UBIR Repository (Univ. at Buffalo)
Not a member yet
5736 research outputs found
Sort by
Supervised-Unsupervised Cancer Subtyping Based on Multi-Task Learning
No full textPh.D.Cancer subtyping has the potential to significantly improve disease prognosis and develop individualized patient management. Traditionally, cancer subtyping can be achieved by supervised or unsupervised strategies. Due to its advantage to discover novel subtypes, unsupervised subtyping has shown great success in many cancer types. However, unsupervised subtyping is usually effected by irrelevant factors and ther is no guarantee that the identified subtypes are cancer-relevant. In order to solve this issue, in this dissertation, we proposed a novel supervisedunsupervised subtyping strategy by incorporating cancer-relevant prior knowledge into the identification of subtypes. To this end, we developed a joint supervised-unsupervised dimension reduction and clustering framework by using the technique of multi-task learning. Specifically, we design a primary task of clustering and an auxiliary task of classification. By training the two tasks together, cancer-relevant subspaces and clusters are simultaneously learned. Within this strategy, we developed two approaches for feature selection and non-linear dimension reduction, respectively. The first method aims to identify a list of cancer-relevant genes and subtypes to facilitate following biological analysis like differential expression analysis or gene set enrichment to discover the mechanism of the subtypes. The second method aims to find compact and well-separated subtypes with great potential in clinical applications. This dissertation is the first attempt of supervised-unsupervised subtyping. With the accumulation of molecular data, we believe that our framework can facilitate the discovery of novel cancer subtypes on higher resolution levels with more clinical and research implications.**To request an accessible version of the file(s) associated with this item, contact [email protected]. Please include the item's persistent URL [http://hdl.handle.net/. . .] in your request.*
Spectatorship and Modernity in American Literature
No full textPh.D.This study aims to trace the multifaceted literary representations of spectatorship in American modernist texts, demonstrating the various ways in which the Emersonian model of the spectator is criticized and reconfigured as the American literary tradition moves from the age of realism to that of modernism. Emerson's ocularcentric model of disembodied observer, romantically envisioned as a "transparent eyeball," provided a foundation for the future development of literary exploration of the modern self as a "spectator," which appears frequently in the texts of American literature from the late 19th through to the 20th century. I argue that the history of modern American literature can be seen as the history of challenging romantic transcendentalism and the assumed authority of the Emersonian model of the spectator. This study focuses on four representative writers: Henry David Thoreau, Theodore Dreiser, Willa Cather, and F. Scott Fitzgerald. Their texts attempt in their own ways to reconfigure spectatorship and explore new modes of representation with a particular emphasis on visual experience. Chapter 1 discusses Thoreau and his deep fascination with a spectatorial mode of existence in the context of the urbanization of 19th century American society. Chapter 2 discusses Dreiser's Sister Carrie, along with some of his short stories, to demonstrate how the novel complicates spectatorship and depicts the protagonist Carrie Meeber as a modern spectator. Chapter 3 deals with Cather and the invention of modern memory, which foregrounds tactility and collective remembrance. Chapter 4 discusses Fitzgerald's novels, particularly The Great Gatsby, to demonstrate how the author, simultaneously fascinated by and obsessed with contemporary visual media, employs visual elements in his linguistic narrative to complicate spectatorship.**To request an accessible version of the file(s) associated with this item, contact [email protected]. Please include the item's persistent URL [http://hdl.handle.net/. . .] in your request.*
Veracity and Vulnerabilities Analysis of Multi-Sourced Data
No full textPh.D.In the last decade, the research community has witnessed the success of machine learning techniques in a large variety of long-standing applications. The success of these techniques is largely driven by the ubiquitous massive data, which are typically collected from multiple data sources. However, the multi-source data usually contains erroneous information and some errors may be caused by intentional manipulation by adversarial attackers. These issues may cause the failure of machine learning models. Thus, it is crucial to analyze the veracity of the multi-sourced data and study the potential vulnerabilities in the multi-sourced data. In this dissertation, we propose: (1) a series of multi-sourced data reliability analysis methods to discover trustworthy information from correlated data and textual data; and (2) multiple data poisoning attack approaches to help understand the impacts of vulnerabilities in the multi-sourced data on real-world machine learning tasks. In multi-sourced data reliability analysis, it is critical to identify reliable sources that provide highly-trustworthy information and leverage the data from these reliable sources to better discover trustworthy information. Existing multi-sourced data reliability analysis methods usually make an independent assumption of the data sources, and are generally designed for structured data. In Part I of this dissertation, we develop multiple