Computer Sciences Commons^™

Probing Representational Emergence In Large Language Models, Shawn Ismail

Master's Theses

This thesis investigates whether abrupt behavioral gains in large language models under scaling are accompanied by systematic changes in internal representations. It combines a behavioral screen of 65 tasks per family with targeted layerwise probing across eight decoder-only, open-weight model families. Behavioral emergence is defined for each family-task trajectory using an empirical jump detector, with segmented regression retained only as a diagnostic. The representational follow-up analyzes 27 selected MMLU subtasks shared across all families, spanning 37 checkpoints and 216 family-task units.

For each follow-up checkpoint, frozen linear probes are trained on every layer's hidden states to measure how much task-relevant …

Deep Learning For Predicting Impact Energy And Compression After Impact Strength Of Composite Materials Using C-Scan Images, K. T. Tan, Jason P. Mack, Faizan Mirza, Zhong-Hui Duan

University Research

Traditional assessment of post-impact performance in carbon fiber reinforced polymer (CFRP) composites often relies on simplified scalar metrics that fail to capture the complex spatial interactions driving failure. This study addresses this limitation by developing an automated, end-to-end deep learning framework that shifts from manual feature extraction to the direct interpretation of raw damage morphology from ultrasonic C-scans. Using a ResNet18-based convolutional neural network (CNN) trained on 1,428 augmented images, the model achieved coefficients of determination (R2) of 0.7948 ± 0.0847 for compression after impact (CAI) strength and 0.9436 ± 0.0098 for impact energy. Beyond prediction, this dual-purpose methodology serves …

Toward Causal Generative Modeling: From Representation Learning To Controllable Generation, Aneesh Komanduri

Graduate Theses and Dissertations

The hallmark of human intelligence is causal reasoning, the ability to infer relationships between causes and effects through observation and intervention. While modern deep learning has excelled at identifying statistical patterns, current generative models often struggle to capture the underlying structural causal mechanisms of the data-generating process, leaving them vulnerable to shortcut learning and spurious associations. To achieve true generalizability and interpretability, artificial intelligence must transition from simple association to higher-level causal reasoning to be capable of scheduling and planning in the real world. This dissertation develops fundamental methodologies for causal generative modeling by integrating Pearl’s Structural Causal Model (SCM) …

The Impact Of Quantum Computing On Ransomware, Byron Denham

Graduate Theses and Dissertations

Ransomware is one of the most pervasive and dangerous threats to cybersecurity today. Attacks involving ransomware have been responsible for the disruption of critical services in many fields including healthcare, education, and government. In the current paradigm, crypto ransomware relies on asymmetric cryptography to perform its operations in a way that ensure the victim’s only chance for file recovery is by paying the attacker’s requested ransom payment. However, advancements in quantum computing threaten the asymmetric cryptography that ransomware relies on. It has been shown that, once developed, a sufficiently powerful quantum computer will have the ability to break asymmetric cryptography …

Efficient Deep Neural Networks For Autonomous Perception, Reeshad Khan

Graduate Theses and Dissertations

Autonomous perception systems must operate reliably under uncertainty arising from noisy observations, incomplete supervision, and hardware constraints. This dissertation investigates the design of efficient deep neural networks for autonomous perception through a unified perspective that treats uncertainty, efficiency, and sensing as interconnected challenges. The first contribution develops adaptive extensions of unbiased risk estimators, including eSURE and ePURE, enabling unsupervised training of deep neural networks for magnetic resonance image denoising under Gaussian and Poisson noise. However, these methods rely on known noise assumptions, which motivates the second contribution: a unified diffusion and Bayesian risk framework that estimates and adapts to unknown …

Genre Prediction Using Rnns And Llm-Enhanced Video Game Review Data, Gabriel Young

Graduate Theses and Dissertations

LLMs (Large Language Models) are powerful tools for engaging with textual data, carrying many advantages over classical NLP (Natural Language Processing) and ML (Machine Learning) approaches. However, a classical ML model can still be faster, more efficient to run, and accessible than an LLM. We seek to gain the benefits of LLM text comprehension and preserve them in a classical ML model, a hybrid approach. The LLM operates on text to surface relevant information and associations in our problem space, then the ML model trains on the LLM output. The model may learn from the LLM and provide a more …

