Doctor of Philosophy (PhD), Ohio University, 2024, Civil Engineering (Engineering and Technology)

Colloidal aggregation is a critical phenomenon influencing various environmental processes. However, limited research has been conducted on the aggregation of particles with heterogeneous physical and chemical properties, which are more representative of practical environmental systems than homogeneous particles. The central hypothesis of this dissertation is that primary particle size polydispersity along with chemical and material heterogeneity of primary particles exert non-trivial effects on the aggregate growth rate and the fractal dimensions of aggregates. In this dissertation, the aggregation and stability of heterogeneous nano- and sub-micrometer particles in aqueous systems were investigated using in situ microscopy and image analysis. Initially, the study examined the growth kinetics and structures of aggregates formed by polystyrene microplastics in mono- and bidisperse systems. Findings indicated that while the primary particle size distribution did not affect the scaling behavior of aggregate growth, it delayed the onset of rapid aggregation. Structural analysis revealed a power law dependence of the aggregate fractal dimension in both mono- and bidisperse systems, with mean fractal dimensions consistent with aggregates from diffusion-limited cluster aggregation. The results also suggested that aggregate fractal dimension was insensitive to shape anisotropy. The dissertation further explored the structure of DLCA aggregates in heterogeneous systems composed of particles with varying sizes, surface charges, and material compositions. The fractal dimensions of DLCA aggregates in these heterogeneous particle systems were similar, ranging from 1.6 to 1.7, and consistent with theoretical predictions and experimental evidence for homogeneous DLCA aggregates. This confirmed the universality of aggregate structures in the DLCA regime, regardless of particle composition. Additionally, a scaling relationship was demonstrated between aggregat (open full item for complete abstract)

Committee: Lei Wu (Advisor); Guy Riefler (Committee Member); Daniel Che (Committee Member); Sumit Sharma (Committee Member); Natalie Kruse Daniels (Committee Member) Subjects: Chemical Engineering; Civil Engineering; Environmental Engineering; Physical Chemistry

PhD, University of Cincinnati, 2024, Engineering and Applied Science: Computer Science and Engineering

The field of medicine has witnessed remarkable progress in recent years, largely owing to the technological advancements in machine learning and deep learning frameworks. The healthcare industry has been a significant contributor to this massive influx of data, generating approximately 30% of the world's data volume. While data mining has been a crucial tool for discovering hidden patterns from data and to extract valuable insights. Effective feature learning, on the other hand, plays an important role in the performance of machine learning models in attaining increased predictive accuracies and learning efficiencies. This research aims to understand and explore the feature selection techniques on both clinical data and medical image analysis, and to attain comprehensive insights into image understanding (IU) by focusing on the segmentation methods for both object recognition and scene classification. The first part of this research studies two feature selection approaches namely, the principal component analysis (PCA) and particle swarm optimization (PSO) on the clinical data using Wisconsin Diagnostic Breast Cancer (WDBC) dataset, to study and extract the top features and evaluate the predictive performances across the five of the most widely used supervised classification algorithms. It is inferred that the study yields significant insights into the effectiveness and efficiency of various classification algorithms for predicting breast cancer type. In the context of PCA, it is imperative to have a good understanding of how features may have a positive/negative impact on the PCs. The study emphasizes the critical role of feature selection in enhancing classification accuracy. The second part of this research delves into IU, as it plays a pivotal role in various computer vision tasks, such as extraction of essential features from images, object detection, and segmentation. At a higher level of granularity, both semantic and instance segmentation are (open full item for complete abstract)

Committee: Anca Ralescu Ph.D. (Committee Chair); Wen-Ben Jone Ph.D. (Committee Member); Chong Yu Ph.D. (Committee Member); Dan Ralescu Ph.D. (Committee Member); Boyang Wang Ph.D. (Committee Member) Subjects: Computer Science

PhD, University of Cincinnati, 2023, Engineering and Applied Science: Computer Science and Engineering

