116964 research outputs found
Sort by
Doctoral thesis recital (clarinet)
No full text[Unidentified work for bass clarinet and piano] -- [Three unidentified works for clarinet and piano]MusicName of supervisor not provide
Massively parallelized multi-task reinforcement learning
No full textMulti-task Reinforcement Learning (MTRL) has emerged as a critical training paradigm for applying reinforcement learning (RL) to a set of complex real-world robotic tasks, which demands a generalizable and robust policy. However, online MTRL has largely been limited to training policies using unstable off-policy algorithms in the slow and CPU-intensive low-parallelization regime. As a result, multi-task approaches for robotics have been dominated by distillation from single-task experts, behavioral cloning from expert demonstrations, or high-capacity sequence models trained with offline RL. This thesis proposes to take advantage of recently popularized massively parallelized GPU-accelerated simulators for MTRL, which significantly accelerates data collection across multiple tasks by simulating heterogeneous scenes in parallel, and, as a result, offers a path to make MTRL a practical technique for multi-task learning. Concretely, we introduce a massively parallelized Multi-Task Benchmark (MTBench), a highly extendable, open-sourced benchmark featuring a broad distribution of 50 manipulation tasks, implemented using the GPU-accelerated simulator IsaacGym. In addition, MTBench includes re-implementations of state-of-the-art MTRL algorithms and architectures, providing a unified framework for evaluating their performance. We perform extensive experiments to confirm whether the reliance on off-policy methods in the low-parallelization regime of existing MTRL literature holds in the massively parallel regime, and then evaluate a suite of MTRL approaches in this new regime using on-policy methods across our evaluation settings, emphasizing their significant speed advantage on MTBench. These experiments reveal key observations on applying existing MTRL approaches to the massively parallelized regime for robotic manipulation tasks, facilitating future directions for algorithmic research in MTRL. Code is available at https://github.com/Viraj-Joshi/MTBenchComputer Scienc
Curation and organization of densely sampled data sources for virtual metrology in semiconductor manufacturing
No full textSemiconductor manufacturing tools now produce an enormous amount of raw sensory trace data. A single etch tool can simultaneously produce over 100 sensor readings, each generating up to 10 data-points per second, often with over 50 such tools operating in a single wafer fabrication facility (fab). Thus, a single fleet of etch tools in a fab can now generate well in excess of 1 TB of data each day. As data density increases, possibilities for quality product prediction based on this data improves, while functional and robust methods for data processing, curation, and organization become increasingly necessary. The scale of these data sources requires data compression for storage, with analysis taking place after the data is decompressed, and informative features for the relevant task are extracted from the decompressed signals. This traditional approach to tool data curation and analysis in a semiconductor fab is time-consuming and requires attention by an engineer or technician for every new recipe, sensor, or tool. Consequently, given those difficulties, analysis is generally done in a fragmented manner, rarely streaming relevant data from multiple sources, and when such aggregation happens, it is also one-of-a-kind, not transferable or comparable to the entirety of data within the fab. This challenge was addressed through a novel approach for compression of raw sensor trace readings into a representation that can be directly mined. This compression was accomplished through a fully automatic parsing and segmentation of raw sensor trace data into an exhaustive set of mutually exclusive segments of transient and steady state behaviors, with steady-state segments being represented via statistics-based metrics and transient portions being compressed into a set of standard transient signal descriptors. Automation was achieved due to a novel Hidden Markov Model (HMM) based formalism, which enabled segment alignment and correction of errors in segmentation, adding a level of robustness required to enable lossy compression of said data at an industrial scale.
