Digital Commons Network^™

Understanding Protein Deep Learning Models Through Explainability, Zahra Fazel

Electronic Thesis and Dissertation Repository

This thesis investigates the application of Explainable Artificial Intelligence (XAI) techniques to deep learning models in the field of protein analysis, specifically targeting protein language models and protein interaction site prediction models. Despite the increasing adoption of these sophisticated deep learning models in bioinformatics, their intrinsic complexity often results in a black-box nature, obscuring the understanding of their decision-making processes.

This research represents a thorough effort to integrate explanation methods within this context. We analyze the resulting interpretations using biological-specific statistical tests to enhance the transparency and interpretability of the models. This work evaluates the efficacy of current XAI methods …

Development And Optimization Of A 1-Dimensional Convolutional Neural Network-Based Keyword Spotting Model For Fpga Acceleration, Trysten E. Dembeck

Masters Theses

Spoken Keyword Spotting (KWS) has steadily remained one of the most studied and implemented technologies in human-facing artificially intelligent systems and has enabled them to detect specific keywords in utterances. Modern machine learning models, such as the variants of deep neural networks, have significantly improved the performance and accuracy of these systems over other rudimentary techniques. However, they often demand substantial computational resources, use large parameter spaces, and introduce latencies that limit their real-time applicability and offline use. These speed and memory requirements have become a tremendous problem where faster and more efficient KWS methods dominate and better meet industry …

Transfer Learning For Predictive Maintenance: A Case Study, Colter A. Swanson

Masters Theses

In light of recent strides in high-performance computing, the concept of transfer learning has emerged as a prominent paradigm within the realm of Artificial Intelligence and Machine Learning methodologies. Analogous to the human brain's capacity to assimilate information across related domains for pattern recognition, transfer learning has swiftly asserted its dominance, particularly in deep learning applications such as image classification and natural language processing. Despite its ascendancy in these domains, there exists a lack of comprehensive investigations in alternative domains, notably those encompassing tabular data formats. This thesis seeks to redress this gap by conducting an empirical examination of transfer …

Context In Computer Vision: A Taxonomy, Multi-Stage Integration, And A General Framework, Xuan Wang

Dissertations, Theses, and Capstone Projects

Contextual information has been widely used in many computer vision tasks, such as object detection, video action detection, image classification, etc. Recognizing a single object or action out of context could be sometimes very challenging, and context information may help improve the understanding of a scene or an event greatly. However, existing approaches design specific contextual information mechanisms for different detection tasks.

In this research, we first present a comprehensive survey of context understanding in computer vision, with a taxonomy to describe context in different types and levels. Then we proposed MultiCLU, a new multi-stage context learning and utilization framework, …

Morp: Monocular Orientation Regression Pipeline, Jacob Gunderson

Master's Theses

Orientation estimation of objects plays a pivotal role in robotics, self-driving cars, and augmented reality. Beyond mere position, accurately determining the orientation of objects is essential for constructing precise models of the physical world. While 2D object detection has made significant strides, the field of orientation estimation still faces several challenges. Our research addresses these hurdles by proposing an efficient pipeline which facilitates rapid creation of labeled training data and enables direct regression of object orientation from a single image. We start by creating a digital twin of a physical object using an iPhone, followed by generating synthetic images using …

Detection And Classification Of Diabetic Retinopathy Using Deep Learning Models, Aishat Olatunji

Electronic Theses and Dissertations

Healthcare analytics leverages extensive patient data for data-driven decision-making, enhancing patient care and results. Diabetic Retinopathy (DR), a complication of diabetes, stems from damage to the retina’s blood vessels. It can affect both type 1 and type 2 diabetes patients. Ophthalmologists employ retinal images for accurate DR diagnosis and severity assessment. Early detection is crucial for preserving vision and minimizing risks. In this context, we utilized a Kaggle dataset containing patient retinal images, employing Python’s versatile tools. Our research focuses on DR detection using deep learning techniques. We used a publicly available dataset to apply our proposed neural network and …

