Browsing by Subject "Deep Learning"
Now showing 1 - 20 of 21
- Results Per Page
- Sort Options
Item Advancing Deep Learning For Scientific Inverse Problems(2023-09) Zhuang, ZhongArtificial intelligence (AI) has ushered in a new paradigm for addressing scientific complexities, harnessing the computational prowess of robust machines and sophisticated algorithms tailored to domain-specific constraints. Across diverse domains encompass- ing scientific and engineering landscapes such as astronomy, biomedical science, and material science, the emergence of inverse problems is inherent. These quandaries are distinguished by the overarching objective of elucidating the reconstruction of meticulously structured entities from observational data, buoyed by a foundational bedrock of prior knowledge. In a generalized sense, the inverse problem is cast as Y = F(x), subject to the constraint G(x) = 0, wherein F denotes the function orchestrating the transformation of object x to observation Y, while G embodies the constraints imposed by pre-existing knowledge. Owing to the intrinsic characteristics of F—often marked by pronounced nonlinearity—inverse problems seldom conform to well-posed paradigms. Herein lies the significance of AI tools that leverage the extraction of latent insights from voluminous empirical data, epitomizing the realm of data-driven AI tools. This innovation extends the purview of inference beyond the bounds of human-established priors and experiential wisdom. However, the scientific realm stands in contrast to the more prolific AI domains, as the availability and fidelity of data are not perpetually assured in the context of inverse problems. Consequently, a promising avenue emerges in the form of singular-instance AI tools, underpinned by the potency of formidable constructs such as deep neural networks (DNNs). Operating autonomously from expansive data repositories, these tools offer an alternative avenue to address inverse problems within the scientific continuum. Within this work, we delineate our recent endeavors directed at catalyzing break- throughs in the resolution of intricate scientific inverse problems. Central to our approach are pragmatic strategies that harmoniously blend the attributes of data-driven and singular-instance AI tools into coherent pipelines. These endeavors culminate in a iiinovel problem-solving landscape that bridges the domains of AI and science, encapsulating the essence of innovation and advancement.Item Artificial Intelligence to Accelerate COVID-19 Identification from Chest X-rays(2021-05) Adila, DyahImportance: Clinical signs and symptoms for COVID-19 remain the mainstay of early diagnosis and initial management in the emergency department (ED) and inpatient setting at many hospitals due to de- lays in obtaining results of PCR testing and limitations in access to rapid antigen testing. The majority of many patients with COVID- 19 will present with respiratory symptoms necessitating a chest x-ray (CXR) as a routine part of screening. An AI-based model to predict COVID-19 likelihood from CXR findings can serve as an important and immediate adjunct to accelerate clinical decision making.Objective: To develop a robust AI-based diagnostic model to identify CXRs with COVID-19 compared with all non-COVID-19 CXRs. Setting: Labeled frontal CXR images (samples of COVID-19 and non-COVID-19) from the M Health Fairview (Minnesota, USA), Va- Valencian Region Medical ImageBank (Spain), MIMIC-CXR, Open-i 2013 Chest X-ray Collection, GitHub COVID-19 Image Data Collection (International). Main Outcome and Measure: Model performance assessed via Area under the Receiver Operating Curve (AUROC) and Area Under the Precision and Recall Curve (AUPRC). Results: Patients with COVID-19 had significantly higher COVID- 19 Diagnostic Scores than patients without COVID-19 on both real-time electronic health records and external (non-publicly available) validation. The model performed well across all four methods for model validation with AUROCs ranging between 0.7 – 0.96 and high PPV and specificity. The model performed had improved discrimination for patients with “severe” as compared to “moderate” COVID-19 disease. The model had unrealistic performance using publicly available databases, reflecting the inherent limitations in many previously developed models relying on publicly available data for training and validation. Conclusions and Relevance: AI-based diagnostic tools may serve as an adjunct, but not replacement, to support COVID-19 diagnosis which largely hinges on exposure history, signs, and symptoms. Future research should focus on optimizing discrimination of “mild” COVID-19 from non-COVID-19 image findings.Item Bridging Mri Reconstruction Across Eras: From Novel Optimization Of Traditional Methods To Efficient Deep Learning Strategies(2024-03) Gu, HongyiMagnetic Resonance Imaging (MRI) has been extensively used as a non-invasive modality for imaging the human body. Despite substantial advances over the past decades, scan duration remains as a principal issue for MRI scanning, requiring novel techniques to accelerate data acquisition. Such techniques are poised to improve clinical patient throughput, reduce motion artifacts, enhance subject comfort, and allow higher resolution imaging in many applications. Several methods have been proposed to accelerate MRI scans. In parallel imaging (PI), k-space data was acquired at a sub-Nyquist rate with with multiple receiver coils, and the redundancy among these coils were used for image reconstruction. Following the clinical impact and success of PI methods, compressed sensing (CS) techniques were developed to reconstruct images by using compressibility of images in a pre-specified linear transform domain. Transform learning (TL) was another line of work that learned the linear transforms from data, while enforcing sparsity as in CS. Recently, deep learning (DL) has shown great promise for MRI reconstruction, especially at high acceleration rates where other traditional methods would fail. Specially, physics-guided DL (PG-DL) unrolls a traditional optimization algorithm for solving regularized least squares for a fixed number of iterations, and uses neural networks to implicitly perform regularization. These unrolled networks are trained end-to-end with large databases, using well-designed loss functions and advanced