When it comes to numerous opportunities to analyze and mine the rich data in modern semiconductor manufacturing, Virtual Metrology (VM) has been one of the most impacting. VM involves predicting physical quality characteristics on the wafer using process data obtained from the tool. In this doctoral work, improvements in VM were facilitated by novel data curation. This improvement was achieved via two approaches. First, as mentioned earlier, previous studies that used trace data have required manual segmentation and alignment of sensors/recipes. Second, through the recently available dense sensing of Optical Emission Spectroscopy (OES) signals with high resolution in time and wavelength. Unlike traditional approaches which focus on one or just a few spectral wavelengths determined based on the relevant process chemistries, this doctoral research used a novel multi-way Singular Value Decomposition (SVD) to perform extraction of OES features based on the analysis of the entire OES signal. Significant improvements in VM performance were gained from one’s ability to mine all wavelengths and all time-samples of OES readings recorded during processing of each wafer.
Lastly, the newly enabled large-scale generation of curated features creates new requirements for industrial data organization and analysis. Appropriate curation of the aforementioned trace and OES signals produced feature sets that may change in size and shape with response to segment type, process recipe, and sensor behavior. In order to engage fab-wide analysis and take actionable decisions on all available information, comparisons across signal features need to identify similarities across variable signal representations. To address this and ensure actionable decision making can be made on the newly curated features, a novel distance metric over sequences was developed, allowing for organization of arbitrary sequences and series. This allows for new organization and exploration of fab-wide data as operators are no longer limited to searching through feature sets of fixed size and shape.Operations Research and Industrial Engineerin
The interplay among sex, hormones, and early immune environment in neurodevelopment
No full textAll multi-cellular organisms develop from a single cell. During development, cells multiply and organize themselves into separate tissues with specialized functioning, establishing critical biological systems for the remainder of an organism’s life. Development, therefore, marks a particularly vulnerable period for the brain as rapid changes in brain cell functioning and structure mean that changes to the brain microenvironment can disrupt typical brain development. Additionally, neurological differences between the sexes are believed to be established during this time, as sexual dimorphism in brain structure and hormonal milieu appear immediately following birth in mammals. In this dissertation I propose that: (1) early life represents a vulnerable time during which an organism is particularly susceptible to alterations in the environment, and (2) the early life microenvironment programs the peripheral and central immune system for the duration of life. In humans, infections during pregnancy increase the risk for developing neurodevelopmental disorders such as autism in offspring. This phenomenon, known as maternal immune activation (MIA), is replicated in animal models, demonstrating evolutionarily conserved mechanisms through which maternal inflammation alters behavioral phenotypes. In Chapter 2 and Chapter 3, I explored how MIA altered behaviors associated with autism, such as sociability and repetitive-like behaviors. MIA resulted in reduced sociability in both sexes and increased repetitive-like behaviors only in males, in line with evidence that males appear more susceptible to MIA compared to females. Additionally, MIA disrupted immune cell populations in the periphery and brain, indicating that MIA led to alterations in peripheral and central inflammation. Notably, most pregnancies exposed to infection do not result in offspring with neurodevelopmental pathology. To determine whether MIA is a disease primer (requiring secondary inflammatory insults to result in pathology in certain offspring), in Chapter 2, I examined whether MIA paired with postnatal immune stimulation led to exacerbated behavioral pathology. Surprisingly, although a single hit of MIA resulted in expected behavioral phenotypes, the combination of the two-hits did not exacerbate, but rather sometimes ameliorated pathology. Additionally, MIA resulted in a range of behavioral phenotypes, with many offspring appearing resilient to behavioral pathology following maternal infection. Therefore, in Chapter 3, I examined immune mediators that may contribute to susceptibility to MIA. As expected, most offspring were resilient to behavioral alterations. However, those that did show behavioral alterations, specifically social deficits, had higher percentages of cytotoxic T cell populations, indicating T cells as a potential contributor to vulnerability to MIA. Many psychiatric disorders including neurodevelopmental disorders, such as autism, have some degree of sex bias, leading to valuable questions about the contributions of sex and hormones to vulnerability for certain diseases. The