80264 research outputs found
Sort by
The Gentrification Balancing Act: Exploring the Nexus of Stigma, Gentrification, and Media Discourse in Third Ward
No full textHouston’s Third Ward is a historically Black neighborhood that has been undergoing gentrification for the past few decades. As a stigmatized place, Third Ward has been represented negatively in the media in prior decades. Now that the neighborhood is experiencing increased investment and development, what does that mean for its stigmatized reputation? As a neighborhood gentrifies and its demographic composition changes, neighborhood discourse changes to reflect a new neighborhood identity. The objective of this study is to understand the association of discourse change and neighborhood demographic change in Third Ward, uncovering how media discourse is involved in this process. Using structural topic modeling, I analyzed discourse changes in Houston Chronicle articles that mention Third Ward between 1985 and 2024. I find that most of the changes in discourse occur before gentrification. Further, discourse change does not align with neighborhood demographic change in expected ways. This study considers the role that discourse plays in how neighborhoods are both symbolically and physically constructed
Reward Design with Program Graphs for Reinforcement Learning Guided Training of Large Language Models for Program Synthesis
No full textIn recent years, Large Language Models (LLMs) have increasingly been utilized to solve the problem of automatic code generation, also known as program synthesis. Frequently termed Code LLMs, these models continue to make headlines as new model architectures, training techniques and benchmark datasets are explored by academic institutions and commercial entities all around the world. For all its strengths and impact, most Code LLMs share one similarity that also happens to be a limitation: the models are trained using standard supervised learning objectives such as next token prediction (NTP). In effect, the models are trained on code data as if it were a natural language text thereby ignoring unique properties of code such as syntax and semantics that have the potential to serve as rich signal to the LLM. To address this limitation, training frameworks that incorporate alternative techniques such as reinforcement learning (RL) have been proposed. However, the introduction of RL to this set up brings an additional challenge: the task of designing a reward function for the Code LLM.
This thesis looks at reward design within the context of RL for Code LLMs. We hypothesize that a better reward signal for Code LLMs undergoing RL-based training will result in a better resulting model, as measured by performance on downstream code generation tasks. To test this hypothesis, we design a model, also a Code LLM, that is able to perform a deep semantic analysis on code in order to assign scores to programs. These scores serve as a measure of code quality, specifically the syntactic and semantic correctness of the generated code in a given context. Consistent with the terminology introduced in an earlier work, we call this model a \emph{discriminator} as its capabilities allow it to distinguish between human-written and machine-generated code. First, we design and build a discriminator and analyze its performance on a standard benchmark dataset. In contrast to existing versions of the discriminator, our proposed framework incorporates signal from the code text as well as the corresponding code graphs, including data flow graphs (DFG) and control flow graphs (CFG). We find that our proposed model is able to significantly outperform existing baselines in the task of distinguishing between human-written and machine-generated programs. Next, we deploy this enhanced discriminator within the context of RL-based training of Code LLMs. We perform a comprehensive analysis of the performance of classic Code LLMs trained using NTP objectives and how these compare against Code LLMs trained using RL, using both the existing discriminator as well as our novel graph-based discriminator. Through these experiments, we explore the role of reward functions in influencing the RL training of Code LLMs and the potential of deploying RL-based techniques within the space of LLMs for code
Persistent tailoring of MSC activation through genetic priming
No full textMesenchymal stem/stromal cells (MSCs) are an attractive platform for cell therapy due to their safety profile and unique ability to secrete broad arrays of immunomodulatory and regenerative molecules. Yet, MSCs are well known to require preconditioning or priming to boost their therapeutic efficacy. Current priming methods offer limited control over MSC activation, yield transient effects, and often induce expression of pro-inflammatory effectors that can potentiate immunogenicity. Here, we describe a ‘genetic priming’ method that can both selectively and sustainably boost MSC potency via the controlled expression of the inflammatory-stimulus-responsive transcription factor IRF1 (interferon response factor 1). MSCs engineered to hyper-express IRF1 recapitulate many core responses that are accessed by biochemical priming using the proinflammatory cytokine interferon-γ (IFNγ). This includes the upregulation of anti-inflammatory effector molecules and the potentiation of MSC capacities to suppress T cell activation. However, we show that IRF1-mediated genetic priming is much more persistent than biochemical priming and can circumvent IFNγ-dependent expression of immunogenic MHC class II molecules. Together, the ability to sustainably activate and selectively tailor MSC priming responses creates the possibility of programming MSC activation more comprehensively for therapeutic applications