Deep learning techniques are being unified for decision support in various applications. However, it remains challenging to train robust deep learning models, due to the inherent insufficient labeled data that is usually time-consuming and labor-intensive. Self-supervised learning is a feature representation learning paradigm to learn robust features from insufficient annotated datasets. It contains two types of task stages, including the pretext task and the downstream task. The model is typically pre-trained with the pretext task in an unsupervised manner, where the data itself provides supervision. Afterward, the model is fine-tuned in a real downstream supervised task. Although self-supervised learning can effectively learn the robust latent feature representations and reduce human annotation efforts, it highly relies on designing efficient pretext tasks. Therefore, studying effective pretext tasks is desirable to learn more effective features and further improve the model prediction performance for decision support. In self-supervised learning, pretext tasks with deep metric/contrastive learning styles received more and more attention, as the learned distance representations are useful to capture the similarity relationship among samples and further improve the performance of various supervised or unsupervised learning tasks. In this dissertation proposal, we survey the recent state-of-the-art self-supervised learning methods and propose several new deep metric and contrastive learning strategies for learning effective features. Firstly, we propose a new deep metric learning method for image recognition. The proposed method learns an effective distance metric from both geometric and probabilistic space. Secondly, we develop a novel contrastive learning method using the Bregman divergence, extending the contrastive learning loss function into a more generalized divergence form, which improves the quality of self-supervised learned feature representation. (open full item for complete abstract)

Committee: Anca Ralescu Ph.D. (Committee Chair); Kenneth Berman Ph.D. (Committee Member); Dan Ralescu Ph.D. (Committee Member); Lili He Ph.D. (Committee Member); Yizong Cheng Ph.D. (Committee Member) Subjects: Computer Science

Master of Science in Mechanical Engineering, Cleveland State University, 2023, Washkewicz College of Engineering

In structural design, the distribution of wind-induced pressure exerted on structures is crucial. The pressure distribution for a particular building is often determined by scale model tests in boundary layer wind tunnels (BLWTs). For all combinations of interesting building shapes and wind factors, experiments with BLWTs must be done. Resource or physical testing restrictions may limit the acquisition of needed data because this procedure might be time- and resource-intensive. Finding a trustworthy method to cyber-enhance data-collecting operations in BLWTs is therefore sought. This research analyzes how machine learning approaches may improve traditional BLWT modeling to increase the information obtained from tests while proportionally lowering the work needed to complete them. The more general question centers on how a machine learning-enhanced method ultimately leads to approaches that learn as data are collected and subsequently optimize the execution of experiments to shorten the time needed to complete user-specified objectives. 3 Different Machine Learning models, namely, Support vector regressors, Gradient Boosting regressors, and Feed Forward Neural networks were used to predict the surface Averaged Mean pressure coefficients cp on Isolated high-rise buildings. The models were trained to predict average cp for missing angles and also used to train for varying dimensions. Both global and local approaches to training the models were used and compared. The Tokyo Polytechnic University's Aerodynamic Database for Isolated High-rise buildings was used to train all the models in this study. Local and global prediction approaches were used for the DNN and GBRT models and no considerable difference has been found between them. The DNN model showed the best accuracy with (R2 > 99%, MSE < 1.5%) among the used models for both missing angles and missing dimensions, and the other two models also showed high accuracy with (R2 > 97%, MSE < 4%).

Committee: Navid Goudarzi (Committee Chair); Prabaha Sikder (Committee Member); Mustafa Usta (Committee Member) Subjects: Artificial Intelligence; Design; Engineering; Urban Planning

Doctor of Philosophy, Case Western Reserve University, 2023, Materials Science and Engineering

