13463 research outputs found
Sort by
Identification of two distinct uveal melanoma subtypes based on the immune cell infiltration of the primary tumour
No full textOur study reveals cellular and immunological heterogeneity within primary UM tumours. The identification of immunologically distinct tumour types, along with aggressive melanocyte subpopulations and stromal interactions, provides insight into UM pathogenesis and supports stratified immunotherapeutic approaches
Effective practices for reducing underachievement and early school leaving: insights from successful learning environments
No full textOver the past two decades, early school leaving (ESL) rates in the European Union have significantly decreased, from 17.6% in 2000 to 9.5% in 2023. ESL refers to individuals aged 18–24 who leave education without upper secondary qualifications. Ireland has excelled by cutting its ESL rate to 4% in 2023 through initiatives like the Delivering Equality of Opportunity in Schools (DEIS) Plan and TUSLA Education Support Services. These programs address inequities by offering guidance, alternative pathways, and promoting attendance and retention. Despite progress, students from disadvantaged backgrounds, including migrants and Travelers, still face poorer educational outcomes. This research examines inclusive learning environments and their role in improving engagement, retention, and progression. Case studies from two high-performing designated DEIS schools illustrate how strong school leadership fosters effective interventions by building partnerships with families and communities. Applying Bronfenbrenner’s ecological systems theory, the study emphasizes the significance of interactions at the microsystem and mesosystem levels, supported by cohesive macro-level policies. The research identifies best practices for reducing inequities and improving outcomes across diverse educational contexts
Extending Lifelog Retrieval to Multi-stream Video Retrieval at the CASTLE Challenge 2025
No full textWe present the DCU team’s system for the CASTLE Challenge at ACM Multimedia 2025, which explores video retrieval and question answering in egocentric, multi-user environments. Our system adapts techniques developed for lifelogging, particularly eventbased semantic retrieval and QA pipelines, to the CASTLE dataset with minimal architectural changes. It combines vision-language embeddings, transcript-based retrieval, and person tracking to support both automatic and interactive search workflows. In the interactive track, we introduce a modular interface for narrative reconstruction and exploratory search. Qualitative results show that the system can generate plausible, evidence-based answers to complex multimodal queries. These findings suggest that lifelog retrieval systems offer a viable foundation for broader egocentric video analysis
Identification of two distinct uveal melanoma subtypes based on the immune cell infiltration of the primary tumour
No full textUveal melanoma (UM) is the most common intraocular malignancy, with poor prognosis in metastatic cases and limited response to conventional therapies. Despite advances in genetic stratification, the immunological landscape of primary UM remains poorly understoo
An exploration of how gender interacts with other factors in predicting achievement in mathematics
No full textGender gaps persist in international assessments of mathematics, and in most cases, differences are in favour of boys. In Ireland, concern has been raised about lower numbers of girls than boys attaining at the highest levels of mathematics. The research described in this paper examined data from a longitudinal government study in Ireland. Relationships were explored between attainment in mathematics at school completion and socioeconomic status (SES), general anxiety, attainment in mathematics at age 9, and enjoyment of mathematics at age 9. Gendered differences were found in the strength of the correlation in all cases. The completion of mathematics-rich subject(s) was also explored as a potential contributing factor to variations in attainment in mathematics, but no gendered difference was found
Tunable Dot Platform for Controlling Electron Flow in Graphene
No full textWe introduce an innovative graphene-based architecture to control electronic current flows. The tunable dot platform (TDP) consists of an array of gated dots, with independently adjustable potentials, embedded in graphene. Inspired by Mie theory, and leveraging multiscattering effects, we demonstrate that tailored current behavior can be achieved due to the variety of possible dot configurations. Optimization is performed using differential evolution, which identifies configurations that maximize specific objectives, such as directing or splitting an electron beam by tuning the angular dependence of scattering. Our results demonstrate the potential of the TDP to provide precise control over induced current flows in graphene, making it a promising component for next-generation electronic and electron optic devices
Aligning Vision and Language: Harnessing Language Semantics for Efficient Vision Models
Get PDFThis thesis explores methods to enhance the efficiency, flexibility, and generalization of vision models by leveraging semantic information from language. Inspired by human multi-modal perception, the research conducted addresses key limitations in current models: high data and compute demands, limited generalization to new concepts, and suboptimal unimodal features. The thesis begins by exploring how
structured semantic information, such as domain-expert knowledge, can enhance the few-shot learning of visual concepts. A novel algorithm, BaseTransformers, is proposed to integrate semantic information, enabling computer vision models to learn new concepts with minimal labeled data by associating them with semantically similar and well-represented base concepts. Extensive evaluations on benchmark
datasets highlight improvements over few-shot vision models that do not leverage semantic information. Recognizing the scalability challenges of curated semantics, this thesis introduces a strategy to leverage large language models (LLMs) as a scalable source of semantic knowledge. The proposed VDT-Adapter learns to dynamically select and aggregate LLM-generated semantic information, supporting zero-shot and
few-shot domain transfer of CLIP models, as validated through evaluations on 12 benchmark datasets. The research further identifies challenges faced by CLIP models regarding compute and data requirements for pretraining, as well as issues with flexibility and generalization due to suboptimal unimodal features in the joint embedding space. To address these challenges, recent advancements in unimodal vision and language encoders are leveraged, with an analysis of these models conducted through the perspective of representational similarity. Motivated by the hypothesis that vision and language encoders model the same physical reality, this thesis studies their semantic similarity, revealing that their representations often share a high