2024 Storm Impacts in Houston and Harris County: A Descriptive Overview
No full textThis study looks at Houston residents' experiences both immediately and in the aftermath of the May derecho and Hurricane Beryl. It also examines the cumulative impact of the weather events, focusing on the proportion of residents affected by more than one storm and how these experiences may have overlapped
A kinetic Monte Carlo simulation of solid-electrolyte interphase formation and dendrite growth during electroplating
No full textThe formation of 3D structures such as dendrite, filament, and moss during electroplating is an obstacle to the development of a number of battery systems vital to a sustainable future, particularly lithium metal batteries. The morphological evolution of lithium metal electrode is strongly affected by the presence of passivating species formed by electrolyte decomposition, known as solid-electrolyte interphase (SEI). A 2D kinetic Monte Carlo (kMC) algorithm on a hexagonal grid was developed to account for the competing effects of deposition, diffusion, and surface passivation, providing an elementary understanding of electrodeposition systems with passivation. Growth from flat electrode and from hemispherical nucleus were both investigated. Morphological information and shape statistics were found to be strongly controlled by both SEI initiation time and current density, and a phase map was constructed over both parameters to demonstrate the distribution of results. Spherical deposits formed at high passivation time and high flux, filaments and whiskers at low flux and high passivation time, and dendrites and mosses at high flux and low passivation time. SEI formation is also observed to exacerbate nascent diffusion instabilities on pristine electrode. In the limit of no cross-SEI diffusion, we obtain a scaling relation of filament lengthscale as flux^(0.35) t_pass^(2.17). When cross-SEI diffusion was considered, a contrast is observed between low and high flux regimes: traditionally, thickness decreases with current, but at high fluxes after SEI cracking, we observe that growth from a single active tip can sustain larger thicknesses with larger flux, shedding light on an SEI-free ultrahigh flux regime. These findings provide foundational yet novel understanding of complex SEI phenomena, potentially streamlining the design of next-generation batteries
LLMs One-shot Learning from Human Demonstration on Inductive Reasoning Tasks
No full textLarge Language Models (LLMs) have shown impressive proficiency across a range of natural language tasks. However, recent research has highlighted their limitations in inductive reasoning, a key aspect of human cognitive ability. While chain-of-thought demonstration have improved LLM performance on various tasks, there has been limited exploration into their effectiveness for inductive reasoning tasks specifically. Since inductive reasoning rules can vary widely, providing a tailored demonstration for every question is impractical. This study aims to investigate how LLMs perform with minimal demonstration and whether they can generalize in different regimes. To support this research, we design a programmable dataset with inputs of varying lengths and complexities to test LLMs' generalizability. Our findings suggest that a single human demonstration can enable LLMs to achieve perfect in-distribution generalization performance in some tasks. LLMs sometimes exhibit comparable or even better out-of-distribution generalization performance relative to their in-distribution performance
Reconceptualizing the Role of L1 in Second Language Pedagogy
No full textThis reflective report aims to reimagine the role of the first language (L1) in the second language (L2) classroom by challenging the prevalent monolingual approach in second language pedagogy. Drawing from personal teaching experiences and recent developments in applied linguistics, I argue for a more nuanced understanding of the L1's potential in the L2 classroom. Following a brief description of the historical context in which the monolingual approach gained prominence, I juxtapose the concepts of Common Underlying Proficiency and translanguaging with the artificial limitations imposed by adhering to a strict monolingual approach. By exploring how strategic L1 use can bridge cognitive-linguistic gaps and empower learners, I propose practical strategies for incorporating L1 into the L2 classroom. This report contributes to the ongoing debate on the effectiveness of Communicative Language Teaching (CLT) and advocates for a more inclusive approach that values learners' full linguistic repertoires
Towards Trapped-Ion Quantum Simulation with Ground-State and Optical qubits
No full textTrapped ions offer a controlled and versatile platform for the simulation of spin and spin-boson quantum systems. The Coulomb interaction between ions trapped in a harmonic potential can be used to mediate interactions between effective spins in the ion chain. Generally, coherent operations on ions are performed on ground-state qubits encoded in magnetically insensitive hyperfine sublevels. In this thesis, we extend this toolbox with the aim of
simultaneously manipulating both ground-state and optical qubits, with the latter encoded in metastable optically excited states. The optical qubit provides another set of coherent and dissipative operations that can be performed on the ions, opening new avenues for quantum computing and simulation.