A deeper understanding of processing-structure relationships has been developed with the goal of building single crystal devices using pulsed laser deposition (PLD) and advancing the application of data science to materials science. The targeted device was a half-cell lithium-ion battery, where strontium ruthenate (SRO) is the current collector, lithium cobalt oxide (LCO) is the cathode, and lithium lanthanum titanate (LLTO) is the electrolyte. These were grown on a strontium titanate (STO) substrate. Through studies of the processing parameters and film characteristics, conditions to grow a single crystal LCO/SRO/STO heterostructure were revealed. While the addition of the electrolyte affected the single crystal structure and interfacial quality, underlying reasons have been illuminated to guide further development of multi-layer oxide heterostructures. An in-situ technique called reflection h igh energy electron diffraction (RHEED) is commonly coupled with PLD to provide information on structure-property relationships by recording the diffraction pattern of the film during growth. Traditionally, a small percentage of the data provided is used in analysis. Here data science techniques are applied, both supervised and unsupervised, to reveal additional information from the full data set. As a result, the sensitivity of the length of diffraction spots over other parameters (e.g., width or intensity) to growth characteristics has been uncovered, especially in later stages of growth where the data is dominated by the reflection from the film. Additionally, through unsupervised learning, a phase shift in the intensity oscillations of different RHEED spots was uncovered. Non-negative matrix factorization among other techniques was used to deconvolute information from different diffraction spots. It was revealed that (01) and (0-1) spots are better indicators of thin film growth characteristics especially in material systems that grow in layer-by-layer or step-flow mechan (open full item for complete abstract)

Committee: Alp Sehirlioglu Dr. (Advisor); Xuan Gao Dr. (Committee Member); Roger French Dr. (Committee Member); Frank Ernst Dr. (Committee Member) Subjects: Chemical Engineering; Chemistry; Computer Science; Engineering; Materials Science; Statistics

Master of Science, The Ohio State University, 2021, Aerospace Engineering

Small unmanned aerial systems (sUAS) pose an ever-increasing threat to low-altitude aircraft, particularly those unequipped with automatic dependent surveillance – broadcast (ADS-B) trackers. Surveillance and tracking systems must be developed to find these low-flying aircraft and alert remote pilots to their location, especially for beyond visual line of sight (BVLOS) operations. Radar is introduced as one such means for this surveillance that is independent of the sUAS and aircraft systems operating within the surveillance volume. The downside to radar at such extremely low-altitude environments, however, is the large volume of radar “clutter” returns that arise from ground vehicles, weather, birds, trees, etc. For radar to be useful as the surveillance method of choice, these clutter returns must be filtered out of the radar feeds so remote pilots have the clearest and most complete picture of the surrounding airspace possible. This work proposes a neural network classifier as the tool of choice in decluttering the radar feeds. A neural network algorithm requires no known models of aircraft and clutter behavior inside the surveillance volume. The only requirement for training the model is a labeled set of radar returns spanning a few days. This thesis describes the process for building this classification model, while also providing a method for labeling radar data, building a feature set for the neural network to learn from, and supplementing training data due to the lack of physical radar tracks for specific use cases. The proposed model in this work is shown to reduce clutter in the radar feeds by over 80% in most cases, while correctly identifying more than 94% of aircraft. This research lays the groundwork for future real-time classification algorithms and provides insight and suggestions for means of future improvement.

Committee: Jeffrey Bons (Advisor); Matthew McCrink (Committee Member); Jim Gregory (Committee Member) Subjects: Aerospace Engineering

MS, University of Cincinnati, 2021, Education, Criminal Justice, and Human Services: Information Technology

Crime hotspot locations identification is a very important endeavor to help ensure public safety. Been able to identify these locations effectively and accurately will help provide useful information to law enforcement bodies to help minimize criminal activities. Considering the limited resources available to law enforcements, a more prudent approach will be to deploy these resources at places that record a considerable higher crime rate. We depart from the traditional “higher than” average thresholds and rather rely on a more pragmatic approach in the analysis. We analyze a five-year crime data from the Cincinnati Police Department using clustering algorithms such K-means, DBSCAN, Hierarchical algorithms, and classification machine learning algorithms such as Random Forest, SVM, Logistic Regression, KNN, and Naive Bayes, on the same dataset. The clustering methods are used as a standalone means of identifying crime hotspots rather than used as a data preprocessing step as done in prior experiments. The results from both approaches are compared using their respective evaluation metrics. From the performances, we find classification performed better than clustering on our dataset. The best performing algorithm is the Random Forest when the number of trees is 30. We also find considerable crime concentration along the hotspot street segments that were identified in the dataset.

