33760 research outputs found
Sort by
Understanding and Modeling Explicit and Implicit Representations of the Visual World
No full textWhile a standard megapixel image might be worth a thousand words, it is at the same time worth more than a million pixels. Understanding the local and global structures of these pixels, and their meaning, has been a core problem since the birth of computer vision, enabling tasks like classification and image generation. Storing such massive amounts of pixels is another issue, with hundreds of million of terabytes of new data created every day motivating the development of powerful compression algorithms. In my research, I aim to tackle both of these problems by modeling the visual world with two main paradigms: training neural networks that model images explicitly with representation vectors, and training networks that model images implicitly in their own weights.
In this thesis, I first discuss my work around explicit unsupervised image representation and multimodal understanding. I describe a benchmark where I compare learned representations in terms of both their downstream task performance as well as the relationships between the embeddings themselves. I create a pipeline for generating synthetic text data to help perform better benchmarking and training of multimodal models for long video understanding.
Second, I investigate diffusion as a way to train a potential unified model for explicit image representations, one that can perform both recognition and generative tasks. I explore the capacity of pre-trained diffusion networks for recognition tasks and present a lightweight, learnable feedback mechanism to improve the performance. I then adapt this feedback mechanism for fast, higher-quality image generation.
Finally, I discuss an alternative paradigm for image understanding -- implicit neural representation. I provide an overview of this area, including my works for video compression. I also present a framework for better understanding what these models learn. I benchmark these models, distill best principles for their design, and propose more optimal designs that prioritize encoding time as well as quality-space tradeoffs. I improve hyper-networks, which predict implicit model weights from video inputs, to enable better real-time video compression
Data Set for Hygroscopicity of Isoprene-Derived Secondary Organic Aerosol Mixture Proxies: Importance of Solute Diffusion and Salting-In Effects
No full textSurface tension measurements were obtained using a pendant drop tensiometer over a length of 300 seconds, using the methodology of Fertil et al., 2025
CCN measurements were obtained using a DMT Cloud Condensation Nuclei Counter using the Scanning Mobility CCN Analysis (SMCA) method (Moore et al., 2010). Data obtained was then analyzed using the PyCAT analysis toolkit available for public use (last access: August 27, 2025; https://doi.org/10.5281/zenodo.6329787; Gohil, 2022; Gohil and Asa-Awuku 2022)Raw data (surface tension and cloud condensation nuclei counter (CCNC) measurements) files for publication "Hygroscopicity of Isoprene-Derived Secondary Organic Aerosol Mixture Proxies: The Importance of Solute Diffusion and Salting-In Effects"NSF AGS #2124489https://doi.org/10.5194/egusphere-2025-193
New problems and algorithms in combinatorial optimization
No full textCombinatorial optimization seeks optimal solutions from a finite set of objects. It is widely applied to many real-world problems, such as logistics, network design, and experimental design. Although the finite search space guarantees that an optimal solution can always be obtained by exhaustive search, for many real-world problems, this is intractable. NP-hardness is one of the key features shared by intractable combinatorial optimization problems.
This dissertation includes the study of three new NP-hard combinatorial optimization problems and/or new algorithms for them. The first problem considers data collection in a humanitarian logistics setting. We formulate a new type of vehicle routing problem, in which vehicles are tasked with data collection, and the objective function measures data quality using a nonlinear, nonseparable experimental design criterion. We create novel exact methods for this problem, and demonstrate their practical potential in a realistic case study using a state-of-the-art earthquake simulator. The second problem considers a more stylized knapsack problem in which each experiment has a known cost, and data collection is subject to a budget constraint. Novel deterministic, polynomial-time approximation algorithms are introduced, extending the classical local search algorithm while providing rigorous performance guarantees. The algorithms outperform the only existing method for this problem from both theoretical and empirical standpoints. The third problem arises from minimizing the latency of telecommunication networks, named the minimum stretch spanning tree problem. We introduce a straightforward and promising carousel greedy algorithm to tackle this challenging combinatorial optimization problem. By numerical experiments, we show that our algorithm significantly outperforms the best-known algorithms in the literature for both unweighted and weighted graphs on both constructed challenging instances and real-world instances, demonstrating superior solution quality with efficient running time
BLOOM - A REGENERATIVE MODEL FOR URBAN DEVELOPMENT
No full textThe buildings we design make up our urban ecosystems, yet it is increasingly clear that these ecosystems need to become more compatible with the natural environment. The climate crisis is evidence that our cities are not performing sustainably on the urban scale. Natural ecosystems manage to be both delicate and resilient by balancing resource allocation optimally. We can learn from the systems that have been developed for billions of years in different plant communities and natural habitats. Natural ecosystem regeneration holds the keys to understanding how our cities can learn to adapt to change. With this set of design principles borrowed from nature, we can create sustainable cities that unite communities as co-contributors to a cohesive ecosystem
Autochtonie et/ou résistance : une identité régionale pour décoloniser les voi(es)x artistiques de Mā'ohi Nui-Polynésie française (1970 à nos jours) ?