optimizers, usually using a reference fully-sampled image for supervised learning. Several approaches have noted the difficulty or impossibility of acquiring fully-sampled data in various MRI applications. Among these, self-supervised learning with data undersmapling (SSDU) was developed to allow training without fully-sampled data, and multi-mask SSDU was subsequently proposed for better reconstruction quality at high acceleration rates. Although PG-DL generally shows strong ability for excellent reconstruction performance, there are concerns for generalizabilty, interpretability and stability issues. In this thesis, we aimed to bridge the gap between traditional and DL methods, while also extending the utility of DL methods for non-Cartesian imaging. We first revisited l1-wavelet CS reconstruction for accelerated MRI by using modern data science tools similar to those used in DL for optimized performance. We showed that our proposed optimization approach improved traditional CS, and further performance boost was observed by incorporating wavelet subband processing and reweighted l1 minimization. The final version reached a performance similar to state-of-the-art PG-DL, while preserving better interpretability by solving a convex optimization problem in inference time. Second, we combined ideas from CS, TL and DL to enable the learning of deep linear convolutional transforms in a format similar to PG-DL. Our proposed method performed better than CS and TL, and gave similar performance as state-of-the-art PG-DL. It used a linear representation of image as regularization at inference time, and enabled convex sparse image reconstruction that may have better robustness, stability and generalizability properties. Third, we adapted a self-supervised PG-DL technique to non-Cartesian trajectories and showed its potential for reconstructing 10-fold accelerated spiral fMRI multi-echo acquisitions. Our proposed approach gave substantial improvements in reconstructed image quality over conventional methods. Furthermore, the blood oxygenation level dependent (BOLD) signal analysis of our proposed method provided meaningful sensitivities, with similar activation patterns and extent to the expected baselines.Item Computational Sleep Science: Machine Learning for the Detection, Diagnosis, and Treatment of Sleep Problems from Wearable Device Data(2017-12) Sathyanarayana, AartiThis thesis is motivated by the rapid increase in global life expectancy without the respective improvements in quality of life. I propose several novel machine learning and data mining methodologies for approaching a paramount component of quality of life, the translational science field of sleep research. Inadequate sleep negatively affects both mental and physical well-being, and exacerbates many non-communicable health problems such as diabetes, depression, cancer and obesity. Taking advantage of the ubiquitous adoption of wearable devices, I create algorithmic solutions to analyse sensor data. The goal is to improve the quality of life of wearable device users, as well as provide clinical insights and tools for sleep researchers and care-providers. Chapter 1 is the introduction. This section substantiates the timely relevance of sleep research for today's society, and its contribution towards improved global health. It covers the history of sleep science technology and identifies core computing challenges in the field. The scope of the thesis is established and an approach is articulated. Useful definitions, sleep domain terminology, and some pre-processing steps are defined. Lastly, an outline for the remainder of the thesis is included. Chapter 2 dives into my proposed methodology for widespread screening of sleep disorders. It surveys results from the application of several statistical and data mining methods. It also introduces my novel deep learning architecture optimized for the unique dimensionality and nature of wearable device data. Chapter 3 focuses on the diagnosis stage of the sleep science process. I introduce a human activity recognition algorithm called RAHAR, Robust Automated Human Activity Recognition. This algorithm is unique in a number of ways, including its objective of annotating a behavioural time series with exertion levels rather than activity type. Chapter 4 focuses on the last step of the sleep science process, therapy. I define a pipeline to identify \textit{behavioural recipes}. These \textit{recipes} are the target behaviour that a user should complete in order to have good quality sleep. This work provides the foundation for building out a dynamic real-time recommender system for wearable device users, or a clinically administered cognitive behavioural therapy program. Chapter 5 summarizes the impact of this body of work, and takes a look into next steps. This chapter concludes my thesis.Item Computer Vision Algorithms for Yield Mapping in Specialty Farms(2019-11) Roy, PravakarPrecision farming and phenotyping technologies have the potential to drastically transform the agricultural landscape. For commodity crops such as maize, wheat and soy recurring farming tasks such as seeding, weeding, irrigation, fertilization, application of pesticides, harvesting, and storage are in the process of being completely automated. Specialty crops (tree fruit, flowers, vegetables, and nuts) are excellent candidates for similar automation as they have high monetary value, high management cost and high variability in growth. An important capability for both precision agriculture and phenotyping is yield mapping. Yield mapping for tree fruit is challenging because it involves solving multiple computer vision (fruit detection, counting, recovering underlying 3D geometry for tracking fruit across different frames in continuously changing illumination) as well as planning problems (path planning for covering all fruit, picking fruit). The main goal of this dissertation is to develop computer vision and deep learning algorithms for yield mapping in specialty farms. The dissertation is divided into three parts. The first part is dedicated to developing practical solutions for yield mapping in specialty farms. We present solutions for fruit detection, counting, recovering the underlying scene geometry and fruit tracking. We integrate these individual solutions in a modular manner and create a flexible framework for complete yield