innate immune cell of the brain, microglia, display sex differences in count, function, and morphology across the lifespan and brain regions. However, the role of early life hormone environment on the appearance of these sex differences is unknown. Therefore, in Chapter 4, I investigated how early life hormone environment affects the morphology and function of microglia across the lifespan. Sex differences in hippocampal microglial morphology appear in adolescence, with male microglia showing increased branching through adulthood before retracting their branches to a more proinflammatory phenotype in aged males. Hormone treatment during the first two days of life recapitulated these effects in both males and females, demonstrating the organizational power of hormones in early life on the neuroimmune environment. Taken together, this work demonstrates that the environment, especially during early life, programs the immune and neuroimmune environments, resulting in lasting consequences on behavior and physiology.Neuroscienc
Evaluating selective reporting methods in meta-analysis : considering dependent effect sizes when estimating adjusted effect sizes
No full textSelective reporting detection tests and adjustment methods evaluate the presence of selective reporting and estimate bias-adjusted average effect sizes to address its impact on meta-analytic results. Available tests and adjustment methods are univariate by design, meaning they only handle a single effect size per primary study, but meta-analyses of education research typically include multiple effect sizes per study. Extending previous methodological studies evaluating the performance of detection tests when dependent effect sizes are present, this study evaluates approaches to estimating bias-adjusted average effect sizes in meta-analyses involving dependent effects. This simulation study examines the performance of three adjustment methods (i.e., Regression-based methods, Trim & Fill, and 3-parameter selection models) to estimate bias-adjusted effect sizes under various conditions. In practice, researchers may or may not report adjusted average effect sizes using either a contingent or non-contingent strategy based on the statistical significance of selection bias detection tests. To this end, this simulation study also compares the accuracy of adjusted effect sizes when they are reported if contingent versus non-contingent on the statistical significance of tests for the presence of selection bias. When selective reporting censoring is high, in general, results indicate poor estimation performance for all selective reporting adjustment methods regardless of approach used to handle dependent effect sizes. Variation by study conditions is reported, including: overall average effect size, amount of between-study heterogeneity, within to between-study heterogeneity ratio, and average sample sizes from primary studies. Performance differences between the contingent and non-contingent strategies for adjusted estimates were primarily observed within the regression-based methods — specifically the aggregating and two modeling approaches — with minimal differences identified in the other estimation methods or approaches to dependency. Results can help guide researchers on the best practices for selective reporting tests and methods to use given certain study conditions. Future research is needed to improve current methods or identify new methods to estimate adjusted effect sizes, particularly when the amount of selective reporting is suspected to be at a higher probability.Educational Psycholog
Tree-based models with basis functions
No full textThe Bayesian Additive Regression Trees prior is a powerful prediction tool, with a wide range of extensions applied to a variety of problems. Some of these extensions, however, lack computational efficiency due to increased complexity. We present three novel approaches: a scalable BART with targeted smoothness, which uses a reduced-rank Gaussian Process approximation to improve scalability; a locally adaptive linear BART, providing smoother predictions for BART models while maintaining most of the computational efficiency from BART; and a scalable Bayesian causal forest for continuous treatments, which utilizes a reduced-rank Gaussian Process approximation to extend the Bayesian causal forest model. Simulations were conducted to evaluate computational efficiency and predictive performance. The methods were also applied to real-world datasets.Information, Risk, and Operations Management (IROM
Characterization, modeling, and application of thin-film freeze-drying
