Search results

Title: Image annotation and tag completion via kernel metric learning and noisy matrix recovery
Creator: Feng, Zheyun
Date: 2016
Collection: Electronic Theses & Dissertations
Description: In the last several years, with the ever-growing popularity of digital photography and social media, the number of images with user-provided tags has increased enormously. Due to the large amount and content versatility of these images, there is an urgent need to categorize, index, retrieve and browse these images via semantic tags (also called attributes or keywords). Following this trend, image annotation or tag completion out of missing and noisy given tags over large scale datasets has...
Show moreIn the last several years, with the ever-growing popularity of digital photography and social media, the number of images with user-provided tags has increased enormously. Due to the large amount and content versatility of these images, there is an urgent need to categorize, index, retrieve and browse these images via semantic tags (also called attributes or keywords). Following this trend, image annotation or tag completion out of missing and noisy given tags over large scale datasets has become an extremely hot topic in the interdisciplinary areas of machine learning and computer vision.The overarching goal of this thesis is to reassess the image annotation and tag completion algorithms that mainly capture the essential relationship both between and within images and tags even when the given tag information is incomplete or noisy, so as to achieve a better performance in terms of both effectiveness and efficiency in image annotation and other tag relevant tasks including tag completion, tag ranking and tag refinement.One of the key challenges in search-based image annotation models is to define an appropriate similarity measure (distance metric) between images, so as to assign unlabeled images with tags that are shared among similar labeled training images. Many kernel metric learning (KML) algorithms have been developed to serve as such a nonlinear distance metric. However, most of them suffer from high computational cost since the learned kernel metric needs to be projected into a positive semi-definite (PSD) cone. Besides, in image annotation tasks, existing KML algorithms require to convert image annotation tags into binary constraints, which lead to a significant semantic information loss and severely reduces the annotation performance.In this dissertation we propose a robust kernel metric learning (RKML) algorithm based on regression technique that is able to directly utilize the image tags. RKML is computationally efficient since the PSD property is automatically ensured by the regression technique. Numeric constraints over tags are also applied to better exploit the tag information and hence improve the annotation accuracy. Further, theoretical guarantees for RKML are provided, and its efficiency and effectiveness are also verified empirically by comparing it to state-of-the-art approaches of both distance metric learning and image annotation.Since the user-provided image tags are always incomplete and noisy, we also propose a tag completion algorithm by noisy matrix recovery (TCMR) to simultaneously enrich the missing tags and remove the noisy ones. TCMR assumes that the observed tags are independently sampled from unknown distributions that are represented by a tag matrix, and our goal is to recover that tag matrix based on the partially revealed tags which could be noisy. We provide theoretical guarantees for TCMR with recovery error bounds. In addition, a graph Laplacian based component is introduced to enforce the recovered tags to be consistent with the visual contents of images. Our empirical study with multiple benchmark datasets for image tagging shows that the proposed algorithm outperforms state-of-the-art approaches in terms of both effectiveness and efficiency when handling missing and noisy tags.
Show less

