130261", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:52", "updated_at": "2026-11-15 16:52:52" }, { "id": 737, "title": "Advancing Computational Pathology with Multimodal Deep Learning", "title_ar": "", "title_en": "Advancing Computational Pathology with Multimodal Deep Learning", "author": "Md. Mamunur Rahaman", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "", "university_source": "", "college": "", "department": "", "country": "Australia", "governorate": "", "publication_year": 2025, "language": "en", "research_domain": "nlp", "research_subdomain": "llm", "categories": "machine-learning; deep-learning; computer-vision; nlp; medical-ai; responsible-ai; few-zero-shot; self-semi-supervised; predictive-analytics; multimodal; llm; medical-imaging; xai", "keywords": "", "abstract": "Histopathological assessment of tissue biopsies remains the cornerstone of cancer diagnosis, yet it is limited by diagnostic subjectivity and an incomplete view of the molecular heterogeneity that drives outcomes. This dissertation presents a suite of artificial intelligence (AI) frameworks that integrate deep learning, computer vision, and multimodal data to enhance diagnostic precision, prognostication, and therapeutic guidance across multiple cancer settings. The research first establishes a method for improving classification robustness through HistopathAI, a framework employing supervised contrastive learning and a hybrid deep feature fusion architecture. This approach yields more discriminative and generalizable feature representations, particularly in the context of class imbalance common in medical datasets. Building on this, the work demonstrates the power of integrating visually interpretable features from the tumor microenvironment. An AI-driven biomarker for Head and Neck Squamous Cell Carcinoma (HNSCC), which quantifies the spatial architecture of immune and stromal cells, was developed, and its prognostic value was validated on independent patient cohorts, confirming its strong potential for clinical risk stratification. A central contribution is the development of “virtual spatial transcriptomics” that bridges histology and genomics. The BrSTNet framework predicts spatial gene expression from routine H&E images in breast cancer, establishing a strong benchmark for image-to-transcriptome inference. This capability is translated to a direct clinical application in prostate cancer with ST-DoxPCa, a predictive signature for docetaxel benefit. The validation strategy was explicitly tailored to disease context: for the RTOG0521 trial (localized disease), a prognostic model trained independently in the Cleveland Clinic prostatectomy cohort (CCF) was applied in a fully blinded manner, without refitting to stratify patients and assess treatment effect; for the CHAARTED trial (metastatic disease), a separate treatment-response model was trained exclusively on the standard-of-care (ADT-only) arm. In both settings, ST-DoxPCa stratified patients into subgroups that either derived a significant survival benefit from docetaxel or showed no advantage, supporting treatment de-escalation for likely non-responders. Finally, to address the critical challenge of data scarcity, the thesis introduces MR-PHE, a novel framework for zero-shot learning in histopathology. By leveraging vision-language models with a multi-resolution, prompt-guided approach, MR-PHE enables accurate classification without task-specific training data, often outperforming fully supervised methods and offering a scalable solution for rare diseases. Collectively, these contributions advance the frontier of computational pathology, delivering a suite of validated, AI-driven tools that translate complex tissue-based data into clinically actionable insights. The methodologies developed herein represent a significant step toward a new paradigm of precision oncology, where patient care is personalized based on a deeper, quantitative understanding of tumor biology derived from standard pathology slides.", "source_url": "https:\/\/openalex.org\/W7116859814", "pdf_url": "", "doi": "10.26190\/unsworks\/31906", "handle": "", "repository_name": "UNSWorks (University of New South Wales, Sydney, Australia)", "source_record_id": "W7116859814", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:52", "updated_at": "2026-11-15 16:52:52" }, { "id": 741, "title": "Exploring Multimodal Deep Learning Approaches for Coronary Artery Disease Diagnosis", "title_ar": "", "title_en": "Exploring Multimodal Deep Learning Approaches for Coronary Artery Disease Diagnosis", "author": "Popov, Maxim", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "", "university_source": "", "college": "", "department": "", "country": "United Arab Emirates", "governorate": "", "publication_year": 2025, "language": "en", "research_domain": "nlp", "research_subdomain": "llm", "categories": "machine-learning; deep-learning; computer-vision; nlp; medical-ai; self-semi-supervised; multimodal; llm; medical-imaging", "keywords": "", "abstract": "Coronary Artery Disease (CAD) is a prevalent cause of death in the modern world, with about 20 million victims every year. Limited annotated data, inter-observer variability, and the intrinsic complexity of angiographic imaging challenge its diagnosis. This thesis presents a novel deeplearning framework to enhance coronary artery disease (CAD) diagnosis through multimodal analysis of Xray coronary angiograms. We develop several interrelated contributions to address these issues that combine advanced contrastive learning and vision-language modeling.", "source_url": "https:\/\/openalex.org\/W7106339970", "pdf_url": "", "doi": "10.82419\/166", "handle": "", "repository_name": "MBZUAI iRep", "source_record_id": "W7106339970", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:52", "updated_at": "2026-11-15 16:52:52" }, { "id": 743, "title": "A Study on Robustness, Adaptation, and Generalization for Visual and Multimodal Learning", "title_ar": "", "title_en": "A Study on Robustness, Adaptation, and Generalization for Visual and Multimodal Learning", "author": "Xiasi Wang", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "", "university_source": "", "college": "", "department": "", "country": "", "governorate": "", "publication_year": 2026, "language": "en", "research_domain": "nlp", "research_subdomain": "llm", "categories": "machine-learning; deep-learning; computer-vision; nlp; responsible-ai; ai-applications; transfer-learning; self-semi-supervised; multimodal; llm; fairness; ai-cybersecurity", "keywords": "", "abstract": "Recent advances in visual and multimodal learning have demonstrated impressive performance on diverse tasks. However, their reliability in real-world scenarios remains challenged by issues of robustness, adaptation, and generalization. In this thesis, we tackle these challenges by establishing four integral pillars of reliable deep learning: (1) robustness against adversarial attacks, (2) adaptation to sequential data streams, (3) resilience to domain shift and unseen categories, and (4) enhanced generalization via self-supervised learning. First, we identify a critical vulnerability in vision-language models (VLMs): their susceptibility to adversarial manipulation of inference efficiency. We propose a black-box evaluation framework to assess this efficiency robustness, establishing a new dimension for VLM safety. Second, we find that VLMs exhibit limited adaptability to evolving data streams. We explore long-tailed class incremental learning within the VLM framework and introduce a method