M.S. - Computer Science

Permanent URI for this collectionhttps://hdl.handle.net/10125/20033

Browse

Now showing 1 - 20 of 48

Optimization of protocols
(University of Hawai'i at Manoa, 2025) Hernandez, Oscar Ivan; Pavlovic, Dusko; Computer Science
The development in capabilities of artificial intelligence brings the increased participation of intelligent machines in the protocols of computer networks and society, playing some roles earmarked for machines and others ripe for deception. This exacerbates existing concerns, and it introduces a new dimension to the problems of privacy \& security. Whereas a cryptographic protocol can be analyzed with formal methods in terms of the properties of traces it produces, the probabilistic protocols involving potentially deceitful AI participants are analyzed in terms of probability distributions over its traces. In contrast with formal specifications of explicit requirements, the requirements for such AI protocols are specified informally and implicitly by reward models trained on data. A mathematical model of protocol post-training is proposed in terms of an objective function defined by such rewards and regularized by statistical distances from the pre-trained behaviors. It is shown that any instance of such a protocol post-training problem admits solutions at a level of generality that does not depend on particular details of algorithms or computational paradigms, thus showing the existence of optimal behaviors that learning algorithms aim to represent in a way that applies to reinforcement learning algorithms and algorithms in any other paradigm of learning. This establishes the proposed model of protocol post-training as a general setting for reasoning about the opportunities and limitations of protocols involving AI actors.
Explainability of multi-modal machine learning and deep learning applications in health
(University of Hawai'i at Manoa, 2025) Nitta, Amanda; Washington, Peter Y.; Computer Science
Artificial Intelligence (AI) continues to be developed and has the potential to be incorporated into health, allowing for more efficient and effective diagnosis. However, many AI models operate as ”black-boxes” in how conclusions are drawn or made, leading to a lack of trust. Explainable AI (XAI) has the potential to enable clinicians and patients to understand why a model made the prediction that it made - either for model debugging or for deriving clinically useful insights. This study proposes a pose-inspired framework for autism-based video behavioral analysis, exploring feature influence scores for a Long-Term Short-Term Memory+Neural Network (LSTM+NN) applied to video data. The potential for applicability of multimodal XAI is further shown on publicly available tabular data using variations of a Random Forest model for diabetes diagnosis.
Accessible Parkinson’s disease detection from a gamified website: Deep learning using mouse trace data
(University of Hawai'i at Manoa, 2025) Shahriar Zawad, Md Rahat; Washington, Peter; Computer Science
Parkinson’s disease (PD) is a progressive neurodegenerative disorder characterized by motor symptoms such as tremor, bradykinesia, rigidity, and postural instability, which emerge only after substantial dopaminergic neuron loss. Early detection is critical for timely intervention, yet current clinical assessments and patient-reported scales are subjective and resource-intensive. To overcome these barriers, we developed a remote platform for structured mouse-tracing data collection through gamified web-based tests that require no specialized hardware. A total of 261 participants: 73 with confirmed PD, 155 non-PD, and 33 individuals with suspected PD completed three line-tracing tasks: straight line, sine wave, and spiral wave. During each task, cursor positions were recorded every 500 ms, along with screen dimensions and an in-target boolean flag. From these data, we engineered features and generated mouse trace images. We built three classes of deep learning classifiers: (1) a feed-forward neural network for engineered features; (2) fine-tuned computer vision models; and (3) multimodal models concatenating feed-forward neural network with computer vision models. Performance was evaluated under three scenarios: (i) 5-fold cross-validation on confirmed PD vs. non-PD controls; (ii) training on confirmed PD and non-PD controls, testing on suspected PD vs. non-PD controls; and (iii) training on suspected PD and non-PD controls, testing on confirmed PD vs non-PD controls. The best-performing models were image-based DenseNet-201 model with an F1 score of 0.9027 ± 0.0332 (i), multimodal ResNet-50 with an F1 score of 0.9353 ± 0.0334 (ii), and multimodal ViT with an F1 score of 0.7619 ± 0.0535 (iii). Feature importance in the best-performing models was evaluated using Gradient Shapley Additive Explanations (GradShap). Image inputs consistently proved to be most predictive. These findings suggest that models trained on confirmed PD diagnoses hold promise for early-stage PD screening.
