Computer Science, 1987-2025

Adaptive bitrate (ABR) algorithms for video streaming aim to maximize user Quality of Experience (QoE) by adjusting video quality to network conditions. Pensieve is a pioneering ABR algorithm that uses deep reinforcement learning (RL) to outperform conventional approaches [1]. However, like other deep RL policies, Pensieve may be vulnerable to adversarial perturbations in its state observations. In this paper, we investigate the robustness of Pensieve under adversarial state perturbations and propose adversarial training to harden it. We consider an adversary that injects small bounded errors into Pensieve’s input state (e.g., throughput history, buffer level, etc.), with the goal of inducing rebuffering events playback stalls) that severely degrade QoE. We develop two attack methods: one based on Bayesian Optimization (BO) to find worst-case perturbations in a black-box manner, and another based on Projected Gradient Descent (PGD) as a white-box attack using Pensieve’s policy network gradients. We then adversarially train Pensieve against these attacks to produce robust models.

We present a comprehensive evaluation using the standard Pensieve simulation environment (with the Mahimahi network trace emulator [2]) to compare the BO and PGD adversaries and the resulting robust policies. Our results show that even small input perturbations (within a maximum norm of 5–10% of feature values) can greatly increase rebuffering time for the original Pensieve (by 5–10×). The BO-based adversary is highly effective, finding perturbations that increase rebuffering by up to 20% more than PGD-based attacks, albeit with more attack queries. Adversarial training with either attack significantly improves Pensieve’s robustness: after training, rebuffering induced by attacks drops by 60–70%. The BO-adversarially-trained model is the most robust, with only a minor QoE degrading scenarios. We discuss the efficiency trade-offs between BO and PGD (BO requires fewer iterations but more simulation time per attack, while PGD is faster per attack but somewhat less optimal), and show that adversarially trained Pensieve generalizes well to unforeseen perturbations. This work demonstrates that adversarial training can substantially bolster the reliability of RL-based streaming algorithms against both malicious attacks and noiselike disturbances, paving the way for safer deployment in real-world networks.

Infinigen Physics augments the procedurally generated 3D scenes produced by the Infinigen system with material physics values. Despite recent advances in visual language models (VLMs) for robotics, physical world understanding remains a notable weak point. To address this, our approach enables the generation of photorealistic 3D scenes alongside physics-encoded ground truth renderings for latent material properties (e.g., density, thermal conductivity). This novel framework has applications in a wide range of 3D vision tasks—such as robotic manipulation, navigation, and augmented reality—that require both material and physical scene understanding, capabilities not available in current physics-vision datasets or dynamic physics-simulation engines like Omniverse and Unreal. Furthermore, we introduce an LLM-powered research agent that leverages web search to retrieve reliable sources of material data, facilitating the mapping of Infinigen shaders to corresponding physical properties. Embracing the scalability of Infinigen, our method automates the assignment of material properties for new shaders across an infinite variety of scenes. We demonstrate the utility of Infinigen Physics by constructing a benchmark dataset of 100 scenes with physics-encoded ground truths, which is used to evaluate an existing VLM on a physical understanding task. Our contributions include: • Developing an LLM-powered research agent for sourcing material properties and mapping Infinigen shaders to these values. • Integrating material-physics-encoded segmentation masks into the Infinigen rendering pipeline. • Creating a benchmark dataset of 100 scenes with physics-encoded ground truths for evaluating VLM performance on physical reasoning tasks.

Map applications, which require low-latency responses for a seamless user global experience are well-suited for cloud deployment. However, their stateful nature and the need for strong consistency present significant challenges in fully leveraging the low-latency advantages of running computations at edge locations, or data centers near users. To overcome these limitations, we introduce ReRoute, a serverless, cloud-based map application that integrates Radical, a system designed to preserve consistency guarantees from primary data centers while enabling reduced latencies at edge locations. Our evaluation demonstrates that integrating with Radical substantially lowers median end-to-end latency, particularly for clients near edge locations, highlighting its potential to enhance the performance of map applications.

