36214 research outputs found
Sort by
Learning Structured Representations of the Visual World
Get PDFThesis (Ph.D.)--University of Washington, 2025Humans develop complex internal models of the world which allow us to generalize remarkably well to new scenarios and tasks. While deep learning has steadily improved in performance through data and scale, it conspicuously lags behind in its generalization to changing data distributions and transfer across tasks when compared to biological intelligence. We argue that one key element absent from current deep learning systems is this internal model of the world to enable efficient transfer of knowledge to new settings and data. In this work, we investigate how aspects of world models such as compositionality and 3D spatial understanding can be learned from visual data and be used to improve the efficiency and robustness of current machine learning systems. We develop new methods and loss objectives for learning structured representations. We demonstrate how learning from more complex visual data such as video, embodied exploration, and 360° video enables learning more structured world models which improves sample efficiency and spatial understanding. In addition, we explore other directions and develop methods to improve the transfer of knowledge between tasks and robustness to shifting data distributions
Numerical Simulation of Liquid Oxygen Droplet Combustion in Hydrogen under Microgravity Conditions
Get PDFThesis (Ph.D.)--University of Washington, 2025This work presents an efficiently lean, custom-built numerical simulator developed to study the combustion of a liquid oxygen (LOX) droplet combusting in Hydrogen gas (H2) under microgravity conditions. Motivated by drop-tower tests conducted at ZARM (The Center for Applied Space Research and Microgravitation) in Bremen, Germany, the quasi-static evaporation framework reproduces key coupled processes– flame dynamics, Stefan flow, droplet regression, and surface ice formation– within a computationally minimalist yet physically faithful model. The governing reaction-diffusion equations were solved using finite-difference methods incorporating time-dependent, spatially homogeneous Stefan velocity fields generated by real-time evaporative feedback from the flame. The simulation achieves strong quantitative agreement with experimental and computational benchmarks, reproducing flame stand-off ratios (F/D ≈ 2–3.5) and peak adiabatic flame temperatures (Tpeak ≈ 3000 K) consistent with previous work. Diffusive heat transfer dominates the total energetic flux, contributing 80–85% of the total heat input (Qmax ≈ 0.3–0.5 W), while radiative effects remain secondary, in accordance with previous estimates. Parametric sweeps over surface ice coverage fraction ψ reveal compensating feedback between evaporative impedance and geometric flame shape contraction. A single predominant global reaction mechanism, augmented by equilibrium radical generation at the reactive flame front, suffices to reproduce thin flame-sheet behavior within the high-Damkoehler limit. The resulting simulator balances interpretability, stability, and physical fidelity, requiring no HPC infrastructure and running interactively accessibly in Google Colab. Beyond LOX–H2 combustion, this framework offers a transparent, extensible platform for general coupled parabolic PDEs, bridging the gap between high-overhead CFD and simplistic static equilibrium tools
Generative Keyframing
Get PDFThesis (Ph.D.)--University of Washington, 2025Keyframing is a fundamental element of animation creation and video editing. It involves defining specific frames, i.e., keyframes, that mark important moments of change and guide how the intermediate frames are filled or interpolated. In early hand-drawn animation, a keyframe is a visual drawing created by animators, with assistants manually drawing the in-between frames. With the advent of digital animation and video editing software, a keyframe became a set of parameters that define the state of the rendered character or object at specific times, with in-between transitions produced by interpolating these parameters. However, such parametric approaches rely heavily on manually designed controls and artist-crafted heuristics, making them difficult to capture complex, nuanced, and realistic motions. Furthermore, they do not naturally generalize to real image and video domains. The rapid progress of visual generative models that are trained on large collections of visual data and capable of learning rich appearance and motion patterns, has made it possible to generate high-fidelity imagery and realistic motion. Building on these advances, this thesis investigates generative keyframing, a data-driven, non-parametric, image-based approach to the keyframing process. To this end, I present a series of works in this thesis that collectively develop and explore this idea. I begin with the basic aspect: using generative models to synthesize transitions directly from images, and even to fully generate in-between motions. I first present a GAN-based technique for smoothing jump cuts in talking head videos, synthesizing seamless transitions between the cuts even in challenging cases involving large head movement. I then introduce a method for generating in-between videos with dynamic motion between more distant key frames by adapting a pretrained large-scale image-to-video diffusion model with minimal fine-tuning effort. Beyond automatically generating transitions between keyframes, I further explore multi-scale keyframing for achieving very deep zoom. Specifically, I introduce a multi-scale joint sampling diffusion approach for generating consistent images (keyframes) across different spatial scales while adhering to their respective input text prompts. This enables deep semantic zoom and a continuous zoom video can be rendered from these images. When working with multiple keyframes, one import question is how they should be ordered in the final video. I address this in the context of dance video generation---specifically, music synchronized and choreography-aware animal dance video---where unordered keyframes representing distinct animal poses are arranged via graph optimization to satisfy a specified choreography pattern of beats that defines the long-range structure of a dance. Finally, I conclude with discussions and directions for future works