Sel & Hil Integrations For Real-Time Cyber Threat Modeling, Risk Assessment, And Result Validation, Zachary Coel Harris

Graduate Theses and Dissertations

The electrical power grid is one of the most critical infrastructure systems, with virtually every other critical system relying on its electricity to operate effectively [1]. One failure in the electrical grid can have cascading effects, reducing, or outright stopping the operation of services vital to people’s lives. Ensuring continuous availability of the electrical grid is vital. This constant need for availability poses a challenge to identify potential cyber-attacks and their impacts. The grid is constantly growing with the addition of DERs (Distributed Energy Resources). Common DERs, such as photovoltaic systems or wind farms, can provide primary or secondary generation …

Solid Embedded Telemetry, Pranav Balasubramanian Natarajan

Graduate Theses and Dissertations

This thesis explores the integration of the Solid decentralized data framework within embedded and legacy energy control systems to enable secure, access-controlled, and interoperable data exchange. Two implementations were developed to evaluate Solid’s practical applicability: an ESP32 client implementing Solid-OIDC authentication and telemetry publication to a Solid Pod, and a Modbus-to-Solid bridge translating Modbus RTU communications into RDF-based data for bidirectional control. Experimental evaluation highlighted the effects of task scheduling and blocking operations on the ESP32, and polling-based communication latency within the Modbus bridge. These findings reveal the architectural and computational trade-offs involved in extending Solid to constrained or protocol-bound …

Spike-Sorting Algorithm For Neuropixel Probes, Luis David Davila

Open Access Theses & Dissertations

Modern extracellular neural probes like Neuropixels, allow for high-density extracellular neural probes to record terabytes of electrophysiological brain data. Spike sorting is the process of isolating individual neurons on this data through various clustering and filtering techniques. Modern spike sorting algorithms are held back by their need for human intervention to separate high-quality from low quality clusters, due to the volume of data that can be collected a completely automated approach is required. To address this requirement, we have developed a framework to preprocess, cluster, grade, and separate high/low quality clustering results using custom models trained on the grades. To …

Generation Of Elaborated, Targeted And Effective Feedback For Novice Programmers Using Llm, Hua Leong Fwa

Research Collection School Of Computing and Information Systems

Programming errors and misconceptions are pervasive in novice programmers which causes difficulty in the learning of computer programming. Large Language Models (LLMs), with their ability to comprehend and generate programming codes have shown promising results in the automatic identification of errors. This can potentially benefit student programmers by providing them with timely formative feedback at efficiencies and scale that were not attainable previously. In this study, we leveraged an LLM - OpenAI o4-mini for the generation of elaborated, targeted feedback for novice programmers across PHP and JavaScript exercises. We contend that the feedback needs to be effective and targeted other …

Selective Concolic Testing, Guofeng Zhang, Zhenbang Chen, Ziqi Shuai, Jun Sun, Weijiang Hong, Yufeng Zhang, Ji Wang, Yang Liu

Research Collection School Of Computing and Information Systems

The principled combination of symbolic execution and random testing lacks a formal foundation, especially in deciding which inputs to symbolize. We propose selective concolic testing, a cost-aware framework that formulates this choice as an optimized policy problem of a MDP (Markov Decision Process). We model program exploration over a finite control-flow graph, where MDP states represent covered statements, actions partition path constraints into symbolic and random fragments, rewards reflect coverage gain, and costs account for SMT solving effort and sampling inefficiency. Our framework yields the first formal characterization of selective symbolization as policy synthesis in a probabilistic system. We prove …

Robots And Ageing Population – Rehabilitation And Social Robots, Yuzhou Qian, Keng Siau

Research Collection School Of Computing and Information Systems

Population ageing is placing increasing pressure on healthcare and caregiving worldwide. Governments in different countries are looking into the use of advanced information technologies such as artificial intelligence and robotics to help in caring for ageing population. Older adults often need both physical assistance and psychological support, and robots have emerged as a promising technological approach to supplement eldercare. This paper provides an overview of the current state of robots in eldercare. It reviews the major definitions, categories, and features of robots used in healthcare and eldercare. The paper also examines the potential benefits of robotic technologies, such as reducing …

Software Integration In Personal Healthcare Devices And The Patient User Experience, Yassine Chahid, Patrick Slattery

