21090 research outputs found
Sort by
Addressing Domain Shifts for Computer Vision Applications via Language
No full textSemantic segmentation is used in safety-critical applications such as autonomous driving and cancer diagnosis, where accurately identifying small and rare objects is essential. However, pixel-level annotations are expensive and time-consuming, and distribution shifts (e.g. daytime to snowy weather in self-driving, color variations between tumor scans across hospitals) between datasets further degrade model generalization capabilities. Unsupervised domain adaptation for semantic segmentation (DASS) addresses this challenge by training models on labeled source distributions and adapting them to unlabeled target domains.
Existing DASS methods rely on either vision-only approaches or language-based techniques. Vision-only frameworks, such as masking and utilizing multi-resolution crops, implicitly learn spatial relationships between different image patches but often suffer from noisy pseudo-labels biased toward the source domain. To mitigate noisy predictions, language-based DASS methods leverage generalized representations from large-scale language pre-training. However, those approaches use generic class-level prompts (e.g., "a photo of a \{class\}") and fail to capture complex spatial relationships between objects, which are key for dense prediction tasks like semantic segmentation.
To address these limitations, we propose LangDA, a language-guided DASS framework that enhances spatial context-awareness by leveraging vision-language models (VLMs). LangDA generates scene-level descriptions (e.g., "a pedestrian is on the sidewalk, and the street is lined with buildings") to encode object relationships. At an image-level, LangDA aligns an image's feature representation with the corresponding scene-level text embedding, improving the model’s ability to generalize across domains. LangDA eliminates the need for cumbersome manual prompt tuning and expensive human feedback, ensuring consistency and reproducibility.
LangDA achieves state-of-the-art performance on three self-driving DASS benchmarks: Synthia to Cityscapes, Cityscapes to ACDC, and Cityscapes to DarkZurich, surpassing existing methods by 2.6\%, 1.4\%, and 3.9\%, respectively. Ablation studies confirm the effectiveness of context-aware image-level alignment over pixel-level alignment. These results demonstrate LangDA’s capability to leverage spatial relationships encoded in language to accurately segment objects under domain shift
Towards the development of an all-optical, non-contact, photon absorption remote sensing (PARS) endomicroscope for blood vasculature imaging
No full textThe need for high-resolution, label-free imaging techniques has spurred the development of advanced endoscopic technologies for real-time tissue characterization. This thesis presents the design, development, and validation of the first forward-viewing, non-contact, all-optical Photon Absorption Remote Sensing (PARS) endomicroscope for in vivo vascular imaging. The proposed system is designed to leverage the endogenous optical absorption of hemoglobin to achieve high-resolution contrast, without the use of exogenous labels or acoustic coupling, addressing longstanding limitations of conventional absorption-based and scattering-based imaging modalities.Two prototype designs were developed using image guide fiber (IGF) technology and achromatic graded-index (GRIN) lenses, with systematic de-risking experiments guiding their evolution. The first prototype (P1) achieved a resolution of ~1 µm and signal-to-noise ratio (SNR) of 22 dB, demonstrating the feasibility of high-fidelity PARS imaging within a 1.6mm outer diameter (OD) device footprint. A second design (P2) was introduced to address constraints in working distance and imaging depth for in vivo use, trading resolution for improved accessibility in biological tissues.
This work establishes a novel platform for PARS miniaturization and integration with widefield endoscopy, positioning the technology for future applications, including real-time, in situ virtual biopsies, blood oxygenation measurement, and surgical guidance within internal bodily cavities. The results represent a foundational advancement in the translation of PARS microscopy to clinical settings and lay the groundwork for real-time, high-resolution endoscopic diagnostics
Contributions to Change Point and Functional Data Analysis
No full textThe advent and progression of computers has led to consideration of data previous considered too unwieldy. So called high-dimensional, or big, data can be considered large in both the size of observations and the number of observations. In this thesis, we consider such data which may be infinite dimensional and is often collected over some dimension, such as time. Methodology for detection of changes and exploration of this information-rich data is explored.
Chapter 1 provides a review of concepts and notation used throughout the thesis. Topics related to time series, functional data, and change point analysis are of particular interest and form the foundation of the thesis. The chapter concludes with an overview of the main contributions contained in the thesis.
An empirical characteristic functional-based method for detecting distributional change points in functional time series is presented in chapter 2. Although various methods exist to detect changes in functional time series, they typically require projection or are tuned to specific changes. The characteristic functional-based approach is fully functional and sensitive to general changes in the distribution of functional time series. Integrated- and supremum-type test statistics are proposed. Theoretical considerations for the test statistics are examined, including asymptotic distributions and the measure used to integrate the test statistic over the function space. Simulation, permutation, and approximation approaches to calibrate detection thresholds for the test statistics are investigated. Comparisons to existing methods are conducted via simulation experiments. The proposed methods are applied to continuous electricity prices and high-frequency asset returns.