to address the dual challenges of catastrophic forgetting and data imbalance. Third, we examine vision models’ failures on out-of-distribution samples. We tackle the open-set domain adaptation challenge, where both distribution shift and unknown classes appear in the target domain, and develop a two-stage framework that adapts the model to classify known classes while detecting unknown ones. Fourth, we revisit self-supervised pretraining in computer vision and design a novel objective to enhance the generalizability of pre-trained models. Collectively, these contributions form a unified framework for advancing robustness, adaptation, and generalization, facilitating more trustworthy visual and multimodal intelligence.", "source_url": "https:\/\/openalex.org\/W7135015825", "pdf_url": "", "doi": "10.14711\/thesis-hdl169527", "handle": "", "repository_name": "", "source_record_id": "W7135015825", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:52", "updated_at": "2026-11-15 16:52:52" }, { "id": 745, "title": "Learning semantic representations of language, vision and code for visualization-oriented dialogue", "title_ar": "", "title_en": "Learning semantic representations of language, vision and code for visualization-oriented dialogue", "author": "Henrik Voigt", "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; computer-vision; nlp; generative-ai; responsible-ai; expert-systems; transfer-learning; self-semi-supervised; transformers; multimodal; llm; fairness", "keywords": "", "abstract": "Data visualization is essential for navigating the vast volume of data that defines our time, letting humans use perception to make sense of complex datasets. Yet creating visualizations requires mastery of specialized programming languages, producing a gap between the analytical intention in the user's mind and its implementation in executable code. This raises the cognitive load on experts, who cannot concentrate fully on visual exploration while also writing the visualization program, and it creates a barrier for non-experts, hindering widespread access to insights from data. Bridging this gap is a central challenge for Visualization-oriented Natural Language Interfaces (V-NLIs), which translate human intent into executable code. This work decomposes the problem into two interrelated challenges: understanding, which builds rich, contextual representations of user intent, and generation, which reliably synthesizes visualization programs under strict real-time constraints. The thesis addresses both through five interconnected contributions. It begins with a survey of 119 papers, revealing that while the community focuses on sophisticated interactive exploration, the underlying technology relies on classical, rule-based methods, motivating a turn to deep learning. KeywordScape, a visual analytics system, makes the contextualized representations of transformer models visible and interactive, showing they can themselves resolve ambiguity and enable fine-grained exploration. Paparazzi extends this to multimodal representations, revealing that large vision-language models such as CLIP exhibit surprising deficiencies in 3D viewpoint understanding, and introduces a data-efficient contrastive fine-tuning method to improve geometric grounding. Finally, VIST5, a practical V-NLI built on a retrieval-augmented architecture around a small language model, shows that complex visualization code can be synthesized in real time while ensuring data privacy and modular extensibility.", "source_url": "https:\/\/openalex.org\/W7170734941", "pdf_url": "", "doi": "10.22032\/dbt.71264", "handle": "", "repository_name": "Common Library Network (Der Gemeinsame Bibliotheksverbund)", "source_record_id": "W7170734941", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:52", "updated_at": "2026-11-15 16:52:52" }, { "id": 753, "title": "Aligning vision and language for image captioning using deep learning", "title_ar": "", "title_en": "Aligning vision and language for image captioning using deep learning", "author": "Cai, Chen", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Nanyang Technological University", "university_source": "repository_host", "college": "", "department": "", "country": "Singapore", "governorate": "", "publication_year": 2024, "language": "en", "research_domain": "deep-learning", "research_subdomain": "transformers", "categories": "machine-learning; deep-learning; computer-vision; nlp; self-semi-supervised; predictive-analytics; rnn; transformers; multimodal; video-analysis", "keywords": "", "abstract": "A longstanding objective in the field of multi-modal research uniting computer vision and natural language processing is to develop models that can comprehend the intricate relationship between vision and language. In recent years, we have witnessed notable developments directed towards this objective, enabling computers to interpret visual information and articulate it through captions. Although significant advances have been made, aligning complex visual scenes and language for image captioning tasks using deep learning approaches remains challenging due to the distinct characteristics between the modalities. In this thesis, we propose four new methods that effectively align vision and language for image captioning-based tasks through deep learning techniques. These methods include integrating visual and semantic attributes for controllable image captioning, aligning visual and object semantics for grounded image captioning, learning the relationship between multi-scale visual features with language for image change captioning, and bridging visual and global temporally caption knowledge for temporal sentence grounding. The major contributions of this thesis are summarized as follows: A new Attribute Controlled Image Captioning (ACC) method has been proposed that seamlessly integrates semantic attributes with visual content, enabling automatic modification of the generated captions in the fashion domain. Our approach utilizes semantic attributes as a control signal, giving users the ability to specify particular fashion attributes and styles to incorporate while generating captions. Furthermore, we clean, filter, and assemble a new fashion image caption dataset to facilitate learning and enable us to investigate the effectiveness of our method. A new one-stage Weakly Supervised Grounded Image Captioner (WS-GIC) is proposed that aligns visual and word representation to perform captioning and grounding at the top-down image level. We introduce a Recurrent Grounding Module (RGM) within the decoder to compute Visual Language Attention Maps (VLAMs) for grounding, where VLAMs indicate the spatial region and location of the generated groundable object words in the caption. In addition, we explicitly inject a relation module into our one-stage framework to encourage the relation understanding. The relation semantics aid the prediction of relation words in the caption. A new Interactive Change-aware Transformer Network (ICT-Net) is proposed to extract and incorporate the most critical change of interest in the image, enhancing the generation of change descriptions for complex remote sensing bitemporal scenes. The proposed framework comprises an Interactive Change-aware Encoder (ICE) to capture the crucial difference between bitemporal image features, an Adaptive Fusion Module (AFM) to adaptively aggregate the relevant change-aware features in the encoder layers while minimizing the impact of irrelevant visual features, and a Cross Gated-Attention (CGA) module in the change