Incorporating spatial and temporal dynamics into evaluations of fish populations and habitat
Get PDFThesis (Ph.D.)--University of Washington, 2025The ocean is rapidly changing, with impacts on both the physical environment and ecological systems. In my dissertation, I seek to understand how variations in fish populations over space and time are driven by environmental conditions in the ocean, and to improve statistical methods to accomodate this variation and thereby contribute to sustainable fisheries. In the following four chapters of my dissertation, I develop, test, and apply improved methodologies that link fish demographics to environmental conditions and address pressing management concerns. A spatio-temporal model of weight-at-age of walleye pollock improved our understanding of the dynamics of local and population-level demographic processes and can be used in future stock assessment models. I developed a statistical model that incorporated a physiological response to temperature and oxygen into distribution modeling to better capture this joint effect, in the context of predicting impacts of climate change on local fish densities. Because spatial statistical models rely on environmental data, I used statistical approaches to expand oxygen data available and test the sensitivity of ecological models to environmental data. Lastly, I applied these improved techniques in retrospective statistical models to evaluate evidence for oxygen limitation on the distribution of 32 groundfish species in the northeastern Pacific Ocean. Overall this dissertation advances statistical solutions for accomodating spatio-temporal data in estimates and predictions of fish ecological responses to environmental change
Retrofitting automated verification to systems code by scaling symbolic evaluation
Get PDFThesis (Ph.D.)--University of Washington, 2025Formal verification is a technique for eliminating classes of bugs insystems software by formally proving that a system's implementation
meets its intended specification. While effective at systematically
preventing hard-to-catch bugs, formal verification demands a
significant effort from developers in the form of manual proofs.
Automated verification techniques reduce the burden of verification by
leveraging automated reasoning to avoid the need for manual proofs.
But as a result, they sacrifice generality and require developers to
build bespoke verification tools and to carefully design systems with
automated verification in mind. This dissertation explores how to make it easier to build andreuse automated verifiers, and how to retrofit systems to automated
verification. To do so, we built Serval, a framework for writing
automated verifiers for systems code. To use Serval, developers write
an interpreter for a language; Serval then leverages the Rosette
programming language to lift the interpreter into a verifier via
symbolic evaluation. Serval also comes with a set of techniques and
optimizations to help overcome verification bottlenecks. We use Serval to develop automated verifiers for RISC-V, x86, Arm,LLVM IR, and BPF. We apply these verifiers to retrofit automated
verification to two existing security monitors previously formally
verified using other techniques: CertiKOS, an OS kernel with strict
process isolation, and Komodo, a monitor that implements secure
enclaves. We port these two systems to RISC-V, modifying their
interfaces for automated verification and to improve security.
We write specifications amenable to automation, and compare our
efforts with that of the original systems. To demonstrate applicability to systems beyond security monitors, weuse Serval to build Jitterbug, a framework for writing and verifying
just-in-time (JIT) compilers for the Berkeley Packet Filter (BPF)
language in the Linux kernel. We develop a specification of compiler
correctness suitable for these JITs. Using this approach, we found
and fixed more than 30 new bugs in the JITs in the Linux kernel and
developed a new, verified BPF JIT for 32-bit RISC-V
Probing and Manipulating Novel Electronic States in Graphene Multilayers
No full textThesis (Ph.D.)--University of Washington, 2025The pursuit of understanding, manipulating, and engineering quantum materials has opened new avenues for exploring exotic states of matter. Among these materials, graphene multilayers represent a rapidly emerging platform in which the electronic properties can be drastically tuned through artificial control of stacking configurations. These systems provide a unique opportunity to investigate a wide range of correlated and topological phases, including (Quantum) anomalous Hall effects, fractional quantum Hall states, and unconventional superconductivity. Their versatility offers an exciting pathway for advancing both fundamental science and future technologies in quantum information and materials engineering. This dissertation presents a comprehensive study of the electronic transport properties of graphene multilayers and the development of novel experimental techniques to manipulate their quantum states. First, we revisit the transport behavior of monolayer, Bernal bilayer, and twisted bilayer graphene across a range of twist angles, uncovering new and unexpected phenomena. Building on these findings, we introduce two complementary tuning approaches designed for cryogenic transport measurements: the application of high pressure using a diamond anvil cell and the precise control of in-plane uniaxial strain. These techniques open access to previously unexplored regimes, enabling detailed studies of symmetry breaking, electronic correlations, and topological transitions. The insights gained from these measurements deepen our understanding of the ground states in graphene multilayers and demonstrate their potential as a versatile platform for exploring strongly correlated and topological quantum phases