probabilistic models to resolve these limitations. Particularly, to handle source correlations, the proposed model takes the estimated source correlations as prior, and models the trustworthiness of each piece of information by fusing the trustworthiness of the information provider and its influencers. To identify the reliability of unstructured data like text, we propose another probabilistic model that jointly infers the key factors in the data and estimates the reliability of different data sources. For both models, we conduct extensive experiments on multiple real-world datasets to demonstrate their usefulness and advantages. Apart from multi-sourced data reliability analysis, we also investigate the impacts of vulnerabilities in the multi-sourced data on real-world machine learning tasks. In Part II of this dissertation, we develop data poisoning attack approaches that inject adversarial samples to multi-sourced data to manipulate the knowledge graph embedding methods, recommendation models, and outcome interpretation methods. Such adversarial analysis can help understand the impact of vulnerabilities on these machine learning models. Specifically, the proposed attack strategies against knowledge graph embedding methods generate data samples that can manipulate the embedding of knowledge graph entities and further influence the plausibility of arbitrary target facts in the knowledge graph. We also propose a general reinforcement learning-based attack framework, which can manipulate the recommendation results of various representative next-item recommendation models. Moreover, we explore the vulnerability of machine learning outcome interpretation models and investigate whether attacks can manipulate the interpretations of target samples by injecting well-crafted samples to the training set. All these attack strategies and frameworks are tested on real-world benchmark datasets. The experimental results clearly demonstrate that attackers can indeed craft vulnerabilities in the multi-sourced data to manipulate the results produced by existing machine learning models.**To request an accessible version of the file(s) associated with this item, contact [email protected]. Please include the item's persistent URL [http://hdl.handle.net/. . .] in your request.*
TCN Architecture for Computer Vision-Based Modal Frequency Detection
No full textM.S.Structural health monitoring (SHM) refers to a damage detection strategy for surveilling engineering structures. Modal analysis, one of the most powerful tools for SHM, is the study of the dynamic properties of structures. Traditional modal analysis methods typically use physically-attached sensors for vibration measurement. However, these traditional methods have some distinct disadvantages that limit their usage in SHM tasks. First, the weight of the sensors changes the structural dynamics of the lightweight structure. Second, the limited number of attached sensors with low resolution limits the Spatio-temporal accuracy of vibration measurement. Finally, the sensors' installation process is also time-consuming, complicated, and costly. The installation errors in sensor placement also affect the precision of the collected data. Non-contact computer vision-based method is an alternative measurement technology that can address these drawbacks. In this thesis, a deep learning method built upon the TCN (Temporal Convolutional Network) model/architecture is introduced for the modal analysis. The key idea is to record the beam vibration videos using a high-speed camera and utilize the non-contact TCN-based deep learning method to detect the natural frequency of the beams. The TCN-based deep learning model takes the video streams of a vibrating structure as input and yields the corresponding natural frequency. The results show that the proposed TCN-based deep learning method is more efficient and accurate when compared with traditional methods like analytical and finite element analysis. The measurement accuracy and robustness of the model are demonstrated by using samples of different materials and sizes.**To request an accessible version of the file(s) associated with this item, contact [email protected]. Please include the item's persistent URL [http://hdl.handle.net/. . .] in your request.*
Building Future + Buildings' Futures: Sustainable Vision Planning and Tranformative Narratives for an Equitable and Just Built Environment Environment
No full textM.Arch/M.UPEcological sustainability and social justice each call for radical transformations away from the status-quo; however, these calls often appear to provide disconnected directions for public policy and other infrastructures. This study examined three sustainability strategies geared towards prolonging the lifespans of existing buildings and their materials: building preservation, deconstruction, and material reuse. The study set out to explore the potential for sustainable vision planning, a form of scenario planning which results in co-created narratives of a desired future state, to act as a lever for social justice while promoting sustainability strategies. This historic residential neighborhood was chosen as a case study since it has seen disproportionate rates of demolition in the past due to major urban renewal plans, in the present due to the city's strategy of demolition for lowering vacancy rates, and if current trends persist, in the future due to encroaching institutional development. The first phase of data collection sought to identify the critical intersections between social and ecological injustice in Buffalo's Fruit Belt neighborhood as it related to existing building and building material management. During a focus group which included