No full textThin-film freeze-drying (TFFD) is a promising alternative to conventional freeze-drying (CFD) where frozen films are lyophilized as a bed of particles rather than a solid frozen mass. Understanding the interplay between heat and mass transfer and product structure is essential for refining TFFD as a scalable and efficient freeze-drying method. Comparative drying studies at multiple scales revealed that TFFD achieves faster sublimation and superior product quality than CFD under aggressive conditions, though it requires higher chamber pressures and shelf temperatures for optimal efficiency. Custom-fabricated trays with enhanced surface area were found to reduce drying times by over 37%, with the greatest increases in efficiency coming at low chamber pressures. To complement these empirical findings, a mechanistic model was developed to simulate mass and heat transfer during primary drying of frozen thin films. Micro-CT-derived structural parameters informed calculations of effective thermal conductivity and mass diffusivity, enabling accurate predictions of drying kinetics based on particle shape and bed structure. Model outputs closely matched experimental data, supporting its application to cycle optimization and scale-up. The utility of TFFD was demonstrated through the development of an inhalable dry powder formulation of D29, a shear-sensitive mycobacteriophage. Using efficient experimental design techniques and multivariate analysis, key formulation components and TFFD processing parameters were optimized to achieve high phage stability and aerosol performance. Powder characterization showed an amorphous phase consisting primarily of trehalose and crystalline phases of mannitol and leucine. Low residual moisture content also helped to increase the aerosol performance. The resulting formulation can deliver greater than 10⁸ pfu per dose via a dry powder inhaler and maintained stability for over nine months under refrigerated conditions. More concentrated phage powders could deliver in excess of 10⁹ pfu per dose with minimal reduction in titer compared to delivery of a liquid phage formulation by nebulization, which drastically reduced the phage activity. Together, these studies present a detailed characterization of the TFFD process and demonstrate its use to prepare inhalable powders of a shear-sensitive biological product.Pharmaceutical Science
Defending Red River : cultural districts, preservation, and live music in 21ˢᵗ century Austin
No full textIt’s likely that most people who live in urban areas are familiar with the term “cultural district.” The phrase may seem self-explanatory on the surface – a specific place with a collection of artistic and other cultural spaces. But it’s also likely that many people could not point to exactly what cultural districts do. Cultural districts are a relatively new policy tool, only coming to prominence in the past 30 years in the U.S. Initially, they were mostly designed to be a tool for economic growth, using cultural assets as a commodity to bring revenue to specific neighborhoods. However, the goals of cultural districts have shifted over time. In the past decade, cultural districts are increasingly being designated as a means of preserving cultural assets in neighborhoods that are under threat of displacement. This is the case with the Red River Cultural District (RRCD) in Austin, which was designated by the City of Austin in 2013 and the state of Texas in 2021. After the closure of music venues in the neighborhood and the threat of more closures looming, business owners banded together with support from City staff to create the district to attempt to prevent more venues from closing. This report examines the ways in which cultural districts in this new form attempt to preserve their cultural assets through historical and policy analysis as well as qualitative research through interviews with RRCD stakeholders.Community and Regional Plannin
MENA at the Threshold? Proliferation Risks and Great Power Competition
No full textThis article situates the Middle East and North Africa (MENA) in the global nuclear order, emphasizing how the region has both challenged and spurred adaptations in international nuclear governance for decades. It then examines two pressing contemporary issues: the uncertain trajectory of Iran’s nuclear program after Israeli and US military strikes in June 2025, and the anticipated expansion of nuclear energy across MENA, which could also result in more countries with capabilities that would be conducive to pursuing the bomb. Both developments underscore the difficulties of managing nuclear latency in a conflict-prone region, where tensions among local actors inflect nuclear decision-making. While there are opportunities to mitigate these challenges, and principles that policymakers should follow in addressing them, nuclear aspirations are likely to remain a prominent feature of MENA’s security landscape so long as underlying tensions between regional actors remain unresolved.LBJ School of Public Affair
US Policy Toward North Korea: Quo Vadis?
No full textAs the Trump administration recalibrates America’s global priorities, containing Pyongyang should be at the top of its agenda. Despite the progress of North Korea’s illicit weapons programs, the United States should still pursue its longstanding goal of Complete, Verifiable, and Irreversible Denuclearization through diplomatic actions, such as coordination with democratic allies in Seoul and Tokyo, as well as with coercive tools, such as unilateral sanctions and the use of military force. In doing so, Washington should not neglect North Korea’s continuing and grave human rights abuses, which Pyongyang is still actively perpetrating against its own people, abetted by Beijing and Moscow. Finally, the United States must take concrete steps to counter the strategic collusion among autocratic regimes in China, Russia, Iran, and North Korea.LBJ School of Public Affair