Title: Data clustering with pairwise constraints
Creator: Yi, Jinfeng
Date: 2014
Collection: Electronic Theses & Dissertations
Description: The classical unsupervised clustering is an ill-posed problem due to the absence of a unique clustering criteria. This issue can be addressed by introducing additional supervised information, usually casts in the form of pairwise constraints, to the clustering procedure. Depending on the sources, most pairwise constraints can be classified into two categories: (i) pairwise constraints collected from a set of non-expert crowd workers, which leads to the problem of crowdclustering, and (ii)...
Show moreThe classical unsupervised clustering is an ill-posed problem due to the absence of a unique clustering criteria. This issue can be addressed by introducing additional supervised information, usually casts in the form of pairwise constraints, to the clustering procedure. Depending on the sources, most pairwise constraints can be classified into two categories: (i) pairwise constraints collected from a set of non-expert crowd workers, which leads to the problem of crowdclustering, and (ii) pairwise constraints collected from oracle or experts, which leads to the problem of semi-supervised clustering. In both cases, the costs of collecting pairwise constraints can be expensive, thus it is important to identify the minimal number of pairwise constraints needed to accurately recover the underlying true data partition, also known as a sample complexity problem.In this thesis, we first analyze the sample complexity of crowdclustering. At first, we propose a novel crowdclustering approach based on the theory of matrix completion. Unlike the existing crowdclustering algorithm that is based on a Bayesian generative model, the proposed approach is more desirable since it only needs a much less number of crowdsourced pairwise annotations to accurately cluster all the objects. Our theoretical analysis shows that in order to accurately cluster $N$ objects, only $O(N\log^2 N)$ randomly sampled pairs should be annotated by crowd workers. To further reduce the sample complexity, we then introduce a semi-crowdsourced clustering framework that is able to effectively incorporate the low-level features of the objects to be clustered. In this framework, we only need to sample a subset of $n \ll N$ objects and generate their pairwise constraints via crowdsourcing. After completing a $n \times n$ similarity matrix using the proposed crowdclustering algorithm, we can further recover a $N \times N$ similarity matrix by applying a regression-based distance metric learning algorithm to the completed smaller size similarity matrix. This enables us to reliably cluster $N$ objects with only $O(n\log^2 n)$ crowdsourced pairwise constraints.Next, we study the problem of sample complexity in semi-supervised clustering. To this end, we propose a novel convex semi-supervised clustering approach based on the theory of matrix completion. In order to reduce the number of pairwise constraints needed %to achieve a perfect data partitioning,we apply a nature assumption that the feature representationsof the objects are able to reflect the similarities between objects. This enables us to only utilize $O(\log N)$ pairwiseconstraints to perfectly recover the data partition of $N$ objects.Lastly, in addition to sample complexity that relates to labeling costs, we also consider the computational costs of semi-supervised clustering.%In addition to sample complexity that relates to the labeling costs, we also consider the computational cost of semi-supervised clustering in the final part of this thesis.Specifically, we study the problem of efficiently updating clustering results when the pairwise constraints are generated sequentially, a common case in various real-world applications such as social networks. To address this issue, we develop a dynamic semi-supervised clustering algorithm that casts the clustering problem into a searching problem in a feasibleconvex space, i.e., a convex hull with its extreme points being an ensemble of multiple data partitions. Unlike classical semi-supervised clustering algorithms that need to re-optimize their objective functions when new pairwise constraints are generated, the proposed method only needs to update a low-dimensional vector and its time complexity is irrelevant to the number of data points to be clustered. This enables us to update large-scale clustering results in an extremely efficient way.
Show less

Title: Natural language based control and programming of robotic behaviors
Creator: Cheng, Yu (Graduate of Michigan State University)
Date: 2019
Collection: Electronic Theses & Dissertations
Description: "Robots have been transforming our daily lives by moving from controlled industrial lines to unstructured and dynamic environments such as home, offices, or outdoors working closely with human co-workers. Accordingly, there is an emerging and urgent need for human users to communicate with robots through natural language (NL) due to its convenience and expressibility, especially for the technically untrained people. Nevertheless, two fundamental problems remain unsolved for robots to working...
Show more"Robots have been transforming our daily lives by moving from controlled industrial lines to unstructured and dynamic environments such as home, offices, or outdoors working closely with human co-workers. Accordingly, there is an emerging and urgent need for human users to communicate with robots through natural language (NL) due to its convenience and expressibility, especially for the technically untrained people. Nevertheless, two fundamental problems remain unsolved for robots to working in such environments. On one hand, how to control robot behaviors in dynamic environments due to presence of people is still a daunting task. On the other hand, robot skills are usually preprogrammed while an application scenario may require a robot to perform new tasks. How to program a new skill to robots using NL on the fly also requires tremendous efforts. This dissertation tries to tackle these two problems in the framework of supervisory control. On the control aspect, it will be shown ideas drawn from dynamic discrete event systems can be used to model environmental dynamics and guarantee safety and stability of robot behaviors. Specifically, the procedures to build robot behavioral model and the criteria for model property checking will be presented. As there are enormous utterances in language with different abstraction level, a hierarchical framework is proposed to handle tasks lying in different logic depth. Behavior consistency and stability under hierarchy are discussed. On the programming aspect, a novel online programming via NL approach that formulate the problem in state space is presented. This method can be implemented on the fly without terminating the robot implementation. The advantage of such a method is that there is no need to laboriously labeling data for skill training, which is required by traditional offline training methods. In addition, integrated with the developed control framework, the newly programmed skills can also be applied to dynamic environments. In addition to the developed robot control approach that translates language instructions into symbolic representations to guide robot behaviors, a novel approach to transform NL instructions into scene representation is presented for robot behaviors guidance, such as robotic drawing, painting, etc. Instead of using a local object library or direct text-to-pixel mappings, the proposed approach utilizes knowledge retrieved from Internet image search engines, which helps to generate diverse and creative scenes. The proposed approach allows interactive tuning of the synthesized scene via NL. This helps to generate more complex and semantically meaningful scenes, and to correct training errors or bias. The success of robot behavior control and programming relies on correct estimation of task implementation status, which is comprised of robotic status and environmental status. Besides vision information to estimate environmental status, tactile information is heavily used to estimate robotic status. In this dissertation, correlation based approaches have been developed to detect slippage occurrence and slipping velocity, which provide grasp status to the high symbolic level and are used to control grasp force at lower continuous level. The proposed approaches can be used with different sensor signal type and are not limited to customized designs. The proposed NL based robot control and programming approaches in this dissertation can be applied to other robotic applications, and help to pave the way for flexible and safe human-robot collaboration."--Pages ii-iii.
Show less