Researchers in the field of linguistics use the International Phonetic Alphabet (IPA) as a universal reference to specific speech sounds in human language, which helps them communicate new findings in cross-linguistic contexts. However, the IPA chart is quite extensive, and learning the symbols on this chart may seem boring or difficult to newcomers. Existing resources for studying the IPA online usually serve as reference sheets or simple quizzes. In this paper, I describe the creation of a prototype for a game which both improves learning outcomes for linguistics learners and motivation for learning. I describe the development and architecture of the prototype, and discuss possible improvements and connections to educational research.

We construct a system for representing graphemes as strings, from which the original grapheme may be reconstructed in its entirety. The utility of the system in representing scripts accurately is demonstrated with the analysis of select graphemes from various scripts. From this annotation, we offer various measures of script complexity and distance. We annotate various writing systems and analyze their script complexities, and compare the performance of various metrics at formalizing the notion of script affinity. Having found statistically significant results in the task of graphemic classification, we conclude that the theory proposed here meaningfully captures information about writing systems.

In the fast-growing and fast-paced sport of collegiate men’s volleyball, a strong defensive system is critical to team success. Central to this defense is the positioning of backcourt players, who attempt to ”dig” the opponent’s attacks and prolong rallies by transitioning possession to their own team. While sports analytics has advanced rapidly in recent decades, men’s collegiate volleyball has seen relatively limited development in data-driven performance analysis, particularly in the area of defensive positioning. This project aims to bridge that gap by using computer vision, specifically object classification models, to detect and map defender positions from match footage. These spatial coordinates are then paired with outcome-based statistics to explore the relationship between positioning and dig success. Using homographic transformation techniques, player locations are projected onto a standardized 2D court model to enable comparison across venues. The resulting dataset is visualized through both frame-level position plots and aggregate heatmaps filtered by team, play outcome, or attack location. The data collection process achieved a usable frame conversion rate of 69.9%, indicating that the proposed methodology is viable for scalable, automated defensive analysis. This work demonstrates the potential of computer vision in volleyball analytics and provides a foundation for future research into tactical trends and optimization.

Large language models (LLMs) have demonstrated remarkable capabilities in code generation and repair, yet their performance is often constrained by the availability of high-quality training data, which is expensive to curate. We present a scalable pipeline to generate synthetic data to augment code repair datasets by using an LLM to perturb instances of correct code. Other similar methods produce unrealistic examples and rely on imprecise, human-defined bug taxonomies. Our method learns a well-defined taxonomy of bugs from existing data and creates perturbations in correct code snippets by retrieving the most relevant bug categories. We use few-shot learning with code differentials to demonstrably improve the realism and complexity of the inserted bugs. When used to train a downstream Llama-3.1-8B-Instruct model for code repair, the synthetic buggy-to-repaired code examples lead to performance that far exceeds the previously published baseline and nearly matches that of a model trained on an equally sized human dataset. Our method is accessible from a cost standpoint where a million example synthetic dataset can be created for under $1000.

What is the difference between and and but? In many sentences, they can be freely interchanged—consider that both ”the weather is sunny and cold” and ”the weather is sunny but cold” are true if and only if the weather is sunny and the weather is cold. What then causes speakers to chose and over but and vice versa? To that end, I gather data from the Corpus of Contemporary American English and investigate properties of the distributions of the two conjunctions. I find that and is more unmarked and neutral, while but is more likely to appear when greater contrasts exist between the two conjuncts themselves, or more broadly in more salient contexts. Along the way, novel analyses for the underlying syntactic structure of certain uses of but are proposed.

This study explores how participatory algorithm design can empower gig workers by introducing a configurable Hungarian matching system on Open Deli, a localized food delivery platform. It implements a web-based prototype that allowed couriers to adjust and visualize preference-based matching configurations in real time. Through live study sessions with potential couriers, we found that preference-based matching (e.g., location, food type, compensation) enhanced perceptions of fairness and alignment, even among participants without algorithmic expertise. The study revealed trade-offs between flexibility and information overload, highlighting the need for intuitive preference elicitation. Overall, our findings suggest that transparent, collaborative systems can better reflect worker values and support more equitable gig work experiences.