estimation. Additionally, we perform an extensive experimental evaluation of the developed system and sub-components. Our algorithms successfully predict 97% of the ground truth yield and outperform all existing state-of-the-art methods. Some of these efforts are now in the process of being commercialized. In the second part of the dissertation, we study a problem where a manipulator equipped with a camera, mounted on a ground robot is charged with accurately counting fruit by taking a minimum number of views. We present a method for efficiently enumerating combinatorially distinct world models most likely to generate the captured views. These are incorporated into single and multi-step planners for accurate fruit counting. We evaluate these planners in simulation as well as with experiments on a real robot. In the third part, we study the problem of realistic synthetic data generation for training deep neural networks. We present a method that jointly translates the synthetic images and their underlying semantics to the domain of the real data so that an adversarial discriminator (a deep neural network) cannot distinguish between the real and synthetic data. This method enables us to stylize the synthetic data to any fruit, lighting condition and environment. It can be applied to a wide variety of domain transfer tasks beyond fruit detection and counting (e.g from Grand Theft Auto (GTA) to Cityscapes for autonomous driving). Additionally, it enables us to perform image to image translation with significant changes in underlying geometry (e.g circles to triangles, sheep to giraffe, etc). These results in this dissertation together present a complete yield monitoring system for specialty crops, view planning strategies for accurate fruit counting and a framework for generating realistic synthetic data. These methods together push the state-of-the-art and take us one step closer toward building a sustainable infrastructure for intelligent integrated farm management.Item Deep reinforcement learning for personalized treatment recommendation(2023-03) Liu, MingyangIn precision medicine, the ultimate goal is to recommend the most effective treatment to an individual patient based on patient-specific molecular and clinical profiles, possibly high-dimensional. To advance cancer treatment, large-scale screenings of cancer cell lines against chemical compounds have been performed to help better understand the relationship between genomic features and drug response; existing machine learning approaches use exclusively supervised learning, including penalized regression and recommender systems. When there is only one time point, it refers to individualized treatment selection, which is employed to maximize a certain clinical outcome of a specific patient based on a patient's clinical or genomic characteristics, given a patients' heterogeneous response to treatments. Although developing such a rule is conceptually important to personalized medicine, existing methods such as the $L_1$-penalized least squares \citep{qian2011performance} suffers from the difficulty of indirect maximization of clinical outcome, while the outcome weighted learning \citep{zhao2012estimating} directly maximizing the clinical outcome is not robust against any perturbation of the outcome. We will first propose a weighted $\psi$-learning method to optimize an individualized treatment rule, which is robust again perturbation of data near decision boundary through the notation of separation. To deal with nonconvex minimization, we employ a difference of convex algorithm to solve the non-convex minimization iteratively based on a decomposition of the cost function into a difference of two convex function. On this ground, we also introduce a variable selection method for further removing redundant variables for higher performance. Finally, we illustrate the proposed method through simulations and a lung health study, and demonstrate that it yields higher performance in terms of accuracy of prediction of individualized treatment. However, it would be more efficient to apply reinforcement learning (RL) to sequentially learn as data accrue, including selecting the most promising therapy for a patient given individual molecular and clinical features and then collecting and learning from the corresponding data. In this way, we propose a novel personalized ranking system called Proximal Policy Optimization Ranking (PPORank), which ranks the drugs based on their predicted effects per cell line (or patient) in the framework of deep reinforcement learning (DRL). Modeled as a Markov decision process (MDP), the proposed method learns to recommend the most suitable drugs sequentially and continuously over time. As a proof-of-concept, we conduct experiments on two large-scale cancer cell line data sets in addition to simulated data. The results demonstrate that the proposed DRL-based PPORank outperforms the state-of-the-art competitors based on supervised learning. Taken together, we conclude that novel methods in the framework of DRL have great potential for precision medicine and should be further studied.Item Deep Z-Learning(2018-05-12) Bittner, NathanIn this thesis, I present advancements in the theory of Z-learning. In particular, I explicitly define a complete tabular Z-learning algorithm, I provide a number of pragmatic qualifications on how Z-learning should be applied to different problem domains, and I extend Z-learning to non-tabular discrete domains by introducing deep network function-approximation versions of Z-learning that is similar to deep Q-learningItem Design The Next-Generation Neuroprostheses: From Statistical Modeling to Artificial Intelligence(2022-10) Luu, Diu KhueAbstract:A prosthetic hand ultimately seeks to replace the essential functions of a lost limb in activities of daily living. Yet most existing prostheses only allow limited movements and cannot provide sensory feedback to the amputee. These limitations make the user experience unnatural and unintuitive. Therefore, the next-generation artificial limb must be equipped with a neural interface that can facilitate bidirectional communication between the user's nervous system and the prosthesis's circuitry to create a substitution that genuinely feels and acts like a real hand. In this dissertation, we investigate diverse strategies to inform new designs of an advanced