University of Illinois at Chicago
University of Illinois at Chicago: UIC INDIGO (INtellectual property in DIGital form available online in an Open environment)Not a member yet
21439 research outputs found
Sort by
Tools and Models for Evaluating Cloud and On-Premises HPC Resources
No full textDeciding between on-premises and cloud resources presents a significant challenge for or- ganizations with diverse computational needs, ranging from small clusters to supercomputers. This thesis offers a comprehensive analysis and develops quantitative models to support this decision-making process. Central to this research is a Total Cost of Ownership (TCO) model that accounts for both capital expenditures (CAPEX) and operational expenditures (OPEX) as- sociated with on-premises HPC facilities. Implemented as a web-based tool, this model provides efficient and detailed cost assessments that complement popular cloud pricing calculators. This integration enables users to perform direct and customizable comparisons between on-premises and cloud-based solutions seamlessly.
However, determining the most suitable infrastructure involves more than identifying the cheaper option. Therefore, this thesis also conducts an in-depth analysis of real-world ap- plications and workloads, with a particular focus on High Performance Computing (HPC). By examining workload-specific requirements and infrastructure considerations, this study provides practical insights into the cost-performance trade-offs between cloud and on-premises deploy- ments. Case studies, including examples from Argonne National Laboratory, demonstrate how workload characteristics and job mixes influence the overall TCO and inform strategic resource planning. The findings of this research aim to guide stakeholders in optimizing cost and performance for future infrastructure investments, supporting the ideal allocation and provisioning of com- putational resources across various organizational needs
Photographic Infrastructures: The Modern School and the Framing of American Architectural Photography
No full text"Photographic Infrastructures: The Modern School and the Framing of American Architectural Photography" traces a global history of architectural photography and education between 1890 and 1950. It maps the movement of architectural, photographic, and pedagogical practices across US empire from its colonized periphery to its metropolitan center and rural fringe. The project draws on three novel photographic archives: surveys of infrastructure around the world made for school children, images of school buildings published in the architectural press, and pictures of learning environments featured in exhibitions meant to instruct the general public on the transformative potential of modern school design. It uncovers the social workings of photography by exposing the ways in which new approaches to teaching and learning in the late-nineteenth and early twentieth-century shaped the photographic representation of space and the built environment. "Photographic Infrastructures" argues that educational ideas were central to the making of architectural photography. The framing, circulation, and display of architectural photographs was tied to pedagogical practices and conceptions of the classroom
A Powerful Pathway to Queering Collaboratively
No full textThis study examines how three middle school literacy educators collaboratively queered their curriculum within a "Community of Practice" (Lave & Wenger, 1991, p. 98), focusing on the challenges, tensions, and transformative learning that emerged from this work. Using a narrative inquiry approach, it captures how teachers engaged in critical reflection, curriculum redesign, and classroom implementation to integrate LGBTQIA+ perspectives into their humanities instruction. Findings highlight five key themes: (1) the shift from token representation to structural curricular change, (2) the role of emotional and professional support in overcoming institutional barriers, (3) the importance of intersectionality in addressing the complexities of identity within literacy education, (4) students’ enthusiastic engagement with queer-inclusive texts and discussions, and (5) teachers' increased confidence and agency in implementing queer pedagogies. This research highlights the power of teacher collaboration in creating more inclusive and affirming learning spaces. It emphasizes the need for ongoing professional development, institutional support, and policies that move beyond performative inclusion toward a sustained commitment to anti-oppressive education. By documenting the lived experiences of teachers engaged in queering literacy curricula, this study contributes to critical conversations on teacher learning, curriculum transformation, and the future of inclusive education
Increasing Access to Information Technology to Promote Impactful Community Integration for Colbert Class Members Who Have Transitioned Out of Institutions
No full textAim of project: To support ongoing community participation by providing access to and education on IT.</p
Relationship between Industry and Occupation with Cancer Using the Illinois BRFSS Survey (2018 to 2022)
No full textOccupational cancers (OCs) are one of the major reasons for the increase in global cancer burden, emphasizing the need for surveillance and prevention. We investigated the association between industry, occupation, and self-reported cancer prevalence among working civilians in Illinois, using Behavioral Risk Factor Surveillance System (BRFSS) data from the years 2018 to 2022. The primary objectives were to analyze the association between occupational status and cancer outcomes and to compare estimates derived from industry/occupation (I/O) re-weighted data with those found using the CDC's standard demographic weighting approach.
The method showed mapping civilians' industry (NAICS 2022) and occupation (SOC 2018) codes to International Agency for Research on Cancer (IARC) classifications to identify exposure to known (Group 1), probable (Group 2A), or possible (Group 2B) carcinogens. Multiple logistic regression models were conducted, considering complex survey design and adjusting for demographic, socioeconomic, and lifestyle factors. A comprehensive case analysis was performed on a final sample of 10,503 individuals.