Title: Learning algorithms for detecting disinformation on social media
Creator: VanDam, Courtland
Date: 2019
Collection: Electronic Theses & Dissertations
Description: Social media has become a widely accessible medium for users to share their opinions and details of their personal lives, including first hand accounts of emerging/disaster events, to a wide audience. However malicious entities may abuse users' trust to disseminate disinformation, i.e. false and misleading information. The disinformation disseminated on social media can have a significant impact offline. For example, fake news is suspected to have influenced the 2016 U.S. political election....
Show moreSocial media has become a widely accessible medium for users to share their opinions and details of their personal lives, including first hand accounts of emerging/disaster events, to a wide audience. However malicious entities may abuse users' trust to disseminate disinformation, i.e. false and misleading information. The disinformation disseminated on social media can have a significant impact offline. For example, fake news is suspected to have influenced the 2016 U.S. political election. Rumors on social media can mislead criminal investigations, e.g. the investigation of the 2013 Boston Bombing. To mitigate such impacts, automated detection of social media disinformation is thus an important research problem. This dissertation proposes algorithms to detect two approaches hackers use to disseminate disinformation-hashtag hijacking and compromising accounts. Hashtags are terms added to social media posts that are used to provide context to the posts, so those seeking to learn more about a given topic or event can search for posts containing related hashtags. However critics and attention-seeking trolls can mislead the public via hashtag hijacking. Hashtag hijacking occurs when one group of users takes control of a hashtag by using it in a different context than what was intended upon creation. Anyone can participate in hashtag hijacking, but to be successful, a coordinated effort among several accounts posting that hashtag is needed. This dissertation proposes HASHTECT, an unsupervised learning framework that uses a multi-modal nonnegative matrix factorization method for detecting hijacked hashtags. Experimental results on a large-scale Twitter data showed that HASHTECT is capable of detecting more hijacked hashtags than previously proposed algorithms. Another approach for disseminating disinformation is by compromising users' accounts. A social media account is compromised when it is accessed by a third party, i.e. hacker, without the genuine user's knowledge. Compromised accounts are damaging to the account holder as well as the accounts audience, e.g. followers. Hackers can damage the user's reputation, e.g. by posting hateful rhetoric. They also disseminate misleading information including rumors and malicious websites, e.g. phishing or malware. In this dissertation, I propose two compromised account detection algorithms, CADET and CAUTE. CADET is an unsupervised multi-view learning framework that employs nonlinear autoencoders to learn the feature embedding from multiple views. The rationale behind this approach is that an anomalous behavior observed in one view, e.g. abnormal time of day, may not indicate a compromised account. By aggregating the data from multiple views, CADET projects the features from all the views into a common lower-rank feature representation and detects compromised accounts in the shared subspace. On the other hand, CAUTE focuses on detecting compromised accounts early, by detecting the compromised posts. Given a user-post pair, CAUTE is a deep learning framework which simultaneously learns the encodings for the user as well as the post to detect whether the post was compromised, i.e. was written by a different user. By training a neural network on the residuals from the post and user encodings, CAUTE can classify whether a post is compromised with higher accuracy than several existing compromised account detection algorithms.
Show less

Title: Synthesizing realistic animated human motion using multiple natural spaces
Creator: Ferrydiansyah, Reza
Date: 2011
Collection: Electronic Theses & Dissertations
Description: When animating virtual human, it is important that the movements created are realistic as well as meet various constraint. One way to create motion, given a starting pose, is to first find an ending pose that meets the various constraints. This is followed by calculating the motion needed from the starting space to the ending space. Traditional inverse kinematics method are able to find poses that meets certain constraints, however these poses are not always natural. Linear interpolation...
Show moreWhen animating virtual human, it is important that the movements created are realistic as well as meet various constraint. One way to create motion, given a starting pose, is to first find an ending pose that meets the various constraints. This is followed by calculating the motion needed from the starting space to the ending space. Traditional inverse kinematics method are able to find poses that meets certain constraints, however these poses are not always natural. Linear interpolation between starting pose and ending pose can be used to create motion. Once again however, the interpolation method does not always create motion that is natural. This thesis proposes the creation of natural space. The natural space is a hyper-dimensional space in which every point in this space describes a natural pose. Motion can be created by traversing over the points in this space. The natural space is created by reducing the dimensionality of motion capture data using Principal Component Analysis (PCA). Points in the reduced space retain the characteristic of the original data. Multiple natural spaces are created on different segment of the human skeleton.This thesis describes a method to generate new constrained natural poses that is natural. The poses synthesized are more natural than traditional inverse kinematics, and single space PCA. Motion is created through a space consisting of pose configurations and angular speed. A method to generate realistic looking motion based on this space is given in this thesis.
Show less