Efficient Scaling of Language Models
Get PDFThesis (Ph.D.)--University of Washington, 2025Large language models (LLMs) are progressively reshaping how humans interact with information, offering increasingly sophisticated access to knowledge through natural language interfaces and advancing reasoning capabilities across diverse domains. Yet, their impressive gains have hinged on a simple recipe: exponentially increasing resources. Over the past years, computational requirements have increased tenfold annually, training costs now reach billions of dollars, and we are rapidly exhausting the internet's high-quality text. This dissertation addresses a critical question: How can we unlock further capabilities from scale while curbing the exponential growth in compute, data, and energy?I present three complementary innovations across the LLM development pipeline that fundamentally improve scaling efficiency. Traditional LLMs depend on tokenization—a preprocessing step that introduces biases and inefficiencies. BLT eliminates this bottleneck by learning directly from raw bytes, dynamically grouping them into larger entropy-adaptive patches via lightweight encoder-decoder modules. This tokenizer-free architecture not only improves robustness to noisy and multilingual inputs but achieves up to 50% reduction in inference FLOPs while matching tokenization-based models' performance. Controlled scaling experiments demonstrate that BLT enables a new scaling dimension with improved scaling trends over current approaches. Beyond model architecture, data quality has a significant impact on model performance. Socratic Pretraining transforms unlabeled documents into richer training signal by masking salient sentences, synthetically generating content questions about missing information, and training models to both pose questions and draft answers. Applied to BART-large, this approach achieves state-of-the-art performance on QMSum and SQuALITY (+1.0 and +0.5 ROUGE-1 over strong baselines), halves labeled data requirements, and improves faithfulness across multiple control interfaces—all with minimal computational overhead. QLoRA combines novel 4-bit quantization with parameter efficient finetuning, reducing memory requirements of supervised learning by 15× without performance degradation. This efficiency enabled comprehensive instruction-tuning studies revealing that data quality, not quantity, drives downstream performance in post-training. Efficient finetuning of the quantized base model also alleviates quantization errors reducing inference memory requirements while enabling full-precision quality. These methods collectively demonstrate that sustainable scaling requires rethinking fundamental assumptions at each pipeline stage. BLT's entropy-adaptive computation, Socratic Pretraining's synthetic supervision, and QLoRA's quantized adaptation each extract more capability per unit resource—whether FLOPs, tokens, or memory. Together, they chart a practical roadmap for continued LLM progress within more sustainable computational bounds, proving that smarter algorithms can bend the scaling laws themselves
Beyond the Mirror: Building Body Trust & Media Literacy
Get PDFMaster of Social Work (MSW)The prevalence of eating disorders (EDs) has increased drastically over the past decade, with social media playing a significant role in their development and ongoing perpetuation. This grant proposal not only examines the rising rates of EDs and the underlying factors contributing to their increase, but also highlights the disparities experienced by different communities and identities. Interviews with individuals who have lived experience with EDs, along with a review of current interventions and resources, were conducted to identify the most necessary intervention. The review reveals significant gaps in existing resources, which are primarily based outside the U.S., often costly, lacking in comprehensiveness and gender inclusivity, and focused more on prevention than on providing support for individuals currently struggling with EDs. Clearly, there is a need for an accessible, inclusive intervention that addresses the multiple factors that contribute to negative body image and disordered eating. My proposed intervention, Beyond the Mirror: Building Body Trust and Media Literacy, is a free online program designed for individuals with EDs or body dissatisfaction of all genders, identities, shapes, and sizes. Beyond the Mirror will offer interactive learning modules on Body Trust and media literacy, as well as an optional peer support group. Beyond the Mirror is built on the Body Trust framework and the Developmental Theory of Embodiment (DTE), guiding the program's approach to challenge harmful societal narratives surrounding weight, health, and beauty. Ultimately, this program will empower individuals to cultivate a healthier relationship with their bodies, while also equipping them with the tools to navigate social media in a way that fosters self-acceptance and resilience against unrealistic beauty standards
Saying "Yes" to PrEP: Examining PrEP acceptance among urban men who have sex with men (MSM) at a Seattle public sexual health clinic.