No full textCette thèse propose un état des lieux des littératures contemporaines écrites par les communautés colonisées de Mā'ohi Nui-Polynésie française de 1970 à aujourd’hui. À travers une analyse diachronique d'un large corpus d’artistes (Henri Hiro, Chantal Spitz, Flora Aurima Devatine, Jean-Marc Tera'ituatini Pambrun, Titaua Peu, Yiling Changues, etc.), trois moments se dessinent. On distingue d’abord l’émergence d’une idéologie et d’un renouveau culturel mā'ohi (1970-1980). L’expression d’autres identités — originellement autochtones (rurutu) ou allochtones (hakka) — qui contestent l’hégémonie de l’État-nation français se développe dans les années 1990-2000. Enfin, à partir de 2010, la littérature est le point de départ d’un laboratoire tant formel (multiplicité des médias, innovations artistiques) que thématique (critique de la colonisation, du système capitaliste) dans les littératures de l’extrême contemporain, nourries par le métissage en cours dans ces territoires connectés. Loin d’être quelques îlots éparpillés, ces littératures proposent, depuis plus d’un demi-siècle, des contre-narrations aux H/histoires officielles. Elles sont également pensées à travers le prisme d’une identité régionale Océanique (selon l’idée d’Epeli Hau'ofa), traduite dans des collaborations artistiques et intellectuelles inter-archipélagiques qui s’émancipent d’une tutelle occidentale et replacent l’Océanie au centre de la carte. Ce travail convoque les outils de l’analyse textuelle, les théories postcoloniales, ainsi que les épistémologies océaniennes décoloniales. Il propose également des ressources en ligne — dont une carte enrichie d’entretiens filmés avec quatre artistes de Tahiti (F. Aurima Devatine, M. Tehei'ura, R. Tepa et O. Marrec), de traductions de textes canoniques océaniens — qui présente quelques-unes de ces collaborations
COVARIABILITY OF WINTER AND SUMMER PRECIPITATION IN CENTRAL AMERICA AND MEXICO WITH TROPICAL SEA SURFACE TEMPERATURES UNDER A WARMING CLIMATE
No full textThis dissertation investigates the influence of tropical sea surface temperatures (SSTs) on precipitation variability and the regional hydrological cycle in Central America and Mexico (CAM), focusing on both winter and summer. As global climate change progresses, the understanding of how tropical SSTs affect regional precipitation patterns has become crucial for improving seasonal climate predictions and assessing future climate risks. The study examines the relationships between CAM precipitation and tropical SSTs, with particular attention to the contributions of the Pacific and Atlantic Oceans. Through a combination of historical observational data, atmospheric model simulations, and climate projections, the research identifies significant covariability between CAM precipitation and tropical SSTs during both the winter and summer seasons. In the winter, ENSO-related SSTs in the Pacific and warming trends in the Atlantic Ocean are shown to modulate precipitation patterns, with both oceans contributing to precipitation variability. In summer, tropical SST warming, particularly in the Atlantic, is found to dominate precipitation variability, while ENSO still plays a secondary role. A key feature of CAM’s seasonal cycle, the mid-summer drought (MSD), is characterized in detail, revealing its sensitivity to both ENSO events and long-term trends. Future projections from climate models suggest a warmer, drier future for the region, with reductions in precipitation, soil moisture, and runoff, exacerbating challenges related to water resources and agriculture. The dissertation concludes that tropical SSTs, especially in the Pacific and Atlantic, are critical drivers of CAM precipitation and provides valuable insights for improving seasonal climate predictions. Further research is recommended to refine regional climate models, explore long-term trends in precipitation patterns, and better understand the socio-economic implications of projected hydrological changes
THE INTOXICATED STAGE: MODERNIST THEATRES OF ADDICTION
No full textThis dissertation assesses the narrative, rhetorical, and aesthetic ways in which addiction appears in modernist theatre. It similarly harnesses post-modernist theatrical intervention tactics to both critique and expand the audience’s perception of addiction as it appears on stage. The Intoxicated Stage: Modernist Theatres of Addiction offers several theoretical lenses through which to read addiction in the context of theatre and performance. Foremost among these is an understanding of addiction as a performative, though this dissertation draws heavily from sociological reports of addiction, Critical Race Theory in the context of US laws toward drug use, and Queer Theory’s contributions to narrative aesthetics. Addiction is foremost a sociopolitical concern and audience perspectives toward addiction are likely to change across different social, political, and aesthetic landscapes. Modernist versus postmodernist theatrical motifs present a key indicator of these shifts in perspective.