Title: Some contributions to dimensionality reduction
Creator: Tong, Wei
Date: 2010
Collection: Electronic Theses & Dissertations
Description: Dimensionality reduction is a long standing challenging problem in the fields of statistical learning, pattern recognition and computer vision. Numerous algorithms have been proposed and studied in the past decades. In this dissertation we address several challenging problems emerged in recent studies of dimensionality reduction. We first explore the dimensionality reduction method for semi-supervised classification via the idea of mixed label propagation in which we attempt to find the best...
Show moreDimensionality reduction is a long standing challenging problem in the fields of statistical learning, pattern recognition and computer vision. Numerous algorithms have been proposed and studied in the past decades. In this dissertation we address several challenging problems emerged in recent studies of dimensionality reduction. We first explore the dimensionality reduction method for semi-supervised classification via the idea of mixed label propagation in which we attempt to find the best one dimensional embedding of the data in which data points in different classes can be well separated and the class labels are obtained by simply thresholding the one dimensional representation. In the next, we explore the dimensionality reduction methods for non-vector data representations. We first look into the problem in which a datum is represented by a matrix. We give a convex formulation to the problem of dimensionality reduction for matrices and developed an efficient approximating algorithm to solve the associated semi-definite programming problem. In the last, we studied the problem ofdimensionality reduction with even more challenging data representation, i.e., each datum is described as a different numberof unordered vectors. We convert the problem into the functional data dimensionality reduction and study it in the context of large scale image retrieval.
Show less

Title: Using dually optimal LCA features in sensory and action spaces for classification
Creator: Wagle, Nikita Nitin
Date: 2012
Collection: Electronic Theses & Dissertations
Description: Over years, a number of pattern recognition methods such as Linear Discriminant Analysis (LDA), Support Vector Machine (SVM), sparse auto-encoders, k-means clustering, etc. have been studied in image matching based on global and local templates of image features. The Developmental Network (DN), which uses Lobe Component Analysis (LCA) features, has been applied to spatiotemporal event detection and recognition in complex, cluttered backgrounds. However, the DN method has not been compared to...
Show moreOver years, a number of pattern recognition methods such as Linear Discriminant Analysis (LDA), Support Vector Machine (SVM), sparse auto-encoders, k-means clustering, etc. have been studied in image matching based on global and local templates of image features. The Developmental Network (DN), which uses Lobe Component Analysis (LCA) features, has been applied to spatiotemporal event detection and recognition in complex, cluttered backgrounds. However, the DN method has not been compared to well-known major techniques in the pattern recognition community for global and local template based matching problems. In this work, the experiments fall into two categories --- global template based object recognition and local template based scene classification. We apply the DN method to these problems and compare them to some widely used techniques in the pattern recognition community. The performance of the DN method is better or comparable to the global template based methods and comparable to some major local template based methods.
Show less