decoder that enhances the modeling of essential relationships between multi-scale features with word representation, thereby improving change captioning generation. A new Temporal Sentence Grounding (TSG) method is proposed to bridge the domain gap between multi-modal features by leveraging extensive temporally global caption knowledge that is sourced from the relevant video and temporally localized text queries. We introduce the Pseudo-query Intermediary Network (PIN) to contrastively align visual features with temporally global textual knowledge to enhance the similarity between visual and language features. Furthermore, we leverage the pseudo queries prompt to propagate the knowledge, enhancing the learning of feature alignment within the multi-modal fusion module for better temporal grounding. Throughout the thesis, we illustrate how each of the proposed methods aligns vision and language for image captioning-based tasks. Experimental results on public datasets indicate that the proposed methods can achieve better performance. These approaches contribute to enhancing vision and language multi-modalities understanding.", "source_url": "https:\/\/openalex.org\/W4405996657", "pdf_url": "https:\/\/dr.ntu.edu.sg\/bitstream\/10356\/181511\/2\/CC_ntu_thesis.pdf", "doi": "10.32657\/10356\/181511", "handle": "", "repository_name": "DR-NTU (Nanyang Technological University)", "source_record_id": "W4405996657", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:52", "updated_at": "2026-11-15 16:52:52" }, { "id": 778, "title": "Contrastive Temporal Explanations: A Novel Explanation Method for Machine Learning-Based Fault Prediction. Master Thesis University Utrecht", "title_ar": "", "title_en": "Contrastive Temporal Explanations: A Novel Explanation Method for Machine Learning-Based Fault Prediction. Master Thesis University Utrecht", "author": "Joris Graff", "supervisor": "", "degree_type": "master", "degree_evidence": "نص السجل: «Master»", "university": "", "university_source": "", "college": "", "department": "", "country": "Netherlands", "governorate": "", "publication_year": 2020, "language": "en", "research_domain": "medical-ai", "research_subdomain": "xai", "categories": "machine-learning; medical-ai; responsible-ai; ai-applications; self-semi-supervised; predictive-analytics; xai; fairness; ai-industry", "keywords": "", "abstract": "Digital twinning has enabled the real-time usage of sensor readings from a variety of systems, such as industrial machines and medical patients.These time-series data offer opportunities to accurately predict the system's future behaviour, e.g. by predicting whether faults are likely to occur within the near future on the basis of the most recent sensor data.Machine learning (ML) algorithms have shown an impressive accuracy in such fault prediction tasks, but suffer from a lack of explainability.This situation problematises the application of such algorithms to real-life problems, where user insight into the prediction model is generally a desideratum.This thesis aims to improve the explainability, and thereby trustworthiness, of fault prediction ML algorithms.In order to have a solution that addresses explainability, the main criteria for successful explanation algorithms have to be identified.A first challenge of this thesis is therefore the formulation of these citeria in a way that answers to theoretical and practical aspects of explainability.A further challenge is to meet these criteria within the domain of fault prediction.The result is a novel solution, called Contrastive Temporal Explanations (CTE), which is shown to outperform state-of-the-art methods on most of the key criteria.First of all, I wish to thank both of my supervisors, Paolo Pileggi from TNO and Hakim Qahtan from the University of Utrecht.Paolo has tirelessly assisted me during the duration of the internship of which this thesis is a product, even when the intervention of a pandemic made it impossible to meet in person.His many phone calls were filled with both encouragement and useful comments.Without his enthusiasm and insights this thesis would not have been possible in the current form.Hakim has helped give structure to the initial literature review stage of this project, and has given me many invaluable suggestions on the content, structure and stylistic aspects of the end product.Without his incidental adjustments I would have repeatedly steered decidedly off course.I also wish to thank Ron Snijders of TNO Groningen for commenting on earlier versions of parts of this thesis, and Sietse Drost and Lars-Uno Axelsson for providing the problem and the data that initially inspired this project, as well as for receiving me and other TNO interns and employees for a tour of the OPRA headquarters.Finally, I am indebted to the researchers, coordinators and administrative and HR officers at the ICT department of TNO The Hague, who have all been crucial in the success of my internship, both in their helpfulness in administrative issues and in creating a welcoming and intellectually stimulating environment that has helped shape many of the ideas presented in this thesis.", "source_url": "https:\/\/openalex.org\/W7132381722", "pdf_url": "https:\/\/publications.tno.nl\/publication\/34637482\/Dhw4Ki\/graff-2020-contrastive.pdf", "doi": "", "handle": "", "repository_name": "TNO Repository", "source_record_id": "W7132381722", "other_sources": "", "verification_status": "verified", "created_at": "2026-11-15 16:52:53", "updated_at": "2026-11-15 16:52:53" }, { "id": 848, "title": "Constrained deep networks for medical image segmentation", "title_ar": "", "title_en": "Constrained deep networks for medical image segmentation", "author": "Hoel Kervadec", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "", "university_source": "", "college": "", "department": "", "country": "Canada", "governorate": "", "publication_year": 2021, "language": "en", "research_domain": "medical-ai", "research_subdomain": "medical-imaging", "categories": "machine-learning; deep-learning; computer-vision; medical-ai; self-semi-supervised; predictive-analytics; cnn; image-segmentation; medical-imaging", "keywords": "", "abstract": "Weakly supervised image segmentation, in the form of partially labeled images, is attracting significant research attention as it can mitigate the need for laborious pixel annotations required by deep learning models. Enforcing high-order, global inequality constraints on the network outputs can leverage unlabeled data by guiding the training with prior knowledge, restricting the search space during training to anatomically feasible solutions. A range of possible values (such as a lower\/upper bounds on the size of a organ) can be very valuable to guide training. However, in the context of deep neural networks, standard Lagrangian optimization has been largely avoided, mainly due to the instability and computational complexity ensuing from alternating explicit dual updates and stochastic optimization. Interior point methods, despite their popularity in convex optimization, are not applicable neither, as they require a feasible starting point, which is itself a difficult constrained problem for deep neural networks. In this thesis, we investigate hard inequality constraints in the context of deep networks with both quadratic penalties and more principled log-barrier extensions. We also investigate methods to mitigate class-imbalance in segmentation problems, such as in brain lesions dataset, by constraining the boundary of the predicted segmentation to match the