80264 research outputs found
Sort by
Neural Network Approaches for Integer and Linear Program Value Functions: Theory and Applications
No full textThis thesis develops neural network methods for approximating the value functions of linear and integer programs, which describe how optimal values change with problem's objectives or right-hand side vectors. For linear programs, we propose a structured neural architecture based on a new representation theorem for the generalized value function. For integer programs, we study the theoretical limits of neural approximations and design learning-based methods to estimate value functions. We further apply these models to solve two-stage large scale problems and generate Chvátal-Gomory cuts. Experiments suggest that our approaches can improve accuracy and efficiency in specific settings, highlighting the potential of combining optimization and machine learning
Sense of Achievement, Belonging, and Connectedness: The Role of Loneliness, Connections, and Social Support in Adults’ Life Satisfaction, Purpose, and Self-Efficacy in the Houston Area
No full textFeelings of being accepted, understood, welcomed, and supported are often referred to as “belonging and connectedness.” These feelings play a well-documented role in a variety of outcomes, contributing to students’ academic success and the independence and well-being of older adults. Belonging and connectedness also contributes to adults’ health outcomes and has an impact on those who move to new communities and countries. In 2023, the U.S. surgeon general released a report focused on the epidemic of loneliness happening in the United States. It examined loneliness and other dimensions of belonging and connectedness, and detailed the many ways these factors were related to people’s outcomes. To explore the contributions these factors make in the lives of Houston-area residents, the Houston Population Research Center (HPRC) launched a study in late 2024 looking at how belonging and connectedness related to adults’ sense of achievement
Efficient Path Planning with Soft Homology Constraints
No full textWe study the problem of path planning with soft homology constraints on a surface topologically equivalent to a disk with punctures. Specifically, we propose an algorithm, named H*, for the efficient computation of a path homologous to a user-provided reference path.
We show that the algorithm can generate a suite of paths in distinct homology classes, from the overall shortest path to the shortest path homologous to the reference path, ordered both by path length and similarity to the reference path.
Rollout is shown to improve the results produced by the algorithm.
Experiments demonstrate that H* can be an efficient alternative to optimal methods, especially for configuration spaces with many obstacles
Network-based Analysis of Alternative Splicing Events in Alzheimer's Disease-associated Cell Types
No full textAlternative splicing (AS) contributes to transcriptomic and proteomic diversity by processing a common pre-messenger RNA into various distinct transcript and protein isoforms. Through generation of multiple protein isoforms with distinct interaction profiles, AS contributes to functional diversity and can remodel protein-protein interactions (PPIs). Dysregulation of AS has been implicated in neurodegenerative diseases, including Alzheimer’s Disease (AD), where specific neuronal populations exhibit selective vulnerability. To investigate how cognitively and physically stimulating enriched environment (EE)— a known modulator of cognitive improvement in AD progression— affects neurons vulnerable and resistant to AD, we leverage various bioinformatics approaches to study tissue and environment dynamics at different levels of biological organization.
In this thesis, we demonstrate cell type-specific transcriptional profiles of neurons selected for their association to Alzheimer’s Disease from mice exposed to baseline, non-enriched environments (NE) as well as EE. We identify distinct, cell type-specific gene regulation in response to EE for our vulnerable cell types of interest, suggesting that EE impacts different mechanisms for each of these cells. We then compare network rewiring changes driven by alternative splicing events across AD-associated cell types using a previously developed method, Splitpea, which was originally designed for cancer patient sample analysis. We adapt the tool to use a mouse PPI reference background and to handle additional upstream differential splicing analysis tool results. As a result, we detect PPI network rewiring events in experimental conditions. By comparing the rewired networks across cell types and environment, we discover EE-associated network rewiring for both vulnerable and resistant neurons, identifying changes in protein interactions and finding distinct, vulnerable-specific and resistant-specific biological processes that confer protective advantages against Alzheimer’s Disease. Overall, our study reveals a functional landscape driven by AS events in response to cognitively and physically enriched environment at the cellular level, and provides insights to protective mechanisms in the brain that can prevent Alzheimer’s Disease pathologies
Pedagogy Meets AI: Challenges and Innovations in LLM-Powered Learning
Get PDFThe rapid advancement of artificial intelligence, particularly large language models, is fundamentally reshaping how we learn and interact with knowledge, offering unprecedented opportunities to develop intelligent systems that enhance human learning at scale. However, realizing this potential requires addressing core technical challenges: optimizing LLMs for pedagogically sound instruction, developing robust cognitive models of student misconceptions, and designing scalable NLP approaches for nuanced, actionable feedback. My talk will focus on how my research addresses these challenges by (1) aligning LLM tutoring systems with pedagogical principles, (2) developing nuanced LLM-based learner models, and (3) developing new automated assessment tools for long, open-ended responses. By integrating insights from natural language processing and learning sciences, this work aims to advance the development of effective, scalable, and pedagogically grounded AI-enhanced educational technologies
Exclusive Vector Meson Photoproduction in Ultraperipheral Heavy-Ion Collisions at the LHC with the CMS Detector
No full textAll nuclear matter consists of tiny particles called quarks and gluons. Gluons become increasingly dominant constituents of nuclear matter when probed at higher energies or smaller Bjorken-x values. A key objective of high-energy nuclear physics is to search for the onset of gluon saturation phenomena in the limit of extreme gluon densities.
