Artificial Intelligence and Robotics Commons^™

A Specification-Driven Framework For Self-Supervised Learning In Specialized Vision Domains, Mahmut S. Gokmen

Theses and Dissertations--Computer Science

Self-supervised learning (SSL) has emerged as a principled approach to visual representation learning that derives supervisory signal directly from unlabeled data, enabling foundation models to be trained at scale without manual annotation. Deployments in medical imaging and biometric recognition have demonstrated the potential of this paradigm, yet the assumptions that make SSL effective on natural image benchmarks fail systematically in specialized domains. Generic SSL pipelines encode a tacit assumption that the most informative correspondence is spatial proximity within a single acquisition. In specialized domains this assumption breaks at the level of the data-generating process: the signal that carries domain-specific information …

Exploiting Artificial Intelligence And Optimization For Smart Agriculture, Jackson K. Butcher

Theses and Dissertations--Computer Science

Dynamic integration of Cyber-Physical Systems (CPS) and Artificial Intelligence (AI) has become a vital component for unlocking the potential of smart agriculture. Currently, limitations such as limited computational resources, poor network connectivity, and rigid treatment strategies stifle optimal agricultural outcomes. This creates a challenge of leveraging the capabilities of modern artificial intelligence to combat the natural and artificial constraints of the smart agriculture environment. The primary contribution of this thesis is the development of frameworks to alleviate the overhead data and computational demand for AI within smart agriculture settings. The first framework, iCrop+, utilizes TinyML and LoRa to guarantee high-precision …

Tiered Coalition Formation Game Variants, Stability, And Simulation, Nathan Arnold

Theses and Dissertations--Computer Science

Tiered coalition formation games (TCFGs) have been proposed for modeling the ordering of power in intransitive structures. Furthering our understanding of the usefulness of this concept requires a close examination of this game and its variants, as well as the delineation between stability concepts and methods of finding stable outcomes. Derived from a simulation of the performance of characters in the games Pokémon Red and Blue Versions, we present an approximation of its power structure found via machine learning. We compare our findings to the community consensus ranking presented on a fan-run website, and further comment on the stability of …

Information-Theoretic Methods For Efficient Training And Robust Evaluations In Self-Supervised Learning, Oscar Skean

Theses and Dissertations--Computer Science

Self-supervised learning (SSL) has become a cornerstone of modern machine learning, offering a scalable alternative to costly human annotation by constructing pretext tasks directly from raw data. While SSL has delivered strong results across vision, language, and multimodal domains, two major limitations persist: (1) SSL methods are often significantly slower to train than supervised counterparts, and (2) evaluation protocols remain narrow, with most studies relying on linear probing accuracy on the pretraining dataset. . These challenges are particularly acute for large language models (LLMs), where training costs and interpretability of intermediate representations are critical concerns.

In this work, we propose …

Flexible Attenuation Fields: Tomographic Reconstruction From Heterogeneous Datasets, Clifford S. Parker

Theses and Dissertations--Computer Science

Traditional reconstruction methods for X-ray computed tomography (CT) are highly constrained in the variety of input datasets they admit. Many of the imaging settings -- the incident energy, field-of-view, effective resolution -- remain fixed across projection images, and the only real variance is in the detector's position and orientation with respect to the scene. In contrast, methods for 3D reconstruction of natural scenes are extremely flexible to the geometric and photometric properties of the input datasets, readily accepting and benefiting from images captured under varying lighting conditions, with different cameras, and at disparate points in time and space. Extending CT …

Language Models For Rare Disease Information Extraction: Empirical Insights And Model Comparisons, Shashank Gupta