ground-truth boundary. This thesis produced five different publications as first author, and four papers as co-author. Our papers received several awards, and we were invited to publish extended versions of our works in two special issues of Medical Image Analysis (MedIA). In our first contribution, we propose to introduce a differentiable penalty, which enforces inequality constraints directly in the loss function, avoiding expensive Lagrangian dual iterates and proposal generation. From constrained-optimization perspective, our simple penalty-based approach is not optimal as there is no guarantee that the constraints are satisfied. However, surprisingly, it yields substantially better results than Lagrangian-based constrained convolutional neural networks, while reducing the computational demand for training. By annotating only a small fraction of the pixels, our approach reaches performances comparable to full supervision, on three separate tasks. While our experiments focused on basic linear constraints such as the target-region size and image tags, our framework can be easily extended to other non-linear constraints, e.g., invariant shape moments and other region statistics. In our second contribution, we propose log-barrier extensions, which approximate Lagrangian optimization of constrained-CNN problems with a sequence of unconstrained losses. Unlike standard interior-point and log-barrier methods, our formulation does not need an initial feasible solution. We report comprehensive weakly supervised segmentation experiments, with various constraints, showing that our formulation outperforms substantially the existing constrained-CNN methods, both in terms of accuracy, constraint satisfaction and training stability. In our third contribution, we enforce constraints on the boundary of predicted segmentation. Widely used loss functions for CNN segmentation, such as Dice or cross entropy, are based on integrals over the segmentation regions. Unfortunately, for highly unbalanced segmentations, such regional summations have values that differ by several orders of magnitude across classes, which affects training performance and stability. We propose a boundary loss, which takes the form of a distance metric on the space of contours, not regions. This can mitigate the difficulties of highly unbalanced problems because it uses integrals over the interface between regions instead of unbalanced integrals over the regions. Furthermore, a boundary loss complements regional information. Inspired by graph-based optimization techniques for computing activecontour flows, we express a non-symmetric L2 distance on the space of contours as a regional integral, which avoids completely local differential computations involving contour points. This yields a boundary loss expressed with the regional softmax probability outputs of the network, which can be easily combined with standard regional losses and implemented with any existing deep network architecture for N-D segmentation. We report comprehensive evaluations on different unbalanced problems, showing that our boundary loss can yield significant increases in performances while improving training stability. In a fourth contribution, we investigates a curriculum-style strategy for semi-supervised CNN segmentation, which devises a regression network to learn image-level information such as the size of the target region. These regressions are used to effectively regularize the segmentation network, constraining the softmax predictions of the unlabeled images to match the inferred label distributions.", "source_url": "https:\/\/openalex.org\/W3215133863", "pdf_url": "https:\/\/espace.etsmtl.ca\/id\/eprint\/2734\/1\/KERVADEC_Hoel.pdf", "doi": "", "handle": "", "repository_name": "Espace ÉTS (ETS)", "source_record_id": "W3215133863", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:53", "updated_at": "2026-11-15 16:52:53" }, { "id": 897, "title": "Transfer learning for language model adaptation", "title_ar": "", "title_en": "Transfer learning for language model adaptation", "author": "Bari M. Saiful", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Nanyang Technological University", "university_source": "repository_host", "college": "", "department": "", "country": "Singapore", "governorate": "", "publication_year": 2023, "language": "en", "research_domain": "nlp", "research_subdomain": "llm", "categories": "machine-learning; deep-learning; computer-vision; nlp; transfer-learning; self-semi-supervised; predictive-analytics; llm; sentiment-analysis; machine-translation", "keywords": "", "abstract": "Language is the pathway to democratize the boundary of land and culture. Bridging the gap between languages is one of the biggest challenges of Artificial Intelligent (AI) systems. The current success of AI systems is dominated by the supervised learning paradigm where gradient-based learning algorithms (i.e., SGD, Adam) are designed to optimize complex high-dimensional planes. These algorithms learn from statistical observations that are typically collected with the intention of a specific task (i.e., product review, sentiment analysis). The use of task-dependent samples makes the learning procedure tedious as it requires manually annotated data. On the contrary, without having a sufficient amount of samples to represent the distribution, deep learning models tend to suffer from the lack of robustness. Due to the natural puzzle of randomness, not all sets of observations are observed in the data collection procedure, thus creating Out-of-Distribution (OOD) problem in learning algorithms. In the search of a generic task agnostic distribution, a large collection of text over various domains can be considered as - Standard Natural Text Distribution (SNTD). The generic idea of Transfer Learning for Natural Language Processing (NLP) is to utilize the SNTD knowledge for any other task-dependent training. Learning SNTD, followed by a task-adaptation method with a smaller volume of annotated data has yielded state-of-the-art (SOTA) results in various supervised NLP tasks. However, annotated data for every task in every language is rare. In language models, there are many kinds of distributional variance. One of the most common ways that distributional variance is encoded into a language model is when it is trained with mono-lingual text and learned dis-jointly. Word embeddings produced from these language models are then used as pre-trained embedding vectors to adapt to a downstream task. We propose adversarial training to project two monolingual distributions into the same space and then improve the robustness of the model by augmented fine-tuning with parameter sharing. By projecting mono-lingual language distribution into the same cross-lingual space, the distribution of the language become aware of each other. These projected distribution are semantically aware of each other in the latent space. Thus when we train one distribution the other distribution automatically adapt to the training data, making it easier to transfer (exchange) knowledge. On top of that, the proposed novel self-training architecture improves cross-lingual transfer by a large margin. Then, we focus on the jointly trained multi-lingual language model, where there is no pre-dominant distributional variance. In the multi-lingual language model, we put more focus on downstream task adaptation. We found that semi-supervised learning with pseudo-augmented data from a pre-trained language model can greatly improve the performance of tasks further down the line. In the