116964 research outputs found
Sort by
Advancing channel coding via deep learning
Get PDFAdvancements in coding theory have played a significant role in the evolution of digital communication, enhancing the reliability of message decoding in noisy environments. Since Shannon’s seminal 1948 work, the field has made tremendous progress, yielding near-optimal codes for well-studied channels such as the Additive White Gaussian Noise (AWGN) channel. However, many practically important channel models still lack efficient coding solutions. This thesis addresses that gap by leveraging deep learning to design practical codes for diverse channel conditions.
Our first contribution focuses on the two-user interference channel, where two transmitters share a medium to communicate with their respective receivers. The mathematical complexity of these channels has led existing solutions to rely on heuristic strategies such as time division or treating interference as noise—methods that perform well only in extreme cases of strong or weak interference. We show that neural codes, guided by principles from network information theory, significantly enhance reliability in moderate interference scenarios. Our interpretability experiments reveal that the network implicitly learns a form of partial time division, effectively exploiting the structure of the interference channel. Furthermore, we demonstrate that this learned strategy remains robust even under practical fading conditions.
In our second contribution, we propose the first family of deep learning-based codes for noisy active feedback channels. We introduce a learning framework that exceeds the performance of existing analytical methods, and then analyze its learned latent features to develop a low-complexity analytical coding scheme.
While these two contributions focus on classical wireless channels, the underlying principle of encoding information against noise and distortion extends naturally to image-based information embedding. In our third contribution, we explore neural cover selection for image steganography, where a secret message is concealed within a cover image. By optimizing the cover via the latent space of a pre-trained generative model, we achieve minimal perceptual distortion while ensuring robust message recovery. An information-theoretic analysis of the encoder reveals similarities with the waterfilling strategy in parallel Gaussian channels.
Building on these ideas, our fourth contribution addresses the watermarking of generative model outputs. We propose an optimization-driven approach for constructing hidden messages to be embedded in the outputs of generative models, enabling both targeted watermarks for individual images and universal watermarks applicable to larger image sets. Our analysis again confirms connections to the waterfilling principle, underscoring how coding-theoretic concepts can guide practical solutions for both wireless communications and image-based embedding tasks.Electrical and Computer Engineerin
A Scalable Machine Learning Workflow for Computing Higher-Order Force Constants with Automatic Differentiation
No full textUnderstanding and predicting thermal transport in crystalline materials requires accurate computation of higher-order derivatives of the total energy with respect to atomic displacements—particularly third-order and fourth-order force constants. Traditional finite-difference approaches, while widely used, suffer from numerical instabilities and scaling issues, especially when applied to complex or low-symmetry materials. In this work, we develop a machine learning-based workflow that leverages automatic differentiation (AD) to compute these higher-order energy derivatives with high fidelity and computational efficiency.
Our method integrates neural network interatomic potentials trained on ab initio datasets with AD-enabled frameworks to compute gradients of arbitrary order analytically. We benchmark the accuracy of this approach against conventional finite-difference methods across a range of training dataset sizes. Importantly, the entire workflow is designed to scale on high-performance computing (HPC) resources, and we utilize the supercomputing infrastructure at the Texas Advanced Computing Center (TACC)—including Frontera, Lonestar6, and Stampede3—to manage the large-scale DFT data generation, neural network training, and AD-based derivative computation.
The use of TACC systems has been essential for parallelizing data generation across large configuration spaces, enabling hyperparameter tuning at scale, and performing memory-intensive AD operations. By exploiting these HPC capabilities, we achieve a more efficient probing of atomic configuration space and an increase in model accuracy and scalability. Our results demonstrate that machine learning combined with AD offers a robust path forward for computing higher-order force constants needed in lattice thermal conductivity calculations and phonon scattering predictions.
Our work highlights how combining machine learning, automatic differentiation, and high-performance computing enables next-generation materials modeling workflows—delivering both improved accuracy and scalability for thermal transport applications.Texas Advanced Computing Center (TACC
Designing Spike-Compatible Neural Networks for Low-Latency Inference on HPC Platforms
No full textSpiking Neural Networks (SNNs) are well-suited for their potential to deliver low-latency and energy-efficient inference, especially on neuromorphic hardware. In contrast to conventional Artificial Neural Networks (ANNs) that use continuous activation functions, SNNs employ discrete spikes to communicate among neurons. Therefore, SNNs provide a computational model that is more biologically plausible.
Despite the potential advantages, the development of SNNs has faced significant challenges. Training these networks is particularly difficult because of the non-differentiable nature of spiking events and the difficulty of mapping continuous-valued ANN activations to discrete spike events.
This work explores a spike-compatible training approach, which includes training ANN architectures (such as VGG-8, VGG-16, and ResNet-18) from scratch using custom spike-compatible activation functions. These activations are quantized to match the behavior of Leaky Integrate-and-Fire (LIF) neurons. The trained ANN is then converted into an equivalent SNN without additional re-training, transferring activation parameters to maintain the original network's scaling.