How the price of bitcoin affects gambling and exchange transactions
(University of Hawai'i at Manoa, 2025) Wilder, Gregory J.; Biagioni, Edoardo; Computer Science
The Bitcoin Protocol incentivizes computation power in a distributed network in order to create a reliable grantor-grantee ledger tracking ownership of Bitcoin on cryptographic wallet addresses, known as the Bitcoin blockchain. The Bitcoin blockchain creates a permanent record of transactions between wallet addresses. Wallet ownership does not typically change (the owner is the one who holds the cryptographic keys to use the Bitcoin in the wallet). Thus, once a wallet address is classified (i.e. a gambling service or an exchange), all transactions involving that wallet address can be studied and analyzed using the wallet’s classification. For this thesis, I classified 43,621,232 wallet addresses by combining four publicly-available databases containing Bitcoin wallet classifications. Using these classifications, I iteratively went through each of the 3.3 Billion transactions on the Bitcoin blockchain between blocks 0 (January 3, 2009) and 500,000 (December 12, 2017), and classified 1.07 Billion transactions (32.22%). I then aggregated the daily totals for each classification and calculated the averages in both volumes of Bitcoin and value in U.S. Dollars for each classification. Finally, I calculated the correlations between gambling and exchange transactions (including the number of transactions, the total volume in Bitcoin, the total value in U.S. Dollars the average transactions sizes in both Bitcoin and U.S. Dollars) and the price of Bitcoin in U.S. Dollars (including the closing price, and the change in price and percentage from one day prior, seven days prior, and thirty days prior). With this data, I attempt to determine the answers to three questions: 1. To what extent does the changing price of Bitcoin, in U.S. Dollars, affect the number and size of gambling transactions? 2. To what extent does the changing price of Bitcoin, in U.S. Dollars, affect the number and size of transactions sent to and from exchanges? 3. What was Bitcoin primarily used for before January 2018, and to what extent were those uses ethically questionable? I find that, until a large price spike occurred for Bitcoin starting around November 30, 2016, people tended to wager fairly constant amounts in Bitcoin rather than constant amounts of their local fiat currency (i.e. U.S. Dollars). Starting around November 30, 2016, the wagers became more closely tied to U.S. Dollars with diminishing amounts being wagered in Bitcoin as the price of Bitcoin increased. I also find the opposite to be true for exchange transactions: people tended overall to buy and sell differing amounts of Bitcoin to reflect fairly constant values in U.S. Dollars during several different time periods, albeit with the average selling transaction in U.S. Dollars slowly increasing over time as the price of Bitcoin rose over the years. Additionally, I find that people appeared to care more about the absolute value change of their investments and to care less about the percentage change of their speculative investments over time, as shown in exchange buying and selling transactions. In other words, if Bitcoin’s price in U.S. Dollars went from $100.00 to $120.00, people would buy and sell fewer times and in smaller amounts than if Bitcoin went from $1,000.00 to $1,200.00, even though the percentages were the same. This shows that people tend to view the price changes of their speculative investments in absolute terms rather than in percentages. Furthermore, I find that during a period of time when Bitcoin was dropping, gamblers increased the number of times they gambled, possibly in an attempt to offset their speculative investment losses. In contrast, exchange buying did not seem to be correlated with the price during this time period, and people tended to continue to buy Bitcoin in the same constant amounts of U.S. Dollars. Finally, I find that during the first 4 years of Bitcoin, it was heavily used for gambling and ethically questionable purposes, and that its use shifted significantly towards being treated primarily as a speculative asset at least through January 2018, but with significant continued ethically questionable uses.