neighborhood advocates and building stock experts, participants identified context-specific systematic barriers and opportunities to building preservation, deconstruction, and material reuse. A grounded theory approach was used to analyze this data, leading to the development of challenge themes relationships among these themes. The second phase of data collection was designed to give residents space to express their values and come up with creative visions of transformation towards a more sustainable, equitable, and just system of building and building material management in the neighborhood. The survey prompts were derived from the challenge themes which resulted from the expert focus group. The survey asked residents to rate the significance of each challenge to their community, rate the perceived effectiveness of challenge-relevant interventions from across the USA if they were to be adapted for the Fruit Belt, and share any of their own unique visions for change in reaction to the challenges. The study results suggest that resident values and visions are more aligned with sustainable development compared to the trajectories of the status-quo. They also suggest that facilitated sustainable vision planning, with locally-relevant visioning prompts defined by experts, can result in vision narratives which promote social justice and ecological sustainability simultaneously. To match the visionary nature of the methodology, I have expressed the results through the medium of a graphic novel. The graphic novel, titled "Building Future + Buildings' Futures in Buffalo's Fruit Belt" speculates on two plausible neighborhood transformations and their associated impacts: one scenario depicts what could happen if business as usual attitudes are allowed to persist, another depicts an alternative vision for the future which prioritizes the values and visions of Fruit Belt residents. The novel features the buildings as the antagonists in the business as usual scenario, and protagonists in the alternative visions scenario. The graphic novel serves to present examples of plausible transformations and their associated social, environmental, and economic impacts. In hope, it can act as a toolkit for advocacy, education, and creative community empowerment, inspire the reconsideration of traditional visualization methods in the fields of architecture and planning, and provoke further discussion and engagement centered around community-driven sustainable vision planning which results in creation of sustainable, equitable, and just systems of building and building material management.**To request an accessible version of the file(s) associated with this item, contact [email protected]. Please include the item's persistent URL [http://hdl.handle.net/. . .] in your request.*
From Occupation to Decolonization: Art, Media, and Film in Contemporary Social Movements
No full textPh.D.From Occupation to Decolonization: Art, Media, and Film in Contemporary Social Movements examines the roles of art, media, and film in social movements in New York City and beyond over the past decade. Using decolonial and abolitionist modes of analysis and research, this practice-based dissertation tracks the work of several collectives of artists and organizers based in New York, including MTL Collective, Global Ultra Luxury Faction, Direct Action Front for Palestine, and Decolonize This Place as they emerged from the possibilities and failures of the Occupy Wall Street movement in 2011. Particular emphasis is given in the first two chapters to the activation of museums and cultural institutions as sites of struggle, ranging from the American Museum of Natural History to the Whitney Museum of American Art, as well as the 2019 strategic pivot taken by artists and organizers towards targeting state agencies, including the MTA and the NYPD. The third chapter considers the role of multiple media forms and tactics in the work of movement building, with special emphasis on my film practice concerning Palestine, and its constant dialectic with my work as an artist and organizer in New York. This chapter concludes with the manifesto "Principles of Decolonial Film," as well as a pedagogical framework for the Decolonial Media Lab, an extra-academic project I am currently bringing to fruition in the context of a new movement space in New York City. The conclusion discusses the prospects of decolonial and abolitionist work in light of the COVID-19 crisis and subsequent uprising, including the mass removal of monuments around the United States and the need to develop cultures of care and accountability within movement spaces. The dissertation ends by proposing feminist and queer ethics of decolonial healing as the core of an emergent artistic, spiritual, and political practice.**To request an accessible version of the file(s) associated with this item, contact [email protected]. Please include the item's persistent URL [http://hdl.handle.net/. . .] in your request.*
Development of Fluorophore-Appended Transition Metal MRI Contrast Agents for Cell Labeling Applications