The methodology is implemented in PyTorch and is compatible with high-performance computing (HPC) platforms such as those at the Texas Advanced Computing Center (TACC). While experimental evaluations are ongoing, the goal is to demonstrate that this spike-compatible training process can produce functional SNNs capable of operating effectively with a small number of time steps, which is important for latency and energy considerations in large-scale computing environments.Texas Advanced Computing Center (TACC
Prevalence and Parenting: Family Practices and Adolescent Mental Health in the U.S. (NSCH 2018–2023)
No full textObjective: To estimate the prevalence of adolescent anxiety, depression, and behavioral problems in the United States and examine their associations with five family practices: sharing ideas, problem-talk, parental aggravation, shared meals, and working together.
Methods: We pooled repeated cross-sectional waves of the National Survey of Children’s Health (2018–2023). Outcomes were coded as ever/never. After cleaning and harmonizing items, we summarized yearly prevalence, computed Spearman correlations between each practice and outcome, and fit logistic regressions with the five standardized practice predictors (per 1-SD increase), adjusting for survey year. To aid interpretation, we (a) compared prevalence across intuitive practice categories (e.g., “Never,” “Sometimes,” “Most of the Time,” “Always”) and (b) assessed practice stability over time using 0–1 rescaled means.
Results: Across 2018–2023, the prevalence of anxiety, depression, and behavioral problems increased. In contrast, the distribution of the five family practices remained largely stable over the same period. At the individual level, lower parental aggravation showed the strongest protective association with all outcomes (adjusted odds ratios ≈ 0.60–0.75 per SD). Sharing ideas, eating meals together, and working together were also protective (≈ 0.70–0.90). “Problem-talk” displayed small odds ratios slightly above 1.0, consistent with reactive conversations after concerns arise. Correlation patterns and effect sizes were broadly consistent across years.
Conclusions: Core family-engagement behaviors are robustly associated with lower odds of adolescent mental-health problems, yet their average levels did not shift materially from 2018 to 2023. Rising population-level prevalence is therefore unlikely to be explained by changes in these practices alone, pointing to broader upstream drivers. These findings highlight family engagement as an actionable target for individual-level prevention while motivating investigation of social and environmental contributors to the secular increase.Texas Advanced Computing Center (TACC
Doctoral thesis recital (saxophone (chamber))
No full textSunburst / Karalyn Schubring -- Gradient. I., Tints / Katahj Copley -- Archangels. II., Raphael ; III., Gabriel / Stacy Garrop -- Submerged / Miguel del Aguila -- Piano quintet, op. 67 II., Adagio expressivo / Amy Beach ; arr. Connor O'Toole.MusicName of supervisor not provide
Volumetric characterization of zircon in support of detrital geochronology
No full textZircon is one of the most studied accessory minerals because of its chemical composition and ability to record events as a geochronometer. It is highly resistant to weathering, which enables it to be transported from source to sink in a sedimentary system. Zircon is a nesosilicate (ZrSiO₄) with a tetragonal crystal system, often with bipyramidal terminations. Zircon morphology may reflect both source rock and transport histories. Zircon often contains substantial uranium, thorium, and hafnium. Uranium and thorium decay causes damage that can lead to metamictization. This project focuses on characterizing zircon grains in 3D using X-ray Computed Tomography (CT), which nondestructively provides data on shape (including euhedrality, roundness, brokenness, etc.), and composition. An overarching goal is to investigate whether the information gained from CT analysis of epoxy mounts used for detrital zircon geochronology justifies the effort and expense by creating new research possibilities. New functionality in the Blob3D software package provides information on crystal form by detecting faces and evaluating their arrangement and symmetry. The zircon samples are from the Amazonian Craton in South America, the San Joaquin River in California, and the Colorado River in Texas. Grain shapes are evaluated manually using a scoring table, using both the 3D CT data and 2D microscopy, to categorize grain geometry, roughness, and brokenness. These data are then compared to automated Blob3D analysis to evaluate how well the software replicates human judgment. We explore a new method to quantify crystal structure by mapping faces as isosurfaces and analyzing their orientations on a stereonet. We find that automated analysis reflect meaningful geometric trends, they exhibit weak to moderate correlation with human judgement. This highlights how subjective classification can affect quantitative methods and provides the potential for improved analysis through machine learning. Crystal form data extracted from CT enhances our ability to interpret the complex histories recorded within zircon grains. In geochronology, this study’s findings will help with nondestructive grain analyses, provenance studies, and crystallization history reconstructions.Earth and Planetary Science
Master's thesis recital (flute)
No full textUnidentified works for: flute and piano ; solo flute ; flute, clarinet and piano.MusicName of supervisor not provide
Decadal Monitoring of Colonias Using Remote Sensing and Machine Learning in Lower Rio Grande Valley (LRGV)
No full textColonias are under-resourced communities along the Texas–Mexico border that lack access to basic infrastructure such as potable water, sewage systems, and paved roads. These areas are highly vulnerable to environmental hazards—particularly flooding—and continue to expand with limited monitoring. The most recent spatial dataset on colonias, published in 2014 by the Texas Office of the Attorney General, leaves a critical information gap for planners and policymakers.
