I2S Masters/ Doctoral Theses

Fine-tuning large language models (LLMs) for domain-specific use cases, such as question answering, offers valuable insights into how their performance can be tailored to specialized information needs. In this project, we focused on the University of Kansas (KU) as our target domain. We began by scraping structured and unstructured content from official KU webpages, covering a wide array of student-facing topics including campus resources, academic policies, and support services. From this content, we generated a diverse set of question-answer pairs to form a high-quality training dataset. LLaMA 3.2 was then fine-tuned on this dataset to improve its ability to answer KU-specific queries with greater relevance and accuracy. Our evaluation revealed mixed results—while the fine-tuned model outperformed the base model on most domain-specific questions, the original model still had an edge in handling ambiguous or out-of-scope prompts. These findings highlight the strengths and limitations of domain-specific fine-tuning, and provide practical takeaways for customizing LLMs for real-world QA applications.

The pursuit of autonomy in cyber-physical systems (CPS) presents a challenging task of real-time interaction with the physical world, prompting extensive research in this domain. Recent advances in artificial intelligence (AI), particularly the introduction of deep neural networks (DNN), have significantly improved the autonomy of CPS, notably by boosting perception capabilities.

CPS perception aims to discern, classify, and track objects of interest in the operational environment, a task that is considerably challenging for computers in a three-dimensional (3D) space. For this task, the use of LiDAR sensors and processing their readings with DNNs has become popular because of their excellent performance. However, in CPS such as self-driving cars and drones, object detection must be not only accurate but also timely, posing a challenge due to the high computational demand of LiDAR object detection DNNs. Satisfying this demand is particularly challenging for on-board computational platforms due to size, weight, and power constraints. Therefore, a trade-off between accuracy and latency must be made to ensure that both requirements are satisfied. Importantly, the required trade-off is operational environment dependent and should be weighted more on accuracy or latency dynamically at runtime. However, LiDAR object detection DNNs cannot dynamically reduce their execution time by compromising accuracy (i.e. anytime computing). Prior research aimed at anytime computing for object detection DNNs using camera images is not applicable to LiDAR-based detection due to architectural differences. This thesis addresses these challenges by proposing three novel techniques: Anytime-LiDAR, which enables early termination with reasonable accuracy; VALO (Versatile Anytime LiDAR Object Detection), which implements deadline-aware input data scheduling; and MURAL (Multi-Resolution Anytime Framework for LiDAR Object Detection), which introduces dynamic resolution scaling. Together, these innovations enable LiDAR-based object detection DNNs to make effective trade-offs between latency and accuracy under varying operational conditions, advancing the practical deployment of LiDAR object detection DNNs.

The absence of a single, comprehensive database or repository cataloging all bioinformatics databases and software creates a significant barrier for researchers aiming to construct computational workflows. These workflows, which often integrate 10–15 specialized tools for tasks such as sequence alignment, variant calling, functional annotation, and data visualization, require researchers to explore diverse scientific literature to identify relevant resources. This process demands substantial expertise to evaluate the suitability of each tool for specific biological analyses, alongside considerable time to understand their applicability, compatibility, and implementation within a cohesive pipeline. The lack of a central, updated source leads to inefficiencies and the risk of using outdated tools, which can affect research quality and reproducibility. Consequently, there is a critical need for an automated, accurate tool to identify bioinformatics databases and software mentions directly from scientific texts, streamlining workflow development and enhancing research productivity. 

The bioNerDS system, a prior effort to address this challenge, uses a rule-based named entity recognition (NER) approach, achieving an F1 score of 63% on an evaluation set of 25 articles from BMC Bioinformatics and PLoS Computational Biology. By integrating the same set of features such as context patterns, word characteristics and dictionary matches into a machine learning model, we developed an approach using an XGBoost classifier. This model, carefully tuned to address the extreme class imbalance inherent in NER tasks through synthetic oversampling and refined via systematic hyperparameter optimization to balance precision and recall, excels at capturing complex linguistic patterns and non-linear relationships, ensuring robust generalization. It achieves an F1 score of 82% on the same evaluation set, significantly surpassing the baseline. By combining rule-based precision with machine learning adaptability, this approach enhances accuracy, reduces ambiguities, and provides a robust tool for large-scale bioinformatics resource identification, facilitating efficient workflow construction. Furthermore, this methodology holds potential for extension to other technological domains, enabling similar resource identification in fields like data science, artificial intelligence, or computational engineering.