Ultraperipheral collisions (UPCs) are collisions of relativistic heavy ions at impact parameters larger than the sum of their nuclear radii. The intense electromagnetic fields generated by relativistic heavy ions can be treated as a flux of linearly polarized quasi-real photons. Photon-induced vector meson production in UPCs provides a unique and powerful probe of the gluon distribution in nuclei, as the cross section is directly sensitive to the nuclear gluon density. The J/ψ meson, a bound state of charm and anticharm quarks, is an ideal probe of the gluon density in the nucleus due to its large mass and small size. However, in symmetric UPCs, a two-way ambiguity in determining the photon emitter and the target prevents the extraction of contributions involving high- and low-energy photon-nucleus interactions. This limitation reduces the capability to probe the small-x regime. The first measurement of coherent charmonium photoproduction, where the two-way ambiguity is resolved using a forward neutron tagging technique in UPC PbPb collisions at 5.02 TeV, unveils a novel behavior of the nuclear gluon density at small-x. The results provide new insights into the gluon saturation regime and the small-x nuclear gluonic structure.
The φ meson lies at the boundary of hard scales between the perturbative and nonperturbative QCD regimes, making it uniquely suited to probe the transition between these two domains. However, the significant challenge of detecting extremely low transverse momentum kaons from coherent φ meson decays has hindered the measurement of its production in UPCs for decades. The first observation and measurement of exclusive φ photoproduction via the φ to KK in PbPb UPCs at 5.36 TeV, using the CMS detector with a new low-pT reconstruction, is presented. The results are compared to various theoretical models and provide new insights into the small-x nuclear gluonic structure at a critical scale
Compression Algorithms for Efficient Inference and Adaptation of Foundation Models
Get PDFFoundation models, including large language models (LLMs) and generative AI (GenAI) models, have demonstrated impressive capabilities and are driving progress across many industries. Despite their potential, their substantial size makes them difficult to deploy at scale and challenging to adapt efficiently. This thesis presents novel algorithms for compressing foundation models, enabling faster and more efficient inference as well as adaptation. By making these models more accessible to both individuals and organizations, this work seeks to democratize their use and accelerate innovation across science and industry.
First, we present Dynamic-Length Float (DFloat11), a lossless compression framework that reduces the size of foundation models by around 30% while preserving bit-for-bit identical outputs. DFloat11 is motivated by the observation that model weights stored in BFloat16 format exhibit low entropy and are highly information-inefficient. By leveraging entropy coding, DFloat11 assigns compact, dynamic-length encodings to weights based on their frequency. To support efficient inference with compressed models, we design a custom GPU kernel that enables low-latency, online decompression using hierarchical lookup tables, a two-phase kernel, and transformer-block-level decompression. Experiments on LLMs and diffusion transformers demonstrate that DFloat11 reduces model size by more than 30% without compromising accuracy and enables significantly higher throughput and longer sequence generation under fixed GPU memory budgets. Notably, DFloat11 is the first framework to efficiently support lossless inference of models as large as Llama 3.1 405B on a single DGX A100 node.
Second, we develop quantization algorithms that significantly reduce the GPU memory requirements of LLMs without compromising model quality. We first introduce LeanQuant, a scalable and accurate post-training quantization framework that improves upon traditional error-based methods by learning loss-aware quantization grids instead of relying on fixed min-max affine grids. This improves quantization quality, generalizes to both uniform and non-uniform schemes, and remains compatible with standard inference frameworks. LeanQuant enables accurate quantization of models as large as Llama 3.1 405B using modest compute resources. To further reduce memory usage during inference, we propose Coupled Quantization, a method for compressing the key/value (KV) cache in transformer models. We observe that KV cache channels exhibit significant redundancy, which existing quantization methods fail to exploit. Coupled Quantization jointly compresses multiple channels by leveraging their interdependence, resulting in more efficient representations. It supports quantization down to 1 bit per activation while maintaining model quality and achieves up to 3.5x throughput improvement.