Theses and Dissertations--Computer Science

End-to-end relation extraction (E2ERE) is a crucial task in natural language processing (NLP) that involves identifying and classifying semantic relationships between entities in text. This thesis compares three paradigms for end-to-end relation extraction (E2ERE) in biomedicine, focusing on rare diseases with discontinuous and nested entities. We evaluate Named Entity Recognition (NER) to Relation Extraction (RE) pipelines, sequence-to-sequence models, and generative pre-trained transformer (GPT) models using the RareDis information extraction dataset. Our findings indicate that pipeline models are the most effective, followed closely by sequence-to-sequence models. GPT models, despite having eight times as many parameters, perform worse than sequence-to-sequence models and …

Tension Control And Interproximation Techniques Forshape Design And Rgb-Depth Segmentation Reconstruction And Modeling, Anastasia Kazadi

Theses and Dissertations--Computer Science

Human eyes possess remarkable capabilities to perceive and interpret a wealth of information about our environment; from discerning colors and depths to identifying object boundaries and navigating obstacles, our eyes serve as invaluable guides in our daily lives. Ongoing research in the fields of computer vision and computer graphics continuously explore the ways to replicate extraordinary human vision abilities in order to develop systems and frameworks which would enable computers to capture, analyze, and act upon discerned information. In this context, this dissertation seeks to investigate and automate various shape control and data processing techniques for 3D modeling and shape …

Finding Hierarchies To Improve Learning In Hierarchical Reinforcement Learning, Roy Mobley

Theses and Dissertations--Computer Science

Reinforcement Learning (RL) is an approach to allowing computer agents to try and learn how to solve problems by learning what actions are best to take in a given situation. RL is effective for learning what to do in an environment, but as the problem grows larger, the amount of information needed grows exponentially, making RL less effective on complex problems. A big challenge, often called the curse of dimensionality, is that the number of states and possible number of actions in an environment can grow too large to sufficiently test every possible combination of state and action. One method …

Practical Ai Value Alignment Using Stories, Md Sultan Al Nahian

Theses and Dissertations--Computer Science

As more machine learning agents interact with humans, it is increasingly a prospect that an agent trained to perform a task optimally - using only a measure of task performance as feedback--can violate societal norms for acceptable behavior or cause harm. Consequently, it becomes necessary to prioritize task performance and ensure that AI actions do not have detrimental effects. Value alignment is a property of intelligent agents, wherein they solely pursue goals and activities that are non-harmful and beneficial to humans. Current approaches to value alignment largely depend on imitation learning or learning from demonstration methods. However, the dynamic nature …

Multi-Domain Adaptation For Image Classification, Depth Estimation, And Semantic Segmentation, Yu Zhang

Theses and Dissertations--Computer Science

The appearance of scenes may change for many reasons, including the viewpoint, the time of day, the weather, and the seasons. Traditionally, deep neural networks are trained and evaluated using images from the same scene and domain to avoid the domain gap. Recent advances in domain adaptation have led to a new type of method that bridges such domain gaps and learns from multiple domains.

This dissertation proposes methods for multi-domain adaptation for various computer vision tasks, including image classification, depth estimation, and semantic segmentation. The first work focuses on semi-supervised domain adaptation. I address this semi-supervised setting and propose …

Machine Learning Framework For Real-World Electronic Health Records Regarding Missingness, Interpretability, And Fairness, Jing Lucas Liu

Theses and Dissertations--Computer Science

Machine learning (ML) and deep learning (DL) techniques have shown promising results in healthcare applications using Electronic Health Records (EHRs) data. However, their adoption in real-world healthcare settings is hindered by three major challenges. Firstly, real-world EHR data typically contains numerous missing values. Secondly, traditional ML/DL models are typically considered black-boxes, whereas interpretability is required for real-world healthcare applications. Finally, differences in data distributions may lead to unfairness and performance disparities, particularly in subpopulations.