Committee: M. Murat Ozer Ph.D. (Committee Chair); Nelly Elsayed Ph.D. (Committee Member) Subjects: Information Technology

Master of Science, The Ohio State University, 2021, Food, Agricultural and Biological Engineering

According to the United States Department of Agriculture, projections show that food production needs to increase by 70-100% from 2010 – 2050 due to population growth in addition to other socioeconomic pressures. New methods are needed to increase the productivity and efficiency of agricultural systems and are critical for mitigating climate change and ensuring food security. Remote sensing (RS) and Unmanned Aerial Systems (UAS) have the potential to allow agricultural researchers to better manage and monitor these complex systems. Foliage temperature is a key variable in biophysical vegetative modeling and has been well documented to be an indicator of crop water stress. The ability to monitor subtle changes in foliage temperature using a calibrated thermal infrared (TIR) camera mounted on a UAS would open avenues for field-based stress monitoring at scales not possible without using airborne systems. However, current approaches to process thermal image data are time-consuming, inaccurate, or not well suited for foliage in field environments. And importantly, methods to extract foliage pixels from the background (i.e. soil, weeds, etc) are needed to remove the influence of background elements that can have dramatically different temperatures from the surrounding plant tissue. This study aims to train and validate a Supervised Machine Learning (SML) algorithm using a dual-camera system to extract foliage temperature in a complex field environment. A UAS campaign focused on a set of maize treatments was conducted at Waterman Farm throughout the summer of 2020, spanning diurnal acquisitions across the growing season. In-situ tower-based sensors were deployed to provide validation of the airborne data. Remotely sensed images, which included red, green, blue, and thermal infrared bands, were used to train an SML algorithm. Our results show that the combination of these four bands can be used to accurately identify foliage pixels within complex field scenes with an accu (open full item for complete abstract)

Committee: Darren Drewry (Advisor); Scott Shearer (Committee Member) Subjects: Agricultural Engineering; Computer Science

Doctor of Philosophy, The Ohio State University, 2020, Political Science

This dissertation leverages developments in machine learning methods to better model networked social processes, with an emphasis on international politics. The first chapter develops a dataset with estimates for every country's level of dissatisfaction with the international system from 1816-2012. The second chapter takes these dissatisfaction measures and uses them as features in a machine learning model which predicts international conflict onset. The third chapter explores spillover effects in social networks, demonstrating how causal forests can be employed to uncover spillover effect heterogeneity. Across these chapters, machine learning techniques are instrumental in modeling outcomes of interest and leveraging information from social networks. In the first chapter, I propose a novel measure of international dissatisfaction spanning from 1816 to 2012 which explicitly operationalizes Gilpin's framework: the difference between a state's expected and actual benefits from the international status quo. I estimate a state's expected international benefits by building upon recent efforts to train machine learning ensembles on war outcomes, which I then use to weight a state's observable material capabilities. I estimate actual international benefits by averaging across a state's centrality in valued international networks. The measure provides multiple advantages over alternative estimates both conceptually and statistically. Beyond its conceptual value, the measure's association with militarized conflict is robust to model specifications, unlike the current go-to measure when modeling country-level sentiments: ideal point estimates from United Nations voting records. The second chapter asks: when do revisionist states occur? Most responses to this question fall under one of three categories: 1) differential growth rates, 2) domestic political changes, or 3) international dissatisfaction. While these arguments are all based on rich research traditions, it is a (open full item for complete abstract)

Committee: Bear Braumoeller (Committee Chair); Skyler Cranmer (Committee Member); Christopher Gelpi (Committee Member); James Wilson (Committee Member) Subjects: International Relations; Political Science; Statistics

PhD, University of Cincinnati, 2020, Engineering and Applied Science: Computer Science and Engineering

