Repository logo

Browsing by Author "King, Emily, committee member"

Now showing 1 - 6 of 6
  • Thumbnail Image
    ItemOpen Access
    Compound-Gaussian-regularized inverse problems: theory, algorithms, and neural networks
    (Colorado State University. Libraries, 2024) Lyons, Carter, author; Cheney, Margaret, advisor; Raj, Raghu G., advisor; Azimi, Mahmood, committee member; King, Emily, committee member; Mueller, Jennifer, committee member
    Linear inverse problems are frequently encountered in a variety of applications including compressive sensing, radar, sonar, medical, and tomographic imaging. Model-based and data-driven methods are two prevalent classes of approaches used to solve linear inverse problems. Model-based methods incorporate certain assumptions, such as the image prior distribution, into an iterative estimation algorithm, often, as an example, solving a regularized least squares problem. Instead, data-driven methods learn the inverse reconstruction mapping directly by training a neural network structure on actual signal and signal measurement pairs. Alternatively, algorithm unrolling, a recent approach to inverse problems, combines model-based and data-driven methods through the implementation of an iterative estimation algorithm as a deep neural network (DNN). This approach offers a vehicle to embed domain-level and algorithmic insights into the design of neural networks such that the network layers are interpretable. The performance, in reconstructed signal quality, of unrolled DNNs often exceeds that of corresponding iterative algorithms and standard DNNs while doing so in a computationally efficient fashion. In this work, we leverage algorithm unrolling to combine a powerful statistical prior, the compound Gaussian (CG) prior, with the powerful representational ability of machine learning and DNN approaches. Specifically, first we construct a novel iterative CG-regularized least squares algorithm for signal reconstruction and provide a computational theory for this algorithm. Second, using algorithm unrolling, the newly developed CG-based least squares iterative algorithm is transformed into an original DNN in a manner to facilitate the learning of the optimization landscape geometry. Third, a generalization on the newly constructed CG regularized least squares iterative algorithm is developed, theoretically analyzed, and unrolled to yield a novel state-of-the-art DNN that provides a partial learning of the prior distribution constrained to the CG class of distributions. Fourth, techniques in statistical learning theory are employed for deriving original generalization error bounds on both unrolled DNNs to substantiate theoretical guarantees of each neural network when estimating signals from linear measurements after training. Finally, ample numerical experimentation is conducted for every new CG-based iterative and DNN approach proposed in this paper. Simulation results show our methods outperform previous state-of-the-art iterative signal estimation algorithms and deep-learning-based methods, especially with limited training datasets.
  • Thumbnail Image
    ItemOpen Access
    GAN you train your network
    (Colorado State University. Libraries, 2022) Pamulapati, Venkata Sai Sudeep, author; Blanchard, Nathaniel, advisor; Beveridge, Ross, advisor; King, Emily, committee member
    Zero-shot classifiers identify unseen classes — classes not seen during training. Specifically, zero-shot models classify attribute information associated with classes (e.g., a zebra has stripes but a lion does not). Lately, the usage of generative adversarial networks (GAN) for zero-shot learning has significantly improved the recognition accuracy of unseen classes by producing visual features on any class. Here, I investigate how similar visual features obtained from images of a class are to the visual features generated by a GAN. I find that, regardless of metric, both sets of visual features are disjointed. I also fine-tune a ResNet so that it produces visual features that are similar to the visual features generated by a GAN — this is novel because all standard approaches do the opposite: they train the GAN to match the output of the model. I conclude that these experiments emphasize the need to establish a standard input pipeline in zero-shot learning because of the mismatch of generated and real features, as well as the variation in features (and subsequent GAN performance) from different implementations of models such as ResNet-101.
