137091 research outputs found
Sort by
EcoPhaze: A Hierarchical Approach to Carbon-Aware Accelerator Pruning for Distributed Deep Learning Systems
No full textThis paper introduces EcoPhaze, a hierarchical approach to carbon-aware accelerator pruning for distributed deep learning systems. Recognizing the escalating environmental impact of large-scale AI, the research addresses the co-optimization of accelerator architecture and distributed training deployment with the explicit goal of reducing carbon footprint and energy consumption. Building upon the Phaze framework for hardware architecture search, EcoPhaze integrates carbon modeling that accounts for both embodied and operational emissions. A novel hierarchical pruning algorithm is proposed, which optimizes for a performance-per-carbon metric, guiding the search towards more sustainable and efficient accelerator designs. Evaluation on BERT-Large demonstrates that EcoPhaze can achieve comparable performance to baseline methods with significantly reduced resource utilization and embodied carbon, highlighting the potential of co-designing hardware and deployment strategies with environmental considerations at the forefront.UndergraduateComputer Scienc
Exploring Dopamine Dynamics of the Basal Ganglia in Freely Moving Mice
No full textDopamine depletion in the basal ganglia is associated with bradykinesia, tremor, and other motor deficiencies commonly seen in Parkinson’s disease. Drugs such as L-dopa increase dopamine levels in this brain region, restoring motor function. Although the importance of dopamine in this system has been firmly established using in vitro and in vivo headfixed methods, the short-term effect of dopamine in vivo in freely moving animals is less understood. Recent technological advances have allowed for dopamine recordings using fiber photometry to be collected. However, there is a lack of published data about the immediate effects of dopamine on the activity of striatal neurons during free behavior in vivo. With concurrent measurements of calcium as a proxy for neuronal firing, we can examine how dopamine affects the activity of neurons. Here, we explore a novel dual color fiber photometry experiment in the striatum. By injecting the mouse striatum with a cocktail of dLight and jRCaMP, we were able to record dopamine and calcium levels simultaneously. This work indicates that such experimental procedures can provide interpretable data in future research.UndergraduateNeuroscienc
Hardware-Software Co-Designed GPU Security Mechanism for Memory Safety
No full textMemory safety has long been a critical challenge in low-level programming languages such as C and C++. Similar concerns have recently emerged in GPU programming, which relies on C/C++-style languages and now plays a vital role in critical systems such as autonomous driving and medical diagnostics. Recent studies have identified vulnerabilities in GPU programs, including buffer overflow attacks that adversaries can exploit to execute malicious code. These exploits pose a significant threat, potentially compromising the integrity of other users' deep learning applications on shared cloud servers.
To address this pressing issue, this dissertation proposes hardware-software co-designed mechanisms to enhance GPU memory safety. By leveraging the massively parallel architecture and unique programming model of GPUs, these mechanisms integrate hardware and software collaboration to deliver highly optimized solutions that surpass the efficiency of directly applying CPU-centric approaches to GPUs.
The first proposed mechanism, GPUShield, introduces region-based bounds checking using pointer tagging, tags stored in the upper bits of a pointer are utilized to retrieve bounds metadata in memory. By exploiting the GPU's specialized load/store instructions and employing static-time filtering, GPUShield achieves significant performance advantages over existing software-based GPU memory safety mechanisms.