Publications and Research

This study examines the current landscape and future direction of medical device hardware and software integration, focusing on how each contributes to patient care. It begins by analyzing hardware focused medical devices, such as implantable tools patients may rely on to assist with their condition, alongside diagnostic and monitoring equipment used to treat conditions in a variety of medical areas (e.g. cardiovascular conditions). It then evaluates how software is currently integrated through embedded systems, data processing, and user interfaces that support real time monitoring and clinical decision making, and how this impacts quality of care for the patient whilst minimizing …

Automated Arctic Coastal Boundary Extraction From Multispectral Satellite Imagery, Harshavardhini Bagavathyraj

Open Access Theses & Dissertations

Arctic coastlines are changing rapidly as permafrost thaw and coastal erosion accelerate under a warming climate, creating a need for accurate and scalable methods to monitor landscape change. Two geomorphically important indicators of this change are the shoreline, which represents the land--water boundary, and the bluff edge, which marks the transition from vegetated tundra to the coastal slope or cliff. However, delineating these features from high-resolution imagery remains challenging because Arctic scenes are spatially extensive and visually complex. In this study, we address these challenges by developing deep learning methods for automated shoreline and bluff-edge extraction from high-resolution satellite imagery …

Bayesian Deep Learning For Photovoltaic Power Forecasting: A Probabilistic Framework For Uncertainty Quantification And Grid Reliability Optimization, Pablo Abraham Bustamante

Open Access Theses & Dissertations

The global energy landscape is undergoing a profound transformation, driven by the urgent need to decarbonize power systems, enhance energy security, and meet growing electricity demands. Solar photovoltaic (PV) power has emerged as a critical component of future energy infrastructure due to its abundance, scalability, and cost-effectiveness. However, PV generation is inherently variable and weather-dependent, introducing significant uncertainty into grid operations and complicating the task of balancing supply and demand. Accurate forecasting of PV power generation-particularly on day-ahead and hour-ahead horizons-has become a strategic necessity for grid stability, economic efficiency, and environmental sustainability. PV output is influenced by numerous factors, …

Prediabetes Prediction Before Disease Onset Using Multimodal Health Data: A Machine Learning Approach, Luisa Veronica Gracia Mazuca

Open Access Theses & Dissertations

Prediabetes is a critical health condition that increases the risk of developing type 2 diabetes. The hemoglobin A1c (HbA1c) test diagnoses patients with prediabetes, but the disease has already caused metabolic alterations. Early detection is essential for timely interventions, and machine learning models offer a promising approach to identify prediabetic individuals through the analysis of biomarkers such as cytokines. We compared four classifiers-logistic regression, decision tree, random forest, and k-nearest neighbors - using cytokines (TNF-α, MCP-1, IL-1β, IL-6, IFN-γ), age, BMI, and waist-to-hip ratio (WHR). Models were evaluated using stratified 5-fold cross-validation and ROC-AUC. K-Nearest Neighbors (k-NN) achieved the highest …

Development Of A High-Throughput Framework For Studying Choice-Movement Dynamics In Freely Moving Rats Across Motivational And Pharmacological States, Atanu Giri

Open Access Theses & Dissertations

This thesis develops and applies high-throughput behavioral assays and analysis pipelines to study how external contingencies and internal state shape rodent decision-making. First, it introduces RECORD (Reward-Cost in Rodent Decision-making), a modular platform that combines 3D-printed arenas, microcontroller-based control, and closed-loop trial structure to deliver graded rewards and costs in a foraging-like environment. RECORD supports scalable data collection across multiple arenas and a software stack for parsing, databasing, and extracting spatiotemporal behavioral features and psychometric choice functions. Using this framework, the thesis quantifies how animals integrate sucrose reward magnitude with aversive light cost, revealing structured individual differences and sex-dependent movement …

Three Essays On Workplace Ethics And Ai: Conceptualizing, Scale Development, And Validation Of Self-Value Actualization, And Ai-Enabled High-Performance Work Systems, Mansura Nusrat

Open Access Theses & Dissertations

This dissertation advances understanding of workplace ethics and responsible AI integration across three interconnected essays unified by social cognitive theory. My first essay introduces self-value actualization (SVA), a novel construct defined as the dynamic, integrated psychological process through which individuals continually strive to align and fulfill their core moral values through observable professional behavior. Drawing on social cognitive theory and emotional intelligence theory, I conceptualize SVA as a higher-order self-regulatory orientation comprising three mutually reinforcing processes: value congruence, moral reflection, and adaptive morality. I developed and validated a 9-item SVA scale across four studies using expert panels, exploratory factor analysis, …