Musicians who perform with electronic synthesizers often adjust synthesis parameters during live performance to achieve more expressive sounds. Enabling the performer to teach a computer to make these adjustments automatically during performance frees the performer from this responsibility, while maintaining an expressive sound in line with the performer's musical vision. We have created a machine learning system called Larasynth that can be trained by a musician to make these parameter adjustments in real-time during live performances. Larasynth is trained using examples in the form of MIDI files created by the user. Learning is achieved using Long Short-Term Memory (LSTM) recurrent neural networks. To accomplish this, we have devised a set of features which capture the state of the synthesizer controller at regular intervals and are used to make regular predictions of parameter values using an LSTM network. To achieve sufficient generalization during training, transformations are applied to the training data set before each training epoch to simulate variations that may occur during performance. We have also created a new lightweight LSTM library suitable for small networks under real-time constraints. In this thesis we present details behind Larasynth's implementation and use, and experiments that were performed to demonstrate Larasynth's ability to learn behaviors based on different musical situations.

Committee: Anca Ralescu Ph.D. (Committee Chair); Yizong Cheng Ph.D. (Committee Member); Chia Han Ph.D. (Committee Member); Mara Helmuth D.M.A. (Committee Member); Ali Minai Ph.D. (Committee Member) Subjects: Computer Science

Master of Computing and Information Systems, Youngstown State University, 2020, Department of Computer Science and Information Systems

Pseudo Random Number Generators (PRNG) are algorithms that help to create randomness in programs, but they are not truly random like the randomness obtain from physical processes. Despite this, PRNGs are widely used in applications that require random numbers, from gaming to security. This research evaluates the use of machine learning algorithms to predict the numbers generated by PRNG. These experiments involve three commonly used PRNG from C++, Python, and Java, and uses two Machine Learning algorithms, Linear Regression and Artificial Neural Networks. The outcome of the research confirms the possibility that machine learning algorithms can be trained to predict certain PRNGs. Even when trained with a small amount of data, there is evidence that machine learning algorithms can be used to predict the values created by pseudorandom number generators. Given that linear regression algorithm and a simple regression neural network were able to produce fairly good predictions with reasonable accuracy in our experiments, it is believable that more complex machine learning algorithms such as deep learning and recurrence algorithms might produce still better results.

Committee: John Sullins PhD (Advisor); Alina Lazar PhD (Committee Member); Abdu Arslanyilmaz PhD (Committee Member) Subjects: Artificial Intelligence; Computer Science; Information Systems; Information Technology

Master of Computer Science, Miami University, 2020, Computer Science and Software Engineering

While classification performance has improved with the adoption of Neural Network models, the cost of acquiring and labeling the data required to outperform other classification methods is often prohibitively high. Semi-Supervised learning attempts to incorporate unlabeled data in the learning process which can improve performance, however such methods assume preexisting, static sets of labeled and unlabeled data, which are often difficult to attain for novel problems. Active learning addresses these problems by determining which unlabeled samples will, when labeled, best improve a supervised model's performance. Existing methods to prioritize samples have primarily been considered in isolation, despite the existing Information Density framework to combine these methods together. We employ this framework to combine the current state of the art uncertainty based method with a novel similarity based method to improve performance. We also extend the framework itself by considering a dynamic combination of these two methods that shifts priority from one to the other. This iterative process of increasing the labeled set with data prioritized by our acquisition function enables the creation of powerful classification models at greatly reduced costs.

Committee: John Femiani (Advisor); Karsten Maurer (Committee Co-Chair); Daniela Inclezan (Committee Member); Zaobo He (Committee Member) Subjects: Computer Science