Chapter 3 is devoted to graph-based change point detection. Graph-based approaches provide another method for detecting distributional changes in functional time series. Four test statistics and their theoretical properties are discussed. Extensive simulations provide context for graph-based tuning parameter choices and compare the approaches to other functional change point detection methods. The efficacy of graph-based change point detection is demonstrated on multi-year pedestrian counts, high-resolution stock returns, and continuous electricity prices.
Despite increased interest in functional time series, available implementations are largely missing. Practical considerations for applying functional change point detection are covered in chapter 4. We present fChange, a functional time series package in R. The package combines and expands functional time series and change point methods into an easy-to-use format. The package provides functionality to store and process data, summarize and validate assumptions, characterize and perform inference of change points, and provide visualizations. The data are stored as discretely observed observations, promoting usability and accuracy. Applications to continuous electricity prices, cancer mortality, and long-term treasury rates are shown.
In chapter 5, we propose novel methodology for analyzing tumor microenvironments (TMEs) in cancer research. TMEs contain vast amounts of information on patient's cancer through their cellular composition and the spatial distribution of tumor cells and immune cell populations. We present an approach to explore variation in TMEs, and determine the extent to which this information can predict outcomes such as patient survival or treatment success. Our approach can identify specific interactions which are useful in such predictions. We use spatial functions to summarize interactions, and then apply a functional random forest-based model. This approach is shown to be effective in simulation experiments at identifying important spatial interactions while also controlling the false discovery rate. We use the proposed approach to interrogate two real data sets of Multiplexed Ion Beam Images of TMEs in triple negative breast cancer and lung cancer patients. The publicly available companion R package funkycells is discussed.
The random coefficient autoregressive model of order 1, RCA(1), is a model well-suited for volatile time series. Detection of changes between stable and explosive regimes of scalar data modeled with the RCA(1) is explored in chapter 6. We derive a (maximally selected) likelihood ratio statistic and show that it has power versus breaks occurring even as close as O(\log \log N) periods from the beginning/end of sample. Moreover, the use of quasi maximum likelihood-based estimates yields better power properties, with the added bonus of being nuisance-free. Our test statistic has the same distribution - of the Darling-Erd\H{o}s type - irrespective of whether the data are stationary or not, and can therefore be applied with no prior knowledge on this. Our simulations show that the test has very good power and, when applying a suitable correction to the asymptotic critical values, the correct size. We illustrate the usefulness and generality of our approach through applications to economic and epidemiological time series.
Chapter 7 provides summaries and discussions on each chapter. Directions for future work are considered. These directions, with the provided commentary, extend the scope of the models and may behoove practitioners and researchers alike
Leveraging Emerging Data Center Technologies to Build High-Performance Data Stores
No full textDistributed in-memory storage systems play a critical role in supporting modern applications and meeting their performance, reliability, and scalability requirements. Current in-memory storage systems adopt three design decisions that limit their performance and efficiency. First, these systems rely on the write-ahead log to guarantee data consistency and tolerate failures. The write-ahead log enforces a strict sequential ordering on operations that is often unnecessary for many applications, introducing a performance bottleneck. Second, these systems are designed for traditional, server-centric hardware, overlooking potential design optimization of emerging hardware capabilities and rendering them incompatible with the recently proposed hardware-disaggregated architecture. Third, their disaster recovery mechanisms are designed under the assumption of complete time asynchrony across machines, resulting in either a large data loss window or a significant performance overhead. This thesis explores a fundamentally different design space for building high-performance, replicated in-memory storage systems.
First, to address the inefficiency of the write-ahead log, this thesis explores a novel system design that forgoes the write-ahead log and builds the Logless, Linearizable Key-Value storage system (LoLKV). By removing the log, LoLKV eliminates the serialization bottleneck and unnecessary memory copy operations, achieving a higher level of concurrency and improving resource utilization. LoLKV relies on one-sided RDMA to efficiently replicate data. Evaluation results demonstrate that LoLKV achieves 1.7–10× higher throughput and 20–92% lower tail latency compared to state-of-the-art RDMA-based systems.
Second, to address the performance challenges of current storage systems on the hardware-disaggregated architecture, I propose SplitKV, a low-latency linearizable key-value store designed for the hardware-disaggregated architecture. SplitKV leverages one-sided RDMA for communication with memory nodes, ensuring that memory nodes remain completely passive. SplitKV co-designs the replication protocol with the data structures of the system to minimize the number of RDMA operations required to process client requests. Evaluation results show that SplitKV achieves 2.6–21× higher throughput and 80–89% lower latency compared to Sift, the state-of-the-art disaggregated key-value store.