Title: Example-Based Parameterization of Linear Blend Skinning for Skinning Decomposition (EP-LBS
Creator: Hopkins, Kayra M.
Date: 2017
Collection: Electronic Theses & Dissertations
Description: This thesis presents Example-based Parameterization of Linear Blend Skinning for Skinning Decomposition (EP-LBS), a unified and robust method for using example data to simplify and improve the development and parameterization of high quality 3D models for animation. Animation and three-dimensional (3D) computer graphics have quickly become a popular medium for education, entertainment and scientific simulation. In addition to film, gaming and research applications, recent advancements in...
Show moreThis thesis presents Example-based Parameterization of Linear Blend Skinning for Skinning Decomposition (EP-LBS), a unified and robust method for using example data to simplify and improve the development and parameterization of high quality 3D models for animation. Animation and three-dimensional (3D) computer graphics have quickly become a popular medium for education, entertainment and scientific simulation. In addition to film, gaming and research applications, recent advancements in augmented reality (AR) and virtual reality (VR) are driving additional demand for 3D content. However, the success of graphics in these arenas depends greatly on the efficiency of model creation and the realism of the animation or 3D image.A common method for figure animation is skeletal animation using linear blend skinning (LBS). In this method, vertices are deformed based on a weighted sum of displacements due to an embedded skeleton. This research addresses the problem that LBS animation parameter computation, including determining the rig (the skeletal structure), identifying influence bones (which bones influence which vertices), and assigning skinning weights (amounts of influence a bone has on a vertex), is a tedious process that is difficult to get right. Even the most skilled animators must work tirelessly to design an effective character model and often find themselves repeatedly correcting flaws in the parameterization. Significant research, including the use of example-data, has focused on simplifying and automating individual components of the LBS deformation process and increasing the quality of resulting animations. However, constraints on LBS animation parameters makes automated analytic computation of the values equally as challenging as traditional 3D animation methods. Skinning decomposition is one such method of computing LBS animation LBS parameters from example data. Skinning decomposition challenges include constraint adherence and computationally efficient determination of LBS parameters.The EP-LBS method presented in this thesis utilizes example data as input to a least-squares non-linear optimization process. Given a model as a set of example poses captured from scan data or manually created, EP-LBS institutes a single optimization equation that allows for simultaneous computation of all animation parameters for the model. An iterative clustering methodology is used to construct an initial parameterization estimate for this model, which is then subjected to non-linear optimization to improve the fitting to the example data. Simultaneous optimization of weights and joint transformations is complicated by a wide range of differing constraints and parameter interdependencies. To address interdependent and conflicting constraints, parameter mapping solutions are presented that map the constraints to an alternative domain more suitable for nonlinear minimization. The presented research is a comprehensive, data-driven solution for automatically determining skeletal structure, influence bones and skinning weights from a set of example data. Results are presented for a range of models that demonstrate the effectiveness of the method.
Show less

Title: Learning from noisily connected data
Creator: Yang, Tianbao
Date: 2012
Collection: Electronic Theses & Dissertations
Description: Machine learning is a discipline of developing computational algorithms for learning predictive models from data. Traditional analytical learning methods treat the data as independent and identically distributed (i.i.d) samples from unknown distributions. However, this assumption is often violated in many real world applications that leading to the challenge of learning predictive models. For example, in electronic commerce website, customers could purchase a product by the recommendation of...
Show moreMachine learning is a discipline of developing computational algorithms for learning predictive models from data. Traditional analytical learning methods treat the data as independent and identically distributed (i.i.d) samples from unknown distributions. However, this assumption is often violated in many real world applications that leading to the challenge of learning predictive models. For example, in electronic commerce website, customers could purchase a product by the recommendation of their friends. Hence the purchasement records of customers are not i.i.d samples but correlated. Nowadays, data become correlated due to collaborations, interactions, communications, and many other types of connections. Effective learning from these connected data not only provides better understanding of the data but also brings significant economic benefits. How to learn from the connected data also brings unique challenges to both supervised learning and unsupervised learning algorithms because these algorithms are designed for i.i.d data and are often sensitive to the noise in the connected data. In this dissertation, I focus on developing theory and algorithms for learning from connected data. In particular, I consider two types of connections: the first type of connection is naturally formed in real wold networks, while the second type of connection is manually created to facilitate the learning process which is called must-and-cannot link. In the first part of this dissertation, I develop efficient algorithms for detecting communities in the first type of connected data. In the second part of this dissertation, I develop clustering algorithms that effectively utilize both must links and cannot links for the second type of connected dataA common approach toward learning from connected data is to assume that if two data points are connected, they are likely to be assigned to the same class/cluster. This assumption is often violated in real-word applications, leading to the noisy connection problems. One key challenge of learning from connected data is how to model the noisy pairwise connections that indicates the pairwise class-relationship between two data points. In the problem of detecting communities in networked data, I develop Bayesian approaches that explicitly model the noisy pairwise links by introducing additional hidden variables, besides community memberships, to explain potential inconsistency between the pairwise connections and pairwise class-relationship. In clustering must-and-cannot linked data, I will try to model how the noise is added into the pairwise connections in the manually generating process. The main contributions of this dissertation include (i) it introduces popularity and productivity for the first time besides the community memberships to model the generation of noisy links in real networks; the effectiveness of these factors is demonstrated through the task of community detection; (ii) it proposes a discriminative model for the first time that combines the content and link analysis together for detecting communities to alleviate the impact of noisy connections in community detection; (iii) it presents a general approach for learning from noisily labeled data, proves the theoretical convergence results for the first time and applies the approach in clustering noisy must-and-cannot linked data.
Show less

