UMBC Faculty Collection

Permanent URI for this collectionhttp://hdl.handle.net/11603/15

Browse

Now showing 1 - 20 of 9768

Memorization Over Reasoning? Exposing and Mitigating Verbatim Memorization in Large Language Models' Character Understanding Evaluation
(2024-12-30) Jiang, Yuxuan; Ferraro, Francis
Recently, Large Language Models (LLMs) have shown impressive performance in character understanding tasks, such as analyzing the roles, personalities, and relationships of fictional characters. However, the extensive pre-training corpora used by LLMs raise concerns that they may rely on memorizing popular fictional works rather than genuinely understanding and reasoning about them. In this work, we argue that 'gist memory'-capturing essential meaning - should be the primary mechanism for character understanding tasks, as opposed to 'verbatim memory' - exact match of a string. We introduce a simple yet effective method to mitigate mechanized memorization in character understanding evaluations while preserving the essential implicit cues needed for comprehension and reasoning. Our approach reduces memorization-driven performance on popular fictional works from 96% accuracy to 72% and results in up to an 18% drop in accuracy across various character understanding tasks. These findings underscore the issue of data contamination in existing benchmarks, which often measure memorization rather than true character understanding.
MedReg-KG: KnowledgeGraph for Streamlining Medical Device Regulatory Compliance
(2024-12-15) Chattoraj, Subhankar; Joshi, Karuna
Healthcare providers are deploying a large number of AI-driven Medical devices to help monitor and medicate patients. For patients with chronic ailments, like diabetes or gastric diseases, usage of these devices becomes part of their daily lifestyle. These medical devices often capture personally identifiable information (PII) and hence are strictly regulated by the Food and Drug Administration (FDA) to ensure the safety and efficacy of the medical device. Medical device regulations are currently available as large textual documents, called Code of Federal Regulations (CFR) Title 21, that cross-reference other documents and so require substantial human effort and cost to parse and comprehend. We have developed a semantically rich framework MedReg-KG to extract the knowledge from the rules and policies for Medical devices and translate it into a machine-processable format that can be reasoned over. By applying Deontic Logic over the policies, we are able to identify the permissions and prohibitions in the regulation policies. This framework was developed using AI/Knowledge extraction techniques and Semantic Web technologies like OWL/RDF and SPARQL. This paper presents our Ontology/Knowledge graph and the Deontic rules integrated into the design. We include the results of our validation against the dataset of Gastroenterology Urology devices and demonstrate the efficiency gained by using our system.
Mapping adaptive capacity and smallholder agriculture: applying expert knowledge at the landscape scale
(Elsevier, 2016-09-26) Holland, Margaret Buck; Shamer, Sierra Zaid; Imbach, Pablo; Zamora, Juan Carlos; Medellin Moreno, Claudia; Hidalgo, Efraín J. Leguía; Donatti, Camila I.; Martínez-Rodríguez, M. Ruth; Harvey, Celia A.
The impacts of climate change exacerbate the myriad challenges faced by smallholder farmers in the Tropics. In many of these same regions, there is a lack of current, consistent, and spatially-explicit data, which severely limits the ability to locate smallholder communities, map their adaptive capacity, and target adaptation measures to these communities. To explore the adaptive capacity of smallholder farmers in three data-poor countries in Central America, we leveraged expert input through in-depth mapping interviews to locate agricultural landscapes, identify smallholder farming systems within them, and characterize different components of farmer adaptive capacity. We also used this input to generate an index of adaptive capacity that allows for comparison across countries and farming systems. Here, we present an overview of the expert method used, followed by an examination of our results, including the intercountry variation in expert knowledge and the characterization of adaptive capacity for both subsistence and smallholder coffee farmers. While this approach does not replace the need to collect regular and consistent data on farming systems (e.g. agricultural census), our study demonstrates a rapid assessment approach for using expert input to fill key data gaps, enable trans-boundary comparisons, and to facilitate the identification of the most vulnerable smallholder communities for adaptation planning in data-poor environments that are typical of tropical regions. One potential benefit from incorporating this approach is that it facilitates the systematic consideration of field-based and regional experience into assessments of adaptive capacity, contributing to the relevance and utility of adaptation plans.
Making the Future with the Nonhuman:Shenzhen, the Greater Bay, and “Made in China Intelligently”
(OpenEdition, 2023-12-01) Yang, Fan
This essay examines two interconnected human-made nonhuman entities stemming from Shenzhen, China’s first special economic zone, that have become dominant figures in mapping the city’s – and by extension, China’s – future: the robot and the drone. I bring an interdisciplinary, cultural studies approach to the multiple meaning-making practices that engage with these two objects; both participate in enacting the vision for the Guangdong-Hong Kong-Macao Greater Bay Area as an extension of the success of Shenzhen. These practices simultaneously normalise aspirations for a future fuelled by the power of nonhuman technological agents while offering glimpses into the uneven power relations between different humans that underpin such future making. At the same time, they also point to the emergent possibilities of meaning-making that conjoin the human and the nonhuman.
