22689 research outputs found
Sort by
Minimizing Language Interference for Multilingual Models
Get PDFAdvancements in machine learning have revolutionized natural language processing, enabling the development of models that can understand and generate multiple languages. Multilingual models leverage a unified architecture to support multiple languages, facilitating cross-lingual knowledge transfer and significantly benefiting low-resource languages. However, these models encounter substantial challenges of language interference, where the performance on some languages deteriorates as more languages are incorporated, resulting in sub-optimal outcomes. This thesis investigates methods to minimize negative language interference, with a focus on multilingual machine translation tasks. The research is organized around three main aspects: modeling, optimization, and adaptation.
In modeling, we explore parameter-efficient methods to enhance multilingual performance by introducing sparsely activated architectures and language-specific modules. Beyond architectural changes, the optimization aspect involves proposing techniques to de-conflict gradients across languages and introducing novel training objectives to optimize language-specific parameter utilization. In adaptation, we adapt to the era of large language models (LLMs) by introducing a new foundational training paradigm for LLM-based translation models and developing robust, massively multilingual LLM-based translation models. The insights and methodologies presented in this thesis primarily focus on improving overall performance across all languages with minimized negative interference and adapting to the evolving architectural frameworks of multilingual LLMs
On the Security and Adaptation of Neural Networks: A Study of Adversarial Robustness, Backdoor Attacks, and Transfer Learning
Get PDFNeural networks have become the heart of modern AI systems, playing a crucial role in both commercial and critical applications. However, these models face two significant challenges: security vulnerabilities and adaptation to new test data and tasks. These models are vulnerable to various adversarial attacks, including evasion, data poisoning, and backdoor attacks, which can compromise their performance and reliability. On the other hand, when these models are deployed as prebuilt solutions, pretrained on tasks different from their intended downstream applications, leading to inferior performance due to domain gaps. Motivated by these challenges, this dissertation investigates the areas of adversarial robustness, backdoor attacks, and transfer learning for neural networks.
We first discuss how to reconstruct adversarial perturbations and classify these reconstructed perturbations based on the algorithm that generated them. This pipeline, REDRL, can detect the attack algorithm used to generate a sample from only the sample itself. We then present a new hidden trigger backdoor attack, Sleeper Agent, which leverages gradient matching, data selection, and target model re-training to craft highly effective poisons. Sleeper Agent is the first hidden trigger backdoor attack to be effective against neural networks trained from scratch. We demonstrate its effectiveness on ImageNet and in black-box settings. Next, we use guided diffusion to synthesize base samples from scratch, utilized for creating poisons and backdoors that are significantly more potent than previous state-of-the-art attacks. Our Guided Diffusion Poisoning (GDP) base samples can be combined with any downstream poisoning or backdoor attack to boost its effectiveness.
In addressing the adaptation challenge, we propose a novel approach within the transfer learning paradigm, where highly informative posteriors are learned from the source task, either through supervised or self-supervised methods, and used as priors for the downstream task. Finally, we introduce Battle of the Backbones (BoB), a comprehensive benchmark that evaluates various popular pretrained checkpoints and randomly initialized baselines across a wide range of downstream tasks, including image classification, object detection, segmentation, out-of-distribution generalization, and image retrieval. BoB sheds light on promising directions for the research community to advance computer vision by illuminating strengths and weaknesses of existing approaches
Molecular simulation studies of extrinsic and intrinsic mechanism of permeation block in eukaryotic potassium channels
Get PDFIon channels are an extremely important class of proteins. They are crucial to every aspect of life, including sensory perception, movement, thought and memory, and heartbeat. Potassium (K+) channels are the most abundant type of ion channel. Mutations in these proteins are linked to a wide array of neurological, cardiovascular, and metabolic diseases in humans. Research into the molecular mechanisms of these proteins is necessary to understand the origin of these diseases and to identify potential therapeutic strategies. This thesis is specifically focused on investigating intrinsic and extrinsic mechanisms of blocking ion permeation in three distinct K+ channels using Molecular dynamics simulations: human TMEM175, which is key in lysosomal homeostasis; Shaker from Drosophila melanogaster, a prototypical member of the family of eukaryotic voltage-gated K+ channels and important in neuronal function; and Kv2.1 from Rattus norvegicus, another member of the family of voltage-gated K+ channels