neural interface. Our approaches range from conventional statistical modeling to deep learning-based artificial intelligence (AI). First, we formulate a statistical model governing the mismatch error in real-world sensors and devices. Specifically, we derive a framework for manipulating random mismatches to achieve super-resolution where the system could gain an effective resolution 500 times more precise than the conventional limitation. This mechanism has been applied to design a super-resolution neurostimulator used by a human amputee in sensory restoring experiments. Second, we design a pseudo-online AI neural decoder to translate the amputee's movement intents from peripheral nerve data. Various decoding strategies, including one-step and two-step approaches and a data representation called feature extraction, are studied to optimize the decoder's performance. We show that utilizing feature extraction could help lower the decoder's complexity substantially, making the design feasible for real-time applications. We then demonstrate an AI neural decoder based on recurrent neural networks (RNN), which outperforms all other classic machine learning techniques in decoding a large nerve dataset. Third, we study the AI neural decoder's real-time performance and long-term longevity with three human amputees. The AI neural decoder allows the amputees to control individual fingers and the wrist of a robotic hand in real-time with 97-98% accuracy. We design a gesture-matching performance test; the amputee can achieve a reaction time of 0.8 seconds and an information throughput of 365 bit-per-minute. We also show that the decoder's predictive performance is robust over a 16-month implant duration. Our study lays the groundwork for the next-generation neuroprostheses that are enabled by a bidirectional, high-bandwidth, and intuitive neural interface. Our comprehensive investigation of various statistical and empirical approaches would inform the design of new neuro-sensors and neural decoders. We envision that AI technology, specifically deep learning, will be at the heart of the next-generation dexterous and intuitive prosthetic hands.Item Efficient Robotic Manipulation with Scene Knowledge(2023-05) Lou, XibaiIn recent years, robots have transformed manufacturing, logistics, and transportation. However, extending the success to unstructured real-world environments (e.g., domestic kitchens, warehouses, grocery stores, etc.) remains difficult due to three key challenges: (1) assumption of structured environments (such as organized bottles in the factories); (2) hand-engineered solutions that are difficult to generalize to novel scenarios; (3) limited flexibility of action primitives, which prevents the robot from reaching target objects. In this thesis, we address these challenges by learning scene knowledge that improves the efficiency of robotic manipulation systems. Grasping is a fundamental manipulation skill that is constrained by the scene arrangement (i.e., the locations of the robot, the objects, and the environmental structures). Understanding scene knowledge, such as the robot's reachability to objects, is crucial to improve the robot's capability. We developed a reachability-aware grasp pose generator that predicts feasible 6-degree-of-freedom (6-DoF) grasp poses (i.e., approaching with an arbitrary direction and wrist orientation). Then, we extended to target-driven grasping in constrained environments and added collision awareness to our scene knowledge. When objects are densely cluttered, we improved the robot's efficiency by employing graph neural networks (GNN) to exploit the underlying relationships in the scene. To accomplish complex manipulation tasks in constrained environments, such as rearranging adversarial objects, we hierarchically integrated a heterogeneous graph neural network (HetGNN)-based coordinator and the 3D CNN-based actors. The system reasons about the relational knowledge between scene components and coordinates multiple robotic skills (e.g., grasping, pushing) to minimize the planning cost. As we anticipate an increase in the number of domestic robots, the robotics community necessitates a framework that not only commands the robot accurately, but also reasons about the unstructured scene to improve robots' efficiency. This thesis contributes to the goal by equipping robotics manipulation with learned scene knowledge. We present 6-DoF robotic systems that can grasp novel objects in dense clutter with reachability awareness, retrieve target objects within arbitrary structures, and rearrange multiple objects into goal configurations in constrained environments.Item Enhancing Visual Perception in Noisy Environments using Generative Adversarial Networks(2018-08) Fabbri, CameronAutonomous robots rely on a variety of sensors – acoustic, inertial, and visual – for intelligent decision making. Due to its non-intrusive, passive nature, and high information content, vision is an attractive sensing modality. However, many environments contain natural sources of visual noise such as snow, rain, dust, and other forms of distortion. This work focuses on the underwater environment, in which visual noise is a prominent component. Factors such as light refraction and absorption, suspended particles in the water, and color distortion affect the quality of visual data, resulting in noisy and distorted images. Autonomous Underwater Vehicles (AUVs) that rely on visual sensing thus face difficult challenges, and consequently exhibit poor performance on vision driven tasks. This thesis proposes a method to improve the quality of visual underwater scenes using Generative Adversarial Networks (GANs), with the goal of improving input to vision-driven behaviors further down the autonomy pipeline. Furthermore, we show how recently proposed methods are able to generate a dataset for the purpose of such underwater image restoration. For any visually-guided underwater robots, this improvement can result in increased safety and reliability through robust visual perception. To that effect, we present quantitative and qualitative data which demonstrates that images corrected through the proposed approach generate more visually appealing images, and also provide increased accuracy for a diver tracking algorithm.Item Improving Automatic Painting Classification Using Saliency Data(2022-10) Kachelmeier, RosalieSince