An Efficient Computational Algorithm for Modeling Slow Soliton Interactions in Microresonators
(Optica, 2024) Akter, Sanzida; Shandilya, Pradyoth; Courtright, Logan; D’Aguanno, Giuseppe; Leshem, Amir; Gat, Omri; Menyuk, Curtis
Standard simulations of microresonator waveforms are limited by the photon lifetime. We describe a computational method that enables simulations on a laboratory timescale and apply this approach to study two-soliton interactions.
Lobbying from the Outside: Foreign Agents, Lobbyists, and Foreign Advocacy in the USA
(Springer Nature, 2024-12-01) Straus, Jacob
In 1938, Congress enacted the Foreign Agent Registration Act (FARA) as a tool to identify and potentially combat Nazi propaganda. In the 1960s, FARA was amended to shift its focus from capturing propaganda to documenting the advocacy relationship between foreign principals (e.g., governments, political parties, and corporations) and the US government. Today, foreign agents—those hired by foreign principals to represent their interests—are required to register with the Department of Justice unless they meet one of the eight statutory exemptions, including one that allows for the registration under the Lobbying Disclosure Act (LDA), a 1995 law that requires lobbyists to register with Congress. This paper provides background on FARA, LDA, and how they interact. Using data from both laws, three logit regression models test the importance of American foreign assistances and international alliance connections on a countries decision to hire an advocate in the USA. The analysis demonstrates that countries that receive have more international alliance connections with the USA are more likely to hire a foreign agent, lobbyist, or both. Those that receive more foreign assistance are more likely to hire a lobbyist or a lobbyist and a foreign agent.
Linking Biodiversity and Human Wellbeing in Systematic Conservation Assessments of Working Landscapes
(MDPI, 2023-06-21) Huber, Patrick R.; Baker, Matthew; Hollander, Allan D.; Lange, Matthew; Miller, Daphne; Quinn, James F.; Riggle, Courtney; Tomich, Thomas P.
Systematic land use planning to address environmental impacts does not typically include human health and wellbeing as explicit inputs. We tested the effects of including issues related to human health, ecosystem services, and community wellbeing on the outputs of a standard land use planning process which is primarily focused on environmental variables. We consulted regional stakeholders to identify the health issues that have environmental links in the Sacramento, California region and to identify potential indicators and datasets that can be used to assess and track these issues. Marxan planning software was used to identify efficient land use patterns to maximize both ecological conservation and human health outcomes. Outputs from five planning scenarios were compared and contrasted, resulting in a spatially explicit series of tradeoffs across the scenarios. Total area required to meet imputed goals ranged from 10.4% to 13.4% of the total region, showing somewhat less efficiency in meeting biodiversity goals when health outcomes are included. Additionally, we found 4.8% of residential areas had high greening needs, but this varied significantly across the six counties. The work provides an example of how integrative assessment can help inform management decisions or stakeholder negotiations potentially leading to better management of the production landscapes in food systems.
Leveraging the Medicaid Current Beneficiary Survey to Estimate Undercounting in Claims Data during the COVID-19 Public Health Emergency
(2024-06-29) Gill, Christine; Blake, Elizabeth; Roll, Melissa
Leveraging State Data Sources to Improve Analysis of Medicaid Participant Race and Ethnicity
(2024-06-29) James, Parker; Woodcock, Cynthia; Middleton, Alice; Idala, David; Smirnow, Alexis
Lack of Premeditation Mediates the Relationship Between Adverse Childhood Experiences and Posttraumatic Stress Disorder in Individuals in Residential Treatment for Substance Use Disorder
(Taylor & Francis, 2024-12-01) Thomas, Julia; Carrano, Jennifer; Schacht, Rebecca; Fishman, Marc; Wenzel, Kevin
Objective: Adverse Childhood Events (ACEs) are associated with increased vulnerability for posttraumatic stress disorder (PTSD), but not everyone who experiences ACEs develops PTSD. Impulsivity has gained interest as a potential mediator between ACEs and PTSD, given that both PTSD and ACEs have been closely related to impulsivity. However, less is known about the relationship within the context of substance use disorder (SUD), a population highly vulnerable to co-occurring PTSD. This study examined whether impulsivity mediates the relationship between ACEs and PTSD symptoms in adults seeking residential treatment for substance use disorder (SUD). Methods: N = 134 consenting adults in residential treatment for SUD completed questionnaires measuring impulsivity (UPPS-P), ACEs (ACEs Scale), and PTSD symptoms (PCL-5). Regression models tested our hypotheses that ACEs would predict PTSD symptoms and that impulsivity would mediate this relationship. Results: Two-thirds of participants met the PCL-5 score threshold for a provisional diagnosis of PTSD. ACEs score and impulsivity were significant direct predictors of PTSD symptoms (coeff = 2.23, p < .001; coeff = 1.03, p < .001). Among UPPS-P subconstructs, only lack of premeditation emerged as a partial mediator (z = 2.14, p = 0.032). Conclusions: Individuals with SUD experience adverse and traumatic events at alarming rates and are at increased risk for PTSD. Our mediation finding suggests that impulsivity and especially lack of premeditation may be clinically relevant in the development or maintenance of PTSD symptoms among individuals with SUD.