Zero-Knowledge Proofs For Verifiable Machine Unlearning, Bidur Niroula

Open Access Theses & Dissertations

Machine unlearning aims to remove the influence of deleted data from a trained machine learning model. This is important because privacy regulations such as GDPR, CCPA, and PIPEDA give individuals the right to request the deletion of their personal data. Since unlearning can be expensive, a dishonest server may skip it or return an incorrect model and this motivates verifiable machine unlearning. Existing cryptographic approaches prove that the server executed a specific training or unlearning algorithm on the committed dataset, so the proof is tied to the algorithm's trajectory. Empirical verifiers avoid this cost but are known to be circumventable …

Efficient Solvers For Phase Field Models, Raymond Obeng

Open Access Theses & Dissertations

Phase field models provide a versatile framework for describing phase transitions and pattern formation in materials, enabling the study of complex phenomena such as crystallization, grain growth, and defect dynamics. Among these models, the Phase Field Crystal (PFC) equation has gained significant attention due to its ability to capture atomic-scale structures while evolving on diffusive time scales. However, the numerical solution of the PFC equation presents substantial challenges arising from its high-order nature and the large-scale, coupled linear systems generated by discretization.

In particular, standard formulations of the discretized PFC system lead to non-symmetric and often ill-conditioned linear systems, which …

Computational Studies Of Alpha-Lactalbumin Binding To Perfluorodecanoic Acid, Randhal Smith Ramirez Orozco

Open Access Theses & Dissertations

This thesis investigates the structural and thermodynamic interactions between Perfluorodecanoic acid (PFDA) and bovine alpha-lactalbumin (ALAC), a critical milk protein essential for infant nutrition. The primary objectives were to computationally determine the most probable binding sites of PFDA to ALAC and to quantify the strength of these interactions alongside the structural changes they induce. This study addresses a critical gap in understanding how persistent environmental contaminants like PFDA are recruited and transported by nutritional proteins within the milk matrix. The methodology integrated high-throughput molecular docking with long-range, molecular dynamics (MD) simulations to provide a dynamic characterization of the ALAC-PFDA complex. …

Penalty Approach To Constrained Optimization Problems In Water Reservoir And Energy Generation Management, Edwin Horacio Trejo

Open Access Theses & Dissertations

The growing demand for resilient and sustainable energy generation has driven interest in Hybrid Floating Photovoltaic-Hydropower (HFPVH) systems. Operating these systems effectively requires making water release decisions that satisfy physical and regulatory constraints while maximizing energy production. Prior work by Vega (2024) developed a Dynamic Outlier Filter Long Short-Term Memory (DOF-LSTM) architecture for forecasting reservoir release patterns. That predictive work is valid and addresses an important component of the HFPVH decision pipeline. However, the constraint system in that work operates externally to the learning process, and the behavior of penalty-based constraint enforcement had not been studied independently in this context. …

Hdra-Fusion: Hybrid Detection With Routed Architecture For Manipulation-Aware Ai Face Forgery Detection, Omar Ebeid

Theses and Dissertations

With the rapid advancements in artificial intelligence-based image generation and manipulation tools, it is extremely difficult to detect if an image is genuine or artificially crafted. Despite extensive research in this area, existing image detection systems suffer from three major problems: suboptimal cross-dataset generalization due to shortcut learning of dataset-specific patterns, unreliable probability estimates due to domain shift, particularly in cross-manipulation evaluation settings, and an inability to detect images manipulated by multiple types of manipulations within a single detection framework. To address these limitations, we propose HDRA-Fusion (Hybrid Detection with Routed Architecture), a framework built on the conclusion that different …

Exploration Of Language Models In Automating System Design Diagrams From Natural Language Specifications, Candice Chambers

Theses and Dissertations

System design is the fundamental process of defining the architecture, components, and interactions of a system to satisfy specified requirements. In modern engineering, stakeholders typically require system engineers to produce System Design Diagrams (SDDs) to clearly visualize a system's functional and non-functional requirements. By leveraging Model-Based Systems Engineering (MBSE) modeling languages, SDDs can be used to employ formal methods for verification and validation early in the system's life cycle, long before production begins. However, manually generating these SDDs remains a persistent challenge in practice, as the process is tedious, error-prone, and highly time-consuming, often requiring deep expertise in specialized modeling …