No full textThesis (Master's)--University of Washington, 2025Pre-exposure prophylaxis (PrEP) is extremely effective at preventing HIV, but uptake has been below targets. PrEP acceptance and uptake are commonly examined through hypothetical acceptability and active use, but the act of accepting a PrEP recommendation from a clinical provider is under-studied. This cross-sectional study used patient survey data and health records from the Public Health—Seattle & King County Sexual Health Clinic in Seattle, WA to examine how PrEP acceptance differed by key covariates among men who have sex with men (MSM). Among 1721 patient visits where a provider recommended PrEP to an MSM, PrEP was accepted at 1168 (68%) visits. Prevalence of PrEP acceptance was lower among patients aged 25 years or older compared to patients under age 25 (PR = 0.92, 95% CI = 0.85, 0.99) and higher among patients who reported greater numbers of recent sexual partners compared to those reporting one or fewer (PR = 1.38 95% CI = 1.13-1.68), but was lower among Black patients compared to White patients (61%, PR = 0.85, 95% CI = 0.76-0.96) and patients identifying as bisexual compared to those identifying as gay (PR = 0.84, 95% CI = 0.75-0.93). According to visit notes recorded by providers, patients who declined commonly believed their risk of HIV exposure was too low to warrant daily PrEP, were concerned about side effects, or preferred to discuss with their primary care provider, while some preferred to abstain from sex altogether following the HIV/STI exposure which brought them to the clinic. To increase PrEP uptake, future work should identify and test strategies to maximally support some MSM who have an indication for but less often accept PrEP
Microdosing Hallucinogens/ Psychedelics: Trends and Practices Among College Students and Young Adults
No full textThesis (Ph.D.)--University of Washington, 2025The emerging behavior of microdosing hallucinogens/psychedelics is, at minimum, broadly defined as the act of taking a low dose of a hallucinogen/psychedelic. Interest in microdosing hallucinogens/psychedelics generally stems from the promising clinical trial results of "full" dose hallucinogen/psychedelic use as well as legalization efforts and corporate investment in hallucinogens/psychedelics. Increasingly, interest in microdosing includes as a wellness practice for enhancing mood and cognition as well as a treatment for mental health and substance use symptoms, however research is sparse. As part of two large ongoing studies, we recruited participants who were generally screened for lifetime microdosing. Through a brief web survey, we aimed to document the frequency of microdosing, including concurrent use of other substances (Aim 1); to examine associations between etiological risk factors (e.g., use patterns, normative perceptions, motivations, expectancies, openness, consequences, and other substance use) as they relate to microdosing frequency and consequences (Aim 2); and to provide preliminary psychometric validation of measures of microdosing expectancies and motives (Aim 3). A purposive sample of survey respondents with variability in constructs of interest to understanding microdosing was then invited to participate in a qualitative interview to provide a more nuanced understanding of patterns of use, perceived benefits, and perceived consequences of microdosing (Aim 4). Results indicate microdosing hallucinogens/psychedelics among college students and young adults is a variable and heterogenous behavior. Increased cannabis use is significantly associated with a slightly higher microdosing frequency, though alcohol and NMPS use are nonsignificant. Both descriptive and injunctive microdosing norms for close friends are associated with greater microdosing frequency as well as injunctive norms for close friends associated with more microdosing consequences. Greater microdosing frequency is significantly associated with more microdosing consequences, and a more positive evaluation of microdosing consequences. Neither microdosing frequency nor consequences are significantly associated with greater openness to experience. Increased insomnia symptoms, described in some qualitative interviews as a stimulant-like energy disrupting sleep, are associated with more microdosing consequences. Increased anxiety symptoms are associated with lower microdosing frequency. There is no significant association with depression symptoms and either microdosing frequency or consequences. Preliminary psychometric validation of two cognitive microdosing constructs, found neither identified factor structure for microdosing motives (Factor 1: "Fun Without Tripping", Factor 2: Perception Shift, Factor 3: Cognitive Enhancement, and Factor 4: Improve Mood/Mental Health) nor expectancies (Factor 1: Creativity and Connectedness, Factor 2: Improving Affect/Reducing Unwanted Affect, and Factor 3: Undesired Outcomes) is a strong fit. Nevertheless, microdosing motives Factors 1-3 are partially associated with microdosing frequency and consequences, yet all associations with microdosing expectancies factors are nonsignificant. In sum, among college students and young adults, findings from this exploratory study suggest microdosing hallucinogens/psychedelics is a low frequency, variable, and heterogenous behavior that may be considered as part of a larger polysubstance context. Microdosing trends in some associations with several etiologic risk factors (i.e., perceptions of social norms, motives, etc.) similar to other forms of substance use, which suggests it is not completely categorically different from other forms of substance use. Microdosing is associated with consequences, varying from its public perception as a risk-free means of enhancing health. Though this study did not recruit a clinical sample, microdosing was associated with some clinical symptoms (i.e., insomnia and anxiety). This exploration of microdosing associations and measurement psychometrics may inform future empirically supported mental health and substance use interventions for college students and young adults during a period of a rapidly shifting hallucinogen/psychedelic landscape