Finally, to address the shortcomings of current disaster recovery mechanisms, I leverage modern data center time synchronization hardware and protocols to build Slogger, a new disaster recovery system. Slogger achieves near-zero data loss and guarantees prefix linearizability at the backup site. Slogger uses continuous asynchronous replication to minimize the overhead on the system. Slogger employs a watermark service to guarantee the linearizability of the backup site while avoiding across-shard coordination. Evaluation experiments show that Slogger reduces the data loss window by 50% compared to the incremental snapshotting approach
AdaptPrompt with Diffusion Set: A Unified Framework for Generalizable Deepfake Detection
No full textDeepfake detection focuses on identifying synthetic media. It has critical applications in cybersecurity, misinformation mitigation, digital forensics, and media authentication. Re- cent developments in deepfake detection have achieved impressive performance, leveraging deep learning models to distinguish between real and synthetic content. However, recent developments in diffusion-based generative models and publicly available tools such as Stable Diffusion and DALL-E pose immediate challenges to existing detection techniques. Diffusion models generate photorealistic and high-resolution content with less observable or detectable artifacts and are, therefore, more difficult to detect using traditional deepfake detection techniques.
In this thesis, we present a comprehensive study of existing deepfake detection tech- niques and adapting large vision-language models, i.e., CLIP, to generalizable deepfake detection. We introduce the Diffusion Set, a new dataset of 100k diffusion-generated fake images and 100k real images. Our experiments reveal that detectors trained on Diffusion Set outperform detectors trained on GAN-based datasets. To further enhance deepfake detection, we introduce a new transfer learning strategy that learns randomly initialized prompts and a lightweight adapter network while having CLIP frozen. Extensive experi- ments confirm its efficiency, and we also investigate the impact of dropping specific CLIP layers on detection accuracy.
Our study utilizes the Diffusion Set for training and evaluates models on 25 unseen test sets, covering images synthesized by GAN-based models, diffusion-based models, and commercially available tools. Beyond large-scale training, we assess model performance in few-shot settings, where models are trained with only a small fraction of the dataset (e.g., 320 real and 320 fake images), providing insights into their adaptability under data constraints.
Additionally, we extend our analysis beyond classification by exploring image attribu- tion and training models in a few-shot setting to attribute images to specific generators such as BigGAN, StarGAN, and Stable Diffusion. Our findings showcase the robustness of CLIP-based models in deepfake detection and their ability to generalize across unseen gen- erative techniques. We also investigated the feasibility of using the same transfer learning strategy for attribution, with experimental results that demonstrate its high effectiveness in closed set attribution
Dual energy CT for more accurate diagnosis and monitoring of early osteoarthritis-related shoulder injuries
No full textThe rapid acceleration of population aging has led to a growing prevalence of age-related musculoskeletal (MSK) conditions, such as osteoarthritis (OA). Dual-energy computed tomography (DECT) is an advanced imaging modality that shows promise in enhancing the diagnosis and characterization of MSK disorders by providing improved visualization of joint and tissue changes after injury. These advancements may support more effective treatment planning and better patient outcomes. The primary aim of this study was to refine input parameters used in DECT imaging and apply them to better understand the relationship between shoulder injury and early osteoarthritic changes over a six month period. This knowledge is expected to improve therapeutic outcomes and support early screening for individuals at high risk of developing OA. DECT was employed to quantify volumetric bone mineral density (vBMD) and to model bone stiffness and loading using finite element modeling (FEM) in both cadaveric specimens and participants. Three anatomical regions of the proximal humerus were assessed: the humeral shaft diaphysis, the articular surface of the humerus (humeral head), and the anatomical neck. Cadaveric scans were performed using both dipotassium phosphate (K2HPO4) and hydroxyapatite (HA) calibration phantoms while participants were scanned using only the HA phantom. Imaging was conducted using both BONE and Standard (STD) reconstruction kernels at three energy pair combinations: 40/90, 90/140, and 40/140 keV. These combinations were chosen to evaluate whether higher energy pairs could help mitigate attenuation issues commonly encountered at lower energy pair combinations. The BONE kernel was selected for its superior bone edge sharpening and contrast, whereas the STD kernel was used to enhance visualization of surrounding soft tissues. Participant imaging occurred at baseline (within six weeks of injury) and again at six-month follow-up. Results of this study demonstrated that both vBMD and FEM-derived stiffness values were significantly higher in the diaphysis when scanned using the BONE kernel at the highest energy pair combination (90/140 keV). In contrast, the anatomical neck consistently showed the lowest vBMD and stiffness values, with no significant differences in vBMD or FEM-derived stiffness values observed within the anatomical neck or humeral head regions under the same parameters. By developing patient-specific, image-based computational models, this study contributes to a deeper understanding of both biomechanical and imaging characteristics of early shoulder OA, potentially informing future diagnostic and therapeutic strategies
Monitoring Ultrafast Lattice Dynamics in 2D NbTe2
No full textThe discovery and control of emergent phenomena in strongly-correlated materials is a
cornerstone of modern condensed matter physics and materials science. Among these
phenomena, charge density waves (CDWs) represent a striking example of how the coupling
between electrons and the atomic lattice can give rise to new properties. Understanding the
microscopic mechanisms behind CDW formation and their dynamical evolution is crucial
not only for fundamental science, but also for the development of ultrafast, energy-efficient
electronic and quantum devices. The idea behind controlling such phenomena has been
propelled by the advent of ultrafast lasers which enables investigation of electron-lattice
interactions and has lead to the realization of many phase transitions.