This dissertation proposes methods to address missing data, interpretability, and fairness issues. The first work proposes an ensemble prediction framework for EHR data with large missing …

The Basil Technique: Bias Adaptive Statistical Inference Learning Agents For Learning From Human Feedback, Jonathan Indigo Watson

Theses and Dissertations--Computer Science

We introduce a novel approach for learning behaviors using human-provided feedback that is subject to systematic bias. Our method, known as BASIL, models the feedback signal as a combination of a heuristic evaluation of an action's utility and a probabilistically-drawn bias value, characterized by unknown parameters. We present both the general framework for our technique and specific algorithms for biases drawn from a normal distribution. We evaluate our approach across various environments and tasks, comparing it to interactive and non-interactive machine learning methods, including deep learning techniques, using human trainers and a synthetic oracle with feedback distorted to varying degrees. …

Structured Attention For Image Analysis, Xin Xing

Theses and Dissertations--Computer Science

Attention mechanism, an approach to maintain the local and global features over the input, is the crucial element of the Transformer. This dissertation explores structured attention for image analysis, proposing attention-based methods for multi-label learning and Alzheimer’s Disease (AD) diagnosis.
For the multi-label learning task, I present two works under the Vision Transformer (ViT) framework. The first work focuses on supervised learning of multi-label classification. I address the problems of the multi-label classification and propose a model named AssocFormer, which adopts the association module to access the objects’ association relationship to improve the model performance. The second work addresses the …

Peer-To-Peer Energy Trading In Smart Residential Environment With User Behavioral Modeling, Ashutosh Timilsina

Theses and Dissertations--Computer Science

Electric power systems are transforming from a centralized unidirectional market to a decentralized open market. With this shift, the end-users have the possibility to actively participate in local energy exchanges, with or without the involvement of the main grid. Rapidly reducing prices for Renewable Energy Technologies (RETs), supported by their ease of installation and operation, with the facilitation of Electric Vehicles (EV) and Smart Grid (SG) technologies to make bidirectional flow of energy possible, has contributed to this changing landscape in the distribution side of the traditional power grid.

Trading energy among users in a decentralized fashion has been referred …

Learning A Scalable Algorithm For Improving Betweenness In The Lightning Network, Vincent Davis

Theses and Dissertations--Computer Science

This paper presents a scalable algorithm for solving the Maximum Betweenness Improvement Problem as it occurs in the Bitcoin Lightning Network. In this approach, each node is embedded with a feature vector whereby an Advantage Actor-Critic model identifies key nodes in the network that a joining node should open channels with to maximize its own expected routing opportunities. This model is trained using a custom built environment, lightning-gym, which can randomly generate small scale-free networks or import snapshots of the Lightning Network. After 100 training episodes on networks with 128 nodes, this A2C agent can recommend channels in the Lightning …

The Performance Optimization Of Asp Solving Based On Encoding Rewriting And Encoding Selection, Liu Liu

Theses and Dissertations--Computer Science

Answer set programming (ASP) has long been used for modeling and solving hard search problems. These problems are modeled in ASP as encodings, a collection of rules that declaratively describe the logic of the problem without explicitly listing how to solve it. It is common that the same problem has several different but equivalent encodings in ASP. Experience shows that the performance of these ASP encodings may vary greatly from instance to instance when processed by current state-of-the-art ASP grounder/solver systems. In particular, it is rarely the case that one encoding outperforms all others. Moreover, running an ASP system on …

Don't Give Me That Story! -- A Human-Centered Framework For Usable Narrative Planning, Rachelyn Farrell

Theses and Dissertations--Computer Science

Interactive or branching stories are engaging and can be embedded into digital systems for a variety of purposes, but their size and complexity makes it difficult and time-consuming for humans to author them. Narrative planning algorithms can automatically generate large branching stories with guaranteed causal consistency, using a hand-authored library of story content pieces. The usability of such a system depends on both the quality of the narrative model upon which it is built and the ability of the user to create the story content library.