at least antiquity, humans have been categorizing art based on various attributes. With the invention of the internet, the amount of art available and people searching for art has grown significantly. One way to keep up with these increases is to use computers to automatically suggest categories for paintings. Building upon past research into this topic using transfer learning as well as research showing that artistic movement affected gaze data, we worked to combine transfer learning with gaze data in order to improve automatic painting classification. To do this, we first trained a model on a large object recognition dataset with synthesized saliency data. We then repurposed it to classify paintings by 19th century artistic movement and trained it further on a dataset of 150 paintings with saliency data collected from 21 people. Training on this was split into two stages. In the first, the final layer of the model was trained on the dataset with the rest of the model frozen. Next, the entire model was fine-tuned on the data using a much lower learning rate. Fifteen trials of this were done with different random seeds in order to decrease any effect that randomness might have. Overall it achieved an accuracy of 0.569 with standard deviation of 0.0228. Comparatively, a similar existing method had an accuracy of 0.523 with standard deviation of 0.0156. This ends up being a statistically significant difference (p = 0.0479), suggesting that when given enough training time a more complex model utilizing saliency data can outperform a simpler model that does not use saliency data when it comes to classifying paintings.Item Improving Signal to Noise Ratio in Ultra High Field Magnetic Resonance Imaging(2021-03) Tavaf, NaderUltra-High Field (UHF) Magnetic Resonance Imaging (MRI) advantages, including higher image resolution, reduced acquisition time via parallel imaging, and better Signal-to-Noise Ratio (SNR) have opened new opportunities for various clinical and research projects, including Functional Magnetic Resonance Imaging (fMRI), brain connectivity mapping, and anatomical imaging. The advancement of these UHF MRI performance metrics, especially SNR, was the primary motivation of this thesis.Unaccelerated SNR depends on receive array sensitivity profile, receiver noise correlation and static magnetic field strength. Various receive array decoupling technologies, including overlap/inductive and preamplifier decoupling, were previously utilized to mitigate noise correlation. In this dissertation, I developed a novel self-decoupling principle to isolate elements of a loop-based receive array and demonstrated, via full-wave electromagnetic/circuit co-simulations validated by bench measurements, that the self-decoupling technique provides inter-element isolation on par with overlap decoupling while self-decoupling improves SNR. I then designed and constructed the first self-decoupled 32 and 64 channel receiver arrays for human brain MRI imaging at 10.5T / 447MHz. Experimental comparisons of these receive arrays with the industry’s gold-standard 7T 32 channel receiver resulted in 1.81 times and 3.53 times more average SNR using the 10.5T 32 and 64 channel receivers I built, respectively. To further improve the SNR of accelerated MRI images, I developed a novel data-driven model using a customized conditional Generative Adversarial Network (GAN) architecture for parallel MRI image reconstruction and demonstrated that, when applied to human brain images subsampled with rate of 4, the GAN model results in a Peak Signal-to-Noise Ratio (PSNR) of 37.65 compared to GeneRalized Autocalibrating Partial Parallel Acquisition (GRAPPA)’s PSNR of 33.88. In summary, the work presented in this dissertation improved the SNR available for human brain imaging and provided the experimental realization of the advantages anticipated at 10.5T MRI. The insights from this thesis inform future efforts to build self-decoupled transmit arrays and high density (i.e. 128 channel) loop-based receive arrays for human brain MRI especially at ultra-high field as well as future studies to utilize deep learning techniques for reconstruction and post-processing of parallel MRI images.Item Looking Ahead By Looking Back(2018-09) Naeem, HammadTo understand the evolution of manufacturing and its future requires multidimensional study of different historical milestones, systems developed over a period of time and some concrete analysis amalgamated with experimental results. This thesis is about identifying major milestones in the manufacturing history and then using this information to understand the evolution of innovation process and innovative models. Further, using knowledge obtained from the study of innovation processes to understand the modern trends in manufacturing industry. Experimental analysis is performed using modern machine learning techniques like deep learning to correctly identify human facial expressions. The increasing utility of artificial intelligence was the driving force to exploit modern machine learning techniques that have proven that now decision power can be carefully delegated or shared with these intelligent systems. The Deep Learning based approach using convolutional neural network is tested on human facial expression recognition and accuracy of over 86% is achieved which is higher than other mathematical based machine learning models. These modern machine learning algorithms are also tested on numerical dataset to prove their flexibility and adaptability for different applications which can be faced in any modern day manufacturing industry. The results from this study show that these modern machine learning algorithms have outperformed old decision making methodologies due to their capacity and intelligence in learning different patterns present in the data and correspondingly helping in correct decision making. As a future recommendation, a hybrid system is proposed which is a combination of predictive as well as corrective maintenance. The proposed system is based on deep learning using convolutional neural network to predict end of life of a part.Item Machine Learning Methods with Emphasis on Cancerous Tissue Recognition(2018-08) Stanitsas, PanagiotisToday, vast and unwieldy data collections are regularly being generated and analyzed in hopes of supporting an ever-expanding range of