Title: Discrete vector and 2-tensor analyses and applications
Creator: Liu, Beibei
Date: 2015
Collection: Electronic Theses & Dissertations
Description: We present novel analysis methods for vector fields and an intrinsic representation of 2-tensor fields on meshes, and show the benefits they bring to discrete calculus, geometry processing, texture synthesis and fluid simulation. For instance, such vector fields and tensor fields in flat 2D space are necessary for example-based texture synthesis. However, many existing methods cannot ensure the continuity automatically or control the singularities accurately. Moreover, extending such analyses...
Show moreWe present novel analysis methods for vector fields and an intrinsic representation of 2-tensor fields on meshes, and show the benefits they bring to discrete calculus, geometry processing, texture synthesis and fluid simulation. For instance, such vector fields and tensor fields in flat 2D space are necessary for example-based texture synthesis. However, many existing methods cannot ensure the continuity automatically or control the singularities accurately. Moreover, extending such analyses to curved surfaces involves several known challenges. First, vectors at different surface points are defined in different tangent planes, so their comparison necessarily involves a concept calledconnection to transport vectors from one tangent plane to another in a parallel way. The few existing approaches for discrete connections offer neither a globally optimal principled definition nor a consistent disretization of differential operators. Second, symmetric 2-tensors, which play a crucial role in geometry processing, are often discretized as components stored in the predefined local frames. There is no convenient way to perform coordinate-independent computations with arbitrary 2-tensor fields on triangulated surface meshes. Finally, the persistent pursue for efficiency in the processing of vector fields in applications such as incompressible fluid simulation often results in undesired artifacts such as numerical viscosity, which prevents a predictive preview for the fine-resolution simulation at coarse spatial and temporal resolutions.We offer solutions to address these issues using our novel representation and analysis tools.First, we present a framework for example-based texture synthesis with feature alignment to vector fields with two way rotational symmetry, also known as orientation fields. Our contribution is twofold: a design tool for orientation fields with a natural boundary condition and singularity control, and a parallel texture synthesis adapted specifically for such fields in feature alignment.Second, we define discrete connection on triangle meshes, which involves closed-form expressions within edges and triangles and finite rotations between pairs of incident vertices, edges, or triangles. The finite set of parameters of this connection can be optimally computed by minimizing a quadratic measure of the deviation from the connection induced by the embedding of the input triangle mesh. Local integrals of other first-order derivatives as well as the L2-based energies can also be computed.Third, we offer a coordinate-free representation of arbitrary 2-tensor fields on triangle meshes, where we leverage a decomposition of continuous 2-tensors in the plane to construct a finite-dimensional encoding of tensor fields through scalar values on oriented pieces of a manifold triangulation. We also provide closed-form expressions of common operators for tensor fields, including pairing, inner product, and trace for this discrete representation, and formulate a discrete covariant derivative induced by the 2-tensors instead of the metric of the surface. Other operators, such as discrete Lie bracket, can be constructed based on these operators. This approach extends computational tools for tensor fields and offers a numerical framework for discrete tensor calculus on triangulations.Finally, a spectral vector field calculus on embeded irregular shape is introduced to build a model-reduced variational Eulerian integrator for incompressible fluid. The resulting simulation combines the efficiency gains of dimension reduction, the qualitative robustness to coarse spatial and temporal resolutions of geometric integrators, and the simplicity of sub-grid accurate boundary conditions on regular grids to deal with arbitrarily-shaped domains. A functional map approach to fluid simulation is also proposed, where scalar-valued and vector-valued eigenfunctions of the Laplacian operator can be easily used as reduced bases. Using a variational integrator in time topreserve liveliness and a simple, yet accurate embedding of the fluid domain onto a Cartesian grid, our model-reduced fluid simulator can achieve realistic animations in significantly less computation time than full-scale non-dissipative methods but without the numerical viscosity from which current reduced methods suffer.
Show less

Title: A probabilistic topic modeling approach for event detection in social media
Creator: VanDam, Courtland
Date: 2012
Collection: Electronic Theses & Dissertations
Description: Social media services, such as Twitter, have become a prominent source of information for event detection and monitoring applications as they provide access to massive volume of dynamic user content. Previous studies have focused on detecting a variety of events from Twitter feeds, including natural disasters such as earthquakes and hurricanes and entertainment events, such as sporting events and music festivals. A key challenge to event detection from Twitter is identifying user posts, or...
Show moreSocial media services, such as Twitter, have become a prominent source of information for event detection and monitoring applications as they provide access to massive volume of dynamic user content. Previous studies have focused on detecting a variety of events from Twitter feeds, including natural disasters such as earthquakes and hurricanes and entertainment events, such as sporting events and music festivals. A key challenge to event detection from Twitter is identifying user posts, or tweets, that are relevant to the monitored event. Current approaches can be grouped into three categories---keyword filtering, supervised classification, and topic modeling. Keyword filtering is the simplest approach but it tends to produce a high false positive rate. Supervised classification approaches apply generic classifiers, such as support vector machine (SVM), to determine if a tweet is related to the event of interest. Their performance depends on the quality of features used to represent the data. Topic modeling approaches such as latent Dirichlet allocation (LDA) can automatically infer the latent topics within the tweets. However, due to the unsupervised nature of the algorithm, they are not as effective as supervised learning approaches. The approach developed in this thesis combines probabilistic topic modeling with supervised classification to leverage the advantages from each approach. This supervised topic modeling approach, called subtopicLDA, utilizes label information to help guide the topic model to select topics that best fit the label information. The model is evaluated for its effectiveness in detecting foodborne illness related tweets.
Show less

