MSU Libraries Digital Repository (Michigan State University)
Not a member yet
106711 research outputs found
Sort by
Learning, Seeing, and Highlighting with Multimodal Models
No full textThesis (Ph.D.)--Michigan State University. Computer Science - Doctor of Philosophy, 2025Multimodal models integrate heterogeneous data sources such as text, images, audio, and video to support tasks that rely on information from multiple modalities. By jointly analyzing these modalities, they achieve more comprehensive and robust performance compared to unimodal approaches. Despite recent progress, several fundamental challenges in multimodal models remain unresolved. First, it is unclear how multimodal models understand and learn across different downstream tasks, and whether their learned representations are fair or can be less biased. Secondly, what does the \u2018world\u2019 imagined by multimodal models look like through generation, and can this synthesized information enhance their multimodal understanding ability? Finally, with the explosive growth of visual content, can multimodal models efficiently highlight useful information from massive data streams?In this thesis, We will investigate how multimodal models learn across different downstream tasks and achieve unbiased representations, see the imagined content before understanding, and highlight important content from extensive vision data.We first explore the learning of existing video language models in localization-related downstream tasks. Existing video temporal grounding (VTG) methods rely on pretrained features from trimmed videos, lacking temporal context. We propose a novel pretraining approach, which conducts an untrimmed pretraining with a Video-Text Similarity-based Grounding Module to enhance robustness against noise. In the second part of the thesis, We explore the fair representation learning in current vision-language models. Large pre-trained vision-language models such as CLIP provide compact and general-purpose representations of text and images. However, due to their pretraining, these models are at risk of both perpetuating and intensifying existing biases, as well as relying on spurious attributes. We propose a general approach by formulating the problem of jointly debiasing CLIP's image and text representations in reproducing kernel Hilbert spaces (RKHSs), which makes zero-shot predictions of CLIP more fair and robust to spurious correlations. In the third and fourth parts of the thesis, We explored the imagination ability of multimodal models. First, We investigates how multimodal models can generate videos. While existing video generation / editing tasks are limited to changes in objects, backgrounds, and styles, the new task aims to predict the necessary modifications to a given video to generate an altered video in response to a question (e.g., ``what if the man was running instead of walking?''). To solve this problem, We propose a comprehensive approach that first understands and reasons about the video content and questions, and then generates the corresponding imaginary video. Beyond imagination, We further investigate whether multimodal models can understanding better after seeing the imagined content. Specifically, existing multimodal visual retrieval tasks typically rely on a combination of visual and textual information. In contrast, We explored a different approach, which first makes the model imagine a pseudo target image based on the visual and textual inputs, and then using this imagined image to guide the multimodal retrieval process. Through seeing the imagined image, the model can perform retrieval more effectively.In the final part, We focus on leveraging multimodal models for long-video understanding tasks. As the scale of video data rapidly increases, the length of videos is no longer confined to just a few minutes but often extends to hours, posing new challenges incomputation and memory. To address this, We propose a semantic learning framework that enables models to identify and highlight contextually rich tokens, which capture the semantic flow and structure of the video over long time spans. This approach facilitates more effective downstream applications such as temporal grounding and video question answering.Description based on online resource. Title from PDF t.p. (Michigan State University Fedora Repository, viewed ).Includes bibliographical references
MANAGEMENT OF AN INVASIVE WASP AND A NATIVE WEEVIL IN COMMERCIAL CHESTNUT ORCHARDS IN MICHIGAN
No full textThesis (M.S.)--Michigan State University. Entomology - Master of Science, 2025Asian chestnut gall wasp (ACGW) (Dryocosmus kuriphilus Yasumatsu), an invasive species from China, was first detected in the US in Georgia in 1975. High densities of galls caused by larval feeding inhibit tree growth, reduce tree vigor, and decrease nut production. A parasitoid of ACGW, Torymus sinensis Kamijo, was imported and released in the US in the 1970\u2019s. This specialized parasitoid has become established in other eastern states, via natural dispersal and additional introductions. In 2015, ACGW was identified in an orchard in Berrien County, Michigan. We have monitored presence and spread of ACGW and the T. sinensis parasitoid in Michigan since 2017. To date, ACGW is established in at least 25 orchards across 14 counties. The T. sinensis parasitoid, first detected in Michigan in 2017, appears to follow ACGW spread, generally lagging 1-3 years behind ACGW establishment. Along with monitoring regional distribution, we are analyzing ACGW spread within individual orchards to assess spatial-temporal dynamics