and also crucial for neuronal function. Investigating the mechanism of inhibition for a known inhibitor, 4-aminopyridine (4-AP) of the non-canonical TMEM175, we learn that 4-AP is recognized by TMEM175 in both a dynamic and specific manner and inhibits permeation in an electrostatic as well as steric manner. We also investigate the recently solved structure of C-type inactivated Shaker and learn that dilation of the filter through repositioning of crucial neighboring residues causes inhibition of rapid ion permeation. Finally, we compare RY785, a potent and specific inhibitor of Kv2.1, to a well characterized inhibitor tetraethylammonium (TEA), and discover that while TEA occludes the pore, RY785 still allows for permeation indicating RY785 facilitates channel closing suggesting an allosteric mechanism of inhibition. The insights gained into mechanisms of intrinsic and extrinsic inhibition can not only aid in the development of more targeted therapeutic approaches and novel drugs for these specific K+ channels but can contribute to a broader understanding of the mechanisms of inhibition and inactivation in other K+ channels, as well as, ion channels in general
UNDERSTANDING THE SKILLS AND COMPETENCIES VALUED BY HIRING MANAGERS OF ENTRY-LEVEL AND EARLY-CAREER RESEARCH ADMINISTRATION POSITIONS
Get PDFOver the past few decades, there has increasingly been a high demand for the services and expertise of research administrators at research institutions across the United States. However, the profession has faced challenges with meeting this growing demand, particularly in entry-level and early-career positions. This is due in part to an aging workforce and difficulties with recruitment and retention. Despite this problem, little research has been conducted on the skills and competencies essential for success in these roles or those valued by hiring managers.
The purpose of this study was to identify the skills and competencies most valued by hiring managers when screening candidates for entry-level or early-career research administration positions, while also examining variations across four main sub-areas: Pre-Award, Post-Award, Research Development & Policy, and Research Integrity & Compliance. To achieve this, a survey was distributed to U.S. research administrators through the Society of Research Administrators International (SRAI) “Community Sections” platform. Survey respondents rated 13 skill and competency groups based on their importance during the hiring process for early-career positions within their respective sub-area offices.
The study found that “Critical thinking / problem solving” and “Interpersonal skills / works well with others” were the highest-rated skills, along with several other soft skills being highly rated. Additionally, "Written communication" was rated significantly higher for Pre-Award research administrators compared to Post-Award roles, and those in Research Integrity & Compliance placed significantly more importance on “Ethical / social responsibility” than individuals in other sub-areas. These findings provide a foundation for future research and can help hiring managers of entry-level or early-career research administrators refine their screening processes
Combating Drug Resistance in Plasmodium falciparum: The Pharmacodynamic Impact of Artemisinin Ring-Stage Resistance and Transmission of Atovaquone Resistance
Get PDFABSTRACT
Statement of problem: Antimalarial chemotherapy and prophylaxis are key strategies in combating malaria, which affects nearly half the world’s population. However, increasing spread of drug resistance threatens treatment efficacy, leading to higher morbidity and mortality. We investigate 4-aminoquinoline drugs and their hemozoin-binding properties against artemisinin-ring-resistant P. falciparum. We further define the hemozoin-binding mechanism by which the heme-artemisinin adduct metabolite inhibits resistant parasites. Finally, we characterize transmission of atovaquone-resistant P. falciparum through Anopheles mosquitos. Our goal is to understand the impact of drug resistance on antimalarial mechanisms of action and drug resistance transmissibility, with the goal of preventing further treatment failures.
Methods: We optimized heme crystal nucleation and extension assays to assess the inhibitory effects of chloroquine and pyronaridine. We pulse-dosed chloroquine-resistant PfCamWT and isogenic artemisinin-ring-resistant PfC580Y parasites to evaluate stage-dependent killing. We used similar methods to study the heme-artemisinin adduct metabolite’s effect on resistant P. falciparum. To create atovaquone-resistant Pfcytb mutants, we cultured PfNF54 under continuous atovaquone pressure. We infected An. gambiae mosquitos with PfNF54-Y268S parasites to assess transmission by measuring parasite load at various stages of the mosquito lifecycle.
Results: Chloroquine and pyronaridine, despite being similar in structure and targeting hemozoin, differ in stage-dependent parasite killing phenotypes. Pyronaridine extends its cytocidal activity to artemisinin-resistant-ring stages by inhibiting hemozoin nucleation, while chloroquine targets trophozoites through heme crystal growth inhibition. The heme-artemisinin adduct also inhibits hemozoin nucleation and extension, causing irreversible damage in resistant parasites. Atovaquone-resistant Y268S parasites fail to transmit in An. gambiae mosquitos, even in mixed infections with wild-type alleles as was observed in An. stephensi.