Land Rights in Sustainability: Rights and Responsibilities
(Elsevier, 2019-10-25) Zahringer, Julie G.; Beale, Colin Michael; Meyer, Carsten; Li, Tania; Ceddia, M. Graziano; Veit, Peter; Knoke, Thomas; Grafton, Quentin; Meinzen-Dick, Ruth; Nkonya, Ephraim; Robinson, Brian; Holland, Margaret B.
Is Regional Electricity Governance Broken?
(I Hate Politics Podcast, 2024-12-27) Dasgupta, Sunil; Charkoudian, Lorig; Hansen, Evan
With news that wholesale prices of electricity are going to rise sharply next year, there is increasing concern about the regional transmission organization called PJM, which covers 13 states from New Jersey to Illinois. Sunil Dasgupta talks with Maryland Delegate Lorig Charkoudian and West Virginia Delegate Evan Hansen about how they see the problem and are working together. Music from Washington DC rock ‘n’ roll band 7 Door Sedan.
Invertebrate herbivores influence seagrass wasting disease dynamics
(Wiley, 2024-12-04) Graham, Olivia J.; Aoki, Lillian R.; Burge, Colleen; Harvell, C. Drew
Although invertebrate herbivores commonly impact terrestrial plant diseasesby facilitating transmission of plant pathogens and increasing host susceptibil-ity to infection via wounding, less is known about the role of herbivores inmarine plant disease dynamics. Importantly, transmission via herbivores maynot be required in the ocean since saline ocean waters support pathogen sur-vival and transmission. Through laboratory experiments with eelgrass (Zosteramarina), we showed that isopods (Pentidotea wosnesenskii) and snails (Lacunaspp.) created grazing scars that increased disease severity and thus indirectlyfacilitated transmission of Labyrinthula zosterae (Lz), a protist that causesseagrass wasting disease. Experiments also quantified different feeding prefer-ences among herbivores: Amphipods (Ampithoe lacertosa) selectively con-sumed diseased eelgrass, while isopods and snails selectively grazedasymptomatic leaves, suggesting different herbivore taxa may have contrastingimpacts on disease dynamics. Our experiments show no sign that herbivoresdirectly vector Lz from diseased to asymptomatic eelgrass. However, we iso-lated live Lz from isopod, amphipod, and snail feces and detected Lz withquantitative polymerase chain reaction in amphipods and snails, suggestingthat herbivores eating diseased eelgrass could pass the live pathogen. Finally,field surveys demonstrated a close association between seagrass wasting dis-ease and invertebrate grazing scars; disease prevalence was 29 ± 4.7% (95% CI)higher on eelgrass leaves with herbivore scars. Collectively, these findingsshow that some herbivores can increase eelgrass disease risk by facilitating thespread of an important pathogen via wounding, but not via direct transmis-sion. Thus, herbivores may play different roles in plant disease dynamics interrestrial versus marine ecosystems depending on the pathogen’s ability tosurvive and transmit without a vector.