Title: Early linguistic environments and language development in children with cochlear implants
Creator: Khalil Arjmandi, Meisam
Date: 2019
Collection: Electronic Theses & Dissertations
Description: Prior research has documented tremendous variability in language outcomes of children with cochlear implants (CIs); despite more than a decade of research, a large portion of this variability remains unexplained. This study characterized the quantity and quality of early linguistic input in naturalistic environments of 14 early-implanted children with CIs to investigate variability across children as a possible source of variation that might help explain language outcome variability. In...
Show morePrior research has documented tremendous variability in language outcomes of children with cochlear implants (CIs); despite more than a decade of research, a large portion of this variability remains unexplained. This study characterized the quantity and quality of early linguistic input in naturalistic environments of 14 early-implanted children with CIs to investigate variability across children as a possible source of variation that might help explain language outcome variability. In Chapter 2, daylong audio recordings from home environments of these children were analyzed to examine individual variability in language input they experienced in terms of lexical, morphosyntactic, and social-pragmatic properties. It was found that the quantity and quality of early language input varies substantially across children with CIs, where the degree of variability was comparable in magnitude to what has been reported between the most-advantaged and least-advantaged typically hearing children. In Chapter 3, estimates of the quantity and quality of language input were adjusted to consider environmental noise and reverberation to better represent the "useable" amount of input experienced by the children. It was found that children with CIs are differentially impacted by noise and reverberation in their naturalistic environments, such that some children are doubly disadvantaged in acquiring spoken language, both due to substantial variability in the amount and quality of linguistic input available to them, as well as due to their exposure and susceptibility to environmental noise and reverberation. In Chapter 4, an initial test was conducted to obtain preliminary results regarding how language input factors might shape development of language outcomes in children with CIs. The preliminary results estimating the contribution of language input measures to language outcomes of the children with CIs suggested that measure of speech intelligibility tailored to children with CIs strongly predicted language outcomes. Overall, this study has provided evidence for substantial individual variability across children with cochlear implants in quantity and quality of their early language experience, which were mainly influenced by factors of child-directed speech and environmental noise and reverberation. This evidence-based knowledge can be used by parents and clinicians to effectively adjust early linguistic environments of children with CIs to maximize the advantage of using CIs.
Show less

Title: Hierarchical learning for large multi-class classification in network data
Creator: Liu, Lei
Date: 2015
Collection: Electronic Theses & Dissertations
Description: Multi-class learning from network data is an important but challenging problem with many applications, including malware detection in computer networks, user modeling in social networks, and protein function prediction in biological networks. Despite the extensive research on large multi-class learning, there are still numerous issues that have not been sufficiently addressed, such as efficiency of model testing, interpretability of the induced models, as well as the ability to handle...
Show moreMulti-class learning from network data is an important but challenging problem with many applications, including malware detection in computer networks, user modeling in social networks, and protein function prediction in biological networks. Despite the extensive research on large multi-class learning, there are still numerous issues that have not been sufficiently addressed, such as efficiency of model testing, interpretability of the induced models, as well as the ability to handle imbalanced classes. To overcome these challenges, there has been increasing interest in recent years to develop hierarchical learning methods for large multi-class problems. Unfortunately, none of them were designed for classification of network data. In addition, there are very few studies devoted to classification of heterogeneous networks, where the nodes may have different feature sets. This thesis aims to overcome these limitations with the following contributions.First, as the number of classes in big data applications can be very large, ranging from thousands to possibly millions, two hierarchical learning schemes are proposed to deal with the so-called extreme multi-class learning problems. The first approach, known as recursive non-negative matrix factorization (RNMF), is designed to achieve sublinear runtime in classifying test data. Although RNMF reduces the test time significantly, it may also assign the same class to multiple leaf nodes, which hampers the interpretability of the model as a concept hierarchy for the classes. Furthermore, since RNMF employs a greedy strategy to partition the classes, there is no theoretical guarantee that the partitions generated by the tree would lead to a globally optimal solution.To address the limitations of RNMF, an alternative hierarchical learning method known as matrix factorization tree (MF-Tree) is proposed. Unlike RNMF, MF-tree is designed to optimize a global objective function while learning its taxonomy structure. A formal proof is provided to show the equivalence between the objective function of MF-tree and the Hilbert-Schmidt Independence Criterion (HSIC). Furthermore, to improve its training efficiency, a fast algorithm for inducing approximate MF-Tree is also developed.Next, an extension of MF-Tree to network data is proposed. This approach can seamlessly integrate both the link structure and node attribute information into a unified learning framework. To the best of our knowledge, this is the first study that automatically constructs a taxonomy structure to predict large multi-class problems for network classification. Empirical results suggest that the approach can effectively capture the relationship between classes and generate class taxonomy structures that resemble those produced by human experts. The approach can also be easily parallelizable and has been implemented in a MapReduce framework.Finally, we introduce a network learning task known as co-classification to classify heterogeneous nodes in multiple networks. Unlike existing node classification problems, the goal of co-classification is to learn the classifiers in multiple networks jointly, instead of learning to classify each network independently. The framework proposed in this thesis can utilize prior information about the relationship between classes in different networks to improve its prediction accuracy.
Show less