Conclusions: Heme-binding drugs which inhibit hemozoin nucleation, can induce irreversible damage in ring- resistant Pfkelch13 mutant parasites which exhibit enhanced cell stress responses to counteract drug pressure Furthermore, once activated, artemisinin does not remain inactive; it effectively blocks hemozoin nucleation and extension, thereby inhibiting all stages of the erythrocytic lifecycle in artemisinin-resistant parasites. In contrast to the rapid spread of chloroquine and artemisinin resistance, the transmission of atovaquone-resistant Pfcytb mutations through mosquitos appears to be unlikely. This suggests that atovaquone can continue to be used effectively as a malaria prophylactic
LEARNING PDE SOLUTION OPERATORS VIA DIFFEOMORPHIC MAPPINGS: APPLICATIONS IN FLUID DYNAMICS
Get PDFApproximating the solution operator for partial differential equations (PDEs) is critical across numerous scientific and engineering disciplines. Neural operators have emerged as a powerful tool for predicting these solution operators when trained on high-quality ground truth data, such as results from numerical simulations. However, their ability to generalize to unseen spatial domains is often hindered by the need for extensive datasets encompassing diverse geometric configurations, which may be infeasible to generate in many scenarios. To address this challenge, we propose training a latent neural operator using solution fields that are diffeomorphically mapped from varying spatial domains to a unified reference configuration. The efficacy of this approach hinges on preserving the differential operator’s properties during the mapping process, which enhances the regularity of the solution fields and reduces the data requirements for training accurate models. Through two numerical experiments, we validate our framework: (1) leveraging the conformal invariance of the Laplacian in 2D doubly-connected domains, we demonstrate that conformal mappings significantly improve learning efficiency compared to other standardization methods; and (2) extending to 3D nonlinear flows around spheroids of varying aspect ratios, we accurately predict surface shear stress distributions using reduced training data. These results underscore the potential of this framework for applications in computational fluid dynamics, including scenarios such as biomedical flows, where data collection is constrained by practical limitations
FROM TRANSMISSION TO TRANSFORMATION: IMPROVING ACADEMIC WRITING AND CRITICAL THINKING PROFICIENCY WITH COLLABORATIVE PROBLEM-BASED EXPERIENTIAL LEARNING, TRANSFORMATIVE EDUCATIONAL TECHNOLOGY, AND CRITICAL PEDAGOGY
Get PDFUndergraduate students internationally struggle with engagement, self-efficacy, and proficiency in academic writing and critical thinking on societal problems impacting people across multicultural contexts. This applied dissertation explores contributing factors to this educational problem using the framework of nested and networked ecological systems theory. Next, the study examines how the problem of practice manifests in a state university in Hungary. The issue is explored through an empirical needs assessment using a mixed methods approach and convergent parallel design with quantitative and qualitative data sets involving a school-wide survey and focus group interviews with teachers and international students studying in English. Based on the literature review, needs assessment results, and extant data exploring similar constructs, the author develops a holistic theory of learning called ETCDC by synthesizing five learning theories into a hybrid holistic model. The framework integrates principles of experiential, transformative, and collaborative learning as well as digital hybrid design and critical theory –– each having positive associations with increasing student engagement, self-efficacy, and proficiency in analytical thinking and writing. For an applied project, the ETCDC model is utilized in conjunction with Universal Design for Learning (UDL) and Design Thinking (DT) to develop (1) a curriculum enrichment framework for innovative critical pedagogy in mobile classrooms, (2) a pilot syllabus, and (3) a business plan for an educational consultancy called Intelligence Motion that will implement the curriculum framework across public and private HEIs in the United States and Europe during phase one of a proof-of-concept pilot study
MURINE GUT MICROBIOTA DYSBIOSIS VIA ENTERIC INFECTION MODULATES THE FOREIGN BODY RESPONSE TO A DISTAL BIOMATERIAL IMPLANT
No full textImplantable biomaterials and medical devices have transformed modern healthcare, but the long-term success of these implants is often limited by complications induced by the foreign body response (FBR). The FBR is an immune-mediated reaction triggered by any foreign material that is implanted in the body. Researchers have explored various material-based strategies to mitigate the FBR and enhance the long-term biocompatibility of medical implants. However, these approaches do not fully address the variability in FBR severity due to patient-to-patient heterogeneity. Specifically, the influence of host factors such as age, sex, and gut microbiota on the FBR remains largely unexplored.