Personalized Mobile Sensing for Predicting Recurrent Stress Events using Self-Supervised Pre-Training
(University of Hawai'i at Manoa, 2024) Islam, Tanvir; Washington, Peter; Computer Science
Stress is widely recognized as a major contributor to a variety of health issues. Stress prediction using biosignal data recorded by wearables is a key area of study in mobile sensing research because real-time stress prediction can enable digital interventions to immediately react at the onset of stress, helping to avoid many psychological and physiological symptoms such as heart rhythm irregularities. Electrodermal activity (EDA) is often used to measure stress. However, major challenges with the prediction of stress using machine learning include the subjectivity and sparseness of the labels, a large feature space, relatively few labels, and a complex nonlinear and subjective relationship between the features andoutcomes. To tackle these issues, I examine the use of model personalization: training a separate stress prediction model for each user. To allow the neural network to learn the temporal dynamics of each individual’s baseline biosignal patterns, thus enabling personalization with very few labels, I pre-train a 1-dimensional convolutional neural network (1D CNN) using self-supervised learning (SSL). I evaluate my method using the Wearable Stress and Affect Detection (WESAD) dataset. I fine-tune the pre-trained networks to the stress prediction task and compare against equivalent models without any self-supervised pre-training. Embeddings learned using pre-training method outperform supervised baselines with signif- icantly fewer labeled data points: the models trained with SSL require less than 30% of the labels to reach equivalent performance without personalized SSL. Apart from this single modality, to make a comparison, I have developed a multi-modal personalized stress predic- tion system using wearable biosignals (EDA), electrocardiogram (ECG), electromyography (EMG), respiration (RESP), core body temperature (TEMP), and three-axis acceleration (ACC) data using the same dataset. This study demonstrates that SSL models outperform non-SSL models while utilizing less than 5% of the annotations while using multi-modal biosignals. This personalized learning method can enable precision health systems which are tailored to each subject and require few annotations by the end user, thus allowing for the mobile sensing of increasingly complex, heterogeneous, and subjective outcomes such as stress.
AI Code Generation Tools - Experience and Perspectives of Computer Science Students
(University of Hawai'i at Manoa, 2024) Anderson, Tyler; Peruma, Anthony; Computer Science
Recent advancements in large language models (LLMs) and generative artificial intelligence (GenAI) have made significant impacts globally. These AI tools have simplified traditionally strenuous and time-consuming tasks, fostering both optimism and concern regarding their use by students in educational settings. This thesis investigates the usage and perceptions of AI code generation tools, particularly ChatGPT, among computer science (CS) students. We conducted a study involving seventy students from varied academic levels who participated in a 45-minute programming activity with the optional use of ChatGPT assistance. This was followed by an extensive online questionnaire designed to explore the AI code generation tools students use, the frequency of their use, the specific tasks for which they employ these tools, and their underlying motivations. Through comprehensive qualitative and quantitative analysis, we discovered that an overwhelming majority of students are not only familiar with ChatGPT, but also use it for over half their academic work. Students expressed a strong interest in leveraging AI tools to enhance their learning experience, citing benefits such as increased efficiency and deeper understanding of complex concepts. However, they also shared concerns similar to those of their instructors, particularly regarding over-reliance on these tools, difficulty comprehending generated code, and lacking the skills necessary for their future careers. Additionally, we observed a wide range of behaviors in how students use ChatGPT, with most not employing advanced techniques like model priming and prompt engineering strategies. Our findings highlight the need for greater awareness and training in the design and behavior of these tools to maximize the benefits of AI assistance. We discuss the implications of our findings for students, educators, and the industry, suggesting strategies to ensure that the integration of AI code generation tools in educational contexts results in a net positive impact for all stakeholders.
