116964 research outputs found
Sort by
Designing the imagined body : embodied costume design and contemporary belly dance
Get PDFThe belly dance costume in popular imagination and within dance subcultures evokes a multitude of meanings. Often these are personal to the individual dancer, who is the craftsperson of their image and how they imagine their body within a complex and evolving artistic heritage. This thesis seeks to explore, through the lens of a costume designer and belly dance practitioner, the means in which costuming acts as a form of dialogue between the body in performance and the imagined body within this artistic heritage. Utilizing the author’s body as site of research and the experiences shared by a small segment of the Central Texas belly dance community, the practice of costuming for performance is tied to both innovation and placing the body within specific associations, categories, and genres.Theatre and Danc
Unraveling the diary of Merer : how project management built the Great Pyramid of Khufu
Get PDFIn 2008, scientists led by French archaeologist Pierre Tallet excavated the ancient harbor complex at Wadi al-Jarf on the Red Sea in Egypt. Inside hidden caves adjacent to the port, Tallet discovered papyrus sheets of hieratic writing documenting important details of Fourth Dynasty construction. The fragmented texts formed a journal of the daily activities of workers under the superintendent Merer, an Egyptian official tasked with coordinating the maritime transportation of the casing stones placed on Khufu’s Pyramid. Tallet’s findings advanced our knowledge of Old Kingdom logistics, but further examination of the diary’s meaning exposes hidden relevance. This thesis asserts that previously established principles in what we now term project management were the most influential methods employed in pyramid construction, with administrative tools, such as writing and analytics, broadly used throughout organizational hierarchies.Architectur
Longing for the break : improvisationality and forms of Indian Anglophone poetry
Get PDFThis dissertation theorizes ‘improvisationality’ as a creative-critical paradigm for making and reading modern and contemporary Indian poetry written in English. Drawing on scholarship on both the Anglo-American lyric and African American poetry and music, this transnational study of South Asian poetry sees in three South Asian poets a distinctive affinity. Like ‘the break’ in a live jazz performance, ‘improvisationality’ occurs before an actual improvisation is chosen and performed. It lies ‘in between’ thought and word. Agha Shahid Ali, Karthika Naïr, and Tishani Doshi provide a range of formal practices that preserve fleeting moments of ‘improvisationality.’ Their poems also form an archive of reflections on how ‘improvisationality’ might be understood as both an aesthetic and an ideological paradigm. Using sound and texture, they play with received structures and effects of spontaneity through strategic uses of oral registers and musical resonances. To craft their own improvisational aesthetics, these poets look for, and create breaks within, the already improvisational formal South Asian vernacular traditions of the ghazal, the epics, Bhakti, and Hindustani music. They engage in a bridging of text and performance, which both revoices South Asian expressive
genres, making them contemporary, and reimagines the English language in subversive ways 8
that demand attention to colonized voices and bodies. Their poems are plural and relational, thereby undoing essentialist understandings of lineage and identity through their demand that we listen and listen again. Ultimately, by foregrounding form itself over authoritative voice, these poets capture suspended moments of potentiality and spontaneity in between the violence of the past and affirmative, even ecstatic, future voicings.Englis
Improving recommendation system efficiency through hardware software co-design
Get PDFRecommendation systems are the core workload behind algorithmic feeds, advertising placement, and music recommendation. Deep recommendation models rely on memory intensive embedding tables to convert categorical context into continuous values. The models spend a significant fraction of their parameter budget on embeddings and a large portion of runtime looking-up and reducing embedding vectors. This work outlines approaches to co-design recommendation models and the systems they operate on in order to improve efficiency.
I demonstrate an effective technique for enhancing embedding training. Our scheme trains embeddings using training-time forced factorization of the embedding (linear) layer, with an inner dimension higher than the target embedding dimension. During inference we convert the trained two-layer embedding into a single-layer, reducing memory footprint. I call this approach Multi-Layer-Embedding Training (MLET). MLET consistently produces better models when tested on multiple recommendation models for click-through rate (CTR) prediction. At constant model quality, MLET allows embedding dimension reduction by up to 16x, and 5.8x on average, across the models.