Current narrative planning algorithms use either a limited or no model of character belief, …

Supporting Stylized Language Models Using Multi-Modality Features, Chengxi Li

Theses and Dissertations--Computer Science

As AI and machine learning systems become more common in our everyday lives, there is an increased desire to construct systems that are able to seamlessly interact and communicate with humans. This typically means creating systems that are able to communicate with humans via natural language. Given the variance of natural language, this can be a very challenging task. In this thesis, I explored the topic of humanlike language generation in the context of stylized language generation. Stylized language generation involves producing some text that exhibits a specific, desired style. In this dissertation, I specifically explored the use of multi-modality …

Smart Decision-Making Via Edge Intelligence For Smart Cities, Nathaniel Hudson

Theses and Dissertations--Computer Science

Smart cities are an ambitious vision for future urban environments. The ultimate aim of smart cities is to use modern technology to optimize city resources and operations while improving overall quality-of-life of its citizens. Realizing this ambitious vision will require embracing advancements in information communication technology, data analysis, and other technologies. Because smart cities naturally produce vast amounts of data, recent artificial intelligence (AI) techniques are of interest due to their ability to transform raw data into insightful knowledge to inform decisions (e.g., using live road traffic data to control traffic lights based on current traffic conditions). However, training and …

Markov Decision Processes With Embedded Agents, Luke Harold Miles

Theses and Dissertations--Computer Science

We present Markov Decision Processes with Embedded Agents (MDPEAs), an extension of multi-agent POMDPs that allow for the modeling of environments that can change the actuators, sensors, and learning function of the agent, e.g., a household robot which could gain and lose hardware from its frame, or a sovereign software agent which could encounter viruses on computers that modify its code. We show several toy problems for which standard reinforcement-learning methods fail to converge, and give an algorithm, `just-copy-it`, which learns some of them. Unlike MDPs, MDPEAs are closed systems and hence their evolution over time can be treated as …

Multi-Stream Longitudinal Data Analysis Using Deep Learning, Sajjad Fouladvand

Theses and Dissertations--Computer Science

Longitudinal healthcare data encompasses all tasks where patients information are collected at multiple follow-up times. Analyzing this data is critical in addressing many real world problems in healthcare such as disease prediction and prevention. In this thesis, technical challenges in analyzing longitudinal administrative claims data are addressed and novel deep learning based models are proposed for multi-stream data analysis and disease prediction tasks. These algorithms and frameworks are assessed mainly on substance use disorders prediction tasks and specifically designed to tackled these disorders. Substance use disorder is a public health crisis costing the US an estimated $740 billion annually in …

Representing And Learning Preferences Over Combinatorial Domains, Michael Huelsman

Theses and Dissertations--Computer Science

Agents make decisions based on their preferences. Thus, to predict their decisions one has to learn the agent's preferences. A key step in the learning process is selecting a model to represent those preferences. We studied this problem by borrowing techniques from the algorithm selection problem to analyze preference example sets and select the most appropriate preference representation for learning. We approached this problem in multiple steps.

First, we determined which representations to consider. For this problem we developed the notion of preference representation language subsumption, which compares representations based on their expressive power. Subsumption creates a hierarchy of preference …

Personality And Emotion For Virtual Characters In Strong-Story Narrative Planning, Alireza Shirvani

Theses and Dissertations--Computer Science

Interactive virtual worlds provide an immersive and effective environment for training, education, and entertainment purposes. Virtual characters are an essential part of every interactive narrative. The interaction of rich virtual characters can produce interesting narratives and enhance user experience in virtual environments. I propose models of personality and emotion that are highly domain independent and integrate those models into multi-agent strong-story narrative planning systems. I demonstrate the value of the strong-story properties of the model by generating story conflicts intelligently. My models of emotion and personality enable the narrative generation system to create more opportunities for players to resolve conflicts …

Revisiting Absolute Pose Regression, Hunter Blanton

Theses and Dissertations--Computer Science

Images provide direct evidence for the position and orientation of the camera in space, known as camera pose. Traditionally, the problem of estimating the camera pose requires reference data for determining image correspondence and leveraging geometric relationships between features in the image. Recent advances in deep learning have led to a new class of methods that regress the pose directly from a single image.