In the results, the analysis showed no statistically significant increase in cancer prevalence among individuals classified as exposed to IARC Group 1, 2A, or 2B carcinogens. In fully adjusted models, particularly the CDC-weighted model, an inverse association was observed (OR = 0.795, 95% CI: 0.655–0.964, p = 0.020), suggesting exposure misclassification. This inverse association was not statistically significant in the I/O-weighted model. Significant demographic predictors of cancer prevalence included increasing age, female gender, and racial/ethnic disparities, with lower odds among Hispanic individuals being consistent with the "Hispanic paradox". Higher carcinogenic occupational exposure rates were consistently observed among males, Hispanics, individuals with lower educational attainment, current smokers, and those with no physical activity, emphasizing persistent disparities in exposure risks.
This study underscored the importance of incorporating detailed occupational data into public health surveillance. It concluded that weighting methods had failed to reveal significant associations due to information bias, particularly when granular data of job tasks or exposure duration is lacking. Future research can focus on longitudinal designs, comprehensive exposure histories, and connections with cancer registries to address these limitations and enhance our understanding of occupational cancer risks
Advancing NLP Frontiers in Information Extraction, Opinion Mining, and Text Synthesis
No full textAdvances in generative AI and transformer architectures have propelled NLP toward systems capable of extracting, interpreting, and synthesizing information with greater accuracy and contextual awareness. This dissertation presents an integrated research program spanning three pillars: information extraction, opinion mining, and text synthesis, where each stage builds upon the previous to form a cohesive pipeline from distilling knowledge to generating coherent, faithful narratives.
In the first pillar, we address keyphrase generation as an information extraction task for capturing the main topics and salient concepts of research papers. To overcome limited context and data scarcity, we introduce FullTextKP, the first large-scale full-text keyphrase dataset, and propose domain-agnostic augmentation methods for low-resource settings.
These advances form the foundation for the second pillar, which targets opinion mining through the introduction of the Target-Stance Extraction (TSE) task to jointly identify targets and detect corresponding stances, even in noisy, real-world scenarios. Building on this, we develop Stanceformer, a target-aware transformer architecture that biases attention toward target terms, achieving state-of-the-art performance and robust cross-domain generalization.
The third pillar extends these foundations to text synthesis via the Scientific Introduction Generation (SciIG) task and benchmark, which evaluate large language models on coherence, faithfulness, content coverage, and citation correctness in scholarly writing. A key innovation is the use of keyphrase-based coverage measures, originating from our first pillar, to assess whether generated introductions preserve essential information without distortion.
Collectively, these contributions deliver datasets, architectures, and evaluation frameworks that advance context-aware, reliable, and scalable NLP systems, demonstrating the transformative potential of generative AI in scholarly and social domains
The Making of U.S. Monetary Policy: A Linguistic Analysis of FOMC Transcripts
No full textThis dissertation uses 152 verbatim FOMC transcripts (2000–2018) to quantify how leadership and gender shape deliberation and to infer individual policy preferences from what participants say behind closed doors. I build speaker-level measures of participation (speech share and frequency) and opinion content, then apply modern NLP with three reasoning-capable LLMs (GPT-4o, Claude 3, Gemini 2.5) to classify belief-marked sentences (e.g., “I think,” “I believe,” “I’m concerned”) as hawkish or dovish, aggregating to a meeting-speaker Hawkish Index. Deliberation decentralizes across chairs, and a gender gap that widened under Bernanke narrows markedly with Janet Yellen’s appointment; a difference-in-differences design attributes roughly 0.6–0.8 percentage-point gains in women’s speaking share to her leadership. Validating the textual measure against individualized behavior, higher hawkishness robustly predicts tighter discount-rate recommendations (a 10-point rise maps to ≈0.7–2.8 bps higher proposals with meeting fixed effects) and a greater probability of dissent (≈1–2 percentage points for a 10-point rise), with consistent directional signals across models. Together, the results show that who speaks, how much, and—critically—how they speak leave measurable fingerprints on the making of U.S. monetary policy, and that carefully identified subjective language provides a scalable, interpretable proxy for policymakers’ underlying stance
Regulation of Expression and Function of Neuronal Nicotinic Receptors by Accessory Subunits