end, we introduce a new novel data augmentation framework that uses the neighboring (vicinal) samples of the original training data without explicitly using any parallel text corpora or machine translation system. Our proposed method performs simultaneous self-training with data augmentation and unsupervised sample selection. It also proposes curriculum strategies for different domain samples. With extensive experiments on three different cross-lingual tasks spanning many language pairs, we have demonstrated the effectiveness of our proposed method. While all of the work done above focused on improving task adaptation in multiple languages without supervision, we further investigate how adding a few samples affects multi-lingual task adaptation. To this end, we leverage a small number of support samples from each language and propose an inference-time transductive nearest neighbor-based approach which utilizes the entropy of the query samples for predictions. We show that our proposed method outperforms full model\/full head fine-tuning as well as cross-task fine-tuning. We also show impressive performance gain (~37x) in computational costs for the full inference prediction. However, as the language models grow bigger, it becomes increasingly difficult to perform efficient inferences, especially for multiple tasks. The jointly optimized multi-lingual distribution helps transfer knowledge from a high-resource language to a low-resource language. While working on transductive nearest neighbor inference, we observe that language models are significantly prone to task distributions. Unless we use an extremely large language model (>100B), a model used for a specific task adaptation cannot be used for another task. In this dissertation, our final proposed method addresses this issue by multitask prompted learning.", "source_url": "https:\/\/openalex.org\/W4386409617", "pdf_url": "https:\/\/dr.ntu.edu.sg\/bitstream\/10356\/169892\/2\/sbmaruf_thesis.pdf", "doi": "10.32657\/10356\/169892", "handle": "", "repository_name": "DR-NTU (Nanyang Technological University)", "source_record_id": "W4386409617", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:54", "updated_at": "2026-11-15 16:52:54" }, { "id": 910, "title": "Learning generalizable and transferable representations across domains and modalities", "title_ar": "", "title_en": "Learning generalizable and transferable representations across domains and modalities", "author": "Donghyun Kim", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Boston University", "university_source": "repository_host", "college": "", "department": "", "country": "United States", "governorate": "", "publication_year": 2022, "language": "en", "research_domain": "medical-ai", "research_subdomain": "medical-imaging", "categories": "machine-learning; deep-learning; computer-vision; medical-ai; responsible-ai; transfer-learning; few-zero-shot; self-semi-supervised; multimodal; video-analysis; medical-imaging; fairness", "keywords": "", "abstract": "While deep neural networks attain state-of-the-art performance for computer vision tasks with the help of massive supervised datasets, it is usually assumed that all train and test examples are drawn independently from the same distribution. However, in real-world applications, dataset bias and domain shift violate this assumption. Test data can be from different domains represented by different distributions, which can seriously affect the model performance. Thus, learning generalizable and transferable representations is important to make a model robust to many different types of distributional shift. Domain transfer such as Domain Adaptation (DA) and Domain Generalization (DG) have been proposed to learn generalizable and transferable features across domains. Domain transfer consists of two steps: 1) pre-training, where a model is first pre-trained on an upstream task with a massive supervised dataset, e.g., ImageNet, and 2) transfer (adaptation), where the model is fine-tuned on downstream multi-domain data. In this thesis, we highlight the limitations of current domain transfer approaches and relax the limitations to produce more practical and diverse domain transfer methods. To be specific, we study: 1) Cross-Domain Self-supervised Learning for Domain Adaptation. Prior DA methods use ImageNet pre-trained models as a weight initialization (i.e., pre-training stage). However, the downstream data can be very different from that of ImageNet. Previous domain adaptation approaches assume there are many labeled data in the source domain. Some applications (e.g., Medical Imaging) may not have enough source labels. We explore the problem of few-shot domain adaptation where we only have a few source labels. In addition, we propose cross-domain self-supervised pre-training, which uses only unlabeled multi-domain data. We show that our method significantly boosts the performance of diverse domain transfer tasks. 2) Pre-training for Domain Adaptation. While many DA and DG methods have been proposed and studied extensively in prior work, little attention has been paid to pre-training for domain transfer. We provide comprehensive experiments and an in-depth analysis of pre-training in terms of network architectures, datasets, and loss functions. Finally, we observe significant improvements from the modern pre-training and propose to modernize the current evaluation protocols. 3) Multimodal Representation Learning for Domain Adaptation. We devise self-supervised formulations for multimodal domain adaptation where we promote better knowledge transfer by aligning multimodal features. We first explore a language-vision task where we align the features of multiple languages and images. Then, we explore video domain adaptation with RGB and Flow modalities and propose a joint contrastive regularization that interplays among cross-modal and cross-domain features. 4) Domain Adaptive Keypoint Detection. Lastly, we explore domain adaptive keypoint detection tasks (e.g., human and animal pose estimation) which are not well explored in prior work. We propose a unified framework for diverse keypoint detection scenarios, where we can encounter different types of domain shifts. To handle these domain shifts, we propose a multi-level feature alignment using the input-level and output-level cues and show that our method generalizes well to diverse domain adaptive keypoint detection tasks.", "source_url": "https:\/\/openalex.org\/W7126339286", "pdf_url": "", "doi": "", "handle": "", "repository_name": "OpenBU (Boston University)", "source_record_id": "W7126339286", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:54", "updated_at": "2026-11-15 16:52:54" }, { "id": 915, "title": "Data-efficient domain adaptation for pretrained language models", "title_ar": "", "title_en": "Data-efficient domain adaptation for pretrained language models", "author": "Xu Guo", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Nanyang Technological University", "university_source": "repository_host", "college": "", "department": "", "country": "Singapore", "governorate": "", "publication_year": 2023, "language": "en", "research_domain": "nlp", "research_subdomain": "llm", "categories": "machine-learning; deep-learning; nlp; responsible-ai; transfer-learning; self-semi-supervised; predictive-analytics; optimization; rnn; transformers; llm; federated-learning", "keywords": "", "abstract": "Recent advances in Natural Language Processing (NLP) are built on a range of large-scale pretrained language models (PLMs), which are based on deep