The second mechanism, LMI, is a fine-grained memory safety solution that employs in-pointer metadata for runtime bounds checking. To mitigate the reduced availability of upper bits for tagging caused by virtual memory expansion and multi-GPU support, LMI efficiently utilizes power-of-two-aligned pointers. This design enables fine-grained memory safety with minimal hardware overhead compared to earlier mechanismsPh.D.Computer Scienc
Efficient Depletion Monte Carlo Methodology with Hybrid Multi-Group and Continuous Energy Resolution
No full textThis thesis focuses on the development of hybrid depletion-based methodologies for Monte Carlo (MC) reactor physics simulations. More specifically, a hybrid depletion MC scheme that leverages both the high-fidelity continuous energy (CE) and quick multigroup (MG) physics approaches is developed here. The novel efficient time integration scheme removes the need to use the expensive CE simulations and replaces them with more frequent and relatively cheap MG simulations. The time integration scheme incorporates a sub-stepping technique, in which both spectrum and spatial updates to the reaction rates are combined using these low-cost MG MC solutions. This is accomplished by generating on-the-fly homogenized groupwise macro- and micro-cross sections for various regions using the high-fidelity CE MC approach. Groupwise microscopic cross sections are then used to perturb the material composition after solving the Bateman equations, and thus also allowing for the update of the groupwise macroscopic cross sections. The updated macroscopic cross sections are then used in a reduced order MG MC simulation for the substeps. Getting the spectrum from the substeps then allows for the one group microscopic cross sections to be updated continuing to the next substep or step. The treatment of how to homogenize a region and perturb the macroscopic cross sections is important and rigorously tested in this thesis.M.S.Nuclear Engineerin
Understanding Traffic Congestion Through Multiscale Analysis: Sensor Deployment, Car-Following and Universal Scaling
No full textTraffic congestion is a persistent and multifaceted problem in urban transportation systems. Its characterization requires understanding not only driver behavior and infrastructure but also the limitations of the data and models used to represent traffic dynamics. This thesis approaches the problem through a multiscale framework, addressing how congestion is measured, modeled, and interpreted from both empirical and theoretical perspectives.
Macroscopic representations of traffic states, such as the macroscopic fundamental diagram (MFD), are widely used to capture the relationship between average flow and density across urban networks. While loop detector data is the most commonly used source for estimating the MFD, its accuracy can be affected by how detectors are positioned along the network. Analytical derivations, empirical analyses, and simulation experiments show that the spatial distribution of detectors introduces systematic bias in the observed MFD. These findings point to structural limitations in how traffic states are aggregated and motivate practical guidance on sensor deployment and interpretation of aggregated traffic data.
At the microscopic level, accurate representation of car-following behavior is essential for reproducing traffic phenomena such as capacity drop and oscillations. The existing stochastic car-following models provide a flexible framework for capturing traffic instabilities such as stop-and-go waves and capacity drop, but its calibration is challenging due to its stochastic formulation and high-dimensional parameter space. This study evaluates numerical and heuristic optimization methods under a maximum likelihood estimation framework to improve parameter robustness. Results show that while traditional methods are sensitive to initial conditions and prone to local optima, global search strategies such as genetic algorithms enhance reliability in the presence of noisy and limited data.
Despite improvements in measurement and modeling, traffic congestion often emerges in unexpected ways, especially in dense networks operating near capacity. To better understand this concepts from statistical physics are applied to investigate whether congestion exhibits underlying critical behavior. High-resolution trajectory data are used to reinterpret traffic density as a growing surface and congestion as a percolating fractal structure, where stop-and-go waves form percolation clusters. This formulation enables the estimation of dynamic and roughness exponents. The results show strong empirical evidence that traffic congestion exhibits criticality and follows scaling behavior consistent with the Kardar-Parisi-Zhang universality class. Recognizing this criticality opens new possibilities for phase-aware traffic control strategies that incorporate fluctuation statistics to detect and mitigate near-critical traffic states.
By integrating measurement, modeling, and theory, this thesis provides a comprehensive view of traffic congestion, from how it is empirically observed, to how individual driving behaviors can be inferred and calibrated under uncertainty, and to how its large-scale dynamics can be understood through universal scaling relationships. The results support more reliable data interpretation, robust calibration methods for stochastic models, and a new theoretical framing of traffic as a critical system with universal features.Ph.D.Civil Engineerin
Anticipatory and Reactive Motion Planning for Close-Proximity Human-Robot Interaction
No full textHuman-Robot Collaboration in close-proximity environments requires precise motion planning for an effective partnership between the human and robot. This thesis introduces the motion planning system Real-time Collaboration via Multi-Objective Trajectory Optimization, which possesses both anticipatory and reactive motion planning traits, effectively making it able to plan in real-time. Our real-time motion-planning framework optimizes on a set of both task-based and human-based cost functions. We then formulate a Nonlinear Model-Predictive Control problem and use trajectory optimization to find local optimal solutions for the objective function to optimize the trajectory for small time horizons. This thesis goes in-depth about how the system separates itself from prior work in the field.UndergraduateComputer Scienc
Evolution of Essential and Ancient Genes: Understanding and Modeling Fitness in Transfer RNAs
No full textA central question in evolutionary biology is how genes evolve over time and how mutations to those genes affect fitness. How a mutation affects fitness may depend in part upon mutations at other sites. For example, the negative impact of a bad mutation may be compensated by other mutations that restore or enhance function. Understanding this dynamic is crucial to elucidating not just the evolution of genes, but the evolution of complex traits, including interactions between genes.