No full textPh.D.Paramagnetic transition metal complexes generate MRI contrast through modification of1H water T1, T2 relaxation rates or chemical exchange saturation transfer (CEST) of ligand 1H resonances. In this thesis, Saccharomyces cerevisiae (Baker’s yeast) or Candida albicans cells are labeled with transition metal complexes including Co(II) complexes as paramagnetic chemical exchange saturation transfer agents (paraCEST) and Fe(III) complexes as relaxivity agents. In addition, β-glucan particles isolated from yeast are studied as delivery vehicles for Fe(III) based MRI contrast agents. Co(II) based complexes containing 1,4,7-triazanonane (TACN) macrocycle with amide appended fluorophores were developed as bimodal paraCEST agents. The complexes exhibited unusual pH dependent paraCEST and water shift properties. The dramatic pH dependence is attributed to the addition of an inner-sphere water upon deprotonation of the amide pendant group. The labeling of S. cerevisiae with these complexes was optimized through fluorescence microscopy and validated by ICP-MS quantitation of cobalt. Weak asymmetry was observed in the Z-spectra of Co(II) complex labeled S. cerevisiae cells towards the development of cellCEST agents. The Fe(III) based analogues of the Co(II) complexes displayed promising T1 relaxivity properties despite the lack of an inner-sphere water. S. cerevisiae cells labeled with these Fe(III) complexes displayed enhanced T1 relaxation rate constants and fluorescence properties. The complex displayed both cytosolic and organeller localization of the Fe(III) complex, as observed through fluorescence microscopy. Yeast cell wall derived β-glucan particles (GP) are micron sized particles with macrophage targeting properties. Strong interactions between coordinatively unsaturated Fe(III) based T1 agents and GP were observed without any physical or chemical modification of the complex. To further explore the binding and release of these coordinatively unsaturated complexes from GP, a fluorescent analogue with dansyl fluorophore was prepared (Fe(TOD)). The Fe(III) complex labeled GP were stable under physiological conditions and displayed quenched T1 relaxation rate constants. Treatment with a small molecule bidentate chelator (Maltol) or under acidic conditions such as found in phagosome of macrophages released the Fe(III) complex from the GP and restored the T1 relaxivity of the complex in solution. Preliminary studies with J774A.1 human macrophage cell line treated with Fe(TOD) labeled GP suggests promising labeling of the cells as observed through fluorescence microscopy. Successful labeling of two types of yeast, S. cerevisiae and pathogenic C. albicans was achieved using coordinatively unsaturated high spin Fe(III) macrocyclic complexes that are effective T1 contrast agents. The labeled yeast cells demonstrated quenched T1 relaxation rate constant and enhanced T2 relaxation rate constant. Cell wall localization of these complexes was proposed based on maltol mediated retrieval of these complexes and change in surface morphology of the Fe(III) complex labeled cells. The hydrophobic nature of the β-glucan based cell wall could potentially diminish the T1 proton relaxation while the rigidity of the cell wall structure could provide the necessary organization to the labeled cells to boost the bulk magnetic susceptibility contribution towards T2 proton relaxation.**To request an accessible version of the file(s) associated with this item, contact [email protected]. Please include the item's persistent URL [http://hdl.handle.net/. . .] in your request.*
Taking Care: Experiences with Self-Care Among Black American Women Working for Social Change
No full textPh.D.Even though self-care has been discussed and written about in printed media by Black women activists, limited research exists about how Black women working for social change practice self-care to manage stress. This study explored the views and experiences of practicing self-care among Black American women working for social change. An exploratory qualitative methodology with a critical-ideological paradigm and reflexive thematic analysis as the data analysis approach were used. Two theories, Black Feminist Thought and intersectionality, were also incorporated with the qualitative methods throughout the study to center Black women’s experiences. 10 Black women working for social change were interviewed about their social change work, how they practice self-care, benefits they receive from self-care, experiences of their bodies when doing self-care, and the barriers to practicing self-care. The nine themes that occurred from the data were: (a) self-care is essential preservation of self against stressors and oppression; (b) self-care is practices focused on internal care; (c) self-care is intentional; (d) self-care community; (e) lack of resources is a barrier to self-care; (f) Strong Black Woman ideal makes self-care challenging; (g) stress as disconnection from the body; (h) mindfulness; and (j) self-care sustains mission of activism. The context of the participants was also described through their stressors and relationship with social change work. The findings revealed that participants stated they received the most benefits to their well-being from self-care practices that focus on internal care. Internal self-care included practices that enhanced participants’ relationship with their own mind, emotions, and body as well as assisted in maintaining healthy interpersonal relationships. Additionally, self-care practices that helped participants reach positive embodiment through feeling connected with their bodies were also reported to elicit more benefits for participants’ well-being. Using the frameworks of Black Feminist Thought and intersectionality, the implications of the findings for Black women working for social change, mental health professionals, and social change organizations are discussed.**To request an accessible version of the file(s) associated with this item, contact [email protected]. Please include the item's persistent URL [http://hdl.handle.net/. . .] in your request.*
Application of Machine Learning to Cue Analysis for Behavioral Modification