challenging sensing applications. Modern inference schemes usually involve millions of parameters to learn complex real-world tasks, which creates the need for large annotated datasets for training. For several visual learning applications, collecting large amounts of annotated data is either challenging or very expensive; one such domain is medical image analysis. In this thesis, machine learning methods were devised with emphasis on Cancerous Tissue Recognition (CTR) applications. First, a lightweight active constrained clustering scheme was developed for the processing of image data which capitalizes on actively acquired pairwise constraints. The proposed methodology introduces the use of the Silhouette values, conventionally used for measuring clustering performance, in order to rank the degree of information content of the various samples. Second, an active selection framework that operates in tandem with Convolutional Neural Networks (CNNs) was constructed for CTR. In the presence of limited annotations, alternative (or sometimes complementary) venues were explored in an effort to restrain the high expenditure of collecting image annotations required by CNN-based schemes. Third, a Symmetric Positive Definite (SPD) image representation was derived for CTR, termed Covariance Kernel Descriptor (CKD) which consistently outperformed a large collection of popular image descriptors. Even though the CKD successfully describes the tissue architecture for small image regions, its performance decays when implemented on larger slide regions or whole tissue slides due to the larger variability that tissue exhibits at that level, since different types of tissue can be present as the regions grow (healthy, benign disease, malignant disease). Fourth, to leverage the recognition capability of the CKDs to larger slide regions, the Weakly Annotated Image Descriptor (WAID) was devised as the parameters of classifier decision boundaries in a multiple instance learning framework. Fifth, an Information Divergence and Dictionary Learning (IDDL) scheme for SPD matrices was developed for identifying appropriate geometries and similarities for SPD matrices and was successfully tested on a diverse set of recognition problems including activity, object, and texture recognition as well as CTR. Finally, a transition of IDDL to an unsupervised setup was developed, dubbed alpha-beta-KMeans, to address the problem of learning information divergences while clustering SPD matrices in the absence of labeled data.Item Machine Vision for Improved Human-Robot Cooperation in Adverse Underwater Conditions(2021-05) Islam, Md JahidulVisually-guided underwater robots are deployed alongside human divers for cooperative exploration, inspection, and monitoring tasks in numerous shallow-water and coastal-water applications. The most essential capability of such companion robots is to visually interpret their surroundings and assist the divers during various stages of an underwater mission. Despite recent technological advancements, the existing systems and solutions for real-time visual perception are greatly affected by marine artifacts such as poor visibility, lighting variation, and the scarcity of salient features. The difficulties are exacerbated by a host of non-linear image distortions caused by the vulnerabilities of underwater light propagation (e.g., wavelength-dependent attenuation, absorption, and scattering). In this dissertation, we present a set of novel and improved visual perception solutions to address these challenges for effective underwater human-robot cooperation. The research outcomes entail novel design and efficient implementation of the underlying vision and learning-based algorithms with extensive field experimental validations and real-time feasibility analyses for single-board deployments. The dissertation is organized into three parts. The first part focuses on developing practical solutions for autonomous underwater vehicles (AUVs) to accompany human divers during an underwater mission. These include robust vision-based modules that enable AUVs to understand human swimming motion, hand gesture, and body pose for following and interacting with them while maintaining smooth spatiotemporal coordination. A series of closed-water and open-water field experiments demonstrate the utility and effectiveness of our proposed perception algorithms for underwater human-robot cooperation. We also identify and quantify their performance variability over a diverse set of operating constraints in adverse visual conditions. The second part of this dissertation is devoted to designing efficient techniques to overcome the effects of poor visibility and optical distortions in underwater imagery by restoring their perceptual and statistical qualities. We further demonstrate the practical feasibility of these techniques as pre-processors in the autonomy pipeline of visually-guided AUVs. Finally, the third part of this dissertation develops methodologies for high-level decision-making such as modeling spatial attention for fast visual search, learning to identify when image enhancement and super-resolution modules are necessary for a detailed perception, etc. We demonstrate that these methodologies facilitate up to 45% faster processing of the on-board visual perception modules and enable AUVs to make intelligent navigational and operational decisions, particularly in autonomous exploratory tasks. In summary, this dissertation delineates our attempts to address the environmental and operational challenges of real-time machine vision for underwater human-robot cooperation. Aiming at a variety of important applications, we develop robust and efficient modules for AUVs to 'follow and interact' with companion divers by accurately perceiving their surroundings while relying on noisy visual sensing alone. Moreover, our proposed perception solutions enable visually-guided robots to 'see better' in noisy conditions, and 'do better' with limited computational resources and real-time constraints. In addition to advancing the state-of-the-art, the proposed methodologies and systems take us one step closer toward bridging the gap between theory and practice for improved human-robot cooperation in the wild.Item Mitigating Adverse Outcomes in Health and Wellness with Data Analytics: Investigation of Medical Device Recalls and Digital Exercise