Chapter One focuses on Lorca’s 1934 play Yerma to indicate how modernist theatre witnessed the invention of the addict as a newfound character trope, using Simon Stone’s 2017 adaptation to demonstrate the impact of twenty-first century gender politics and the perception of addiction as a form of consumer deviance. By detaching Long Day’s Journey into Night from Eugene O’Neill’s own biography, Chapter Two points to a crucial but conflicting dichotomy between addiction as a moral issue and addiction as a criminal concern as per my respective arguments about Harold Clurman and Robert O’Hara’s stagings of O’Neill’s play in 1965 and 2022. Chapter Three strives to unpack the naturalization of addiction with a focus on aesthetics in Bertolt Brecht’s 1918 play Baal and considering how this character evolves into a society that is paradoxically more “addictified” yet open to new ways of approaching substance use disorders. Chapter Four pinpoints a historical paradox related to addiction at the cusp of postmodernism, noting that The Living Theatre’s 1959 production The Connection was as ironically rooted in commercialization as it was to artistic radicalism
PHOTONIC RESERVOIR COMPUTING BASED ON OPTOELECTRONIC OSCILLATORS
No full textOptoelectronic Oscillator (OEO) is a device widely used in modern electronic and photonic systems. Its applications include but not limited to chaos communication systems, random number generation, chaotic radar and lidar, ultra-pure microwave generation, and sensing systems. Moreover, it has rich nonlinear dynamics suitable for studying phenomena like period-one (P1) oscillation, period doubling (P2), and chaos.Reservoir computing (RC) is a machine learning (ML) algorithm that breaks the traditional von Neumann computational paradigm. It is able to learn directly from data and perform both the regression and the classification tasks. Comparing to other machine learning algorithms like recurrent neural network (RNN), it has a much simpler structure that allows it to avoid computationally expensive back -propagation (BP) algorithm while maintains a competitive performance.
Recently, as an intrinsic time-delay system, optoelectronic oscillator has been introduced as a machine learning platform to perform reservoir computing. Its ability to accept both the radio-frequency (RF) signal and the optical signal make it an ideal machine learning platform for optical fiber communication systems and radio-frequency communication systems.
In Ch. 1, we start the thesis with a general review of the history of artificial intelligence (AI) and the recent development of microwave photonics. Then, we discuss the most recent advances that use photonic devices for machine learning (ML) - one of the most important branches in the field of photonic AI research.
In Ch. 2, we introduce different types of OEOs, namely the broadband OEO and narrowband OEO, and how to mathematically model them. Specifically, we will derive the time-delayed equations that govern their behaviors. For the narrowband OEO, we introduce the derivation of an envelope equation which is suitable for the numerical simulations.
Following the introduction of mathematical models of OEOs, we delve into our research work of the narrowband OEO-based RC, which is the first time a narrowband OEO is introduced into the field of machine learning for a time-delay reservoir computer implementation in Ch. 3. In this chapter, we develop the mathematical model required for studying the narrowband OEO-based RCs and introduce the concepts of OEO-based RCs. We also numerically simulate this narrowband OEO-based RC and demonstrate its suitability for processing IQ-modulated signals. Lastly, we train and test the narrowband OEO-based on IQ-modulation classification tasks, which reaches state-of-the-art performance with a reduced training set size.
In Ch. 4, we further extend the narrowband OEO-based RC to the field of RF fingerprinting - a technology that is widely used to identify RF transmitters. In this chapter, we thoroughly evaluate the performance of the narrowband OEO-based RC across a wide range of benchmark datasets. Our simulation results demonstrate the suitability of the narrowband OEO-based RC for RF fingerprinting. Moreover, we once again show that the narrowband OEO not only requires significantly less training data for the IQ modulation classification task, as presented in Ch. 3, but also maintains excellent performance under limited-resource conditions, extending its effectiveness to RF fingerprinting.