Title: Towards interpretable face recognition
Creator: Yin, Bangjie
Date: 2019
Collection: Electronic Theses & Dissertations
Description: Deep CNNs have been pushing the frontier of visual recognition over past years. Besides recognition accuracy, strong demands in understanding deep CNNs in the research community motivate developments of tools to dissect pre-trained models to visualize how they make predictions. Recent works further push the interpretability in the network learning stage to learn more meaningful representations. In this work, focusing on a specific area of visual recognition, we report our efforts towards...
Show moreDeep CNNs have been pushing the frontier of visual recognition over past years. Besides recognition accuracy, strong demands in understanding deep CNNs in the research community motivate developments of tools to dissect pre-trained models to visualize how they make predictions. Recent works further push the interpretability in the network learning stage to learn more meaningful representations. In this work, focusing on a specific area of visual recognition, we report our efforts towards interpretable face recognition. We propose a spatial activation diversity loss to learn more structured face representations. By leveraging the structure, we further design a feature activation diversity loss to push the interpretable representations to be discriminative and robust to occlusions. We demonstrate on three face recognition benchmarks that our proposed method is able to achieve the state-of-art face recognition accuracy with easily interpretable face representations.
Show less

Title: Automated Speaker Recognition in Non-ideal Audio Signals Using Deep Neural Networks
Creator: Chowdhury, Anurag
Date: 2021
Collection: Electronic Theses & Dissertations
Description: Speaker recognition entails the use of the human voice as a biometric modality for recognizing individuals. While speaker recognition systems are gaining popularity in consumer applications, most of these systems are negatively affected by non-ideal audio conditions, such as audio degradations, multi-lingual speech, and varying duration audio. This thesis focuses on developing speaker recognition systems robust to non-ideal audio conditions.Firstly, a 1-Dimensional Convolutional Neural...
Show moreSpeaker recognition entails the use of the human voice as a biometric modality for recognizing individuals. While speaker recognition systems are gaining popularity in consumer applications, most of these systems are negatively affected by non-ideal audio conditions, such as audio degradations, multi-lingual speech, and varying duration audio. This thesis focuses on developing speaker recognition systems robust to non-ideal audio conditions.Firstly, a 1-Dimensional Convolutional Neural Network (1D-CNN) is developed to extract noise-robust speaker-dependent speech characteristics from the Mel Frequency Cepstral Coefficients (MFCC). Secondly, the 1D-CNN-based approach is extended to develop a triplet-learning-based feature-fusion framework, called 1D-Triplet-CNN, for improving speaker recognition performance by judiciously combining MFCC and Linear Predictive Coding (LPC) features. Our hypothesis rests on the observation that MFCC and LPC capture two distinct aspects of speech: speech perception and speech production. Thirdly, a time-domain filterbank called DeepVOX is learned from vast amounts of raw speech audio to replace commonly-used hand-crafted filterbanks, such as the Mel filterbank, in speech feature extractors. Finally, a vocal style encoding network called DeepTalk is developed to learn speaker-dependent behavioral voice characteristics to improve speaker recognition performance. The primary contribution of the thesis is the development of deep learning-based techniques to extract discriminative, noise-robust physical and behavioral voice characteristics from non-ideal speech audio. A large number of experiments conducted on the TIMIT, NTIMIT, SITW, NIST SRE (2008, 2010, and 2018), Fisher, VOXCeleb, and JukeBox datasets convey the efficacy of the proposed techniques and their importance in improving speaker recognition performance in non-ideal audio conditions.
Show less

MSU Libraries

Digital Repository

Enabled Filters

MSU Libraries

Digital Repository

You are here

Search results