Program Churns(2024-03) Zhu, YiThe adverse outcomes in health and wellness, such as the failure of medical devices and the prevalence of human sedentary lifestyles, pose significant risks to patient safety, public health, and the financial stability of healthcare institutions and firms. To enhance proactive measures against these adverse outcomes and mitigate their potential harm, this dissertation employs empirical methodologies to investigate two types of adverse outcomes in health and wellness. Moreover, it also develops algorithmic approaches for predicting these outcomes. Specifically, this dissertation comprises three essays: the first essay addresses medical device recalls by proposing a design-science-based framework for predicting such recalls. This framework demonstrates superior performance compared to traditional predictive models and offers various insights for improving medical device safety regulations through different prediction problem setups. The second and third essays delve into human physical activity behaviors as recorded by digital exercise platforms. They explore the social contagion effect of churn digital exercise programs, examine the variations in the contagion effects based on individuals’ characteristics, and interpret these effects using complex contagion theory. An advanced deep learning approach is introduced to forecast exercise patterns indicative of potential digital activity program churn, leveraging the relationships between different exercise types and measures. The effectiveness of this approach is validated through experiments with both simulated and actual data, showcasing its advantages over traditional models. This dissertation provides substantial practical implications and contributes to the advancement of knowledge across the domains of information systems, healthcare, and wellness. Insights from the first essay provide critical policy recommendations for enhancing medical device safety, augmenting the understanding of predictive health information system design, and fostering data-driven methodologies for early detection of product recalls. The findings from the second and third essays emphasize the significance of leveraging social strategies to reduce churn in digital exercise programs and highlight the advantages of recommending users personalized exercise programs based on more accurate predictions of exercisers’ activities. These tailored programs aim to boost exercise adherence, thereby enhancing the overall personal wellness of exercisers. Overall, this dissertation enhances our understanding and prediction of medical device recalls and digital exercise program churn, elevating personal health and wellness outcomes.Item Muon Neutrino Disappearance in NOvA with a Deep Convolutional Neural Network Classifier(2016-03) Rocco, DominickThe NuMI Off-axis Neutrino Appearance Experiment (NOvA) is designed to study neutrino oscillation in the NuMI (Neutrinos at the Main Injector) beam. NOvA observes neutrino oscillation using two detectors separated by a baseline of 810 km; a 14 kt Far Detector in Ash River, MN and a functionally identical 0.3 kt Near Detector at Fermilab. The experiment aims to provide new measurements of $\Delta m^2_{32}$ and $\theta_{23}$ and has potential to determine the neutrino mass hierarchy as well as observe CP violation in the neutrino sector. Essential to these analyses is the classification of neutrino interaction events in NOvA detectors. Raw detector output from NOvA is interpretable as a pair of images which provide orthogonal views of particle interactions. A recent advance in the field of computer vision is the advent of convolutional neural networks, which have delivered top results in the latest image recognition contests. This work presents an approach novel to particle physics analysis in which a convolutional neural network is used for classification of particle interactions. The approach has been demonstrated to improve the signal efficiency and purity of the event selection, and thus physics sensitivity. Early NOvA data has been analyzed (2.74$\times10^{20}$ POT, 14 kt equivalent) to provide new best-fit measurements of $\sin^2(\theta_{23}) = 0.43$ (with a statistically-degenerate compliment near 0.60) and $|\Delta m^2_{32}| = 2.48\times10^{-3}~\text{eV}^2$.Item Robustness in Deep Learning: Single Image Denoising using Untrained Networks(2021-05) Singh, EshaDeep Learning has become one of the cornerstones of today’s AI advancement and research. Deep Learning models are used for achieving state-of-the-art results on a wide variety of tasks, including image restoration problems, specifically image denoising. Despite recent advances in applications of deep neural networks and the presence of a substantial amount of existing research work in the domain of image denoising, this task is still an open challenge. In this thesis work, we aim to summarize the study of image denoising research and its trend over the years, the fallacies, and the brilliance. We first visit the fundamental concepts of image restoration problems, their definition, and some common misconceptions. After that, we attempt to trace back where the study of image denoising began, attempt to categorize the work done till now into three main families with the main focus on the neural network family of methods, and discuss some popular ideas. Consequently, we also trace related concepts of over-parameterization, regularisation, low-rank minimization and discuss recent untrained networks approach for single image denoising, which is fundamental towards understanding why the current state-of-art methods are still not able to provide a generalized approach for stabilized image recovery from multiple perturbations.Item Self-Supervised Physics-Guided Deep Learning for Solving Inverse Problems in Imaging(2022-03) Yaman, BurhaneddinInverse problems in computational imaging seek to recover an unknown image of interest from observed measurements acquired using a known forward model. These inverse problems are often ill-conditioned, requiring some form of regularization. The corresponding objective function for inverse problems in computational imaging can often be solved using iterative optimization approaches that alternate between two sub-problems that enforce data consistency and promote the regularization approach, respectively. Such inverse problems