Ai And Statistics: An Examination Of The Capabilities Of Ai Models To Perform Data Science, Erik Rosine

Undergraduate Theses and Capstone Projects

Artificial intelligence has become one of the fastest growing modern technologies, with publicly available AI chatbots becoming particularly popular since the 2022 launch of OpenAI’s ChatGPT. Due to their seemingly endless applications and unrivalled ease of use, these AI models present a new and often highly efficient method of approaching problems and increasing efficiency. However, with so many marketed models available, it is challenging to know which is the best one to use. This experiment tests five of the most popular AI models at performing data science tasks with given datasets, a common application of AI. These models are OpenAI’s …

Evaluating Modern Neural Network Architectures For Suicide Prediction, Kyle Brown

Master's Theses

Suicide remains a leading cause of death among adolescents despite more access to healthcare information than ever before. Medical professionals struggle to make accurate diagnoses and catch warning signs with the overwhelming amount of data available. Machine learning algorithms, including neural networks, have previously been employed for this task, yet it remains an understudied domain.

This research aims to evaluate the capabilities of Multi-Layer Perceptron (MLP) and a selection of its successors, ResNet and MLP with a category embedding layer, at the task of predicting suicidal ideation among high-school students. This research finds ResNet to be the most capable at …

Vision-Language Guided Quadruped Navigation With Reinforcement Learning Control, Chen Yuan Wang

Theses and Dissertations

Large Language Models and their multimodal variants, Vision-Language Models (VLMs), can interpret natural language instructions and visual context. Reinforcement Learning (RL) methods such as Proximal Policy Optimization (PPO) have been shown to successfully learn stable quadruped locomotion. This work investigates whether a VLM integrated with an RL controller can control a quadruped robot to perform instruction-following navigation to a predefined target using text instructions, egocentric vision, and proprioceptive state.

Most VLM-based robotics work directly target prediction of robot actions from multimodal inputs. This approach often depends on large amounts of expert demonstration data, generated by controllers such as trained RL …

Accurate Polyp Segmentation With Visual Mamba, Diego Adame

Theses and Dissertations

Medical image segmentation is a fundamental task in computer-aided diagnosis because it enables precise delineation of anatomical structures and pathological regions from clinical images. While convolutional neural networks and Transformer-based models have achieved strong performance on many medical segmentation benchmarks, CNNs struggle to model long-range dependencies and Transformers often incur high computational complexity for high-resolution medical images. Recently, state space models have emerged as an efficient alternative for dense prediction tasks due to their ability to capture long-range dependencies with linear complexity. However, existing Mamba-based segmentation models still rely on pixel-wise raster scanning that disrupts spatial locality and use simple …

Ai Institute Summer Camp Webinar Summary, Dora Nguyen

Paul English Applied Artificial Intelligence (AI) Institute Publications

This report summarizes the AI Institute Summer Camp informational webinar hosted by the Paul English Applied Artificial Intelligence Institute (PEAAII) at the University of Massachusetts Boston. The webinar introduced the structure, curriculum, learning objectives, and student outcomes associated with the 2026 AI Institute Summer Camp. In addition to summarizing the webinar content, this report analyzes participant questions, engagement trends, and areas of audience interest. Findings indicate strong interest in coding accessibility, mentorship opportunities, student outcomes, research experiences, and program flexibility. The report also identifies opportunities for improving future webinar delivery, including expanded eligibility guidance, pre-camp learning resources, enhanced presentation of …

Learning Global Context For Sparse Activity Recognition In Lengthy Recordings With Limited Dataset Size, Zeyu Tang

All Dissertations

This dissertation describes methods to analyze lengthy recordings of data in order to detect sparsely occurring activities. The narrative below describes the progression of research that led to the development of these methods and their generalization into a unified framework. My research started with designing models for dietary monitoring, including detecting meals from day-long recordings and detecting intake gestures from meal-length recordings. Both tasks share some common characteristics: (a) the target event takes only a small portion of data recordings, and (b) there is global context within full-length data recordings that can help a model make better decisions. After finishing …