transformer neural networks. These PLMs simultaneously learn contextualized word representations and language modeling by training the entire model on massive unlabeled corpora using self-supervised learning techniques, bringing about a paradigm shift that moves our focus from customizing different models for different tasks to adapting one PLM to all tasks. Studying how to adapt a general-purpose PLM to a specific domain of interest is of great significance to the deployment of PLMs. The mainstream practice is to finetune a PLM with a task-specific head on a labeled dataset from the target domain. However, for most target applications, labeled data is limited and even scarce in many low-resource scenarios. The huge number of parameters in a PLM often leaves those small datasets struggling to harness the power of the language priors. As a result, even under the same task, when a PLM finetuned on one dataset is applied to another dataset with some domain gap, it sometimes encounters performance degradation due to overfitting the previous training set. This phenomenon hinders the wide adoption of PLMs in practice, particularly in the face of new domains, calling for approaches to enhance the generalization performance of PLMs during adaptation without requesting more labeled data. Early domain adaptation methods, which leverage similar source domains to boost model performance on the target domains, are developed based on customized models using traditional neural networks such as LSTMs. These models are shallow, require longer training time to converge, and have no prior knowledge compared to PLMs. Studies show that some popular domain adaptation methods can even harm the generalization performance of PLMs on the target domains. The unique characteristics of PLMs such as unprecedented scales, rich language priors, and many hitherto underexplored skills could be uncontrollable factors that make them exhibit different learning behaviors compared to traditional models. To this end, there is a need to develop algorithms for PLMs to enhance their domain adaptation performance, thereby accelerating their wide adoption in real-world scenarios. This thesis aims to explore techniques that can efficiently make use of the target domain labeled data and better adapt a given PLM to the target domains of interest by effectively transferring knowledge from similar source domains to the target domains. To achieve this goal, I conduct research from three perspectives throughout a machine learning pipeline, each assuming only specified locations can be updated with available computing resources. That is, we keep all other conditions fixed and only make updates to the input data, model representations, and output predictions respectively. We show how to achieve better generalization performance with limited labeled data from the target domains under each scenario. To sum up, we propose a new algorithm to generate adversarial perturbations using the domain adaptation objective to enhance the transferability of soft prompt tuning in low-resource scenarios, a new model optimization algorithm that takes into account the next-step gradients of adversarial domain discriminator when optimizing the task classifiers to accommodate competing losses and a new federated learning framework that calibrates the conditional probability distribution to adapt the same PLM to multiple domains under different label distributions. We present the specific problems, related works, detailed methods, extensive experiments, and thorough discussions in the following chapters, and shed light on how to base on traditional machine learning methods while catering to newly emerging learning paradigms.", "source_url": "https:\/\/openalex.org\/W4378983312", "pdf_url": "https:\/\/dr.ntu.edu.sg\/bitstream\/10356\/167965\/2\/PhD_Thesis_GuoXu.pdf", "doi": "10.32657\/10356\/167965", "handle": "", "repository_name": "DR-NTU (Nanyang Technological University)", "source_record_id": "W4378983312", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:54", "updated_at": "2026-11-15 16:52:54" }, { "id": 944, "title": "Stance detection enhanced by large language models", "title_ar": "", "title_en": "Stance detection enhanced by large language models", "author": "Nikesh Gyawali", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Kansas State University", "university_source": "repository_host", "college": "", "department": "", "country": "United States", "governorate": "", "publication_year": 2025, "language": "en", "research_domain": "nlp", "research_subdomain": "llm", "categories": "machine-learning; nlp; medical-ai; responsible-ai; self-semi-supervised; llm; sentiment-analysis", "keywords": "", "abstract": "Stance detection, a subfield of opinion mining, is an important task in Natural Language Processing (NLP) that involves determining the attitude or position expressed by an author of a text towards a particular topic or claim, generally referred to as target.More specifically, the task involves automatically determining if an author of a text is In-Favor (Positive), Against (Negative), or Neutral towards a given target.Accurate stance detection is crucial for understanding complex social issues, guiding policy decisions, and shaping effective interventions.While the field of NLP has seen significant progress, capturing nuanced stances from diverse and often ambiguous data remains a challenge.Recent developments in Large Language Models (LLMs) have transformed the field by offering unprecedented capabilities in interpreting subtle linguistic cues and context-dependent meanings.This dissertation advances the field of stance detection by utilizing Large Language Models (LLMs) to improve methodologies across multiple critical and socially impactful domains.In this dissertation, we first explore LLM-enhanced approaches to stance detection in the socially controversial and polarizing topics of gun regulation and vaccines.We introduce the GunStance dataset, consisting of social media posts from X (formerly Twitter) posted by X users after major mass shooting events in the United States.By integrating labeled and unlabeled posts, this dataset allows comprehensive exploration of a semi-supervised learning framework in the context of stance detection.We propose a novel hybrid model that combines semi-supervised techniques with LLMs and show that our approach significantly outperforms traditional stance detection approaches.Furthermore, we assemble a large dataset of social media posts from X, capturing the vaccine discourse over a decade.The dataset includes seven years before COVID-19, as well as three COVID-19 years.Leveraging LLMs, social cognition theories and emotional dynamics, we analyze the vaccine dataset to capture the evolving public attitudes towards vaccines before and during the COVID-19 pandemic.Our study reveals increasing polarization and heightened emotional engagement, with a notable rise in vaccine skepticism amid the global health crisis.Expanding into the less controversial but important financial domain, we construct a financial stance detection corpus from annual 10-K reports filed to U.S. Securities and Exchange Commission (SEC) and earnings call transcripts (ECT) by extracting short text fragments relevant to key financial metrics, such as debt, earnings per share (EPS), and sales, and annotating them using LLM-driven methodologies with strict human validation.This financial stance detection corpus facilitates extensive evaluation of LLMs' ability to detect subtle stances towards financial metrics, a