Evaluation Of An End-To-End Radiotherapy Treatment Planning Pipeline For Prostate Cancer, Mohammad Daniel El Basha, Court Laurence, Carlos Eduardo Cardenas, Julianne Pollard-Larkin, Steven Frank, David T. Fuentes, Falk Poenisch, Zhiqian H. Yu

Dissertations & Theses (Open Access)

Radiation treatment planning is a crucial and time-intensive process in radiation therapy. This planning involves carefully designing a treatment regimen tailored to a patient’s specific condition, including the type, location, and size of the tumor with reference to surrounding healthy tissues. For prostate cancer, this tumor may be either local, locally advanced with extracapsular involvement, or extend into the pelvic lymph node chain. Automating essential parts of this process would allow for the rapid development of effective treatment plans and better plan optimization to enhance tumor control for better outcomes.

The first objective of this work, to automate the treatment …

Modeling The Spatiotemporal Variations Of The Magnetic Field In Active Regions On The Sun Using Deep Neural Networks, Godwill Asare Mensah Mensah

Open Access Theses & Dissertations

Solar active regions are areas on the Sun's surface that have especially strong magnetic fields. Active regions are usually linked to a number of phenomena that can have serious detrimental consequences on technology and, in turn, human life. Examples of these phenomena include solar flares and coronal mass ejections, or CMEs. The precise predictionof solar flares and coronal mass ejections is still an open problem since the fundamental processes underpinning the formation and development of active regions are still not well understood. One key area of research at the intersection of solar physics and artificial intelligence is deriving insights from …

Robust And Trustworthy Deep Learning: Attacks, Defenses And Designs, Bingyin Zhao

All Dissertations

Deep neural networks (DNNs) have achieved unprecedented success in many fields. However, robustness and trustworthiness have become emerging concerns since DNNs are vulnerable to various attacks and susceptible to data distributional shifts. Attacks such as data poisoning and out-of-distribution scenarios such as natural corruption significantly undermine the performance and robustness of DNNs in model training and inference and impose uncertainty and insecurity on the deployment in real-world applications. Thus, it is crucial to investigate threats and challenges against deep neural networks, develop corresponding countermeasures, and dig into design tactics to secure their safety and reliability. The works investigated in this …

Deep Learning In Indus Valley Script Digitization, Deva Munikanta Reddy Atturu

Theses and Dissertations

This research introduces ASR-net(Ancient Script Recognition), a groundbreaking system that automatically digitizes ancient Indus seals by converting them into coded text, similar to Optical Character Recognition for modern languages. ASR-net, with an 95% success rate in identifying individual symbols, aims to address the crucial need for automated techniques in deciphering the enigmatic Indus script. Initially Yolov3 is utilized to create the bounding boxes around each graphemes present in the Indus Valley Seal. In addition to that we created M-net(Mahadevan) model to encode the graphemes. Beyond digitization, the paper proposes a new research challenge called the Motif Identification Problem (MIP) related …

Advancing Compact Modeling Of Electronic Devices: Machine Learning Approaches With Neural Networks, Mixture Density Networks, And Deep Symbolic Regression, Jack Robert Hutchins

Masters Theses

This thesis pioneers the integration of deep learning techniques into the realm of compact modeling, presenting three distinct approaches that enhance the precision, efficiency, and adaptability of compact models for electronic devices. The first method introduces a Generalized Multilayer Perception Compact Model, leveraging the function approximation capabilities of neural networks through a multilayer perception (MLP) framework. This approach utilizes hyperband tuning to optimize network hyperparameters, demonstrating its effectiveness on a HfOx memristor and establishing a versatile modeling strategy for both single-state and multistate devices.

The second approach explores the application of Mixture Density Networks (MDNs) to encapsulate the inherent stochasticity …

Crash Detecting System Using Deep Learning, Yogesh Reddy Muddam

Electronic Theses, Projects, and Dissertations

Accidents pose a significant risk to both individual and property safety, requiring effective detection and response systems. This work introduces an accident detection system using a convolutional neural network (CNN), which provides an impressive accuracy of 86.40%. Trained on diverse data sets of images and videos from various online sources, the model exhibits complex accident detection and classification and is known for its prowess in image classification and visualization.