Accuracy differences in aboveground woody biomass estimation with terrestrial laser scanning for trees in urban and rural forests and different leaf conditions
(Springer, 2023-06-01) Arseniou, Georgios; MacFarlane, David W.; Calders, Kim; Baker, Matthew
Both rural and urban forests play an important role in terrestrial carbon cycling. Forest carbon stocks are typically estimated from models predicting the aboveground biomass (AGB) of trees. However, such models are often limited by insufficient data on tree mass, which generally requires felling and weighing parts of trees. In this study, thirty-one trees of both deciduous and evergreen species were destructively sampled in rural and urban forest conditions. Prior to felling, terrestrial laser scanning (TLS) data were used to estimate tree biomass based on volume estimates from quantitative structure models, combined with tree basic density estimates from disks sampled from stems and branches after scanning and felling trees, but also in combination with published basic density values. Reference woody AGB, main stem, and branch biomass were computed from destructive sampling. Trees were scanned in leaf-off conditions, except evergreen and some deciduous trees, to assess effects of a leaf-separation algorithm on TLS-based woody biomass estimates. We found strong agreement between TLS-based and reference woody AGB, main stem, and branch biomass values, using both measured and published basic densities to convert TLS-based volume to biomass, but use of published densities reduced accuracy. Correlations between TLS-based and reference branch biomass were stronger for urban trees, while correlations with stem mass were stronger for rural trees. TLS-based biomass estimates from leaf-off and leaf-removed point clouds strongly agreed with reference biomass data, showing the utility of the leaf-removal algorithm for enhancing AGB estimation.
Infrared Band Strengths and Other Properties of Three Interstellar Compounds—Amorphous Isocyanic Acid, Formaldehyde, and Formic Acid
(IOP Science, 2024-12-09) Hudson, Reggie L.; Yarnall, Yukiko; Gerakines, Perry A.
Infrared (IR) spectral features of interstellar and solar system ices have been attributed to solid organic and inorganic compounds for over 50 yr, but in many cases the laboratory IR data needed to fully quantify such work have never been published, forcing researchers to rely on assumptions about gas- or liquid-phase measurements to interpret data for ices. Here, we report the first mid-IR intensity measurements for isocyanic acid (HNCO) ices that are free of such assumptions, providing new results for use by both observational and laboratory astrochemists. We also report similar new IR data for both formaldehyde (H₂CO) and formic acid (HCOOH), which have been discussed in the astrochemical literature for decades, but again without adequate laboratory data to help quantify observational results. Densities and refractive indices of HNCO, H₂CO, and HCOOH as amorphous ices also are reported. Two applications of the new H₂CO work are presented, the first vapor-pressure measurements of solid H₂CO, along with an enthalpy of sublimation, at 100 to 109 K and a set of IR intensities of H₂CO in H₂O + H₂CO ices. Band strengths, absorption coefficients, and optical constants are calculated for all three compounds. Extensive comparisons are made to older results, which are not recommended for future use.
Influence of Bulk Compressibility on Tides in Rocky and Icy Worlds
(2024-12-12) Renaud, Joe P.; Wagner, Nick; Cascioli, Gael; Goossens, Sander J.; Henning, Wade
The tidal response of terrestrial planets or icy moons is a complex interplay between the interior and orbit of the world. A significant factor in this response comes from the thermal and viscoelastic state of the deep interior. Suggesting that if measurements of the tidal and mass-loading (if applicable) response are obtained with sufficient precision, then tides can provide a probe of these properties that are otherwise hidden from other measurement techniques. This has been utilized across the Solar System to constrain questions such as, does Mercury or Venus have a solid core? Does our Moon have significant partial melting in its mantle? Or do icy moons in the outer Solar System contain large, and astrobiologically relevant, liquid oceans? New missions like BepiColombo, VERITAS, Europa Clipper, and more will provide high-precision measurements of tidally-relevant properties such as the Love Number k₂ (and perhaps h₂ and their phase lags). However, linking these observations to the interior requires proper modeling of all the major factors that drive changes in these values.In this study, we look at an often-omitted concept in tidal modeling: Bulk Compressibility, which is the inclusion of bulk modulus in the viscoelastic-gravitational equations used to calculate planetary Love numbers. Including bulk modulus can be mathematically and computationally complex so it is not uncommon to ignore it. This is occasionally justified by the misnomer of “compressibility” being only relevant to large bodies. However, bulk compressibility is not directly dependent on the size or mass of a planet. Instead, it provides another degree of freedom for the world to respond to tidal or loading stresses.In this presentation, we will discuss the background and theory behind bulk compressibility and showcase that its addition can have a measurable impact on tidal and loading Love numbers. For example, work by this team has shown that it can lead to >20% difference in the tidal and loading Love number k₂ and up to 40% difference for h₂. Furthermore, it can significantly alter the phase lag of these numbers which is an important probe of thermal and rheological properties. Finally, we will discuss when bulk compressibility can be ignored and what sort of measurement precision is most affected.