arise in a multitude of imaging modalities, in particular in magnetic resonance imaging (MRI), which is the main application area for this thesis. Lengthy scan times remain a challenge in MRI, thus accelerating MRI scans has remained an open research problem over decades. Conventional accelerated MRI techniques acceleration rate is limited by noise amplification and residual artifacts. Recently, deep learning has emerged as an alternative approach for accelerated MRI. Among deep learning techniques, physics-guided deep learning (PG-DL) has drawn great interest, as it incorporates the known physical forward model into the network architecture. PG-DL unrolls a conventional iterative algorithm for solving a regularized least squares problem for a fixed number of iterations, and replaces the proximal operation corresponding to the regularizer implicitly with neural networks. These unrolled networks are trained end-to-end, with the goal of minimizing the difference between the network output and the corresponding reference data. Most of the existing deep learning approaches in MRI reconstruction are based on supervised learning, which requires ground-truth/ fully-sampled data for training. However, acquisition of fully-sampled data is infeasible in many applications due to physiological constraints, such as organ motion, or physical constraints, such as signal decay. In several other scenarios, such as high-resolution anatomical brain imaging, it is impractical to acquire fully-sampled datasets as the scan time becomes extremely lengthy. Therefore, enabling the training of PG-DL reconstruction without fully-sampled data is essential for the integration of deep learning reconstruction into clinical MRI practice. The present thesis introduces novel frameworks to enable the training of deep learning reconstruction methods for inverse imaging problems in the absence of ground-truth/fully-sampled data. First, we introduce self-supervised learning via data undersampling (SSDU) approach to enable database training without fully-sampled data. Succinctly, SSDU partitions available measurements into two disjoint sets. One of these sets is used in the data consistency units of the unrolled network, while the other is used to define the loss in the measurement domain. Subsequently, we extend SSDU for processing 3D datasets and provide solutions for GPU memory constraints and data scarcity issues faced in 3D processing. To cope with potential performance degradation at very high acceleration rates, we develop a multi-mask self-supervised learning approach, which retrospectively splits available measurements into multiple 2-tuples of disjoint sets to perform training and define a loss function. Furthermore, we introduce a zero-shot self-supervised learning approach to enable training from a single scan/sample without any external training databases. ZS-SSL partitions the available measurements from a single scan into three disjoint sets. Two of these sets are used to enforce data consistency and define loss during training for self-supervision, while the last set serves to self-validate, establishing an early stopping criterion. Finally, we introduce a self-supervised learning algorithm for referenceless image denoising. Self-supervised deep learning algorithms split the pixels for each image into two disjoint sets to perform training and defining loss. In existent self-supervised denoising approaches which are purely data-driven, the set of pixels used as input to the network is not re-utilized in the end-to-end training since the network is only comprised of a neural network. Reusing the pixels within the network would promote consistency with acquired measurements, thus leading to a more robust and improved denoising performance. To tackle this challenge, we build upon existent self-supervised learning algorithms and recast the denoising problem into a regularized image inpainting framework which allows use of algorithm unrolling for denoising.Item Towards Human-Like Machine Intelligence: Generalizability, Transferability, and Trustworthiness(2022-06) Luo, YanAcquiring human-like machine intelligence is a long-standing goal of machine learning. Thanks to the availability of large-scale datasets and the GPU acceleration, modern learning methods achieve remarkable success. Although it surpasses humans on several tasks, e.g., the game of go, there is still a gap between machine intelligence and human intelligence. The reasons are two-fold. Firstly, how the human brain produces intelligence is still little-known, and how to apply the mechanisms that are discovered in the research of neuroscience to machine intelligence remains unclear. Secondly, human intelligence has been proven to be versatile to a wide variety of capacities, e.g., abstraction, logic, understanding, self-awareness, learning, emotional knowledge, reasoning, planning, creativity, critical thinking, problem-solving, etc. It is unclear how to comprehensively measure human intelligence. There is no evidence thus far that machine intelligence can be a replacement for human intelligence in a wide range of real-world applications. Instead of diving into human brain neurons, Piaget studies human intelligence from the aspect of cognitive development along the key periods of growth. In Piaget's theory, two processes are closely related to human intelligence, that is, assimilation and accommodation. Assimilation aims to fit new information into existing cognitive schemas, while accommodation aims to take new information in one's environment and alter existing schemas to fit in the new information. We focus on three specific characteristics, i.e., generalizability, transferability, and trustworthiness, that center around assimilation and accommodation. Specifically, generalizability is an important yet generic concept in machine learning. Instead, we study the generalizability that takes place in the process of fitting new information associated to unknown classes into the knowledge w.r.t. known classes. Secondly, we explore how to transfer the knowledge learned from the source domain samples to the target domain with very few target-domain examples. Last but not least, there is still a gap between state-of-the-art learning-based approaches and a perfect one. Therefore, there is a critical need to understand the trustworthiness of machine intelligence.