Finally, we introduce SketchTune, a compressive adaptation framework for LLMs that unifies compression and fine-tuning into a single streamlined process. In contrast to traditional parameter-efficient fine-tuning (PEFT) methods that rely on low-rank constraints or require separate inference paths, SketchTune compresses model weights into compact, differentiable sketches that can be directly fine-tuned. This design removes the need for low-rank matrices to capture model updates, offering greater expressiveness while significantly reducing memory usage and training costs. Supported by theoretical insights into when sketching outperforms low-rank approximations, SketchTune delivers superior empirical performance compared to leading PEFT approaches such as LoRA, DoRA, S2FT, and LoftQ, despite using smaller base models and fewer trainable parameters.
Together, these contributions address key challenges in the efficient use of foundation models by enabling lossless compression, accurate quantization, and resource-efficient adaptation. This work not only advances the state of the art in model compression but also provides practical solutions for democratizing access to powerful foundation models, accelerating innovation and societal progress
Extensions of finitely generated Veech groups
No full textGiven a closed surface S, a subgroup G of the mapping class group of S has an associated extension group Γ, which is the fundamental group of an S-bundle with monodromy an isomorphism to G. A general problem is to infer features of Γ from G: In this thesis, G is assumed to be a finitely generated Veech group and Γ is shown to be hierarchically hyperbolic. This is a generalization of results from Dowdall, Durham, Leininger, and Sisto regarding lattice Veech groups. The focus of this defense is constructing a hyperbolic space Ê on which Γ acts nicely (isometrically and cocompactly). This example contributes to the growing evidence of a good notion of “geometric finiteness” for subgroups of mapping class groups
Attitudes Toward Mass Deportation and Immigration Policy Preferences: Insights From the Greater Houston Area
Get PDFWhile political party divisions on immigration policy preferences are well documented, attitudes are less understood among individuals who do not align closely with either major political party. To better understand the attitudes of politically moderate residents in the Houston area, nearly 10,000 residents were surveyed regarding their preferences for various immigration policies, including mass deportation, as well as their political ideology (i.e., conservative, moderate, or liberal). Analyses highlight the attitudes of politically moderate individuals to provide insight into how they perceive different strategies for addressing illegal immigration. Results point to a nuanced set of preferences, including many more residents wanting to see increased pathways to citizenship than those who want mass deportation
Solar-driven thermal desalination in off-grid applications for water purification
No full textEconomic, societal, and political consequences are some of the repercussions of water scarcity. It permeates virtually all aspects of life, demonstrated by the fact that 80% of jobs are reliant on water and that disease control is achieved through prevention which highly correlates with access to quality water. Because of this, it is imperative to develop processes that generate drinking water from alternative water sources, guaranteeing constant and safe access to this vital liquid.
Chapter 1 analyses the viability of desalination methods for a given region. To determine this, several factors need to be considered, such as consumption patterns, variability in water availability, and water stress. On the technical side, some are its target production capacity, feed stream quality, energy consumption, and its source. For low-scale applications with production in the tens or hundreds of liters, it is fundamental to have low-maintenance systems that are not dependent on the electrical grid, particularly in remote communities. While it is important to improve the efficiency of systems, it is also to reduce their carbon footprint by migrating to the use of alternative energy sources. This work analyses key components of water desalination technologies from 2017 to 2022 such as their efficiency, size, and used materials.
Chapter 2 presents a solar-driven, membrane-less, and robust thermal system, known as Solar Thermal Resonant Energy Exchange Desalination (STREED). Its purpose is to alleviate challenges of water desalination systems, for example, a high maintenance burden due to the use of membranes, low efficiency, and high electricity input. Through the use of the concept of Resonant Energy Transfer, the efficiency of the system can be maximized by reusing the energy from the enthalpy of evaporation. This section also analyzes the production of water under a realistic solar energy profile, considering the scenario where the flow rate is constant, and the one where it is dependent on solar intensity. It was observed that by adjusting the flow rate to the present available sunlight power, water production increased by 70% for a representative week of the summer in Houston, Texas. STREED’s adaptability to an energy source that varies through time, coupled with its robustness makes it the ideal candidate to install in remote communities or developing countries