This study investigates national playing style identity in professional soccer by applying unsupervised machine learning techniques to match statistics from the 2018 and 2022 FIFA World Cups. Motivated by countries like Spain and Brazil with well-known, signature playing styles, we aim to explore whether other countries exhibit national playing styles in the World Cup and to what extent these styles have cultural and historical ties. Our study uses a dataset of 200 match performances from 24 countries with 21 features that represent in-depth match statistics relating to possession, passing, defensive actions, goalkeeping, and shooting from FBRef.com. We implement four variations of k-means clustering assisted by principal component analysis (PCA), t-distributed stochastic neighbor embedding (t-SNE), and uniform manifold approximation and projection (UMAP) for clustering and visualization. Our results show seven clusters in the match data, each corresponding to a well-known playing style or strategic approach. We find that countries with strong national soccer identity more frequently use one playing style while other countries vary their playing styles between matches. While we observe some correlation between chosen playing style and geopolitical factors like income, population size, and geographical region, the globalization of soccer markets appears to have diminished these effects. This study demonstrates how national playing styles can be quantitatively identified and used to understand how countries express their identity through professional soccer.

Despite the impressive capabilities of large language models, low-resource languages (LRLs) such as Swahili, Haitian Creole, and Yoruba remain significantly underserved due to a lack of training data. This thesis explores a text-only approach to addressing this gap by leveraging back-translation to generate synthetic data. Using pre-trained multilingual models like mT5, original sentences in each target language are translated into English and then back into their original form to produce varied and contextually rich text pairs. These pairs are used to fine-tune LLMs, enhancing their fluency and generalization in low-resource settings. The results show measurable improvements in output diversity and translation quality, demonstrating that synthetic data augmentation can play a key role in advancing equitable language technology.

How can data deepen our understanding of Broadway musicals, their history, their actors, and their writers? This project is a data-driven digital humanities approach to understanding the Broadway musical and its related genres on Broadway. Through data collected from the Internet Broadway Database (IBDB), this project reveals the current limitations of digital theatre records and data, which in turn demonstrate the definitional challenges of this form that make it both difficult to work with and fascinating to parse. The data are analyzed to explore the development of the musical genre, the history of screen musicals, and historical demographic and career information about Broadway musical actors and writers. These analyses employ quantitative reasoning to understand theatre narratives in terms of the writers that defined the genre, examine structural gender inequality in the theatre industry, and support and question historical accounts of the Broadway musical. Furthermore, these analyses emphasize the need for further data-driven theatre research and interdisciplinary approaches to musical theatre scholarship.

This thesis presents Remixer Reloaded, a novel digital interface designed to let listeners interact with the author’s album at a granular level. Inspired by traditional digital audio workstations (DAWs) like Logic Pro, but reimagined for non-musicians, the tool enables users to explore individual audio and MIDI tracks from the album How to Reignite Suns through a web interface. Built with React and Tone.js, the application visualizes notes, waveforms, and timing grids, offering intuitive tools like zooming, track isolation, and note labeling. Unlike standard DAWs, the project emphasizes read-only exploration over editing, centering accessibility and music education. Through user testing, the tool was refined for clarity, performance, and engagement. This project merges the author’s identities as a musician and computer scientist to create an original, listener-centered musical experience.

The introduction of the pitch clock, along with other new rules such as larger bases and pickoff limits, has altered the pace and dynamics of Major League Baseball. This thesis examines the statistical impact of the pitch clock and associated rule changes on gameplay and fan engagement. By leveraging historical and current data from Baseball Reference, Statista, and FanGraphs, I use machine learning models to analyze trends in player performance, isolating the effects of these rule changes. I also test whether a player’s natural pitch tempo before the implementation of the pitch clock had any effect on their performance after. Finally, I analyze injury statistics to see if the pitch clock has had any affect on player health. Interviews with MLB personnel further add to these findings. Ultimately, this research aims to offer insights to teams and the league on how to adapt in this new era of baseball, while also predicting the sport’s future trajectory.

