137091 research outputs found
Sort by
Energy-Efficient Hardware Acceleration of Transformer-Based Models
No full textThis research presents a software-hardware co-optimization framework for energy-efficient deployment of transformer models on FPGAs. It introduces novel quantization techniques tailored for both BERT and generative LLMs. For weight quantization, the proposed Hessian-based parameter-wise method assigns optimal bit precision based on sensitivity analysis, while the proposed row-wise quantization enhances hardware efficiency by converting mixed-precision matrices into two uniform-precision blocks. For attention activations, the proposed Weight-Hessian-aware KV cache quantization applies intra-layer mixed-precision using precomputed sensitivities, eliminating runtime overhead. To further improve hardware efficiency, the proposed Query-Key coupled scheme aligns bit precision within each outer product pair, reducing implementation complexity. A concurrent quantization approach jointly optimizes row-wise weight and Query-Key activation precision using multi-precision formats, improving both compression and energy efficiency. These techniques are implemented on a novel multi-precision FPGA accelerator for BERT and GPT-2, capable of handling both power-of-two and non-power-of-two bit-widths. With optimized dataflow, the design minimizes off-chip memory access and significantly outperforms prior solutions in both energy efficiency and inference performance.Ph.D.Electrical and Computer Engineerin
Distributed Optimization Architectures for Large-Scale Decision-Making
No full textAs the scale and complexity of modern decision-making systems are rapidly increasing in autonomy, robotics, artificial intelligence (AI), and various other domains, there is an emerging interest in developing scalable and effective algorithmic frameworks capable of handling large-scale, networked and uncertain environments. Several fundamental challenges must be addressed, including scalability, complex dynamics, robustness under uncertainty, interpretability and generalizability. This thesis presents a series of novel distributed algorithmic architectures at the intersection of optimization, control theory and deep learning, designed to address these challenges.
The main contributions of this thesis are summarized as follows: (i) Two novel distributed dynamic optimization architectures for large-scale multi-agent control are introduced, providing state-of-the-art scalability for optimal control in autonomous systems and demonstrating their effectiveness through hardware experiments. (ii) A family of scalable decentralized methods for distribution steering in multi-agent stochastic systems is presented, providing a trade-off between safety capabilities and computational efficiency, accompanied with convergence guarantees. (iii) A model predictive control version of the decentralized distribution steering framework is then proposed, enabling its application to real-world multi-agent systems operating under uncertainty. (iv) A hierarchical distribution control architecture is presented for very-large-scale clustered multi-agent systems, which exploits underlying hierarchies to achieve improved scalability and robustness, as demonstrated by experiments involving up to millions of agents. (v) Finally, a deep learning-aided distributed optimization architecture for large-scale quadratic programming (QP) is developed by unfolding a new consensus-based variant of the Operator Splitting QP (OSQP) solver into a supervised learning framework. Experiments demonstrate improved performance over traditional optimization, strong generalization to larger problems, and PAC-Bayes guarantees on the expected optimality gap for unseen instances.Ph.D.Machine Learnin
Development of a Novel Molten Salt Corrosion Coupon Sampling Measurement Apparatus
No full textThis thesis aims to present a novel design for a corrosion coupon sampling apparatus for use in the extreme environment of a molten salt reactor (MSRs), thus providing corrosion qualification capabilities for the new Molten Salt Research Reactor (MSRR) located in Abilene Christian University (ACU). The purpose of this device is to accurately measure the reactor's corrosion rate while remaining within an inert atmosphere. Once removed from the reactor's cover gas (helium), the coupon samples react with nearby oxygen. Prior works were adopted as inspiration from Oak Ridge National Labs (ORNL) in the 1960s with their Molten-Salt Reactor Experiment (MSRE). They were examined for the potential for utilization in modern reactors. This work prioritizes reactor operation, safety, and coupon data viability, aiding further progress of this technology to a higher implementation standard. Five designs are iterated upon from these historical