of this invader in Michigan. Chestnut growers in Michigan, the leading producer of commercial chestnuts in North America, face increasing pressure from native and invasive insect pests, reflecting the expansion of this relatively young industry. Lesser chestnut weevil (Curculio sayi Gyllenhal), a native species which originally infested American chestnut (Castanea dentata), is a particularly serious problem. Female weevils lay eggs through burs into developing nuts in late summer. Larval feeding can destroy entire harvests. Our research objectives included identifying effective pre- and post-harvest tactics to minimize reduce weevil damage and yield loss and determining effective trapping and scouting techniques to monitor adult weevils in chestnut orchards.Description based on online resource. Title from PDF t.p. (Michigan State University Fedora Repository, viewed ).Includes bibliographical references
MECHANISTIC STUDIES OF CONTROLLED CATIONIC POLYMERIZATIONS OF VINYL ETHERS
No full textThesis (M.S.)--Michigan State University. Chemistry - Master of Science, 2025Controlled cationic polymerizations are chain-growth processes involving electron-rich vinyl monomers that utilize an equilibrium between dormant and active states to enable precise control over molecular weight and minimize chain transfer and termination events. This mechanistic assumption implies that molecular weight dispersity (\u110) directly correlates with the concentration of the dormant adduct, but preliminary experimental data from the Chiu group challenges this assumption. This observation suggests a more complex reaction mechanism and highlights the need for systematic studies. This work aims to determine reaction orders for all components in two controlled cationic polymerization initiator-activator systems utilizing variable time normalization analysis (VTNA) with isobutyl vinyl ether (IBVE) as the monomer. Two systems were studied: one using an IBVE-HCl adduct as an initiator, with ethyl acetate and diethyl ether as dual bases, and ZnCl2 as the activator, while the other system utilized an IBVE-thiophene-carboxylate adduct as an initiator, diethyl ether as a base, and ZnCl2 as the activator. The reaction orders identified through VTNA disagree with the previously assumed mechanism and suggest there are multiple equilibria between dormant multiple dormant state, which supports the previous observation that there is a more complex mechanism. Both sets of results provide valuable insights into the field of controlled cationic polymerizations to further understand the reaction mechanism.Description based on online resource. Title from PDF t.p. (Michigan State University Fedora Repository, viewed ).Includes bibliographical references
TOWARDS DATA PRIVACY AND ROBUSTNESS : MEASUREMENT, PROTECTION, AND DEFENSE IN AI MODELS
No full textThesis (Ph.D.)--Michigan State University. Computer Science - Doctor of Philosophy, 2025With the rapid growth of AI usage in real-world applications, from large language models that generate text to computer vision systems that process images and graph-based models that power recommendation or social platforms, data safety and privacy have become urgent concerns. These models are trained on massive datasets that may contain sensitive, private, or copyrighted information.As a result, they not only risk memorizing and exposing training data, but also enable the misuse of user-generated content and face adversarial vulnerabilities in deployment. Addressing these challenges is essential for building trustworthy AI systems that can be safely integrated into society. My work focuses on tackling these issues across different modalities. As for privacy measurement, I study memorization in large language models, quantifying how training data can be remembered and later extracted, thereby highlighting privacy risks. Regarding the copyright protection problem, I propose a noise-based perturbation method for images that prevents unauthorized style transfer, offering a practical defense for content creators against infringement. Moreover, adversarial robustness is an important aspect of my work in AI safety. I develop Deeprobust, an open-source platform that systematizes adversarial attacks and defenses on graphs and images, enabling reproducible evaluation and advancing robustness research. Together, these contributions offer new insights, practical defenses, and community tools that advance our understanding of AI privacy risks and help build more secure and reliable machine learning systems.Description based on online resource. Title from PDF t.p. (Michigan State University Fedora Repository, viewed ).Includes bibliographical references
A MULTI-ISOTOPIC APPROACH TO HUMAN-ENVIRONMENT DYNAMICS IN THE CENTRAL ANDES, PERU