Doctor of Philosophy, The Ohio State University, 2019, Political Science

When modeling interstate military alliances, scholars make simplifying assumptions. However, most recognize these often invoked assumptions are overly simplistic. This dissertation leverages developments in supervised and unsupervised machine learning to assess the validity of these assumptions and examine how they influence our understanding of alliance politics. I uncover a series of findings that help us better understand the causes and consequences of alliances. The first assumption examined holds that states, when confronted by a common external security threat, form alliances to aggregate their military capabilities in an effort to increase their security and ensure their survival. Many within diplomatic history and security studies criticize this widely accepted "Capability Aggregation Model", noting that countries have various motives for forming alliances. In the first of three articles, I introduce an unsupervised machine learning algorithm designed to detect variation in how actors form relationships in longitudinal networks. This allows me to, in the second article, assess the heterogeneous motives countries have for forming alliances. I find that states form alliances to achieve foreign policy objectives beyond capability aggregation, including the consolidation of non-security ties and the pursuit of domestic reform. The second assumption is invoked when scholars model the relationship between alliances and conflict, routinely assuming that the formation of an alliance is exogeneous to the probability that one of the allies is attacked. This stands in stark contrast to the Capability Aggregation Model's expectations, which indicate that an external threat and an ally's expectation of attack by an aggressor influences the decision to form an alliance. In the final article, I examine this assumption and the causal relationship between alliances and conflict. Specifically, I endogenize alliances on the causal path to conflict using supe (open full item for complete abstract)

Committee: Skyler Cranmer (Committee Chair); Box-Steffensmeier Janet (Committee Member); Braumoeller Bear (Committee Member); Gelpi Christopher (Committee Member) Subjects: Artificial Intelligence; Behavioral Sciences; Computer Science; International Relations; Military History; Peace Studies; Political Science; Statistics; World History

MS, University of Cincinnati, 2018, Engineering and Applied Science: Computer Science

Polypharmacy is the simultaneous combination of two or more drugs at a time, which is unavoidable in the elderly population as they often suffer from multiple complex conditions. A drug-drug interaction (DDI) is a change in the effect of a drug due to polypharmacy. Identifying and characterizing the DDIs is important to avoid hazardous complications and also would help reduce development costs for de novo drug discovery. An in-silico method to predict these DDIs a priori using the existing drug profiles can help mitigate not only the DDI-related adverse event risks but also reduce health care costs. In this thesis, drug related feature data such as pathways, targets, SMILES, MeSH, Indications, adverse events, and contraindications are collected from various sources. Drug-drug similarity for individual feature is calculated and integrated along with DDI labels collected from Drugs.com for 10,67,991 interactions. To handle the high imbalance of labels in the dataset, the Synthetic Minority Over-sampling Technique (SMOTE) is applied. Then using the final dataset, a computational machine learning framework is developed to evaluate the classifier performance across multiple datasets and identify the best performing classifier. Random Forest is identified as the best predictive model in this thesis when compared with 5 other classifiers using 5-fold stratified cross-validation. DDI severity characterization is performed using Random Forest for multi-class classification where the labels are safe, minor, moderate and major DDI. The results show that the framework can identify the DDIs and characterize the severity of pairwise drug feature-similarity data, and can therefore be useful in drug development and pharmacovigilance studies.

Committee: Anil Jegga D.V.M. (Committee Chair); Raj Bhatnagar Ph.D. (Committee Member); Ali Mina Ph.D. (Committee Member) Subjects: Computer Science

Doctor of Philosophy (PhD), Ohio University, 2018, Electrical Engineering & Computer Science (Engineering and Technology)

In many machine learning and data mining algorithms, pairwise distances (or dissimilarities) among data samples are computed based on the Euclidean metric, where all feature components are treated equally and assigned with the same weight. Learning a customized metric from the input data can often significantly improve the performance of the algorithms. In this dissertation, we propose two nonlinear distance metric learning (DML) frameworks to boost the performance of semi-supervised learning (SSL) and unsupervised learning (USL) algorithms, respectively. Formulated under a constrained optimization framework, our proposed SSL-DML method learns a smooth nonlinear feature space transformation that makes the input data samples more linearly separable in Laplacian SVM (LapSVM). Our USL-ML solution, on the other hand, aims to increase data's linear separability for k-means. A geometric model called Coherent Point Drifting (CPD) is utilized in both frameworks to move data points towards more desirable locations. The choice of CPD is with two considerations: 1) its remarkable capability in generating high-order yet smooth deformations; and 2) the available mechanism within CPD for assigning different levels of smoothness to data points. Application-wise, we apply our SSL-DML to predict the conversion of Alzheimer's Disease (AD) from its early stage: Mild Cognitive Impairment (MCI). The proposed USL-DML solution is utilized to improve the patient clustering. Using neuroimage data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database, we evaluate the effectiveness of the proposed frameworks. The experimental results demonstrate the improvements over the state-of-the-art solutions within the same category.