works, concluding with a final prototype for study in simulation within SolidWorks fluid simulations for coupon performance and possible system impacts. Later, to further the viability of the design, principal static mechanisms were explored to identify possible safety sealing mechanisms through compound thermal expansion driven by a more expansive internal core material. With this final design, three sample integrations were prototyped and experimented on for possible failure modes through high-temperature testing with inductive heating and physical testing. While failure modes were identified, the thermal expansive sealing function proved not viable for this specific application. The cumulative result of the work provides extensive insight into the current design and produces a clear plan for the implantation of this technology into an MSR.M.S.Nuclear Engineerin
A Survey of Non-English Parallel Corpora for Text Simplification
No full textRecent advancements in high-quality, large-scale English resources have pushed the frontier of English Automatic Text Simplification (ATS) research. However, less work has been done on multilingual text simplification due to the lack of a diverse evaluation benchmark that covers complex-simple sentence pairs in many languages. This thesis performs a comprehensive survey of 27 resources in 12 distinct languages spanning over 1.7 million complex-simple sentence pairs. My experiments and analysis of these datasets with pre-trained multilingual language models reveal exciting performance improvements from multilingual training in non-English settings. I observe strong performance from Russian in zero-shot cross-lingual transfer to low-resource languages. I further show that few-shot prompting with BLOOM-176b achieves comparable quality to reference simplifications outperforming fine-tuned models in most languages.UndergraduateComputer Scienc
Uncertainty Analysis for Hybrid Electric Propulsion in NASA EPFD Vehicles
No full textPublished in: 2025 IEEE/AIAA Transportation Electrification Conference and Electric Aircraft Technologies Symposium (ITEC+EATS)
Date of Conference: 18-20 June 2025
Date Added to IEEE Xplore: 05 August 2025
DOI: 10.1109/ITEC63604.2025.11098006
Publisher: IEEE
Conference Location: Anaheim, CA, USAThe National Aeronautics and Space Administration (NASA)'s Electrified Powertrain Flight Demonstration (EPFD) program aims to advance hybrid-electric propulsion (HEP) as part of the aviation industry's decarbonization efforts. However, the feasibility and performance of hybrid-electric aircraft are highly dependent on the development of key electrical components such as batteries, electric machines, and power converters, whose future capabilities remain uncertain. This paper presents an uncertainty quantification analysis of two hybrid-electric aircraft architectures: a turboprop aircraft and a turbofan aircraft. Using our established uncertainty propagation framework, we assess the impact of technological uncertainty on the design and performance of both vehicle configurations. The turboprop vehicle analysis builds on prior work, incorporating the latest iteration of the vehicle model, which was a retrofit study, while the turbofan analysis extends the framework to a revised model - incorporating an updated thermal management system along with a new operational mode for electric taxiing. The results indicate that hybridization offers promising benefits for the turboprop vehicle, whereas the potential advantages for the turbofan architecture appear less optimistic. The results indicate that hybridization offers promising benefits for the turboprop vehicle, whereas the potential advantages for the turbofan architecture appear less optimistic. These findings highlight the importance of tailored hybrid-electric strategies and demonstrate the utility of uncertainty quantification in guiding technology-informed design decisions
Prolonged Load Carriage During Walking Induces Fatigue and Redistributes Lower Limb Muscle Effort
No full textLoad carriage during walking (e.g., a backpack) is prevalent in everyday tasks as well as in industrial and military settings. During loaded walking, the need to support additional weight and maintain dynamic balance induces immediate changes in joint mechanics and lower body muscle activation patterns. The manifestation of acute changes in workload among lower-limb muscles during prolonged, fatiguing walking bouts remains unclear. Based on evidence that load carriage redistributes the mechanical demands toward distal muscles (e.g., ankle plantar flexors) and away from proximal muscles (e.g. hip extensors), we hypothesize that distal muscles would fatigue at a faster rate than proximal muscles. To test this hypothesis, we recruited 8 young healthy adults and compared the