task that requires complex reasoning.Our findings demonstrate the effectiveness of LLMs in performing accurate stance detection without extensive labeled data, showcasing their potential for real-world financial analysis applications.Building upon these insights, we also introduce the Modular Prompt Optimization for Stance Detection (MoPrO-SD) framework.This framework utilizes the prompt optimization capabilities of LLMs by breaking down the complex stance detection prompt into modular, optimizable components.Each module is iteratively refined using LLMs as prompt optimizers, leading to an improved prompt that outperforms human-crafted prompts on several stance detection benchmarks.Collectively, this dissertation advances the field of stance detection by providing comprehensive evidence on the use of LLMs to enhance the performance, adaptability, and efficiency of stance detection methodologies across social media posts and financial documents, offering an analytical and scalable framework for informed and nuanced decision-making in an increasingly digital and interconnected world. Stance detection enhanced by large language models by", "source_url": "https:\/\/openalex.org\/W7218185954", "pdf_url": "https:\/\/krex.k-state.edu\/bitstreams\/ffe8f4b0-48a3-4b3c-8de7-53c0428e8d47\/download", "doi": "", "handle": "", "repository_name": "K-State Research Exchange (Kansas State University)", "source_record_id": "W7218185954", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:55", "updated_at": "2026-11-15 16:52:55" }, { "id": 1169, "title": "Intuitive Explainable Artificial Intelligence for Molecular Design", "title_ar": "", "title_en": "Intuitive Explainable Artificial Intelligence for Molecular Design", "author": "Alec Lamens", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "University of Bonn", "university_source": "repository_host", "college": "", "department": "", "country": "Germany", "governorate": "", "publication_year": 2026, "language": "en", "research_domain": "medical-ai", "research_subdomain": "xai", "categories": "machine-learning; deep-learning; medical-ai; responsible-ai; expert-systems; self-semi-supervised; predictive-analytics; xai", "keywords": "", "abstract": "With the onset of artificial intelligence (AI), machine learning (ML) approaches have become increasingly embedded into the drug discovery pipeline with the aim of accelerating research and reducing cost. Exemplary applications include quantitative structure-property relationship modeling and de novo design of candidate molecules. However, the practical impact of these ML models has been limited by the opaque nature of their predictions, termed the black-box problem. The proliferation of large-scale public chemical datasets combined with advances in high-performance computing has driven a shift from traditional ML approaches toward more complex deep learning (DL) methods. While powerful, these models further exacerbate the inability to rationalize model outcomes. Therefore, the field of explainable AI (XAI) has attracted substantial interest to improve the transparency of model decisions. Given the wide range of available XAI approaches and lack of established practices, these methods are often applied without careful consideration for their suitability to the task at hand. This represents a problem in drug discovery where interpretable explanations are critical to guide experimental follow-up. To address this issue, this dissertation proposes the adaptation of established XAI approaches to operate on the level of chemical structure and facilitate interpretable ML for molecular design. The first study demonstrates the use of Shapley additive explanations (SHAP) to analyze feature sets driving the correct prediction of compound promiscuity. This is followed by a benchmarking study comparing several methods for the calculation of feature importance values, revealing inconsistencies across approaches that complicate the interpretation of feature-importance explanations. To support chemically meaningful interpretability, SHAP values are then combined with counterfactual reasoning, yielding the SHAP-CF methodology. Thereafter, a novel counterfactual method is introduced that generates candidate counterfactuals through the iterative recombination of molecular cores and substituents. This method is evaluated on a challenging multi-class kinase inhibitor prediction task, demonstrating its capacity to generate large numbers of counterfactuals, enabling in-depth analysis. Next, a closely related XAI concept, contrastive explanations, is adapted for molecular design by introducing rational chemical modifications to test compounds through the exchange of scaffolds or substituents with structural analogues, termed MolCE. This approach is shown to provide chemically relevant contrastive explanations that enable direct causal insights into the structural factors that drive model predictions. Finally, MolAnchor is introduced, a domain-adapted variant of the rule-based Anchors method. By conforming the principles underlying the Anchors framework to operate on retrosynthetically meaningful molecular fragments, MolAnchor produces interpretable fragment-level decision rules. A follow-up study showcases how the if-then formulation of the fragment rules is particularly complementary to causal reasoning. Taken together, the studies presented hereindemonstrate the utility of XAI methods that have been augmented with domain-specific information to provide chemically intuitive explanations for predictive models applied to molecular design.", "source_url": "https:\/\/openalex.org\/W7164705685", "pdf_url": "", "doi": "", "handle": "", "repository_name": "bonndoc (University of Bonn)", "source_record_id": "W7164705685", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:59", "updated_at": "2026-11-15 16:52:59" }, { "id": 1196, "title": "Deep learning for automobile predictive maintenance under Industry 4.0", "title_ar": "", "title_en": "Deep learning for automobile predictive maintenance under Industry 4.0", "author": "Chong Chen", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "Cardiff University", "university_source": "repository_host", "college": "", "department": "", "country": "United Kingdom", "governorate": "", "publication_year": 2020, "language": "en", "research_domain": "deep-learning", "research_subdomain": "rnn", "categories": "machine-learning; deep-learning; ai-applications; self-semi-supervised; predictive-analytics; rnn; ai-cybersecurity; ai-industry; ai-smart-cities", "keywords": "", "abstract": "Industry 4.0 refers to the fourth industrial revolution, which has boosted the development of the world. An important target of Industry 4.0 is to maximize the asset uptime so to improve productivity and reduce the production and maintenance cost. The emerging techniques such as artificial intelligence (AI), industrial Internet of things (IIoT) and cyber-physical system (CPS) have accelerated the development of data-orientated application such as predictive maintenance (PdM). Maintenance is a big concern for an automobile fleet management company. An accurate maintenance prediction can be helpful to avoid critical failure and avoid further loss. Deep learning is a type of prevailing machine learning algorithm which has been widely used in big data analytics. However, how to establish a maintenance prediction model based on historical maintenance data using deep learning has not been investigated. Moreover, it is worthwhile to study how to build a prediction model when the labelled data is