In this thesis, the ultrafast lattice dynamics of the layered quasi-two-dimensional material
niobium ditelluride (NbTe2) are explored, a system known to host a robust CDW phase.
By employing both time-resolved transient reflectivity (TR) and ultrafast electron diffrac-
tion (UED), the femtosecond response is revealed from two different perspectives. These
techniques enable direct observations of the dynamical structural distortion and coherent
phonon generation with sub-picosecond temporal resolution. These findings reveal a rapid,
photoinduced suppression of the CDW order within 200 femtoseconds, followed by coherent
lattice oscillations that reflect the material’s transient structural state. UED measurements
quantify a transient 1.3% CDW order suppression, while TR data show fluence-dependent
modulations of phonon frequencies and lifetimes, highlighting the complex nature of the
lattice response. At high fluence, the CDW order of NbTe2 approaches a complete melting
along with an irreversible tellurium crystallization on the sample surface—a phenomenon
characterized by Raman spectroscopy and interpreted through density functional theory
(DFT)-based calculations.
Beyond characterizing the behavior of NbTe2, this thesis establishes a broader experimental
framework for investigating symmetry-breaking transitions and metastable states in low-
dimensional quantum materials. The work highlights the power of ultrafast techniques
for unveiling non-equilibrium phenomena and offers insights into how light can be used to
engineer and manipulate material properties on demand
Asymptotics of the number of lattice points in the transportation polytope via optimization on Lorentzian polynomials.
No full textWe formally extend the theory of polynomial capacity to power series and totally uni-
modular matrices. Using these results, we prove the log-asymptotic correctness of bounds by
Brändén, Leake, and Pak developed through the use of Lorentzian polynomials ([BLP23])
under certain conditions, and provide a counterexample where these bounds are not log-
asymptotically correct, even when symmetry exists
Changing Education One Story at a Time
No full textHigher education, traditionally founded on white epistemologies and philosophies, promotes Standard Language Ideology resulting in a linguistic hierarchy in which white English is the benchmark expected in the classroom while other varieties of English occupy a lower position in this hierarchy. The language we speak shapes how we perceive and navigate through the world since “language is a carrier of culture” (Ngugi wa Tiong’o). As such, enforcing a Standard Language Ideology mutes and undervalues the socio-cultural and linguistic traditions of Peoples of Colour and pushes them to the margins.
According to Paulo Freire, the aim of education is to free people, not enslave them. The raison d'etre driving my dissertation is to foreground rather than elide the lived experiences of BIPOC speakers and writers of World Englishes, to critique mainstream writing pedagogies that participate in that elision, and to theorise a translingual and code-meshing pedagogy that provides safe and open spaces for the identities, languages, epistemologies, and discourses of BIPOC to prevail in North American writing classrooms and writing centres. I do this by demonstrating my own indoctrination into whiteness and its effects on me as student, writing teacher and writing program administrator. I also trace my journey of decolonising of self which encompasses my ongoing efforts to foreground and amplify voices of People of Colour in education especially in first year composition demonstrating a commitment to adopting a trans- epistemic and translingual philosophy of education. I conclude with a call to all peoples of colour to start telling our complex stories to counter the single story being told in education and offer some suggestions for future opportunities and research inspired by this dissertation
Nature connection across the curriculum: Resources for post-secondary educators
No full textCopyright © 2025 by Steffanie Scott and Jenny Fu. All Rights Reserved.This resource has been designed for educators in any discipline who are looking to either begin or further explore nature connection in post-secondary environments. With contextual and academic background on related topics, the authors layer in interview testimony from real educators to showcase the diversity of approaches and experiences. The practical component of the resource outlines activity facilitation based on key educational themes related to nature. We hope this resource can support you in your journey of sharing nature with your students for learning beyond the classroom