No full textThesis (Ph.D.)--Michigan State University. Anthropology - Doctor of Philosophy, 2025The western Central Andes of South America comprise a mosaic of landscapes shaped by gradients in elevation, latitude, and climate. Archaeological evidence shows that ecological diversity has long sustained human and animal communities, and that interconnected environments have been a defining feature of Andean society from the Pleistocene to the present. Yet this same ecosystem diversity complicates the application of stable isotope analysis to archaeological questions. This dissertation examines the theories, methods, and best practices of isotopic research in archaeological contexts through four case studies from southern Peru. The first article evaluates the use of stable oxygen and hydrogen isotopes to track mobility in western Peru, identifying key limitations and emphasizing the need for alternative tracing methods. The second article develops a deep-time faunal baseline for the high Andes, demonstrating how stable sulfur isotopes can more securely distinguish mobility among ecozones. The third article analyzes new stable carbon, nitrogen, and sulfur isotope data from Late Holocene marine birds and maize to explore how human-introduced marine inputs complicate both paleodietary reconstructions and radiocarbon chronologies. A fourth article engages social theory to assess the methodological and ethical significance of environmental baselines for interpreting isotopic data from human remains. Finally, a concluding, public-facing article considers the enduring importance of Andean resources for human communities and the climate crisis confronting mountain landscapes today. Together, these studies advance isotopic approaches in the Central Andes by addressing equifinality, refining baseline datasets, and foregrounding the social and theoretical dimensions of the archaeological sciences.Description based on online resource. Title from PDF t.p. (Michigan State University Fedora Repository, viewed ).Includes bibliographical references
LIGAND-FIELD PHOTOPHYSICS OF COBALT(III) POLYPYRIDYL COMPLEXES
No full textThesis (Ph.D.)--Michigan State University. Chemistry - Doctor of Philosophy, 2025Fundamental research has consistently driven major technological advancements since the early twentieth century. A notable example is the photophysical investigation of transition metal complexes, which has led to breakthroughs in solar energy conversion technologies, including dye-sensitized solar cells and photocatalysis. Over the past two decades, d6 second- and third-row transition metal complexes have served as workhorses for light-driven applications due to their long-lived charge-transfer excited states and synthetic tunability. However, recent focus has shifted toward earth-abundant first-row transition metals, which not only offer a more sustainable alternative route for efficient energy conversion but also present unique opportunities to uncover unprecedented photophysical and photochemical behavior.First-row transition metal complexes often suffer from rapid deactivation of charge-transfer excited states to lower-energy ligand-field states due to their weaker ligand fields. This deactivation has limited their utility in applications that require long-lived charge-separated excited states. Among low-spin d6 first-row metals, Fe(II) has been extensively studied, but direct observation of ultrafast ligand-field crossover remains challenging due to the short lifetimes of their metal-to-ligand charge-transfer (MLCT) states in addition to inability to directly access their ligand-field states. Isoelectronic Co(III) complexes, in contrast, offer the ability to directly photoexcite into low-energy, spin-allowed ligand-field excited states owing to their energetically demanding ligand-to-metal charge-transfer (LMCT) transitions. This dissertation leverages this unique property of Co(III) polypyridyl complexes to elucidate their excited-state dynamics within the ligand-field manifold after photoexcitation into the lowest energy spin-allowed ligand-field band, bypassing charge-transfer excitation. This thesis takes a systematic molecular approach to investigate ligand-field excited state deactivation mechanism and ways to control energy flow in such excited states through employing a wide array of tools including ultrafast optical spectroscopy, synthesis, computational modeling, multi-dimensional NMR, and chemical intuition. The findings presented throughout the dissertation advance fundamental understanding of ligand-field photophysics and highlight their potential in future light-driven technologies.Description based on online resource. Title from PDF t.p. (Michigan State University Fedora Repository, viewed ).Includes bibliographical references
SUPRAMOLECULAR STRUCTURE AND DYNAMICS OF A CLASS OF HYDROGEN BONDING LIQUIDS : MONOHYDROXY ALCOHOLS
No full textThesis (Ph.D.)--Michigan State University. Chemical Engineering - Doctor of Philosophy, 2025Hydrogen bonding is omnipresent such as DNA, proteins, cellulose, ammonia, etc. The highlighting feature of the hydrogen bonds is their directionality and additivity (reversibility). Interatomic interaction between the hydrogen bonding liquids is relatively weak hence they can break and recombine under experimental time scale, making it difficult to determine the supramolecular structures of these liquids. Investigation of inter or intra molecular H-bonding and its variation with the different molecular structures has been under focus to determine the structure-property relationship. One of simple liquids to show tendency of hydrogen bonding is monohydroxy alcohols (MAs) \u2013 alcohols containing one hydroxyl group. A typical dielectric response of monohydroxy alcohol reveals the presence of the two processes namely the Debye process, whose molecular origin remains a topic of active discussion for over a century and the structural relaxation process. Through a combination of rheology and dielectric spectroscopy (also referred to as rheo-dielectric spectroscopy) we have unraveled several interesting features of these monohydroxy alcohols. Rheo-dielectric spectroscopy shows a clear shear-induced shift in the dielectric as well in the rheological spectra (i.e. shift to a high frequency) with different shifts observed for the Debye process and the structural relaxation process, inconsistent with previous mechanisms for the Debye process. These