This project presents an automated weather classification system utilizing transfer learning with pre-trained convolutional neural networks (CNNs) such as VGG19, InceptionV3, and ResNet50. Designed to classify weather conditions—sunny, cloudy, rainy, and sunrise—from images, the system addresses the challenge of limited labeled data by applying data augmentation techniques like zoom, shear, and flip, expanding the dataset images. By fine-tuning the final layers of pre-trained models, the solution achieves high accuracy while significantly reducing training time. VGG19 was selected as the baseline model for its simplicity, strong feature extraction capabilities, and widespread applicability in transfer learning scenarios. The system was trained using the Adam optimizer and evaluated on key performance metrics including accuracy, precision, recall, and F1 score. To enhance user accessibility, a Flask-based web interface was developed, allowing real-time image uploads and instant weather classification. The results demonstrate that transfer learning, combined with robust data preprocessing and fine-tuning, can produce a lightweight and accurate weather classification tool. This project contributes toward scalable, real-time weather recognition systems that can integrate into IoT applications, smart agriculture, and environmental monitoring.

Driver inattention and human error are the primary causes of traffic crashes. However, little is known about the relationship between driver aggressiveness and safety. Although several studies that group drivers into different classes based on their driving performance have been conducted, little has been done to explore how behavioral traits are linked to driver behavior. The study aims to link different driver profiles, assessed through psychological evaluations, with their likelihood of engaging in risky driving behaviors, as measured in a driving simulation experiment. By incorporating psychological factors into machine learning algorithms, our models were able to successfully relate self-reported decision-making and personality characteristics with actual driving actions. Our results hold promise toward refining existing models of driver behavior by understanding the psychological and behavioral characteristics that influence the risk of crashes.

Artificial Intelligence (AI) is being widely used in diverse domains such as industrial automation, traffic control, precision agriculture, and smart cities for major heavy lifting in terms of data analysis and decision making. However, the AI life- cycle is a major source of greenhouse gas (GHG) emission leading to devastating environmental impact. This is due to expensive neural architecture searches, training of countless number of models per day across the world, in-field AI processing of data in billions of edge devices, and advanced security measures across the AI life cycle. Modern applications often involve multitasking, which involves performing a variety of analyzes on the same dataset. These tasks are usually executed on resource-limited edge devices, necessitating AI models that exhibit efficiency across various measures such as power consumption, frame rate, and model size. To address these challenges, we introduce a novel neural network architecture model that incorporates a layer sharing principle to optimize the power usage. We propose a novel neural architecture, Layer Shared Neural Networks that merges multiple similar AI/NN tasks together (with shared layers) towards creating a single AI/NN model with reduced energy requirements and carbon footprint. The experimental findings reveal competitive accuracy and reduced power consumption. The layer shared model significantly reduces power consumption by 50% during training and 59.10% during inference causing as much as an 84.64% and 87.10% decrease in CO2 emissions respectively. 

The accurate prediction of small molecule binding affinity and toxicity remains a central challenge in drug discovery, with significant implications for reducing development costs, improving candidate prioritization, and enhancing safety profiles. Traditional computational approaches, such as molecular docking and quantitative structure-activity relationship (QSAR) models, often rely on handcrafted features and require extensive domain knowledge, which can limit scalability and generalization to novel chemical scaffolds. Recent advances in language models (LMs), particularly those adapted to chemical representations such as SMILES (Simplified Molecular Input Line Entry System), have opened new ways for learning data-driven molecular representations that capture complex structural and functional properties. However, achieving both high binding affinity and low toxicity through a resource-efficient computational pipeline is inherently difficult due to the multi-objective nature of the task. This study presents a novel dual-paradigm approach to critical challenges in drug discovery: predicting small molecules with high binding affinity and low cardiotoxicity profiles. For binding affinity prediction, we implement a specialized graph neural network (GNN) architecture that operates directly on molecular structures represented as graphs, where atoms serve as nodes and bonds as edges. This topology-aware approach enables the model to capture complex spatial arrangements and electronic interactions critical for protein-ligand binding. For toxicity prediction, we leverage chemical language models (CLMs) fine-tuned with Low-Rank Adaptation (LoRA), allowing efficient adaptation of large pre-trained models to specialized toxicological endpoints while maintaining the generalized chemical knowledge embedded in the base model. Our hybrid methodology demonstrates significant improvements over existing computational approaches, with the GNN component achieving an average area under the ROC curve (AUROC) of 0.92 on three protein targets and the LoRA-adapted CLM reaching (AUROC) of 0.90 with 60% reduction in parameter usage in predicting cardiotoxicity. This work establishes a powerful computational framework that accelerates drug discovery by enabling both higher binding affinity and low toxicity compounds with optimized efficacy and safety profiles. 