PaceVision addresses a gap that runners face in accessing real-time performance metrics without disturbing running form. This project presents an augmented reality (AR) running assistant that utilizes Engo 2 AR sunglasses to display real-time pace data in the runner’s field of view. The system’s key innovation is a dynamic pacing line that provides visual feedback on the current pace in relation to the target pace, thus removing the need to glance at wrist-worn devices. PaceVision utilizes adaptive algorithms that balance responsiveness and stability in pace calculations from GPS data. Furthermore, the system includes an interval training mode that automatically manages workout segments with visual cues. Evaluation results show that the system achieved an average response time under 3 seconds for pace changes. Although traditional watch methods provide slightly better average pace accuracy, PaceVision provides a convenience factor for the runners. Across controlled one-mile trials there was a deviation of 4.4 versus 6.6 seconds; in long-distance runs 7.5 versus 8.6 seconds, and during interval workouts 6.4 versus 7.0 seconds. This research demonstrates the effectiveness of AR for enhancing a user’s running experience.

This thesis presents and evaluates RadSched, a latency-optimizing scheduler designed for stateful, per function execution in a serverless edge computing environment. In distributed systems, where data consistency and latency vary by time and location, selecting the optimal edge location for function execution becomes a complex decision. Built on top of the Radical framework, RadSched maintains records of network conditions and learns from past data consistency outcomes to automate routing function requests to the optimal edge location. The system employs an ϵ-greedy exploration strategy to adapt to shifting network conditions and data availability, thereby ensuring responsiveness. Through empirical evaluation across multiple AWS regions, this thesis demonstrates that RadSched maintains comparable median latency to baseline systems in a stable environment, though with higher tail latency – a tradeoff that allows the system to route functions to a shifting optimal edge in volatile environments. Ultimately, by abstracting edge selection away from the client, RadSched both improves performance and simplifies developer interaction with stateful server-less functions on the edge.

Superconducting high-entropy alloys (HEAs) have recently garnered significant attention across numerous fields due to their unique blend of properties such as increased mechanical strength, structural stability, and tunable electronic properties. These attractive features thus position HEAs as a strong candidate for multiple real-world applications, especially as next-generation superconducting qubit materials considering their robust performance under extreme conditions such as low temperatures and high magnetic fields. However, the creation of HEAs consists of a vast compositional space, enabling researchers to choose from a great range of elements in different proportions heated at multiple cycles. In order to navigate this complex field, this study utilizes Bayesian optimization as a data-driven strategy to expedite the process of discovering and optimizing HEAs with high superconducting performance. Due to the high cost and time often associated with carrying out experiments in laboratories, this approach of iteratively updating a probabilistic model with an initial set of training data proves to be beneficial in focusing efforts on only the most promising configurations. It is also crucial to note that this research study is the first in literature to explore and computationally optimize a novel composition of seven specific elements of Gold, Tin, Antimony, Palladium, Silver, Tellurium, and Indium. This combination of Bayesian optimization and superconducting HEAs demonstrates a dynamic convergence between machine learning and materials innovation, broadening research horizons for quantum technology and engineering.

This paper hypothesizes that machine learning models can predict future stock price movements of a company with higher accuracy if we give them access to news sentiment scores of that company’s major suppliers and customers. The paper tests this idea on 38 companies across 10 different economic sectors. The paper finds that prediction accuracy improves by as much as 6.1% with this idea, and improvements are especially consistent among companies in sectors with high economic elasticity.

Automated video editing has the potential to streamline content creation by intelligently selecting and synchronizing video clips with music. This project presents a video editing assistant that takes raw footage as input, analyzes its visual content using histogram-based scene segmentation, and applies K-Means clustering to identify the most representative clips. The system then aligns selected clips with predefined music segments, ensuring a structured and rhythmically cohesive final edit. By assuming that beat detection is handled externally, the system focuses on optimizing clip selection and sequencing, providing an efficient and adaptable approach to music-driven video editing. This research intends to contribute to the field of automated media production by enhancing creative workflow efficiency while maintaining user control over the editing process.