Incorporating Land Tenure Security into Conservation
(Wiley, 2017-06-11) Robinson, Brian E.; Masuda, Yuta J.; Kelly, Allison; Holland, Margaret B.; Bedford, Charles; Childress, Malcolm; Fletschner, Diana; Game, Edward T.; Ginsburg, Chloe; Hilhorst, Thea; Lawry, Steven; Miteva, Daniela A.; Musengezi, Jessica; Naughton-Treves, Lisa; Nolte, Christoph; Sunderlin, William D.; Veit, Peter
Insecure land tenure plagues many developing and tropical regions, often where conservation concerns are highest. Conservation organizations have long focused on protected areas as tenure interventions, but are now thinking more comprehensively about whether and how to incorporate other land tenure strategies into their work, and how to more soundly ground such interventions on evidence of both conservation and human benefits. Through a review of the literature on land tenure security as it relates to conservation practice, predominantly in the tropics, we aim to help conservation practitioners consider and incorporate more appropriate land tenure security interventions into conservation strategies. We present a framework that identifies three common ways in which land tenure security can impact human and conservation outcomes, and suggest practical ways to distill tenure and tenure security issues for a given location. We conclude with steps for considering tenure security issues in the context of conservation projects and identify areas for future research.
Increasing Visual Literacy With Collaborative Foraging, Annotation, Curation, and Critique
(ACM, 2024-12-05) Williams, Rebecca M.; Syed, Afrin Unnisa; Kurumaddali, Krishna Vamsi
Students today are facing information overload, contamination, and bloat from dubious sources: AI-generated content, masqueraded influencer opinions, context-less listicles, and consumer manipulation - frequently heralded by graphs and charts to bolster the argument. Because this information firehose presents as technical visual communications, the overload is both cognitive and perceptual, potentially causing more insidious misperceptions than text alone. In addition to consuming such media, students in computing fields work with data to produce graphs and charts themselves, including assignments, academic research, and personal projects/blog posts/tweets. Depending on visual literacy (VL) and prior data analysis instruction, many students inadvertently code misleading, unethical, or biased visualizations, potentially contributing to the dark corpus already festering online. Prior research on misconceptions in visualization pedagogy suggests students benefit from repeated opportunities to forage, curate and critique examples, discussing and debating with peers and instructors. Inspired by these findings, we incorporated a visual curation + annotation platform into a Data Visualization Computer Science course, enabling students to participate in processes of searching for and curating found examples of misleading visualizations, collaborative annotation + critique of examples, and structured self-evaluation of misleading elements in their own work. We assess our interventions with pre-/post-course Visualization Literacy Assessment Tests, qualitative evaluation of student reflections, taxonomic evaluation of formative student-produced visualizations, and post-course exit surveys. Post-course, students' VL increased significantly, and the number and severity of misleading visualizations they created decreased. Students also reflected that they gained increased confidence in spotting visual disinformation online, and in avoiding its creation in software.
Improving Gamma Imaging in Proton Therapy by Sanitizing Compton Camera Simulated Patient Data using Neural Networks through the BRIDE Pipeline
(UMBC High Performance Computing Facility, 2024) Chen, Michael O.; Hodge, Julian; Jin, Peter L.; Protz, Ella; Wong, Elizabeth; Cham, Mostafa; Gobbert, Matthias; Barajas, Carlos A.
Precision medicine in cancer treatment increasingly relies on advanced radiotherapies, such as proton beam radiotherapy, to enhance efficacy of the treatment. When the proton beam in this treatment interacts with patient matter, the excited nuclei may emit prompt gamma ray interactions that can be captured by a Compton camera. The image reconstruction from this captured data faces the issue of mischaracterizing the sequences of incoming scattering events, leading to excessive background noise. To address this problem, several machine learning models such as Feedfoward Neural Networks (FNN) and Recurrent Neural Networks (RNN) were developed in PyTorch to properly characterize the scattering sequences on simulated datasets, including newly-created patient medium data, which were generated by using a pipeline comprised of the GEANT4 and Monte-Carlo Detector Effects (MCDE) softwares. These models were implemented using the novel ‘Big-data REU Integrated Development and Experimentation’ (BRIDE) platform, a modular pipeline that streamlines preprocessing, feature engineering, and model development and evaluation on parallelized GPU processors. Hyperparameter studies were done on the novel patient data as well as on water phantom datasets used during previous research. Patient data was more difficult than water phantom data to classify for both FNN and RNN models. FNN models had higher accuracy on patient medium data but lower accuracy on water phantom data when compared to RNN models. Previous results on several different datasets were reproduced on BRIDE and multiple new models achieved greater performance than in previous research.
IN ISLANDIA, VERITAS: CHARACTERIZATION OF VENUS SURFACE ANALOGS
(USRA, 2024) Nunes, D. C.; Smrekar, S. E.; Hensley, S; Adeli, S.; Alemanno, G.; Cascioli, Gael; Zebker, H.

Browse

Recent Submissions