EXPLORING SECURITY VULNERABILITIES IN FHIR SERVER IMPLEMENTATIONS: A CASE STUDY ON IBM’S FHIR SERVER IN THE CONTEXT OF THE 21ST CENTURY CURES ACT
(University of Hawai'i at Manoa, 2024) Opie, Craig Adam; Seidel, Peter-Michael; Computer Science
The 21st Century Cures Act[1], enacted in 2016, marked a pivotal shift in healthcare technology by mandating interoperability and patient access to health data. Central to this transformation is the utilization of Application Programming Interfaces (API), which play a critical role in the seamless exchange of health information. The Interoperability and Patient Access final rule[9], stemming from this Act, delineates a clear roadmap for healthcare data standards, with a particular emphasis on Health Level 7 (HL7) Fast Healthcare Interoperability Resources (FHIR)[26] standards. This rule also introduces the consumer app API rule, designed to enhance data exchange among diverse health stakeholders. These advancements are instrumental in fostering interoperability and actively engaging patients in their healthcare journey. This thesis examines the pressing need for robust security measures in the rapid implementation of FHIR servers, highlighted by the Act’s urgent compliance deadlines which may have inadvertently led to potential security compromises, with a particular emphasis on International Business Machines’ (IBM) FHIR Server. This research is anchored on three pivotal questions: 1. Identifying common security vulnerabilities in FHIR server implementations, specifically IBM’s FHIR Server, and understanding how these vulnerabilities vary across different deployment configurations and usage scenarios. 2. Recommending best practices for enhancing the security of IBM’s FHIR Server based on penetration testing outcomes, while addressing potential challenges in implementing these enhancements. 3. Assessing the impact of FHIR server security vulnerabilities on compliance with healthcare regulations such as Health Insurance Portability and Accountability Act (HIPAA) and General Data Protection Regulation (GDPR), and evaluating the role of penetration testing in ensuring regulatory compliance. This investigation employs empirical security assessments to explore the vulnerabilities inherent in current FHIR server deployments and proposes a series of best practices to mitigate these issues. The findings highlight the critical need for incorporating robust security measures at the early stages of FHIR server implementation to safeguard patient data and comply with legal standards. By detailing the vulnerabilities and offering mitigation strategies, this thesis contributes to the ongoing discussion on securing digital health infrastructures and underscores the importance of rigorous security practices in the rapidly evolving healthcare technology landscape.
DOCUSAGE: HARNESSING HIERARCHICAL CLUSTERING IN SALIENCE-DRIVEN NARRATIVE SYNTHESIS
(University of Hawai'i at Manoa, 2024) Sadmanee, Akib; Belcaid, Mahdi; Computer Science
Text summarization remains a crucial yet challenging task in natural language processing, especially as the volume of text data grows exponentially. This thesis introduces Sumsage, a new optimization-based text summarization method that synthesizes concise yet informative summaries. Our work presents several notable contributions to the field. We developed the Syn-D-sum dataset from the CNN/DailyMail dataset, creating a robust resource for training and evaluating summarization models. We also propose the Sumsage algorithm, which leverages hierarchical clustering to extract key sentences and construct coherent summaries, closely emulating human summarizers. Additionally, we designed two new evaluation methods: the Symphony penalty and the Captured Importance Quantification scores, which assess the quality of generated summaries by considering both narrative structure and sentence order. Sumsage’s dynamic tree structure and hierarchical clustering approach enable efficient and scalable summarization while maintaining contextual relevance and minimizing hallucination. Additionally, our experiments show that Sumsage yields superior performance over GPT-3.5-turbo, generating summaries similar to those written by humans and capturing more essential information. Sumsage represents a novel advancement in text summarization, offering a robust and interpretable method for generating high-quality summaries. This approach not only addresses current challenges but also lays the foundation for future innovations in narrative synthesis and evaluation.
USING GENERATIVE MODELS TO CREATE SYNTHETIC MEDICAL IMAGING DATA TO BOOST MODEL PERFORMANCE ON SPARSE DEMOGRAPHIC GROUPS
(University of Hawai'i at Manoa, 2024) Fitzpatrick, Ryan; Washington, Peter Y.; Computer Science
Recent advancements in deep generative models, specifically Generative Adversarial Networks (GANs) and Diffusion Models, have introduced new methods for generating synthetic medical imaging data. These models can produce highly realistic images, particularly for underrepresented demographic groups, thereby enhancing the diversity of existing medical datasets without incur- ring significant data collection costs. This research utilizes GANs and Diffusion Models to generate synthetic medical images, aiming to address biases in diagnostic models towards majority pop- ulations. The study assesses the impact of these augmented datasets on the performance of a conventional Convolutional Neural Network (CNN) by comparing its classification accuracy on the original imbalanced dataset with that on a dataset enhanced with synthetic images. Specifically, the performance metrics focus on the accuracy of classifying conditions such as edema, no find- ings, and pneumonia across combined minority groups versus each majority demographic group. The results from experimentation demonstrated that CNNs trained on the supplemented datasets did not achieve a higher Area Under the Receiver Operating Characteristic (AUROC) curve score compared to those trained on the original dataset. We fail to reject the null hypothesis and can- not say that the CNNs trained on the supplemented data have a significant difference in AUROC performance.