In Ch. 5, we further push the narrowband OEO-based RC to an extreme by further limiting its computational source and training data size. We experimentally test the narrowband OEO-based RC’s ability in such scenarios. Meanwhile, we propose a new metric named NET to measure ML platforms’ performances across different algorithms that takes into account the complexity of the algorithm, the performance, and data size required for training. Lastly, we also show that this metric is useful to quantitatively analyze the diminished returns of scaling the number of parameters in machine learning algorithms.
It is known that nearly all OEO-based RCs require an expensive and bulky external modulator, which poses a significant obstacle to their practical applications. To resolve this issue, in Ch. 6, we, for the first time, introduce the concept of the directly laser-modulated OEOs for RCs (DL-OEO based RC). In our proposed scheme, we cancel the external modulator and replace it by directly modulating the laser, which not only significantly reduces the system size, but also the total budget for implementation an OEO-based RCs. This implementation makes our DL-OEO based RC one of the simplest OEO-based RCs ever known. First, we start by deriving mathematical models for this proposed scheme. Then, we show our numerical results of using the DL-OEOs for reservoir computing. Lastly, we discuss the usefulness of the nonlinearity introduced by the DL-OEO and how it could contribute a significant improvement in performance in comparison with traditional OEO-based RCs.
In Appx. I, we revisit the concepts of time-division multiplexing, time-delay differential equation, and how to use them to numerically simulate the OEO-based RC.
In Appx. II, for the first time, we show the relationship between the number of cavity modes N_CM supported by OEO and the total number of virtual nodes N required by OEO-based RCs. We conduct extensive simulations of both types of OEO-based RCs (including narrowband and wideband configurations) across a diverse set of datasets. Our simulation shows that the total number of virtual nodes N required for sufficiently good performance of OEO-based RC can be as less as the total number of cavity modes N_CM
Measurement of Dissolved Beta Decay Sources in the LUX-ZEPLIN Experiment
No full textWeakly Interacting Massive Particles (WIMPs) are one of the most well-motivated dark matter candidates. Liquid Xe (LXe) time projection chambers (TPC) are the leading technology for probing the WIMP mass range. The LUX-ZEPLIN (LZ) experiment is a 10 tonne LXe TPC searching for the elusive WIMP dark matter. As detector masses increase to maximize sensitivity, background sources in the Xe, such as Kr and Ar, must be reduced to meet background requirements. These species are more difficult to quantify and require more stringent monitoring compared to previous generations of experiments. In this work, the techniques used to monitor Kr and Ar during commissioning and through LZ's WIMP Search 2024 science run (WS2024) are discussed. A mass spectrometry technique is used to measure ppq (g/g) (Kr/Xe) during WS2024. A complementary measurement of an excited state decay of Kr found a Kr equivalent concentration of ppq (g/g) (Kr/Xe). The measured WS2024 Ar concentration is ppb (g/g) (Ar/Xe). LZ set a world leading limit on the interaction cross section across the expected WIMP mass range, with a minimum of cm at 36 GeV/c
MULTILAYER ELECTRONIC STRUCTURES PRODUCED BY ADDITIVE CODEPOSITION OF METAL AND CERAMIC
No full textConventional manufacturing processes for electronics are wasteful and expensive. Co-depositionadditive manufacturing offers a promising approach for fabricating high-temperature electronic
circuits by integrating ceramic and metallic materials. This study focuses on the dual deposition
of alumina (Al₂O₃) and silver-tin-molybdenum (Ag-Sn-Mo) transient liquid phase sintering
(TLPS) paste to create functional structures for electronic applications. This process of co-
deposition combines conductive and insulating materials in one process resulting in an expedited
operation, with significantly reduced waste compared to regular manufacturing. Direct ink
writing has received attention for metallizing electronic substrates because the printing can be
performed on low-cost equipment. The biggest challenge in direct ink writing is in the ink itself.
Printing on electronic substrates typically requires pressure less sintering, and many available
inks and pastes produce porous structures when sintered without pressure. Additionally, the
elevated temperature results in challenges such as thermal expansion mismatch, cracking, and
shrinkage during the sintering process. These issues are exacerbated by the temperature needed
for processing the alumina paste. This research investigates the material compatibility,
deposition techniques, and processing parameters to optimize the co-deposition of these
materials. Additionally, it evaluates the impact of the sintering profile, and the paste formulation
of the TLPS on the mechanical integrity and interfacial adhesion of the layers