CNN ensures better accident detection in various scenarios and road conditions. This example shows its adaptability to a real-world accident scenario and enhances its effectiveness in detecting early events. A key …

Variational Bias Sampling For Collaborative Filtering Recommender Systems, Prisca Stephens

Theses and Dissertations

Advancements in digitalization has yielded enormous growth of data on online platforms, overwhelming users with multitude of options to choose from. Recommender systems narrow down these options to a few relevant ones thereby facilitating the decision-making processes for users. This study presents a framework for integrating variational bias sampling into model-based collaborative filtering techniques for recommender systems. Variational bias sampling is a novel and unique way to account for random factors that affect explicit ratings in collaborative filtering recommender systems. A Gaussian distribution is used to model all the possible random factors that could affect ratings. Sampling user and item …

Representation Learning For Generative Models With Applications To Healthcare, Astronautics, And Aviation, Van Minh Nguyen

Theses and Dissertations

This dissertation explores applications of representation learning and generative models to challenges in healthcare, astronautics, and aviation.

The first part investigates the use of Generative Adversarial Networks (GANs) to synthesize realistic electronic health record (EHR) data. An initial attempt at training a GAN on the MIMIC-IV dataset encountered stability and convergence issues, motivating a deeper study of 1-Lipschitz regularization techniques for Auxiliary Classifier GANs (AC-GANs). An extensive ablation study on the CIFAR-10 dataset found that Spectral Normalization is key for AC-GAN stability and performance, while Weight Clipping fails to converge without Spectral Normalization. Analysis of the training dynamics provided further …

Pedestrian Pathing Prediction Using Complex Contextual Behavioral Data In High Foot Traffic Settings, Laurel Bingham

All Graduate Theses and Dissertations, Fall 2023 to Present

Ensuring the safe integration of autonomous vehicles into real-world environments requires a comprehensive understanding of pedestrian behavior. This study addresses the challenge of predicting the movement and crossing intentions of pedestrians, a crucial aspect in the development of fully autonomous vehicles.

The research focuses on leveraging Honda's TITAN dataset, comprising 700 unique clips captured by moving vehicles in high-foot-traffic areas of Tokyo, Japan. Each clip provides detailed contextual information, including human-labeled tags for individuals and vehicles, encompassing attributes such as age, motion status, and communicative actions. Long Short-Term Memory (LSTM) networks were employed and trained on various combinations of contextual …

Learning Scene Semantics For 3d Scene Retrieval, Natalie Gleason

Honors Theses

This project presents a comprehensive exploration into semantics-driven 3D scene retrieval, aiming to bridge the gap between 2D sketches/images and 3D models. Through four distinct research objectives, this project endeavors to construct a foundational infrastructure, develop methodologies for quantifying semantic similarity, and advance a semantics-based retrieval framework for 2D scene sketch-based and image-based 3D scene retrieval. Leveraging WordNet as a foundational semantic ontology library, the research proposes the construction of an extensive hierarchical scene semantic tree, enriching 2D/3D scenes with encoded semantic information. The methodologies for semantic similarity computation utilize this semantic tree to bridge the semantic disparity between 2D …

Exploring Tokenization Techniques To Optimize Patch-Based Time-Series Transformers, Gabriel L. Asher

Computer Science Senior Theses

Transformer architectures have revolutionized deep learning, impacting natural language processing and computer vision. Recently, PatchTST has advanced long-term time-series forecasting by embedding patches of time-steps to use as tokens for transformers. This study examines and seeks to enhance PatchTST's embedding techniques. Using eight benchmark datasets, we explore explore novel token embedding techniques. To this end, we introduce several PatchTST variants, which alter the embedding methods of the original paper. These variants consist of the following architectural changes: using CNNs to embed inputs to tokens, embedding an aggregate measure like the mean, max, or sum of a patch, adding the exponential …

Mitigating Safety Issues In Pre-Trained Language Models: A Model-Centric Approach Leveraging Interpretation Methods, Weicheng Ma