An Exploratory Study on Students' Perceptions and Experiences of Assertion Messages in Test Cases
(University of Hawai'i at Manoa, 2024) Takebayashi, Taryn; Peruma, Anthony; Computer Science
Unit testing, or the process of evaluating code at its smallest functional level, is an essential step in ensuring the overall quality of software projects. Core to unit testing are assertion statements developers use to verify that an expected condition has been met (i.e., check a unit of code for correctness.) JUnit, one of the most popular unit test and assertion libraries for Java projects, offers multiple assertion method types to easily and repeatedly check the code. The reason for the failure of a test can be found in the error log report within the integrated development environment (IDE) or from the command line. Furthermore, developers can add their own custom error messages to explain a more detailed cause for the failure of the test. While there is prior research on unit testing and assertions, there is a lack of research on custom assertion error messages.This study investigates the impact of assertion messages on university students' abilities to troubleshoot bugs in common Java string utility methods. Eighty-seven students participated in the study and were randomly assigned to one of three study groups: Group A (no messages), Group B (detailed messages of failure), and Group C (messages that were intended to be less descriptive than those of Group B). The study found that the presence of messages had little impact on the students' abilities to fix the code. Most students could fix all code errors, regardless of academic standing. Students with at least a moderate familiarity in unit testing were significantly more likely to fix more than 4 tests. These findings indicate that familiarity with unit testing had some impact on students' ability to solve the errors in the code. Students were able to assess the quality of messages and students noticed a difference in the quality of the messages between Groups B and C. This study reports features and potential patterns in messages perceived by students to be more or less informative. This study also highlights a discrepancy between understandability and helpfulness, where students could find a message understandable but not necessarily useful in understanding the purpose of the test case.
An Investigation of Identifier Naming Practices and Challenges among University Students
(University of Hawai'i at Manoa, 2024) Huo, Timothy; Peruma, Anthony S.; Computer Science
Identifier names play a pivotal role in software maintenance as they contribute significantly to program comprehension and readability. They provide valuable context for the functionality or purpose of the program. Existing research explores different aspects of identifier names, including length, styling, and structure, and how they impact developers. Consequences of poorly named identifiers have been shown to compromise code quality and maintenance, highlighting the importance of education and identifier naming practices. However, there remains a significant gap in our understanding of naming practices among students in higher education. This study explores the practices and challenges associated with identifier naming among students. The study involved 120 student participants tasked with identifying poor-quality names in a predefined code snippet and providing alternative names. Participants also completed a detailed questionnaire about their experience with identifier naming best practices, including their academic learning in this area. The study found that student participants agree on the crucial role of identifier names in detecting defects and code readability. Most participants highlighted that instructor guidelines and assignment instructions emphasize identifier names. However, insufficient resources and a lack of feedback and communication between instructors and students may contribute to many participants (85.84\%) having only partial or no familiarity with the guidelines or recommendations for crafting high-quality identifier names. These findings either support or expand upon existing research on identifier names and provide insights into the motivations behind naming and renaming practices.
AN EXPLORATORY EXAMINATION OF SOFTWARE VULNERABILITY CLASSIFICATION USING LARGE LANGUAGE MODELS
(University of Hawai'i at Manoa, 2024) Oliveira Araujo, Ana Catarina; Peruma, Anthony; Computer Science
Software vulnerabilities are critical weaknesses that can compromise the security of a system. While current research primarily focuses on automating the classification and detection of them using a range of machine learning models, there remains a notable gap in integrating ontologies like the Vulnerability Description Ontology with Large Language Models (LLMs) for enhanced classification accuracy. Our study utilizes the National Vulnerability Database (NVD) and the National Institute of Standards and Technology’s Vulnerability Description Ontology framework to enhance the clas- sification of these vulnerabilities. The methodology involves an in-depth analysis of NVD data and an investigation of the effectiveness of various LLMs to analyze vulnerability descriptions across 27 vulnerability categories in 5 noun groups. Our findings reveal that LLMs, particularly BERT and DistilBERT, demonstrate stronger performance when compared to traditional machine learn- ing models and entropy-based methods. Moreover, while expanding the dataset aims to capture a broader range of vulnerabilities, its effectiveness varies, highlighting the crucial role of annotation quality. This research emphasizes the importance of advanced machine learning techniques and quality data annotation in optimizing vulnerability assessment processes in cybersecurity.