This project leverages the computational power of TACC to develop a scalable remote sensing and machine learning framework for monitoring colonia development over the past decade. Focusing on the Lower Rio Grande Valley (LRGV), including Hidalgo, Starr, Cameron, and Willacy Counties, we integrate high-temporal-resolution PlanetScope imagery (daily/monthly/quarterly) with high-spatial-resolution NAIP aerial imagery (0.6–1 m).
To address the spatial–temporal tradeoff between these datasets, we use deep neural networks trained on TACC systems to simulate NAIP-like outputs from PlanetScope data. These enhanced images are combined with building footprint datasets (OpenStreetMap, Microsoft, and Overture Maps) to train a second neural network model that maps buildings consistently from 2014 to the present.
TACC’s high-performance computing infrastructure enables efficient model training, inference, and management of large-scale geospatial archives. The resulting monthly and yearly building maps provide new insights into colonia expansion and infrastructure deficits.
Field validation supports remote sensing analysis, focusing on resident needs in flood-prone areas. This integrated, scalable framework directly aligns with TACC’s mission to advance science and society through high-performance computing and is transferable to other regions facing informal urban growth and environmental vulnerability.Texas Advanced Computing Center (TACC
Mechanism of SNARE-Mediated Membrane Fusion
No full textNeuronal communication occurs when synaptic vesicles fuse with the plasma membrane, leading to neurotransmitter release. The core fusion machinery consists of the SNARE proteins syntaxin1 and SNAP25 in the presynaptic membrane, and synaptobrevin-2 in the vesicle membrane. These proteins assemble into a tight four-helix bundle called the SNARE complex through their SNARE motifs, which brings the membranes into close proximity because the SNARE motifs of synaptobrevin and syntaxin-1 are linked to transmembrane domains (TMDs) via short juxtamembrane (jxt) sequences. However, the mechanism by which the SNAREs induce membrane fusion remains enigmatic.
Based on molecular dynamics simulations and physiological data, we postulate that hydrophobic groups from the SNARE jxt linkers and TMDs act as a “local detergent”, facilitating encounters of the hydrophobic acyl chains from both bilayers in the polar membrane-membrane interface. These encounters nucleate formation of a hydrophobic seed that undergoes a rapid expansion, forming stalk-like structures that progressively develop into a fusion pore.
We are testing this hypothesis by quantitatively characterizing lipid behavior and lipid-SNARE interactions in all-atom molecular dynamic simulations of SNARE-mediated fusion. We found that, although most lipid acyl chains of a flat bilayer that mimics the plasma membrane are buried, they can sparsely populate conformations where the acyl chain is located close to the membrane surface. Strikingly, the SNARE complex dramatically increases the propensity of long-lived “trapping” events of these acyl chains at the bilayer surface, increasing the probability of an encounter with an acyl chain from the opposing membrane that gives rise to a hydrophobic nucleus and leads vesicle fusion. These results provide an atomic basis for the mechanism of initiation of SNARE-mediated membrane fusion. This mechanism is attractive because it has a natural physicochemical basis and may be shared by all membrane fusion proteins.Texas Advanced Computing Center (TACC
The mechanics of madness : Lovecraftian horror, player agency, and procedurality
No full textThis thesis examines how video games uniquely express Lovecraftian horror through procedural systems, character customization, and non-player character (NPC) behavior. While cosmic horror in literature and films relies on atmosphere, narrative, and implication to evoke existential dread, video games offer an interactive medium where the core themes of Lovecraftian horror are enacted through play. Drawing on Ian Bogost’s concept of procedural rhetoric, this study argues that horror in these specific games emerge from the rules, mechanics, and constraints that structure player experience. Focusing on Bloodborne (2015) and Cult of the Lamb (2022) as case studies, the thesis explores how procedurality undermines traditional notions of player agency. In Bloodborne, systems like “Insight”, the “rally” mechanic, and the hub area serve to take away control and destabilize perception. In Cult of the Lamb, I analyze how the game’s management simulation systems, such as resource allocation, follower behavior, and “ritual” mechanics, initially grant the player the role of a powerful cult leader. However, these systems ultimately undercut that authority by binding the player to the will of The One Who Waits, a mysterious cosmic entity whose influence limits player autonomy and makes sure that all progress serves a larger, unknowable agenda. These games, aesthetically, narratively, and mechanically different from each other in every way, are still procedural in nature. Both try to replicate the dread of Lovecraftian fiction while simultaneously reimagining how that dread is experienced through interactive systems. This project also examines how character customization deepens horror by creating attachment to the avatar, which is later destabilized by loss of agency. NPCs further this effect by acting as both narrative anchors and procedural entities whose behavior can shift unpredictably, reinforcing the sense of a world indifferent to the player’s actions. Ultimately, this project argues that video games are particularly well-suited to adapting Lovecraftian or “Cosmic” horror. Through procedural systems that take away player agency and force players to confront their own limitations, these video games transform cosmic horror from a narrative theme into an experiential condition.Radio-Television-Fil