Committee: Jundong Liu (Advisor) Subjects: Computer Science; Electrical Engineering

Master of Science, The Ohio State University, 2018, Computer Science and Engineering

Information extraction forms a very large and important component of NLP research which aims at extracting varying nature of information from a text corpus. This information could vary from (named) entities and their inter-relationships in sentences to facts which could later be used for different tasks like search engine retrieval, question answering, etc. Most of these tasks and their associated (primarily) machine learning based solutions ultimately hit a roadblock due to the lack of manually labeled data complemented by an expensive and laborious annotation task. While unsupervised/semi-supervised methods can be developed for these tasks, their effectiveness and usability could be compromised. For the task of relation extraction, the distant supervised paradigm has been shown to have enormous potential in providing a relatively very large amount of training data, at the cost of label noise. Prior efforts have proposed a variety of solutions to reduce the impact of label noise both at an architectural level, as well as by adding a small amount of manual supervision. However, we aim to explore a different relation extraction paradigm - can distant supervision help to improve supervised neural relation extraction? This thesis focuses on exploring various strategies such that a supervised relation extraction model, when supplemented with distant supervision is able to perform better at test time. While we are unable to successfully use approaches based on an attention driven subspace alignment and adversarial training for our goal, a simple distillation based approach can result in an improvement in the model's performance.

Committee: Huan Sun (Advisor); Alan Ritter (Committee Member) Subjects: Artificial Intelligence; Computer Engineering; Computer Science; Language; Linguistics

Doctor of Philosophy (PhD), Ohio University, 2018, Experimental Psychology (Arts and Sciences)

An intriguing and unsolved problem in cognitive science concerns the nature of and the relationship between unsupervised and supervised categorization behavior. The former refers to assessing how observers naturally sort multidimensional objects into groups and investigating whether they can learn more complicated groupings without external feedback from the environment. Conversely, the latter refers to experimental investigations aiming to predict and explain how observers inductively learn a predetermined grouping of stimuli upon receiving “correct” or “incorrect” feedback after each classification response. Although these approaches are very different, a few attempts have been put forth with the goal of connecting behavioral outcomes between the two tasks. In general, these investigations implement both types of tasks and seek to explain the results under a common theoretical or formal framework. Although the results are promising, there is a lack of consensus regarding which theoretical or formal approach best accounts for the data. Following this tradition of integration, we present a novel attempt at connecting unsupervised and supervised categorization behavior. We employ generalized invariance structure theory (GIST; Vigo 2013, 2014), generalized representational information theory (GRIT; Vigo 2011, 2012, 2014), and their associated formal models to predict and explain results from two separate experiments. For the first set of experiments, we assessed unsupervised categorization and associated learning behavior by employing a “construction” task previously implemented by the authors (Doan & Vigo, 2016). Importantly, we modified the procedure in accord with similar techniques as those found in prior investigations to facilitate establishing the connection between unsupervised and supervised learning behavior. We replicated Doan and Vigo (2016) and also observed a decrease in response times for each of the three sub experiments, suggesting participants (open full item for complete abstract)

Committee: Ronaldo Vigo PhD (Committee Chair); Keith Markman PhD (Committee Member); Kimberly Rios PhD (Committee Member); Robert Briscoe PhD (Committee Member); Chao-Yang Lee PhD (Committee Member) Subjects: Behavioral Sciences; Cognitive Psychology; Experimental Psychology; Information Science

Master of Sciences, Case Western Reserve University, 2017, EMC - Mechanical Engineering

This thesis focuses on deep-learning approaches to recognition and pose estimation of graspable objects using depth information. Recognition and orientation detection from depth-only data is encoded by a carefully designed 2D descriptor from 3D point clouds. Deep-learning approaches are explored from two main directions: supervised learning and semi-supervised learning. The disadvantages of supervised learning approaches drive the exploration of unsupervised pretraining. By learning good representations embedded in early layers, subsequent layers can be trained faster and with better performance. An understanding of learning processes from a probabilistic perspective is concluded, and it paves the way for developing networks based on Bayesian models, including Variational Auto-Encoders. Exploitation of knowledge transfer--re-using parameters learned from alternative training data--is shown to be effective in the present application.