rate of change in the mean power frequency (rMPF), a common biomarker of muscle fatigue, for key ankle and hip muscles over a 30 minute walk during both loaded and unloaded conditions. As expected, load carriage caused an immediate increase in recruitment of active muscle volume that was larger for the ankle than for hip muscles. However, contrary to our hypothesis, we found a larger negative rMPF (i.e. more fatigue) for hip extensors (e.g., biceps femoris = -.29 Hz/min ) than for the ankle extensors (e.g., soleus = .032 Hz/min) during loaded walking (p = 0.0022). This finding suggests the possibility for a motor control strategy that acts to prioritize reliance on proximal muscles during long, highly demanding walking bouts in order to limit fatigue in key distal muscles that are important for efficient propulsive power output.M.S.Bioengineerin
Design, Modeling, Imaging, and Control of a Robotically Steerable Transcatheter Delivery System
No full textMitral regurgitation affects 1.7% of the general population and 10% of those over 75, representing the most prevalent valvular heart disease. Transcatheter mitral valve repair (TMVr), increasingly preferred for nearly half of MR patients who are non-surgical candidates, currently utilizes manually operated devices that expose clinical staff to increased radiation, require experienced operators to facilitate intuitive compensations of the joints, and offer limited precision compared to potential robotic systems. This work introduces a robotically steerable transcatheter delivery system for treating mitral regurgitation, demonstrating effective positioning and orientation of the implant with respect to the mitral valve leaflets and extensive testing to validate its clinical relevance. Over five generations of the robotic transcatheter system, the steerable end tip design is optimized for effective implant deployment and the actuation system is refined for adequacy and compactness.
Relevant clinical features are incorporated to enable intravascular surgery, intuitive control is demonstrated via joystick, and precise implant delivery is validated using existing TMVr procedure imaging modalities. The kinematics of the steerable end are derived, and a control system is presented. Since pure kinematics does not capture all the physical phenomena of a long catheter system, research has focused on characterizing tendon elongation, tendon sheath friction, and the influence of catheter configuration and tendon pre-tension. Furthermore, stiffness optimization of the bending joint with super elastic reinforcements minimizes deflection in pulsatile flow while ensuring compliant manipulation by the steerable outer sheath. Moreover, the study demonstrates the integration of a new real-time pose estimation technique utilizing ultrasound imaging for joint space control, and accurate positioning in conjunction with fluoroscopic imaging along with automatic trajectory generation. This is aimed at reducing trauma to the heart tissue during implant deployment. Extensive feasibility testing across phantom benchtop experiments, ex vivo heart models, and in vivo studies confirms the system’s efficacy, highlighting its readiness for clinical adoption and significant advancement in robotic catheter technology.Ph.D.Robotic
Identifying Meal Preparation Difficulties Faced By People With MCI And Their Care Partners
No full textMild Cognitive Impairment (MCI) is a neurological disease that primarily impacts older adults. MCI makes many aspects of life more challenging, including meal preparation, which is a crucial part of independent living. In my research, I aim to gather primary data for the purpose of creating better Artificial Intelligence technology solutions to empower people with MCI and their care network. To gather this primary data, I run kitchen studies with a pair of participants, one with MCI and one that acts as their care partner. They are given three recipes to create, with ingredients provided, and they divide the work up from there. During a study, we collect video and accelerometer data, as well as detailed after-study surveys. These studies are analyzed and coded for key actions, so that other research groups creating technological solutions for people with MCI can have detailed data of what people with MCI and their care partners actually need to alleviate challenges while performing meal preparation. Some challenges that we’ve identified so far include the fact that people with MCI often struggle with remembering long recipe instructions, the issue that picky eaters may struggle to adequately adapt a recipe, and the problem that signifiers on appliances like fridges to close the door may not always work or be noticed. In the future, I hope to run more primary studies with people with MCI and their care partners, as well as better analyze the already-run studies and collaborate with more research teams to get data for their specific projects.UndergraduateComputer Scienc