This thesis proposes methods for absolute camera pose regression. Absolute pose regression estimates the pose of a camera from a single image as the output of a fixed computation pipeline. These methods have many practical benefits …

Multi-Modal Medical Imaging Analysis With Modern Neural Networks, Gongbo Liang

Theses and Dissertations--Computer Science

Medical imaging is an important non-invasive tool for diagnostic and treatment purposes in medical practice. However, interpreting medical images is a time consuming and challenging task. Computer-aided diagnosis (CAD) tools have been used in clinical practice to assist medical practitioners in medical imaging analysis since the 1990s. Most of the current generation of CADs are built on conventional computer vision techniques, such as manually defined feature descriptors. Deep convolutional neural networks (CNNs) provide robust end-to-end methods that can automatically learn feature representations. CNNs are a promising building block of next-generation CADs. However, applying CNNs to medical imaging analysis tasks is …

Estimating Free-Flow Speed With Lidar And Overhead Imagery, Armin Hadzic

Theses and Dissertations--Computer Science

Understanding free-flow speed is fundamental to transportation engineering in order to improve traffic flow, control, and planning. The free-flow speed of a road segment is the average speed of automobiles unaffected by traffic congestion or delay. Collecting speed data across a state is both expensive and time consuming. Some approaches have been presented to estimate speed using geometric road features for certain types of roads in limited environments. However, estimating speed at state scale for varying landscapes, environments, and road qualities has been relegated to manual engineering and expensive sensor networks. This thesis proposes an automated approach for estimating free-flow …

Deep Neural Architectures For End-To-End Relation Extraction, Tung Tran

Theses and Dissertations--Computer Science

The rapid pace of scientific and technological advancements has led to a meteoric growth in knowledge, as evidenced by a sharp increase in the number of scholarly publications in recent years. PubMed, for example, archives more than 30 million biomedical articles across various domains and covers a wide range of topics including medicine, pharmacy, biology, and healthcare. Social media and digital journalism have similarly experienced their own accelerated growth in the age of big data. Hence, there is a compelling need for ways to organize and distill the vast, fragmented body of information (often unstructured in the form of natural …

Temporal Data Extraction And Query System For Epilepsy Signal Analysis, Yan Huang

Theses and Dissertations--Computer Science

The 2016 Epilepsy Innovation Institute (Ei2) community survey reported that unpredictability is the most challenging aspect of seizure management. Effective and precise detection, prediction, and localization of epileptic seizures is a fundamental computational challenge. Utilizing epilepsy data from multiple epilepsy monitoring units can enhance the quantity and diversity of datasets, which can lead to more robust epilepsy data analysis tools. The contributions of this dissertation are two-fold. One is the implementation of a temporal query for epilepsy data; the other is the machine learning approach for seizure detection, seizure prediction, and seizure localization. The three key components of our temporal …

Image-Based Roadway Assessment Using Convolutional Neural Networks, Weilian Song

Theses and Dissertations--Computer Science

Road crashes are one of the main causes of death in the United States. To reduce the number of accidents, roadway assessment programs take a proactive approach, collecting data and identifying high-risk roads before crashes occur. However, the cost of data acquisition and manual annotation has restricted the effect of these programs. In this thesis, we propose methods to automate the task of roadway safety assessment using deep learning. Specifically, we trained convolutional neural networks on publicly available roadway images to predict safety-related metrics: the star rating score and free-flow speed. Inference speeds for our methods are mere milliseconds, enabling …

A Machine Learning Approach To Artificial Floorplan Generation, Genghis Goodman

Theses and Dissertations--Computer Science

The process of designing a floorplan is highly iterative and requires extensive human labor. Currently, there are a number of computer programs that aid humans in floorplan design. These programs, however, are limited in their inability to fully automate the creative process. Such automation would allow a professional to quickly generate many possible floorplan solutions, greatly expediting the process. However, automating this creative process is very difficult because of the many implicit and explicit rules a model must learn in order create viable floorplans. In this paper, we propose a method of floorplan generation using two machine learning models: a …