observations lead to a novel understanding of the dynamics of MAs. In particular, (i) An interesting relationship between the structural relaxation time, t_\u3b1, and the Debye time, t_D, with t_D^2/t_\u3b1 following an Arrhenius temperature dependence. (ii) The presence of an intermediate relaxation process with characteristic time, t_m, between t_\u3b1, and t_D of MAs that is both dielectric and rheology active. (iii) t_m agrees excellently with hydrogen bonding exchange time of MAs from NMR measurements. All these observations help us in the development of theoretical understanding on the dynamics of MAs i.e. the living chain model (LCM) which allows us to obtain relationships consistent with the experimental observations, ultimately leads to us develop a molecular mechanism of the Debye process. LCM successfully predicts the average supramolecular chain size, hydrogen bonding lifetime, activation energies for association and dissociation of hydrogen bonding within different MAs irrespective of their molecular structure (primary as well as secondary alcohols). LCM can also successfully explain the dynamics of MAs on dilution of hydrogen bonds with polar as well as non-polar solvents. Finally, these understandings can be converted into understanding the dynamics of associative polymers containing hydrogen bonded stickers. In this work, we will discuss the findings and the details of the model development.Description based on online resource. Title from PDF t.p. (Michigan State University Fedora Repository, viewed ).Includes bibliographical references
\u201cI DON\u2019T KNOW, IT\u2019S JUST WHAT I AM\u201d : NEGOTIATING IDENTITY, COMMUNITY, AND (UN)BELONGING IN ASIAN AMERICAN ADOPTEE SPACES
No full textThesis (Ph.D.)--Michigan State University. Anthropology - Doctor of Philosophy, 2025Since the 1953 ceasefire of the Korean War, over a quarter-million people have been adopted out of Asia to the United States, most of whom were sent to white households. While touted as an act of humanitarian goodwill and a symbol of a post-racial future, adoption that crosses national, cultural, and racial lines has historically been used as a tool to support U.S. empire building projects and to elide the ongoing realities of racism and xenophobia in the U.S. and abroad. The material inequalities that facilitate a child's perceived adoptability, as well as the potential implications of their geographic displacement and racial isolation are generally overlooked in favor of presenting adoption as apolitical and mutually beneficial. Children adopted from Asia are frequently raised in predominately white communities, often with little exposure to other nonwhite people and without sufficient support to engage with their racial and cultural identities. As a result, transnational Asian American adoptees who were raised in racial isolation often feel a sense of un-belonging in both white and Asian social spaces and thus, may seek out adoptee-specific communities. Adoptee communities have proven to be instrumental in providing pathways for adoptees to contend with the nuances of their identities, unlearn dominant adoption narratives, and find representation. However, over time such spaces can produce unique sets of social expectations and negotiations that end up furthering adoptees\u2019 feelings of un-belonging, leaving many disappointed. This dissertation presents my ethnographic study of adoptee (un)belonging vis \ue1 vis identity and community formation using four primary themes. The first is Journeys which addresses the avenues my participants took to seek out a unified sense of self and belonging. Second is Pressure which identifies the internal and external pressures felt by adoptees to adhere to certain social and cultural scripts in and outside of adoptee communities. Nuance and Reclamation are the final two themes as they explore the possibilities for belonging that emerge when adoptees reject their sociopolitical isolation and instead position themselves within the political and historical contexts of the adoption industry. To accomplish this, I drew upon three virtual ethnographic methods: 1) semi-structured interviews, 2) participant observation of public-facing social media platforms, and 3) a community-engaged photovoice project. By incorporating public social media and my participants\u2019 photography into my interviews, I combine in-depth personal anecdotes, public-facing counternarratives, and visual storytelling to grapple with, and make meaning of the ambiguities of (un)belonging as an adopted person. By drawing from Asian American adoptees\u2019 experiences with navigating identity, belonging, and community within and beyond adoptee spaces, I argue that adoptees use their social positioning to engage in strategic acts of refusals that discursively reject inaccurate or harmful notions of what it means to be adopted. In doing so, they become active agents in establishing their own sense of belonging through constructing flexible, and sociopolitcally grounded identities for themselves. This project thus offers insight into the into the ways searches for identity and community can produce feelings of belonging that are maintained through adoptees\u2019 sense of social responsibility and collective desires for social change.Description based on online resource. Title from PDF t.p. (Michigan State University Fedora Repository, viewed ).Includes bibliographical references
ASPECTS OF COMBINATORIAL SPECTRAL THEORY AND COMMUTATIVE ALGEBRA