Early Breast Cancer Diagnosis via Breast Ultrasound and Deep Learning
(University of Hawai'i at Manoa, 2023) Bunnell, Arianna; Sadowski, Peter; Computer Science
Low- and middle-income countries, such as the U.S.-Affiliated Pacific islands, suffer from much higher advanced stage breast cancer (Stages III and IV) rates than high-income countries, especially where mammography services do not exist or have low accessibility. Examples include Palau (77% of breast cancer cases diagnosed at an advanced stage), American Samoa (72%), and the Federated States of Micronesia (82%). Portable, handheld, AI-enabled breast ultrasound devices operated by a local healthcare worker could greatly reduce advanced stage cancer rates in the U.S.-Affiliated Pacific Islands by making screenings drastically more accessible. In this work, we have explored AI models for both breast lesion detection and breast density estimation from clinical breast ultrasound. Breast density assessment and lesion detection and diagnosis were trained and evaluated on task-specific datasets collected from clinical breast imaging centers across Hawaiʻi, available through the Hawaiʻi Pacific Island Mammography Registry. The results of the breast lesion detection task show that diagnosis of breast lesions is possible on ultrasound with concurrent classification of lesion descriptors for explainability, achieving 0.39 average precision. Precise delineation and classification of breast lesions is possible with AI applied to breast ultrasound. We expect performance to increase as more data become available. The typical performance across the breast lesion detection literature for non-explainable methods is 0.7 mean average precision. The breast density model is the first application of deep learning to predicting the BI-RADS mammographic breast density category from clinical breast ultrasound (inter-modality) and achieves 0.69 mean one-vs.-rest AUROC on a held-out test set. There is signal detectable by AI which relates mammographic breast density to breast ultrasound images. Methods for intra-modality classification of mammographic breast density with deep learning achieve approximately 0.93 mean one vs. rest AUROC on an internal test set.
Enhancing Automatic Emotion Recognition for Clinical Applications: A Multimodal, Personalized Approach and Quantification of Emotional Reaction Intensity with Transformers
(University of Hawai'i at Manoa, 2023) Qian, Yang; Washington, Peter; Computer Science
In the realm of artificial intelligence, Automated Emotion Recognition (AER) has emerged as a pivotal research area, intersecting computer vision, natural language processing, and human- computer interaction. This research is particularly relevant to fields such as healthcare, education, and entertainment. This thesis is primarily concerned with enhancing Facial Expression Recogni- tion (FER), a crucial aspect of AER. In contrast to typical multimodal and Transformer models’ methodologies, this work explores the potential of personalization and the quantification of emo- tional reaction intensity in the pursuit of improving AER.This research draws upon the encouraging advancements of deep learning techniques, such as Convolutional Neural Networks (CNNs) and Attention Mechanisms, to improve FER. A com- prehensive review of the literature is presented, encompassing topics like emotion recognition, affective computing, face detection, and the role of deep learning in multimodal communication. The research methodology and experimental design, which involve the use of emotion recognition datasets and integrated network methodologies such as CNN-LSTM (Long Short-Term Memory) and CNN-Transformers, are delineated in subsequent chapters. In the methodology chapter, a suite of independent experiments is designed to probe different facets of emotion recognition. The first experiment investigates the model architecture, feature ex- traction, and data preprocessing techniques for FER. A comparative analysis is conducted among traditional CNN, transfer learning with ImageNet pre-trained models, and Vision Transformers (ViT) for FER tasks, with the goal of deciphering the causes of their performance differences. This research further investigates the potential of personalization in emotion recognition to en- hance AER performance. This is demonstrated by developing a personalized CNN-LSTM emo- tion recognition model trained on individual-specific data. Additionally, an Emotional Reaction Intensity Estimation experiment is performed, utilizing CNN-Transformer model approaches. The results reveal that a focus on personalization and quantification of emotional reaction intensity contributes to significant improvements in emotion recognition. Notably, the performance metrics recorded a marginal gain of 1.5% and 1% in accuracy and an 8% improvement in the Pearson Correlation Coefficient, respectively. These findings underscore the relevance of personalization and emotional reaction intensity quantification in AER, highlighting the necessity for more precise and robust emotion recognition systems.