degree of alignment, comparable to that of aligned vision-language encoders. Building on this insight, a lightweight framework that aligns pre-trained, strong unimodal encoders using simple projection transformations is developed. This approach is significantly more compute/data efficient while outperforming CLIP on 0-shot domain
transfer to classification/retrieval tasks. Furthermore, the framework’s flexibility and generalization across diverse tasks like multi-lingual retrieval/classification, 0-shot localization, and long-context retrieval are demonstrated. The findings pave the way for flexible, efficient, and generalizable solutions for open-world understanding, contributing to broader applications of multi-modal systems. Finally, the conclusion
of this thesis summarizes the contributions and future research direction
Optimizing the Post-Mining Process in Association Rule Mining: Data Structures, Substitution Item Mining, and Visualization
Get PDFIn Association Rule Mining (ARM), the generation of large volumes of associ- ation rules from complex datasets often presents challenges in terms of scalability, efficiency, and interpretability. This thesis addresses these challenges by developing novel methodologies and data structures tailored to improve the post-mining phase of ARM. Our approach begins with the creation of a specialized data structure to efficiently store and retrieve association rules, enhancing memory efficiency and pro- cessing speed. This data structure is then leveraged to design a robust methodology for substitute item mining, an emerging area that enables the identification of alter- native items based on observed patterns, with potential applications in areas such as inventory management and consumer behavior analysis. Furthermore, to improve the interpretability of ARM results, we propose advanced visualization techniques that utilize the developed data structure, allowing users to effectively explore and understand complex relationships within large rulesets. The effectiveness of these methodologies was evaluated through surveys and case studies, demonstrating sig- nificant improvements in both cognitive load for visualization and alignment with consumer preferences in substitute item identification. This research contributes to the broader field of ARM by providing tools that enhance scalability, interpretabil- ity, and practical applicability, paving the way for more efficient knowledge discovery and decision-making in data-rich environments
Optimizing Feature Prioritization in Software Development Through Structured Usage Analytics: A Data-Driven Approach to Enhancing Development Efficiency
Get PDFIn the rapidly evolving field of software development, the ability to efficiently prioritize and enhance software features based on user feedback is crucial for maintaining competitiveness and development efficiency. Deciding which features to develop or update is often complex and inefficient. Developers and product managers typically use traditional feedback methods, which are slow, subjective, and hard to prioritize. The main issue is the overwhelming amount and subjective nature of feedback from surveys, bug reports, and customer interactions. This often leads to analysis paralysis, where teams struggle to determine the most critical issues to address. This thesis proposes a structured approach to usage analytics aimed at addressing these challenges by helping developers identify and prioritize features that have the most significant impact on users. This method enables developers to understand and assess the impact of updates and new features on user engagement. By creating a more efficient and responsive feedback loop, usage analytics can help teams prioritize development tasks based on actual user interactions rather than subjective feedback, reduce delays in addressing critical issues, and improve overall development efficiency. Case studies demonstrate the practical benefits of usage analytics in software development. The first study on the IBM Academic Cloud project identified challenges in feature prioritization, such as unclear feature definitions and time-consuming data preparation. The second study, on IBM Watson Workspace, revealed issues in data identification, analytics metrics, and feature-action mapping, stressing the need for systematic data collection and iterative testing. The third study applied usage analytics to Odoo Notes, showing the impact of changes on user behavior and helping prioritize future features. These studies demonstrate usage analytics offers actionable insights for data-driven feature prioritization, enhancing decision-making and improving software to better meet user needs
Understanding play from children’s perspectives in order to provide authentic and fully inclusive opportunities for playful learning in education
No full textPlay is a fundamental right of all children and critical for wellbeing and development. Indeed, play is widely recognised as central to children’s educational experiences and a valuable medium for supporting the inclusion of all learners. This is important for autistic children given reported challenges surrounding their inclusive educational experiences. Therefore, initial investigations of this doctoral research examined the potential of play as a key context in supporting autistic children in education. This involved a systematic review of play-based interventions in supporting the social communication skills of autistic children across educational contexts. However, this study called for a consistent and coherent understanding of play to inform future empirical investigations surrounding the role of play in children’s lives, learning and development.
Therefore, subsequent investigations examined a child-centred understanding of play based on a scoping review of the literature on children’s perspectives of play, co-constructed with 49 children (aged 5 to 11 years). In response to identified gaps surrounding the limited research on children’s perspectives on play, subsequent consultations were conducted with autistic and non-autistic children (aged 5 to 13 years) across mainstream (n=68) and Autism classes (n=19). A range of participatory-based methods were used to elicit children’s views of what play is and why it is important alongside how researchers should consult with children on play.
Across these investigations, children consistently highlighted the significance of within-child qualities in play including fun and enjoyment and autonomy and agency. Further, children emphasised meaningful social connections in play including the importance of adults as valuable play partners.
This research extends our understanding of play from the perspectives of the players themselves and highlights characteristics that must be recognised by researchers and educators in understanding play. This has important implications in upholding children’s right to play and in providing quality and authentic play experiences in education