No full textThesis (Ph.D.)--Michigan State University. Mathematics - Doctor of Philosophy, 2025This dissertation advances algebraic and topological methods for data science through four lines of work.The first part introduces the path Dirac and hypergraph Dirac operators together with their persistent counterparts, and investigates their ability to capture harmonic and non-harmonic spec- tra while revealing informative subcomplex structure. Their sensitivity to filtration is analyzed, demonstrating how these operators adapt to topological changes, and their behavior is illustrated across diverse examples. A central application is to molecular science: strict preorders derived from molecular structure generate graphs and digraphs with rich path architecture, and the resulting path complexes encode information depth that varies with the underlying preorder classes. The second part develops Mayer Dirac operators on ?-chain complexes. These operators link an alternating sequence of Mayer Laplacians and generalize the classical identity ?2 = ?. An explicit Laplacian for ?-chain complexes induced by vertex sequences on finite sets is derived, and weighted Mayer Laplacian and Dirac operators are introduced to capture physical attributes more effectively. A generalized factorization of Laplacians as an operator product with its adjoint is also established. Persistent Mayer Dirac operators and extensions are applied to biological and chemical data, where they demonstrate practical utility. The third part establishes a persistent Stanley\u2013Reisner theory that connects commutative algebra with combinatorial algebraic topology, machine learning, and data science. The framework defines persistent h-vectors, persistent ? -vectors, persistent graded Betti numbers, persistent facet ideals, and facet persistence modules. Stability analysis confirms that these algebraic invariants are robust under geometric perturbations, and their predictive value is demonstrated on molecular datasets. The final part proposes Commutative Algebra k-mer Learning (CAKL), a nonlinear algebraic framework for comparative genomics that builds upon persistent Stanley\u2013Reisner theory. CAKL in- tegrates commutative algebra, algebraic topology, combinatorics, and machine learning to address genetic variant identification, phylogenetic tree inference, and viral genome classification. Across eleven datasets, CAKL outperforms five state-of-the-art sequence analysis methods\u2014particularly in viral classification\u2014and maintains stable predictive accuracy as dataset size increases, highlight- ing scalability and robustness. Collectively, these contributions provide new operators, invariants, and learning paradigms that unify algebraic, topological, and combinatorial perspectives on discrete structures and real-world data, yielding great performance in molecular science and genomics.Description based on online resource. Title from PDF t.p. (Michigan State University Fedora Repository, viewed ).Includes bibliographical references
A Novel Framework for Machine Learning : Parametric Matrix Models and Their Application in Nuclear Physics, Scientific Computing, and General AI
No full textThesis (Ph.D.)--Michigan State University. Physics - Doctor of Philosophy, 2025Complex nuclear systems present formidable challenges, including the exponential growth of Hilbert-space dimensions, the slow convergence of perturbative expansions, and the difficulty of extracting excited-state information from limited-order data. Addressing these intricate many-body interactions and the need for highly accurate and interpretable computational frameworks necessitates innovative algorithmic solutions that can efficiently learn governing equations and provide robust predictions.This thesis introduces and extensively develops Parametric Matrix Models (PMMs), a novel class of machine learning algorithms specifically designed to address these challenges. PMMs fundamentally differ from conventional neural network and deep learning inspired approaches by emulating physical systems through matrix equations. This design allows PMMs to learn the underlying governing equations directly from empirical data, offering an interpretable and efficient computational framework capable of input feature extrapolation.Building upon the foundational theory of PMMs, this research significantly extends their applicability. We generalize PMMs to effectively model complex physical phenomena, including state-vector evolution and nonlinear dynamics. The primary focus of this work lies in the application of PMMs to imaginary time evolution data within Quantum Monte Carlo lattice Effective Field Theory (EFT). Here, PMMs are demonstrated to accurately resum perturbation theory from first-order approximations, enabling robust predictions for experimental observables such as ground state energies, charge radii, and first excited states in nuclear systems. Furthermore, this thesis explores the versatility of PMMs by presenting a tailored Navier-Stokes PMM, designed to learn and analyze data from Computational Fluid Dynamics (CFD) simulations, their application to general machine learning problems, including regression and image classification.By leveraging the inherent mathematical structures of physics, this approach offers a powerful and innovative algorithmic solution. This work underscores the transformative potential of physics-inspired machine learning for advancing scientific discovery and addressing critical challenges across a wide range of scientific and general machine learning problems.Description based on online resource. Title from PDF t.p. (Michigan State University Fedora Repository, viewed ).Includes bibliographical references