Improving Efficiency in Data Wrangling With Semantic Type Detection
(University of Hawai'i at Manoa, 2023) Yu, Andy; Belcaid, Mahdi; Computer Science
This thesis presents SLED (Semantic LLM Enrichment of Data), a Python library that leverages Large Language Models (LLMs) to automate essential tasks in data wrangling and management, with a focus on adherence to FAIR (Findable, Accessible, Interoperable, Reusable) data principles. SLED is designed to enrich data through three means: contextualization, documentation, and validation. At its core, SLED performs automatic semantic type detection to contextualize datasets, offering a deeper and more nuanced understanding of primitive data types. This contextualization is vital for facilitating data wrangling, enhancing documentation and improving data validation. To do so, SLED fine tunes an open-source LLM to perform accurate semantic type detection, using real and synthetic training data. Our fine-tuned LLM demonstrates significant improvement in semantic type classification over the base model. SLED significantly streamlines data management tasks while enriching and improving data accessibility. This aligns with the overarching goals of open science and effective data analysis. By aiming to reduce the complexities associated with data science, SLED makes data analysis more approachable and efficient for everyone. Additionally, it ensures compliance with FAIR data principles, reinforcing its commitment to open data science.
A CONCEPTUAL FRAMEWORK FOR SMARTER INTERDISCIPLINARY HYBRID COLLABORATION ON AN INFINITE CANVAS
(University of Hawai'i at Manoa, 2023) Rogers, Michael; Belcaid, Mahdi; Computer Science
In response to the heightened demand for advanced digital collaboration tools in hybrid work environments post-COVID-19, this thesis introduces “Dao,” a conceptual framework that adds to the Smart Amplified Group Environment (SAGE3) ecosystem. Dao integrates data science and artificial intelligence into SAGE3, catering to the evolving needs of the science, technology, engineering, mathematics, and related fields (STEM+) community. It enhances SAGE3’s existing functionalities with a stateful Application Programming Interface (API) and a novel 2D layout algorithm, enriching the platform’s infinite canvas and relaxed “What You See Is What I See” (WYSIWIS) paradigm for more effective spatial organization and data interaction. Dao seamlessly blends vital components such as the Foresight Engine and Seer with SAGE3, fostering an environment conducive to intricate, data-driven collaboration. This integration streamlines interdisciplinary projects, enabling efficient co-programming and dynamic content management. The Dao framework represents a significant leap in collaborative digital tools, emphasizing the role of artificial intelligence and data science in enhancing efficiency and innovation in collaborative research and development, particularly suited for hybrid workplace settings.
DEEPFAKE DETECTION GENERALIZATION VIA KNOWLEDGE DISTILLATION
(University of Hawai'i at Manoa, 2023) Flores, Cristian; Baek, Kyungim; Computer Science
Two ongoing challenges in the field of deepfake detection are a lack of generalized modelsand a loss in performance when analyzing highly compressed videos. This work attempts to address these problems, through the use of knowledge distillation (KD) using heterogeneous teachers. In a typical class setting, a student learns from several teachers, each an expert in their domain. This process is analogous to the use of KD with heterogeneous teachers. A contribution of this work is the creation of a KD pipeline that can effectively utilize the knowledge of heterogeneous teachers to train a student model. This pipeline introduces the Winner-takes-all method for utilizing the knowledge of all teachers during training time. The three primary goals of this work are to show that a student model trained utilizing the proposed KD pipeline is a generalized learner, demonstrate that knowledge from its teachers was retained, and maintain a sufficiently high level of performance despite using a shallower, more compressed model compared to its teacher(s). The results indicate that all three goals were met, with the benefits of using the KD pipeline varying from marginal to moderate.