Compiler optimizations are critical for performance and have been extensively studied, especially for C/C++ language compilers. Our overall goal in this thesis is to investigate and compare the properties and behavior of optimization passes across multiple contemporary, state-of-the-art (SOTA)  C/C++ compilers to understand if they adopt similar optimization implementation and orchestration strategies. Given the maturity of pre-existing knowledge in the field, it seems conceivable that different compiler teams will adopt consistent optimization passes, pipeline and application techniques. However, our preliminary results indicate that such expectation may be misguided. If so, then we will attempt to understand the differences, and study and quantify their impact on the performance of generated code.

In our first work, we study and compare the behavior of profile-guided optimizations (PGO) in two popular SOTA C/C++ compilers, GCC and Clang. This study reveals many interesting, and several counter-intuitive, properties about PGOs in C/C++ compilers. The behavior and benefits of PGOs also vary significantly across our selected compilers. We present our observations, along with plans to further explore these inconsistencies in this report. Likewise, we have also measured noticeable differences in the performance delivered by optimizations across our compilers. We propose to explore and understand these differences in this work. We present further details regarding our proposed directions and planned experiments in this report. We hope that this work will show and suggest opportunities for compilers to learn from each other and motivate researchers to find mechanisms to combine the benefits of multiple compilers to deliver higher overall program performance.

Stimming refers to repetitive actions or behaviors used to regulate sensory input or express feelings. Children with developmental disorders like autism (ASD) frequently perform stimming. This includes arm flapping, head banging, finger flicking, spinning, etc. This is exhibited by 80-90% of children with Autism, which is seen in 1 among 36 children in the US. Head banging is one of these self-stimulatory habits that can be harmful. If these behaviors are automatically identified and notified using live video monitoring, parents and other caregivers can better watch over and assist children with ASD.

Classifying these actions is important to recognize harmful stimming, so this study focuses on developing a deep learning-based approach for stimming action recognition. We implemented and evaluated four models leveraging three deep learning architectures based on Convolutional Neural Networks (CNNs), Autoencoders, and Vision Transformers. For the first time in this area, we use skeletal joints extracted from video sequences. Previous works relied solely on raw RGB videos, vulnerable to lighting and environmental changes. This research explores Deep Learning based skeletal action recognition and data processing techniques for a small unstructured dataset that consists of 89 home recorded videos collected from publicly available sources like YouTube. Our robust data cleaning and pre-processing techniques helped the integration of skeletal data in stimming action recognition, which performed better than state-of-the-art with a classification accuracy of up to 87%.

In addition to using traditional deep learning models like CNNs for action recognition, this study is among the first to apply data-hungry models like Vision Transformers (ViTs) and Autoencoders for stimming action recognition on the dataset. The results prove that using skeletal data reduces the processing time and significantly improves action recognition, promising a real-time approach for video monitoring applications. This research advances the development of automated systems that can assist caregivers in more efficiently tracking stimming activities.
 

Just as the internet-of-things leverages machine learning over a vast amount of data produced by an innumerable number of sensors, the Internet of Catalysis program uses similar strategies with catalysis research. One application of the Internet of Catalysis strategy is treating research papers as datapoints, rich with text, figures, and tables. Prior research within the program focused on machine learning models applied strictly over text.This project is the first step of the program in creating a machine learning model from the images of catalysis research papers. Specifically, this project creates a dataset of faradaic efficiency graphs using transfer learning from pretrained models. The project utilizes FasterRCNN_ResNet50_FPN, LayoutLMv3SequenceClassification, and computer vision techniques to recognize figures, extract all graphs, then classify the faradaic efficiency graphs.

Downstream of this project, researchers will create a graph reading model to integrate with large language models. This could potentially lead to a multimodal model capable of fully learning from images, tables, and texts of catalysis research papers. Such a model could then guide experimentation on reaction conditions, catalysts, and production.