Committee: Wyatt Newman PhD (Advisor); M. Cenk Çavusoglu PhD (Committee Member); Roger Quinn PhD (Committee Member) Subjects: Computer Science; Medical Imaging; Nanoscience; Robotics

Doctor of Philosophy, The Ohio State University, 2016, Computer Science and Engineering

Supervised learning algorithms have achieved significant success in the last decade. To further improve learning performance, we still need to have a better understanding of semi-supervised learning algorithms for leveraging a large amount of unlabeled data. In this dissertation, a new approach for semi-supervised learning will be discussed, which takes advantage of unlabeled data information through an integral operator associated with a kernel function. More specifically, several problems in machine learning are formulated as a regularized Fredholm integral equation, which has been well studied in the literature of inverse problems. Under this framework, we propose several simple and easily implementable algorithms with sound theoretical guarantees. First, a new framework for supervised learning is proposed, referred as Fredholm learning. It allows a natural way to incorporate unlabeled data and is flexible on the choice of regularizations. In particular, we connect this new learning framework to the classical algorithm of radial basis function networks, and more specifically, analyze two common forms of regularization procedures for RBF networks, one based on the square norm of coefficients in a network and another one using centers obtained by the k-means clustering. We provide a theoretical analysis of these methods as well as a number of experimental results, pointing out very competitive empirical performance as well as certain advantages over the standard kernel methods in terms of both flexibility (incorporating unlabeled data) and computational complexity. Moreover, the Fredholm learning algorithm could be interpreted as a special form of kernel methods using a data-dependent kernel. Our analysis shows that Fredholm kernels achieve noise suppressing effects under a new assumption for semi-supervised learning, termed the "noise assumption". We also address the problem of estimating the probability density ratio function q/p, which could be used for s (open full item for complete abstract)

Committee: Mikhail Belkin (Advisor); Wang Yusu (Committee Member); Wang DeLiang (Committee Member); Lee Yoonkyung (Committee Member) Subjects: Computer Science

Doctor of Philosophy, The Ohio State University, 2014, Computer Science and Engineering

In real-world environments, speech often occurs simultaneously with acoustic interference, such as background noise or reverberation. The interference usually leads to adverse effects on speech perception, and results in performance degradation in many speech applications, including automatic speech recognition and speaker identification. Monaural speech separation and processing aim to separate or analyze speech from interference based on only one recording. Although significant progress has been made on this problem, it is a widely regarded challenge. Unlike traditional signal processing, this dissertation addresses the speech separation and processing problems using machine learning techniques. We first propose a classification approach to estimate the ideal binary mask (IBM) which is considered as a main goal of sound separation in computational auditory scene analysis (CASA). We employ support vector machines (SVMs) to classify time-frequency (T-F) units as either target-dominant or interference-dominant. A rethresholding method is incorporated to improve classification results and maximize hit minus false alarm rates. Systematic evaluations show that the proposed approach produces accurate estimated IBMs. In a supervised learning framework, the issue of generalization to conditions different from those in training is very important. We then present methods that require only a small training corpus and can generalize to unseen conditions. The system utilizes SVMs to learn classification cues and then employs a rethresholding technique to estimate the IBM. A distribution fitting method is introduced to generalize to unseen signal-to-noise ratio conditions and voice activity detection based adaptation is used to generalize to unseen noise conditions. In addition, we propose to use a novel metric learning method to learn invariant speech features in the kernel space. The learned features encode speech-related information and can generalize to unseen noise (open full item for complete abstract)

Committee: DeLiang Wang (Advisor); Eric Fosler-Lussier (Committee Member); Mikhail Belkin (Committee Member) Subjects: Computer Science