No full textNicotine binds to nicotinic acetylcholine receptors (nAChRs) and perturbs the biogenesis and trafficking of these channels. The α4β2 and α3β4 subtypes are the most abundantly expressed nAChRs in the central and peripheral nervous system, respectively, and are key mediators of nicotine dependence and withdrawal. These nAChRs are heteropentameric ligand-gated ion channels that form into two stoichiometries in the absence of other subunits: 3α:2β and 2α:3β; each of these contains two canonical ligand binding sites formed by the primary surface of α-subunit and an adjacent β-subunit. Nicotine exposure leads to the upregulation of the 2α:3β stoichiometry only. The only structural difference between the two stoichiometries is the subunit at the accessory position - either α or β. However, it is unknown if this structural nuance yields any biophysical differences to the overall channel function and expression. We aimed to investigate the mechanism by which nicotine modulates differential expression of α4β2 nAChRs in mammalian cells. Our studies revealed that nicotine promotes the surface expression of (α4)2(β2)3 by overcoming the effects of a native chaperone protein that otherwise facilitates (α4)3(β2)2 trafficking. Furthermore, we hypothesized that in the (α3)3(β4)2 stoichiometry, the α3 accessory subunit forms a third ligand binding site as observed in the paralogous α4β2 nAChR. To study this, we engineered a tandem dimer of α3β4 to restrict the stoichiometry and probe separable ligand binding properties using a substituted-cysteine accessibility method followed by covalent modification. This approach identified a third ligand binding site unique to the (α3)3(β4)2 stoichiometry at the α3 accessory interface that is crucial for channel activation. Understanding the stoichiometric differences in nAChRs will inform future precision drug design to target nicotine withdrawal, with the goal of alleviating symptoms while preserving the normal functions
When Feeds Become “Facts”: Memory and Use of Misinformation from Social Media
No full textThis dissertation investigates how features of social media influence memory and use of misinformation. Across four experiments, participants read short posts based on an established general knowledge norm paradigm that varied in factual framing (accurate, neutral, misleading), item familiarity, and either presentation format (Study 1: scrolling vs. page-based) or content relatedness (Study 2: single vs multi-topic). Outcomes were measured either using a general knowledge test or a cued recall task to distinguish acquisition of misinformation from its later retrieval. Misleading framing consistently increased misinformation use, and high familiarity improved gist recall. Yet presentation and relatedness effects reversed original predictions: coherent contexts (page-based and single topic) enhanced memory for both accurate and inaccurate details, while fragmented contexts (scrolling and multi-topic) reduced recall overall. These results reveal a coherence trade-off: conditions that promote integration and fluency strengthen all traces, whereas less coherent conditions weaken encoding but also limit misinformation uptake
Towards Trustworthy Learning in Temporal Learning Environments
No full textThis thesis explores how to build trustworthy machine learning systems that learn and adapt over time. As machine learning moves beyond static benchmarks and into real-world settings, where data distributions shift, environments change, and objectives evolve, it is more important and difficult to ensure consistent and reliable behavior. Yet, most existing methods for safety and robustness are designed for models trained on fixed datasets that do not account for the dynamics of temporal learning, where models must continually update, interact with changing environments, or operate under evolving constraints.
To address this gap, the thesis investigates three representative settings of temporal learning: continual learning, reinforcement learning, and constraint-driven optimization. It begins by showing how continual learning systems that rely on generative replay can be subtly and persistently compromised. A new data poisoning technique is introduced that embeds backdoors into training data, taking advantage of the limited capacity of generative models. The attack causes models to forget previously learned tasks over time, while maintaining strong performance on current tasks. This underscores a critical vulnerability of continual learners: they can be silently compromised, with the effects only emerging after task transitions—when it is too late to recover.
The thesis then turns to reinforcement learning, focusing on a practical variant known as offline-to-online RL. In this setting, agents are first trained on pre-collected data and later fine-tuned online in real environments. A novel attack is presented that perturbs reward signals in the offline data just enough to remain undetected during offline training, but causes significant performance degradation once the agent is deployed online. This exposes a blind spot in current evaluation practices, which often equate strong offline performance with real-world reliability.
Finally, the thesis explores how adaptive data selection can enhance learning in systems governed by universal constraints. It introduces a reinforcement learning-based framework that dynamically selects training inputs in response to the evolving state of the model. This approach is applied to Lyapunov neural networks—aiming to certify the stability of dynamical systems, and physics-informed neural networks—enforcing physical consistency through partial differential equations. By tailoring the input distribution to the model’s learning progress, the method improves both training efficiency and constraint satisfaction compared to traditional static or heuristic sampling strategies.
Taken together, these contributions reveal how temporal learning systems can fail in subtle, time-dependent ways, and how their performance and reliability can be improved through adaptive, temporally-aware methods. Rather than relying solely on static safeguards, the thesis emphasizes the need to understand and anticipate how learning unfolds over time. It offers both a critique of the limitations in current trustworthiness frameworks and a path forward for designing more resilient and responsive models capable of operating effectively in dynamic, real-world environments
Data sources:
Access Repository Dashboard
Do you manage Open Research Online? Become a CORE Member to access insider analytics, issue reports and manage access to outputs from your repository in the CORE Repository Dashboard! 👇