Dartmouth College Ph.D Dissertations

Pre-trained language models (PLMs), like GPT-4, which powers ChatGPT, face various safety issues, including biased responses and a lack of alignment with users' backgrounds and expectations. These problems threaten their sociability and public application. Present strategies for addressing these safety concerns primarily involve data-driven approaches, requiring extensive human effort in data annotation and substantial training resources. Research indicates that the nature of these safety issues evolves over time, necessitating continual updates to data and model re-training—an approach that is both resource-intensive and time-consuming. This thesis introduces a novel, model-centric strategy for understanding and mitigating the safety issues of PLMs by …

Statistically Principled Deep Learning For Sar Image Segmentation, Cassandra Goldberg

Honors Projects

This project explores novel approaches for Synthetic Aperture Radar (SAR) image segmentation that integrate established statistical properties of SAR into deep learning models. First, Perlin Noise and Generalized Gamma distribution sampling methods were utilized to generate a synthetic dataset that effectively captures the statistical attributes of SAR data. Subsequently, deep learning segmentation architectures were developed that utilize average pooling and 1x1 convolutions to perform statistical moment computations. Finally, supervised and unsupervised disparity-based losses were incorporated into model training. The experimental outcomes yielded promising results: the synthetic dataset effectively trained deep learning models for real SAR data segmentation, the statistically-informed architectures …

Enhancing Scanning Tunneling Microscopy With Automation And Machine Learning, Darian Smalley

Graduate Thesis and Dissertation 2023-2024

The scanning tunneling microscope (STM) is one of the most advanced surface science tools capable of atomic resolution imaging and atomic manipulation. Unfortunately, STM has many time-consuming bottlenecks, like probe conditioning, tip instability, and noise artificing, which causes the technique to have low experimental throughput. This dissertation describes my efforts to address these challenges through automation and machine learning. It consists of two main sections each describing four projects for a total of eight studies.

The first section details two studies on nanoscale sample fabrication and two studies on STM tip preparation. The first two studies describe the fabrication of …

Ai-Based Defect Detection In Aerospace Ultrasonic Signals, Rami Issac Lake

Graduate Research Theses & Dissertations

Ensuring the safety and integrity of materials and structures throughout the manufacturing cycle is a critical concern across various industries, including aerospace, automotive, oiland gas, and civil engineering. Non-Destructive Inspection (NDI) techniques allow for the examination of materials without causing damage or alteration, enabling the early detection of potential issues before materials are utilized in the field. The inspection of fuselage composites presents a particular challenge due to their complex structures, diverse materials, and differences in thickness, making defect detection a challenging yet crucial task. Moreover, defects of various types and causes can emerge across all depths of the material …

Deep Learning Applications On Ionospheric Studies, Yang Pan

Physics Dissertations

Machine learning techniques, particularly deep learning techniques, have been vigorously pursued to tackle space physics problems and achieved some impressive results recently. The growth of deep learning technologies in different domains enables innovative solutions to those problems compared to conventional methods. Filling data gaps in instrumental observations is among the demanding issues, which benefits space physicists to study ionospheric phenomena with complete data coverage. Global total electron content (TEC) and regional ionospheric electron density (Ne) are among important physical parameters in ionospheric studies. Due to the limited coverage of global navigation satellite system (GNSS) ground receivers and sporadic …

Leveraging Machine Learning & Deep Learning Methodologies To Detect Deepfakes, Aniruddha Tiwari

All Graduate Theses, Dissertations, and Other Capstone Projects

The rapid evolution of deep learning (DL) and machine learning (ML) techniques has facilitated the rise of highly convincing synthetic media, commonly referred to as deepfakes. These manipulative media artifacts, generated through advanced artificial intelligence algorithms, pose significant challenges in distinguishing them from authentic content. Given their potential to be disseminated widely across various online platforms, the imperative for robust detection methodologies becomes apparent. Accordingly, this study explores the efficacy of existing ML/DL-based approaches and aims to compare which type of methodology performs better in identifying deepfake content. In response to the escalating threat posed by deepfakes, previous research efforts …