insufficient. Furthermore, surrounding factors which may impact automobile lifecycle have not been concerned in the state-of-the-art. Hence, this thesis will focus on how to pave the way for automobile PdM under Industry 4.0. This research is structured according to four themes. Firstly, different from the conventional PdM research that only focuses on modelling based on sensor data or historical maintenance data, a framework for automobile PdM based on multi-source data is proposed. The proposed framework aims at automobile TBF modelling, prediction, and decision support based on the multi-source data. There are five layers designed in this framework, which are data collection, cloud data transmission and storage, data mapping, pre-processing and integration, deep learning for automobile TBF modelling, and decision support for PdM. This framework covers the entire knowledge discovery process from data collection to decision support. Secondly, one of the purposes of this thesis is to establish a Time-Between-Failure (TBF) prediction model through a data-driven approach. An accurate automobile TBF iv Abstract prediction can bring tangible benefits to a fleet management company. Different from the existing studies that adopted sensor data for failure time prediction, a new approach called Cox proportional hazard deep learning (CoxPHDL) is proposed based on the historical maintenance data for TBF modelling and prediction. CoxPHDL is able to tackle the data sparsity and data censoring issues that are common in the analysis of historical maintenance data. Firstly, an autoencoder is adopted to convert the nominal data into a robust representation. Secondly, a Cox PHM is researched to estimate the TBF of the censored data. A long-short-term memory (LSTM) network is then established to train the TBF prediction model based on the pre-processed maintenance data. Experimental results have demonstrated the merits of the proposed approach. Thirdly, a large amount of labelled data is one of the critical factors to the satisfactory algorithm performance of deep learning. However, labelled data is expensive to collect in the real world. In order to build a TBF prediction model using deep learning when the labelled data is limited, a new semi-supervised learning algorithm called deep learning embedded semi-supervised learning (DLeSSL) is proposed. Based on DLeSSL, unlabelled data can be estimated using a semi-supervised learning approach that has a deep learning technique embedded so to expand the labelled dataset. Results derived using the proposed method reveal that deep learning (DLeSSL based) outperforms the benchmarking algorithms when the labelled data is limited. In addition, different from existing studies, the effect on algorithm performance due to the size of labelled data and unlabelled data is reported to offer insights for the deployment of DLeSSL. Finally, automobile lifecycle can be impacted by surrounding factors such as weather, traffic, and terrain. The data contains these factors can be collected and processed via geographical information system (GIS). To introduce these GIS data into automobile TBF modelling, an integrated approach is proposed. This is the first time that the surrounding factors are considered in the study of automobile TBF modelling. Meanwhile, in order to build a TBF prediction model based on multi-source data, a new deep learning architecture called merged-LSTM (M-LSTM) network is designed. Abstract v Experimental results derived using the proposed approach and M-LSTM network reveal the impacts of the GIS factors. This thesis aims to research automobile PdM using deep learning, which provides a feasibility study for achieving Industry 4.0. As such, it offers great potential as a route to achieving a more profitable, efficient, and sustainable fleet management.", "source_url": "https:\/\/openalex.org\/W3168907897", "pdf_url": "https:\/\/orca.cardiff.ac.uk\/137968\/1\/2021ChenCPhD.pdf", "doi": "", "handle": "", "repository_name": "ORCA Online Research @Cardiff (Cardiff University)", "source_record_id": "W3168907897", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:52:59", "updated_at": "2026-11-15 16:52:59" }, { "id": 1236, "title": "Explorations of In-Context Reinforcement Learning", "title_ar": "", "title_en": "Explorations of In-Context Reinforcement Learning", "author": "Ethan Brooks", "supervisor": "", "degree_type": "unspecified", "degree_evidence": "", "university": "University of Michigan", "university_source": "repository_host", "college": "", "department": "", "country": "United States", "governorate": "", "publication_year": 2024, "language": "en", "research_domain": "nlp", "research_subdomain": "llm", "categories": "machine-learning; deep-learning; nlp; reinforcement-learning; self-semi-supervised; rnn; transformers; llm", "keywords": "", "abstract": "In-Context Learning describes a form of learning that occurs when a model accumulates information in its context or memory. For example, a Long Short-Term Memory unit (LSTM) can rapidly adapt to a novel task as input\/target exemplars are fed into it. While in-context learning of this kind is not a new discovery, recent work has demonstrated the capacity of large \"foundation\" models to acquire this ability \"for free,\" by training on large quantities of semi-supervised data, without the sophisticated (but often unstable) meta-objectives proposed by many earlier papers #cite(< finn2017model>)#cite(< stadie2018some>)#cite(< rakelly2019efficient>). In this work we explore several algorithms which specialize in-context learning based on semi-supervised methods to the reinforcement learning (RL) setting. In particular, we explore three approaches to in-context learning of value (expected cumulative discounted reward). The first of these methods demonstrates a method for implementing policy iteration, a classic RL algorithm, using a pre-trained large language model (LLM). We use the LLM to generate planning rollouts and extract monte-carlo estimates of value from them. We demonstrate the method on several small, text-based domains and present evidence that the LLM can generalize to unseen states, a key requirement of learning in non-tabular settings. The second method imports many of the ideas of the first, but trains a transformer model directly on offline RL data. We incorporate Algorithm Distillation (AD) #cite(< laskin2022context>), another method for in-context reinforcement learning that directly distills the improvement operator from data that includes behavior ranging from random to optimal. Our method combines the benefits of AD with the policy iteration method proposed in our previous work and demonstrates benefits in performance and generalization. Our third method proposes a new method for estimating value. Like the previous methods, this one implements a form of policy iteration, but eschews monte-carlo rollouts for a new approach to estimating value. We train a network to estimate Bellman updates and iteratively feed its outputs back into itself until the estimate converges. We find that this iterative approach improves the capability of the value estimates to generalize and mitigates some of the instability of other offline methods.", "source_url": "https:\/\/openalex.org\/W4398230562", "pdf_url": "", "doi": "10.7302\/23097", "handle": "", "repository_name": "Deep Blue (University of Michigan)", "source_record_id": "W4398230562", "other_sources": "", "verification_status": "needs_review", "created_at": "2026-11-15 16:53:00", "updated_at": "2026-11-15 16:53:00" } ] }