In order to achieve maximum model throughput, high effective embedding lookup bandwidth is essential. In distributed settings, achieving high bandwidth requires accesses to each distributed memory to be load balanced. Load balancing embedding access is difficult because embeddings are accessed with a power law distribution. While some existing tools can generate load balanced embedding placement plans, said tools do not consider the network bandwidth constraints or are limited in the placement types they can perform. I present Narwhal, an embedding partitioning tool that simultaneously optimizes embedding placement in memory and reduction placement on cores. Our approach uses a novel methodology that models network traffic and leverages historic embedding access patterns in order to effectively balance requests on distributed DRAM accelerator architectures. I compare to existing state-of-the-art embedding partitioners and show that Narwhal provides 1.10x average speedup with large scratchpads Additionally, we show that Narwhal can produce these solutions faster than SOTA solvers.Electrical and Computer Engineerin
Quarantine in motion : pandemic prevention at the intersection of human mobility, epidemiology, and AI
No full textThe rapid spread of airborne infectious diseases, such as COVID-19, has highlighted the need for more effective and scalable disease mitigation strategies beyond traditional manual contact tracing and exposure notification applications. This dissertation explores the novel paradigm of “Quarantine in Motion”, which enables individuals to maintain mobility while reducing disease transmission risk. By leveraging graph neural networks (GNNs), multi-agent reinforcement learning (MARL), and real-world human mobility data, this work develops an integrated framework for predicting high-risk locations, identifying transmission pathways, and optimizing mobility strategies to mitigate outbreaks. To extend disease contact tracing from reactive to proactive risk management, we introduce risk-informed exposure prediction, where GNNs process Foursquare location data to forecast hourly disease hotspots. Simulating over 36,000 risk-aware agents in Austin, TX, we show that even after 50% of the population has been infected, individuals can still maintain mobility while reducing new infections by 13%. Further, we propose a network science-based approach to dynamically construct and prune contact networks for recurring interactions, which significantly improves the accuracy of individual-level epidemic predictions. Using mobility data from 1.3 million devices, we validate this framework across two major U.S. cities, Austin and New York City, demonstrating improved outbreak curve estimation with reduced model uncertainty. To enhance the effectiveness of automated contact tracing, we introduce Infectious Path Centrality, a novel network metric that enables graph learning-based edge classification, achieving a 94% F1-score in identifying key transmission events. We further demonstrate that bidirectional contact tracing, which retroactively and proactively isolates potential infections, reduces the effective reproduction rate by 71%, outperforming traditional forward tracing. Finally, we present an online automated disease-aware navigation system that dynamically infers health states and deploys mobility-aware agents, achieving a 92% backwards-tracing F1-score and reducing disease spread by 29%, even under conditions of probabilistic testing and social hesitancy. By integrating graph learning, reinforcement learning, and network-based disease modeling, this dissertation provides a robust framework for real-time epidemic mitigation. The proposed methods bridge the gap between individual-level risk management and population-scale disease control, offering scalable, data-driven solutions for future pandemics.Electrical and Computer Engineerin
Towards automatic migration of sequential kernels : Numba to PyKokkos
Get PDFHigh performance computing (HPC) frameworks have notably been useful for scientific communities, however, for an average user, leveraging their power can prove to be challenging. Using these highly performant frameworks requires knowledge about computer architectures, familiarity with parallel programming, and getting across a steep learning curve. Recent work on PyKokkos has bridged this gap considerably by bringing HPC to Python. While PyKokkos provides remarkable performance, switching to it is non-trivial as PyKokkos follows a programming model that is different for an average Python user. It requires type annotations in kernels, and has a rigid resemblance to its parent, Kokkos's programming patterns. Numba, on the other hand, is a well known just-in-time (JIT) compiler for Python that is widely used. This work improves the usability and migration cost to PyKokkos by introducing a new tool -- Caramba. Caramba enables automatic PyKokkos kernel generation from Numba - instantly opening doors for hardware optimized parallel performance. We also add type propagation for kernel arguments in PyKokkos to allow Caramba to function statically. Finally, we evaluate the overhead and performance trade-off with the aforementioned features and show that it is negligible.Electrical and Computer Engineerin
Where and How to Publish Research Data
Get PDFThere are a wide range of platforms available to researchers to share their data, and it can be overwhelming trying to find the best home for your data (including which platforms to avoid). This workshop will provide an overview on how to pick a repository for publishing your data, with an emphasis on generalist (non-discipline-specific) platforms, and provide best practices for sharing high-quality metadata and data that maximizes their accessibility and reuse potential and that complies with federal agencies' policies.UT Librarie
Shear stress evaluation and limits for the prestressed concrete girders end region
Get PDFThis dissertation focuses on developing novel analytical models to capture the shear response of end regions of various types of pretensioned concrete girders. The proposed analytical models account for differences in geometry and boundary conditions across girder types. The research comprises experimental testing, finite element analysis, database analysis, and analytical evaluations of shear stress. Four full-scale box beams are tested under shear loading to analyze stress flow variations at the end block resulting from one versus two supports. Nonlinear finite element analysis further examines internal stress distributions to inform evaluation model development. Additionally, the Prestressed Girder Shear Database (PGSD) is formed by incorporating 113 new experiments from the University of Texas Prestressed Concrete Shear Database (UTPCSDB) and removing outliers using clustering techniques. The proposed evaluation procedure incorporates strut-and-tie modeling and confirms additional failure modes like anchorage and horizontal shear. Tie forces are calculated at critical sections to improve accuracy over previous strut-and-tie approaches. Furthermore, nodal strength, web capacity, and confinement strength are checked through specialized nodes. Verification using 85 experiments proves enhanced predictive capability over current code provisions. Finally, evaluations of Texas standard prestressed girders indicate suitable shear stress limits depending on the type and size of the beams. For Tx-girders, limits between 0.20-0.23 f’ [subscript c] shear stress limits are recommended based on size, while limits need not relax for other girder types. In conclusion, the analytical approach enables extensive shear stress limit relaxation to permit efficient, economic girder design.Civil, Architectural, and Environmental Engineerin
Exploring the frontiers of AI with PaddlePaddle : an AI course for young Chinese students
Get PDFWith the continuous development of artificial intelligence (AI) technology, it has become more critical to popularize knowledge about AI through courses so that more people can understand it. However, in Chinese education, most courses on AI teach people how to develop AI programs, and there are few courses that introduce people to basic knowledge about AI. This report presents a course plan to introduce AI to Chinese 8th-grade students using the 5E teaching model. This course uses examples from the AI platform PaddlePaddle developed by Baidu and the AI models provided by PaddleX to introduce students to the application of AI in daily life, the specific functions and applications of computer vision, and the basic process of developing AI programs using PaddleX models. By studying these contents, students can have a basic knowledge of AI and understand the principles of AI applications in real life. More importantly, it can help to inspire students' interest in AI and encourage them to learn about relevant information actively.Curriculum and Instructio
Master's thesis recital (cello)
No full textSonata for solo cello, op. 8 / Zoltán Kodály -- Sonata no. 5 in D major, op. 102, for cello and piano / Ludwig van Beethoven.MusicName of supervisor not provide