In our experiment, we use 171Yb+ and 172Yb+ ions to encode ground-state and optical qubits. We discuss the laser setup needed to address the narrow quadrupole transition from 2S1/2 →2 D3/2 of 3.02 Hz linewidth, and we describe spectroscopy techniques to study this
transition in both the isotopes. We describe how the low linewidth allows resolved sideband cooling on the shared motional modes in an ion chain, enabling sympathetic cooling with one ion, 172Yb+ , while simultaneously performing coherent operations on the 171Yb+ ion. Finally, we discuss the prospects of using this tool for light-shift entangling gates, showing the flexibility of this trapped-ion system for simulating coherent as well as dissipative quantum
systems
Leveraging Multipath to Increase Radar Field-of-View and Sensing Performance
No full textRadars are an indispensable sensing modality for autonomous navigation, vehicular networking and beyond, with features complementary to visible light sensing systems. Traditional radar signal processing estimates the range and radial velocity of objects in direct line-of-sight to the radar, i.e., objects directly illuminated by the radar that scatter the illumination back to the radar. However, line-of-sight signal processing limits radar performance in three ways. First, in radar systems with highly directional signal transmissions, e.g., those in the millimeter-wave and terahertz frequency bands, line-of-sight processing limits the field-of-view over which objects can be detected/sensed. Second, real-world signal propagation is rarely limited to line-of-sight propagation, and signals undergo significant multipath due to secondary reflections in the environment. Line-of-sight processing in presence of multipath results in the formation of false targets, a.k.a. ``ghosts,'' at physically incorrect locations, degrading accurate target detection and localization capabilities. Third, line-of-sight Doppler processing prevents radars from estimating the tangential velocities of moving objects, making it challenging to distinguish between objects that are stationary versus those that are moving tangentially to the radar.
This thesis tackles all three limitations by rethinking the role of multipath in radar signal processing. The three parts of this thesis demonstrate how the three limitations can be overcome by treating multipath as an opportunity to leverage - by explicitly incorporating multipath into radar signal processing - rather than as a nuisance. The first part of this thesis theoretically demonstrates that leveraging multipath for radar imaging can improve radar resolution when multipath provides new ``looks'' of the imaging scene beyond those provided by line-of-sight, effectively forming a multi-``look'' synthetic aperture without requiring any physical aperture extension. The second and third parts of this thesis translate this theoretical idea into practice. The second part of this thesis utilizes the additional ``looks" provided by multipath to sense beyond-field-of-view objects that are imperceptible with line-of-sight processing, e.g., objects behind the radar or around-corners, without having to contend with the problem of multipath ``ghosts''. The final part of this thesis in turn uses multipath from static features in the environment (building pillars, walls, etc.), that may be known a-priori or estimated via beyond-field-of-view processing, to estimate both the tangential and radial velocities of line-of-sight moving objects.
Overall, this thesis advocates for novel modeling and signal processing approaches to improve and unlock new sensing capabilities with existing radar systems. The methods proposed in this thesis are implementation-agnostic and are compatible with existing radar sensing and communication pipelines across different waveform choices and frequency bands. Hence, the results presented in this thesis are applicable to multiple use-cases, such as autonomous navigation, vehicular networking, emergency services, spatial computing, joint radar sensing and cellular communication, etc
Hydrologic Controls on Riverine Fluxes of Dissolved Inorganic Carbon
No full textRivers serve as an aquatic nexus for the exchange of carbon (C) between the atmosphere, biosphere, geosphere, and the oceans. The transfer of dissolved inorganic carbon (DIC) from terrestrial landscapes to the oceans, carried by rivers, is thought to play a critical control on the atmospheric CO2 concentrations through time. Most rivers are fed by water that has infiltrated through the subsurface through various flowpaths and acquired DIC by several chemical reactions. In this dissertation, I explore how riverine DIC fluxes are affected by the spatial configuration of hydrologic flowpaths and the travel time of water along these flowpaths.
Tropical watersheds are expected to be hotspots of chemical weathering because of elevated temperatures and rainfall rates. However, weathering rates are limited when rainfall takes shallow flowpaths that bypass weatherable minerals in the deeper subsurface. We predict that this process may serve as a “speed limit" on chemical weathering in the tropics: in a warmer and wetter world, weathering rates would not increase substantially because more water is routed along shallow flowpaths. The precise mechanism for limitation of weathering is difficult to assess because shallow flowpaths are also fast flowpaths. That is, weathering may be limited because there are no weatherable minerals along that flowpath, or that there is not enough time for the weathering reaction to proceed substantially. We compute the transit time distribution for one tropical watershed and calibrate a model to describe the progress of a weathering reaction over time. We simultaneously invoke mineral availability and kinetic limitation to explain the weathering limitation in the watershed, and additionally provide evidence for the decoupling of dissolution reactions and precipitation reactions for secondary phases.
While alkalinity generation via chemical weathering serves as a primary regulator of atmospheric CO2 over geologic time, the return of dissolved CO2 to the atmosphere is an important, yet largely unconstrained, component of the modern C cycle. CO2 evasion from rivers is thought to be driven by turbulence, so that steep headwater streams are the biggest emitters. We explore the magnitude of CO2 evasion in the Little Deschutes River, Oregon, where steep headwater streams contribute to a gently sloping mainstem reach. When applying an isotope-enabled stream network model, we find that the mainstem reach emits more CO2 than the headwater streams. Ultimately, the flux of alkalinity through the outlet of the watershed is more substantial than the CO2 flux upstream of the outlet, and this is true for a number of other watersheds. For this reason, the riverine flux of alkalinity is especially important in regulating the C export from terrestrial ecosystems