A New Search for Transient Sources of Astrophysical Neutrinos
No full textThe emerging field of high-energy neutrino astronomy has seen many advances in the last few years, led by the IceCube collaboration, including the identification of several neutrino source candidates. However, the sources that contribute to the astrophysical diffuse neutrino flux are still largely unknown. This dissertation discusses three efforts to identify these astrophysical neutrino sources. The first is an upgrade to IceCube’s processing and filtering pipeline to improve the future collection of track like data with good angular resolution. The filter I was responsible for, known as the Muon Filter, is the single most used data stream in IceCube, and I was responsible for modernizing and improving this filter for the first time since it was finalized in 2013. The second is a search for neutrino emission from GRB 221009A, the brightest GRB of all time, in the 10-1000 GeV range based on prior work conducted by Dr. Chujie Chen. This work set upper limits on neutrino emission from this GRB. Thanks to this GRB’s exceptional brightness, I was able to use these upper limits to place strong constraints on GRB 221009A’s baryon loading. A high baryon loading is needed if GRBs are a source of the highest-energy cosmic rays. The final effort was to identify neutrino transients across the entire sky. IceCube has conducted several such searches in the past, however I implement several improvements in sensitivity including the Pass2 data reprocessing effort, the improved point-source reconstruction methods, the KDE parametrization of the likelihood, and the expectation maximization method for picking flare durations, which combined give a ∼ 30% improvement in discovery potential compared to previous results. This work searches for evidence of transient neutrino emission from three catalogs of sources, including Seyfert galaxies, blazars, and a generic catalog of gamma-ray bright sources based on prior evidence of neutrino emission from the blazar TXS 0506+056 and the Seyfert-II galaxy NGC 1068. No significant evidence of neutrino emission is observed from any single source, and upper limits are placed on the most significant sources in each catalog for several time windows. However, evidence of excess neutrino emission at the 3.3σ level is identified for a sub-population of three Seyfert galaxies: NGC 7469, NGC 4151, and NGC 1068. This work is currently being prepared for future publication, which will also include a search for transient neutrino emission across the entire northern sky.Ph.D.Physic
Examining Individual Differences in Human Metacognition for Multi-choice Versus Two-choice Tasks
No full textMetacognitive ability, or the capacity to evaluate the accuracy of one’s own decision, is a critical faculty that has large implications in the effectiveness of many cognitive functions. Recent investigations have found that individual differences in depressive/anxious traits show a decrease in average confidence ratings on task performance (Benwell et al., 2022; Rouault et al., 2018). However, examinations of metacognition with 2-choice tasks have generally shown poor reliability (Guggenmos, 2021; Kopčanová et al., 2023; Rahnev, 2023), and individual difference research requires methods of investigation that produce high test-retest reliability to be considered valid. Participants completed 2-choice, 4-choice, and 8-choice perceptual discrimination over two days, and the Day1/Day2 correlations were used to determine test-retest reliability of metacognitive measures in the three conditions. Self-report questionnaires assessing trait levels of anxiety, depression, perception of control and self-esteem were collected. We found that, although response choice conditions showed comparatively sufficient test-retest reliability for confidence and metacognitive ability, the 4-choice and 8-choice conditions were slightly more sensitive in detecting a significant relationship between anxiety and average confidence, and between self-esteem and reaction time during the confidence judgement period. Perception of control significantly predicted metacognitive ability regardless of response choice condition. This effect occurred despite high association with the other three self-report measures (anxiety, depression and self-esteem), suggesting that the degree to which one views their capability of controlling outcomes and events in their daily life (i.e. perception of control) is a unique trait individual difference predictor of metacognitive ability that warrants further targeted study into its explanatory power.Ph.D.Psycholog