This thesis is structured into three related projects that examine the following fundamental questions: 1. How do genes evolve, and what does the compensatory evolution pattern of single-copy genes look like? Specifically, I explored mutations in mitochondrial tRNAs (mt-tRNAs) in Caenorhabditis nematodes to identify patterns of compensatory evolution over short and intermediate evolutionary timescales, including interactions between mt-tRNAs and associated factors encoded in the nuclear genome. 2. To what extent can computational techniques enhance our understanding of the impact of mutations, particularly on overall fitness? In this project, I utilized experimentally derived fitness estimates for thousands of S. cerevisiae allelic variants of a nuclear-encoded arginine tRNA (tRNACCU). I assessed how well computational inferences from sequence information, such as secondary structure prediction and the minimum free energy folding score, explained fitness variation. 3. Can machine learning improve upon computational fitness prediction and address the limitations of existing prediction tools? In this project, I developed an inclusive machine learning model that integrates multiple features to estimate fitness from sequence data. Together, these studies aim to provide a more comprehensive understanding of the dynamics of tRNA evolution by combining advanced computational techniques with concepts of evolutionary biology.Ph.D.Biolog
High efficiency broadband power amplifiers using SiGe HBT BiCMOS technology
No full textThe objective of this research is to investigate the design challenges of high efficiency broadband power amplifiers (PAs) using silicon germanium (SiGe) heterojunction bipolar transistor (HBT) and establish design methodologies to address these challenges. Broadband amplification with high efficiency is becoming crucial for emerging wireless broadband applications such as 5G/6G wireless communications, secured military/satellite communications, phased array radars, and high-resolution imaging. To support the system level requirement for low cost and high integration capability, SiGe HBT PAs can be a viable solution. To overcome an inherent efficiency and output power limitations of conventional distributed amplifiers for wideband applications, one part of this research provides two design examples of SiGe HBT non-uniform distributed power amplifier (NDPA) and develops novel and systematic design approaches to improve efficiency over wide bandwidth (BW). To meet the demand for highly efficient PAs with higher peak-to-average power ratio and wider channel BW for modern 5G NR-U/WLAN applications, the other part of this research establishes design guidelines for four types of series-combined transformer (SCT)-based Doherty output matching networks and provides a design example of SiGe HBT SCT-based Doherty PA for higher power back-off (PBO) efficiency enhancement.Ph.D.Electrical and Computer Engineerin
Identifying Barriers to STEM in Underrepresented Groups
No full textThe persistent underrepresentation of minority groups, women, and individuals from disadvantaged socio-economic backgrounds in STEM fields, particularly physics, remains a significant challenge. This thesis aims to delve into the multifaceted barriers that perpetuate this issue, ranging from societal stereotypes and environmental factors to the pervasive imposter syndrome.
Through the lens of Critical Race Theory (CRT), which posits that the overlap of race and property rights can illuminate social biases, this research seeks to identify and analyze the systemic inequities within educational institutions that hinder the participation of underrepresented groups in physics and other STEM disciplines. Additionally, other contributing factors, such as the lack of qualified physics educators in high schools, further dissuade individuals from pursuing degrees in this field.
Utilizing a combination of qualitative methods including literature reviews and interviews with undergraduate physics majors at Georgia Tech and K-12 educators from Gwinnett County, this study aims to uncover the root causes of the lack of diversity in STEM fields. The narratives provided by the participants offer valuable insights into the challenges and obstacles faced by individuals aspiring to enter these fields and successes that could aid in viable solutions.
By shedding light on these barriers and their systemic origins, this thesis hopes to contribute to the ongoing discourse on promoting diversity and equity in STEM education. It underscores the importance of addressing these issues at various levels, from policy-making to institutional practices, to foster a more inclusive and equitable environment for all individuals interested in pursuing careers in STEM.UndergraduatePhysic