No full textM.S.Behavioral researchers work to develop methods to modify the practices of individuals who exhibit poor behaviors, and note that they often also exhibit impulsive decision-making. For these individuals, making changes is especially challenging when they are suffering from obesity, smoking, pathological gambling, or substance abuse, for example. Research has shown that the ability to delay gratification may be an indicator of one's ability to modify these behaviors successfully. If researchers can then develop techniques to use with their subjects that ultimately helps delay gratification, then researchers claim their subjects will be more likely to be successful in their behavior modification. One way researchers attempt to help delay gratification is using a method referred to as Episodic Future Thinking (EFT). EFT is a process where one generates cues during intertemporal decision making that paint a positive picture of a future event. At the same time, researchers often use a quantitative model referred to as Delay Discounting (DD) to measure the ability to delay gratification. The measure can be thought of as one's willingness to wait for a longer time to receive a greater reward, vs. those who would prefer a smaller reward immediately. Researchers have found that EFT and the quantitative effect it has on DD, has a positive impact on decision-making, and helps achieve positive behavioral changes in impulsive decision-makers. In our research, we are using machine learning to analyze the textual cues generated by subjects during EFT to understand their practical value in DD. We seek to answer two primary questions. First, can we train a machine-learning model to predict which of two cues will have a greater positive effect on DD? Second, can we analyze the cues to understand which words have the most significant effect? To accomplish this, researchers at UB ran a set of clinical studies to generate cues and measure their effect on delayed discounting. From those cues and associated discounting measures, we developed an approach to normalizing and defining a cues relative Impulsivity Effect (IE). We then used a language model to extract a contextualized information from cues. The resulting features were then fed to classifiers, including Logistic Regression, Random Forest, and two layers ANN, to predict each cue's impulsivity effect. We evaluated the results using the macro average F1 score and obtained a top score of 95%. To identify high and low impact words, we introduce probabilistic term frequency-inverse cue frequency weighting and identify those words that have to occur most frequently in the highest impact cues. Our research is a step towards helping to automate the cue creation process and understand how to encourage subjects to generate the most useful cues.**To request an accessible version of the file(s) associated with this item, contact [email protected]. Please include the item's persistent URL [http://hdl.handle.net/. . .] in your request.*
Building Information Modeling Applications in Construction Management
No full textM.S.Construction projects are complex in nature because multiple user purposes have to be satisfied, while intensive works must be conducted via a one-time collaboration partnership. The information of construction is fragmented and thus hard to be handled among the various project stakeholders. Building information modeling (BIM) provides a platform that facilitates construction information utilization. It remains challenging that the employment of BIM concerning construction management modules delivers desirable outcomes. In this research, several construction management modules were conducted fulfilling and balancing project duration and cost as well as quality. This study has contributed to the BIM applications in construction in terms of elements for threefold purpose-orientated integration. Firstly, the Greiner Hall, recently built at the University at Buffalo, was used as an example project, whose BIM model has been built streamlining construction information for time, cost, and project monitoring. Next, construction information requirements and BIM model integration as well as utilization have been depicted in detail among modeling processes and construction management modules. The applications of BIM in construction management have been exemplified for construction management modules including scheduling, cost estimating, and performance monitoring. The Greiner BIM model was built in compliance with construction management requirements by assigning as-planned parameters to each element. A project schedule containing activities and sequences has been arranged hierarchically supporting the integration of BIM into construction management modules. Both sorts of structural elements and architectural elements were modeled so as to resemble the variety of activities during project construction. In addition to element counts, parameters such as element height, width, depth, area, and volume were extracted via the model featuring as quantity candidates. These elements were associated with project activities through quantity reasoning processes which determining dependable element quantities while evaluating productivity unit factors at a time. Activity durations would be derived in accordance with the relationship between quantity and productivity of elements resulting the updated project schedule. Cost data were referenced and consolidated into the element hierarchy of project schedule. Three key project performance measures would therefore be imposed emulating project evaluation while augmenting the hierarchy of project schedule by element information.**To request an accessible version of the file(s) associated with this item, contact [email protected]. Please include the item's persistent URL [http://hdl.handle.net/. . .] in your request.*