Analysis Of Twitter Data For Public Health Surveillance And Precision Diagnostics Of Autism Spectrum Disorder
(University of Hawaii at Manoa, 2023) Jaiswal, Aditi; Washington, Peter; Computer Science
The healthcare industry is a prolific source of data, with every patient record, clinical trial, drug test, and medical research generating copious amounts of information. Consequently, the interest in using machine learning algorithms in healthcare applications has increased dramatically, with numerous breakthroughs being made. One such application is using social media to study and understand public health. With millions of users sharing their thoughts, exchanging ideas, and providing health-related information on various social media platforms, researchers and clinicians can conduct studies on diseases and associated symptoms in natural settings by establishing digital phenotypic biomarkers. Twitter is one such platform that has proven to be an exceptional source of health-related information from both public and health officials. In this study, we aim to mine data related to "autism" from " #ActuallyAutistic" tweets on Twitter. The textual differences in social media communications can help identify various behavioral symptoms, which can be used to distinguish an autistic individual from their typical peers. We were able to scrape a total of 6,469,994 tweets from approximately 70,000 individual users. We illustrate the usefulness of the dataset through simple applications such as: sentiment analysis, text classification and topic modeling. Our classifier achieved an accuracy of 73%, which is consistent with previous studies published in Nature Digital Medicine journal where using different modality data such as eye gazing and facial expressions, the authors achieved a similar accuracy. The collected data will also be released publicly to help accelerate scientific research in this field. This sharing of data and interdisciplinary research can enhance its analytical capacity and enable medical practitioners to make more informed decisions.
Machine Learning Based Statistical Downscaling for Rainfall on Hawaiian Islands
(University of Hawaii at Manoa, 2022) Hatanaka, Yusuke Matthew; Sadowski, Peter; Computer Science
Long-term rainfall prediction on Hawaiian islands in the scale of up to decades is a crucial task for water resource management. The current physics based climate models only produce coarse outputs, which are not suitable to the islands due to high rainfall gradient. Statistical downscaling is a method of learning a model to perform super-resolution on weather and climate variables; predicting local weather and climate from coarse resolution variables. This project focuses on rainfall data, and aims at building a framework for statistical downscaling using historical reanalysis data in coarse resolution. Statistical downscaling is typically done using linear regression models. Here we test the use of machine learning methods such as decision trees and neural networks, which are underutilized for this application. Given a set of coarse inputs, non-linear machine learning models are trained to make rainfall predictions. In this study, we compare machine learning methods for statistical downscaling on a large historical dataset for Hawaiʻi's rainfall. In Chapter 2, the dataset used for this project is explained. In Chapter 3, explanations on each method are provided. Chapter 4 iterates the result on feature selection and experiment on site-specific models. It also has a followup on the site-specific experiment, where the effect of sample size on machine learning methods is examined. Our results show that neural networks are able to improve upon linear regression prediction. However, while this is true in aggregate, there are some cases where linear regression is superior to neural networks, typically when there is not much data. Overall, this project provides a demonstration of the capabilities and limitations of non-linear machine learning methods, establishing the initial milestone on improvement on statistical downscaling research to follow.
Articulate+: An Always Listening Interface for Creating Data Visualizations
(University of Hawaii at Manoa, 2022) tabalba, roderick S.; Leigh, Jason; Computer Science
Digital assistants are becoming more frequently used in our daily lives. We use Alexa to turn off the lights, Siri to play music, and Cortana to check the weather. These systems allow users to complete daily tasks, but are we utilizing these systems to their full potential? Current research and technology in Voice User Interfaces (VUIs) are constantly improving the way we interact with digital assistants, but are required to only start listening once the user starts addressing the system. This requirement potentially limits what these systems can accomplish for the user. We should be including these systems in our conversation so that they can start participating in our discussions and meetings by suggesting new ideas or proposing new topics just as a human assistant would do. This would enable digital assistants to become collaborators and co-pilots in our life, rather than merely tools. In my thesis, I will discuss my investigation into collaborative digital assistants by examining Natural Language Interfaces (NLIs) that facilitate data exploration tasks. I created Articulate+, an always listening system that generates data visualizations all through voice commands. Articulate+ presents a digital persona of itself, called Arti, designed to act as a collaborator. I developed an always listening method that leverages conversational data to generate informative data visualizations. I conducted a user study in order to evaluate Articulate+ and investigate the benefits of my always listening method. My contributions in this thesis will bring us one step closer to a future where AI will be able to participate in meetings and discussions just as a human would.

Browse

Recent Submissions