In this doctoral dissertation, we investigate how gut microbiota dysbiosis via enteric infection can impact the immunological and fibrotic response to a distal biomaterial implant. We demonstrated that enteric infection induced systemic changes in immunity and distal tissue function. Subsequently, we established several experimental models coupling an enteric infection with a distal biomaterial implant, testing different variables such as time of infection, type of biomaterial, age of mice, method of implantation, and type of infection. Across all variables, we demonstrated that enteric infection impacts the immune response to a biomaterial implant, including implant-associated inflammation and immune cell infiltration. However, this largely did not translate to changes in fibrosis around the implant. Collectively, this work establishes that there is immune-mediated communication between the gut microbiota and a distal implant site, suggesting that the gut microbiota may be a potential therapeutic target to mitigate the negative immunological consequences of the FBR
DESIGN AND OPTIMIZATION OF ENHANCED MICROELECTRODES TOWARDS ENGINEERING PRIMARY CELLS
No full textThe advent of recombinant gene expression into mammalian cells has revolutionized biomedical research, advancing our understanding of gene functions and advanced disease mechanisms to foster the development of targeted therapeutic strategies. Primary cell gene transfection is critical to advance our understanding of gene functions and advanced disease mechanisms to foster the development of targeted therapeutic strategies. Streamlined workflows for analyzing primary cells are critically needed for the advancement of personalized medicine given the complexity in processing biofluid samples, which involve intricate pre-purification steps and manual transfers between purification and analysis stages. Thus, the focus of this dissertation is to present a microfluidic chip designed to integrate cell purification and biomolecule delivery to realize the efficiency and practicality advantages of a seamless workflow for primary cell assays. Our approach utilizes a microscale electrode layer in conjunction with vortex cell purification technology for sequential electroporation-mediated transfection of cells trapped in vortex. An automative solution exchange system enables multiplexing a variety of biomolecule cargo delivery for a versatile range of downstream cell assays.
We first demonstrate the feasibility of the workflow for clinical applications through pilot-scale combinatorial drug testing onto drug-resistant cancer cells isolated from blood samples. We then expanded the microelectrode circuitry to match the electroporation capacity for an adapted version of the ultra-high throughput vortex cell purification technology, and validated patient-derived cell purification from metastatic breast cancer liquid biopsy samples and transfection through membrane-impermeable molecule delivery. Moreover, we extended our chip applicability in complex gene expression assays by delivering genetic materials such as deoxyribonucleic acid (DNA) and messenger ribonucleic acid (mRNA) into primary cells. By enhancing diagnostic precision and assisting in therapeutic development, our integrated microfluidic tool holds significant promise for advancing personalized medicine
Translational Regulation During Oxidative Stress in Haloarchaea
Get PDFOxidative stress induces a wide range of cellular damage, often causing disease and cell death. While many organisms are susceptible to the effects of oxidative stress, haloarchaea have adapted to be highly resistant. Several aspects of the haloarchaeal oxidative stress response have been characterized, however little is known about the impacts of oxidative stress at the translation level. Using the model archaeon Haloferax volcanii, I performed RNA-seq and ribosome profiling (Ribo-seq) to characterize the global translation land scape during oxidative stress. I identified 281 genes with differential translation efficiency (TE). Downregulated genes were enriched in ribosomal and translation proteins, in addition to peroxidases and genes involved in the TCA cycle. I also observed upregulated TE for several transporters and membrane-bound proteases, highlighting the importance of membrane dynamics during oxidative stress. I additionally identified 42 small noncoding RNAs (sRNAs) with ribosome occupancy. Size distributions of ribosome footprints revealed distinct patterns for coding and noncoding genes, with 12 sRNAs matching the pattern of coding genes, and mass spectrometry confirming the presence of seven small proteins originating from these sRNAs. However, the majority of sRNAs with ribosome occupancy had no evidence of coding potential. Of these ribosome-associated sRNAs, 12 had differential ribosome occupancy or TE during oxidative stress, suggesting that they may play a regulatory role during the oxidative stress response. I also identified 221 novel genes, of which 189 were determined to have coding potential. High confidence structural predictions were generated for 40 putative proteins encoded by novel genes. These novel proteins were then characterized using both sequence and structural homology, revealing a putative novel transcriptional regulator. I also found that over half of the novel genes identified were differentially regulated at the translation level during oxidative stress. In combination with the evidence of translational regulation during oxidative stress, this demonstrates the complexity of gene regulation in response to stress