  • Thumbnail Image
    ItemOpen Access
    Perception systems for robust autonomous navigation in natural environments
    (Colorado State University. Libraries, 2022) Trabelsi, Ameni, author; Beveridge, Ross J., advisor; Blanchard, Nathaniel, committee member; Anderson, Chuck, committee member; King, Emily, committee member
    As assistive robotics continues to develop thanks to the rapid advances of artificial intelligence, smart sensors, Internet of Things, and robotics, the industry began introducing robots to perform various functions that make humans' lives more comfortable and enjoyable. While the principal purpose of deploying robots has been productivity enhancement, their usability has widely expanded. Examples include assisting people with disabilities (e.g., Toyota's Human Support Robot), providing driver-less transportation (e.g., Waymo's driver-less cars), and helping with tedious house chores (e.g., iRobot). The challenge in these applications is that the robots have to function appropriately under continuously changing environments, harsh real-world conditions, deal with significant amounts of noise and uncertainty, and operate autonomously without the intervention or supervision of an expert. To meet these challenges, a robust perception system is vital. This dissertation casts light on the perception component of autonomous mobile robots and highlights their major capabilities, and analyzes the factors that affect their performance. In short, the developed approaches in this dissertation cover the following four topics: (1) learning the detection and identification of objects in the environment in which the robot is operating, (2) estimating the 6D pose of objects of interest to the robot, (3) studying the importance of the tracking information in the motion prediction module, and (4) analyzing the performance of three motion prediction methods, comparing their performances, and highlighting their strengths and weaknesses. All techniques developed in this dissertation have been implemented and evaluated on popular public benchmarks. Extensive experiments have been conducted to analyze and validate the properties of the developed methods and demonstrate this dissertation's conclusions on the robustness, performance, and utility of the proposed approaches for intelligent mobile robots.
  • Thumbnail Image
    ItemOpen Access
    Persistence stability for metric thickenings
    (Colorado State University. Libraries, 2021) Moy, Michael, author; Adams, Henry, advisor; King, Emily, committee member; Ben-Hur, Asa, committee member
    Persistent homology often begins with a filtered simplicial complex, such as the Vietoris–Ripscomplex or the Čech complex, the vertex set of which is a metric space. An important result, the stability of persistent homology, shows that for certain types of filtered simplicial complexes, two metric spaces that are close in the Gromov–Hausdorff distance result in persistence diagrams that are close in the bottleneck distance. The recent interest in persistent homology has motivated work to better understand the homotopy types and persistent homology of these commonly used simplicial complexes. This has led to the definition of metric thickenings, which agree with simplicial complexes for finite vertex sets but may have different topologies for infinite vertex sets. We prove Vietoris–Rips metric thickenings and Čech metric thickenings have the same persistence diagrams as their corresponding simplicial complexes for all totally bounded metric spaces. This immediately implies the stability of persistent homology for these metric thickenings.
  • Thumbnail Image
    ItemOpen Access
    Subspace and network averaging for computer vision and bioinformatics
    (Colorado State University. Libraries, 2023) Mankovich, Nathan J., author; Kirby, Michael, advisor; Peterson, Chris, committee member; King, Emily, committee member; Anderson, Charles, committee member
    Finding a central prototype (a.k.a. average) from a cluster of points in high dimensional space has broad applications to complex problems like action clustering in computer vision or gene co-expression module representation in bioinformatics. A central prototype of a set of points may be cast as the solution to an optimization problem that either minimizes distance or maximizes similarity between the prototype and each point in the cluster. In this dissertation we offer four novel prototypes for a cluster of points: the flag median, maximally correlated flag, cluster expression vector and eigengene subspace. We will formalize the flag median and the maximally correlated flag using subspace representations for data, specifically the Grasmann and flag manifolds. In addition to introducing these prototypes, we will derive a novel algorithm which can be used to calculate subspace prototypes: FlagIRLS. The third and fourth prototypes, the cluster expression vector and eigengene subspace, are inspired by problems involving gene cluster (e.g., pathway or module) representations. The cluster expression vector leverages connections within networks of genes whereas the eigengene subspace is computed using Principal Component Analysis (PCA). In this work we will explore the theoretical under-pinnings of these prototypes, find algorithms to compute and them to computer vision and biological data sets.
  • Thumbnail Image
    ItemOpen Access
    Utilizing network features to detect erroneous inputs
    (Colorado State University. Libraries, 2020) Gorbett, Matthew, author; Blanchard, Nathaniel, advisor; Anderson, Charles W., committee member; King, Emily, committee member
    Neural networks are vulnerable to a wide range of erroneous inputs such as corrupted, out-of-distribution, misclassified, and adversarial examples. Previously, separate solutions have been proposed for each of these faulty data types; however, in this work I show that the collective set of erroneous inputs can be jointly identified with a single model. Specifically, I train a linear SVM classifier to detect these four types of erroneous data using the hidden and softmax feature vectors of pre-trained neural networks. Results indicate that these faulty data types generally exhibit linearly separable activation properties from correctly processed examples. I am able to identify erroneous inputs with an AUROC of 0.973 on CIFAR10, 0.957 on Tiny ImageNet, and 0.941 on ImageNet. I experimentally validate the findings across a diverse range of datasets, domains, and pre-trained models.