Doctor of Philosophy (PhD), Wright State University, 2020, Computer Science and Engineering PhD

Deep Neural Networks (DNNs) are powerful tools blossomed in a variety of successful real-life applications. While the performance of DNNs is outstanding, their opaque nature raises a growing concern in the community, causing suspicions on the reliability and trustworthiness of decisions made by DNNs. In order to release such concerns and towards building reliable deep learning systems, research efforts are actively made in diverse aspects such as model interpretation, model fairness and bias, adversarial attacks and defenses, and so on. In this dissertation, we focus on the research topic of DNN interpretations for visual intelligence, aiming to unfold the black-box and provide explanations for visual intelligence tasks in a human-understandable way. We first conduct a categorized literature review, systematically introducing the realm of explainable deep learning. Following the review, two specific problems are tackled, explanations of Convolutions Neural Networks (CNNs), which relates the CNN decisions with input concepts, and interpretability of multi-model interactions, where an explainable model is built to solve a visual inference task. Visualization techniques are leveraged to depict the intermediate hidden states of CNNs and attention mechanisms are utilized to build an instinct explainable model. Towards increasing the trustworthiness of DNNs, a certainty measurement for decisions is also proposed as an extensive exploration of this study. To show how the introduced techniques holistically realize a contribution to interpretable and reliable deep neural networks for visual intelligence, further experiments and analyses are conducted for visual entailment task at the end of this dissertation.

Committee: Derek Doran Ph.D. (Advisor); Michael Raymer Ph.D. (Committee Member); Tanvi Banerjee Ph.D. (Committee Member); Pascal Hitzler Ph.D. (Committee Member) Subjects: Artificial Intelligence

Master of Sciences (Engineering), Case Western Reserve University, 2017, EECS - Computer and Information Sciences

The use of multiple sensors in modern day vehicular applications is necessary to provide a complete outlook of surroundings for advanced driver assistance systems (ADAS) and automated driving. The fusion of these sensors provides increased certainty in the recognition, localization and prediction of surroundings. A deep learning-based sensor fusion system is proposed to fuse two independent, multi-modal sensor sources. This system is shown to successfully learn the complex capabilities of an existing state-of-the-art sensor fusion system and generalize well to new sensor fusion datasets. It has high precision and recall with minimal confusion after training on several million examples of labeled multi-modal sensor data. It is robust, has a sustainable training time, and has real-time response capabilities on a deep learning PC with a single NVIDIA GeForce GTX 980Ti graphical processing unit (GPU).

Committee: Wyatt Newman Dr (Committee Chair); M. Cenk Cavusoglu Dr (Committee Member); Michael Lewicki Dr (Committee Member) Subjects: Artificial Intelligence; Computer Science

Master of Science in Computer Engineering, University of Dayton, 2022, Electrical and Computer Engineering

Image compression algorithms are the basis of media transmission and compression in the field of image processing. Decades after their inception, algorithms such as the JPEG image codec continue to be the industry standard. A notable research topic gathering momentum in the field of compression is deep learning (DL). This paper explores the opti- mization of DL models for ideal image compression and object detection (OD) applications. The DL model to be optimized is based upon an existing compression framework known as the CONNECT model. This framework wraps the traditional JPEG image codec within two convolutional neural networks (CNNs). The first network, ComCNN, focuses on com- pressing an input image into a compact representation to be fed into the image codec. The second network, RecCNN, focuses on reconstructing the output image from the codec as similarly as possible to the original image. To enhance the performance of the CONNECT model, an optimization software called Optuna wraps the framework. Hyperparameters are selected from each CNN to be evaluated and optimized by Optuna. Once the CONNECT model produces ideal results, the output images are applied to the YOLOv5 OD network. This paper explores the impact of DL hyperparameters on image quality and compres- sion metrics. In addition, a detection network will provide context to the effect of image compression on computer vision applications.

Committee: Bradley Ratliff (Committee Chair); Eric Balster (Committee Member); Barath Narayanan (Committee Member) Subjects: Computer Engineering

Master of Science (M.S.), University of Dayton, 2018, Electrical Engineering

This thesis proposes an optimized convolutional neural network architecture to improve homography estimation applications. The parameters and structure of the CNN including the number of convolutional filters, stride lengths, kernel size, learning parameters, etc are altered from previous implementations. Multiple modifications of the network are trained and evaluated until a final network yields a corner pixel error of 4.7 which is less than a network proposed in previous literature's.

Committee: Eric Balster (Advisor); Yakov Diskin (Committee Member); Tarek Taha (Committee Member) Subjects: Computer Engineering; Electrical Engineering

Master of Sciences (Engineering), Case Western Reserve University, 2017, EECS - Electrical Engineering

This thesis explores the use of deep learning for robot localization with applications in re-localizing a mislocalized robot. Seed values for a localization algorithm are assigned based on the interpretation of images. A deep neural network was trained on images acquired in and associated with named regions. In application, the neural net was used to recognize a region based on camera input. By recognizing regions from the camera, the robot can be localized grossly, and subsequently refined with existing techniques. Explorations into different deep neural network topologies and solver types are discussed. A process for gathering training data, training the classifier, and deployment through a robot operating system (ROS) package is provided.

Committee: Wyatt Newman (Advisor); Murat Cavusoglu (Committee Member); Gregory Lee (Committee Member) Subjects: Computer Science; Electrical Engineering; Robotics

Doctor of Engineering, Cleveland State University, 2024, Washkewicz College of Engineering

This dissertation research investigated how to use the blockchain technology to secure sensor data processing and logging. The research was done in three phases. First, to ensure the legitimate of the sensor to log data into Blockchain, sensor identifcation and authentication mechanism is used where only the defned sensors sensing data are accepted. Second, to minimize the throughput demand on large public blockchain such as Bitcoin and Ethereum and the fnancial cost of using blockchain services, only a small amount of raw sensing data are placed on the blockchain through an aggregation process, where a group of raw sensing data is converted to one condensed data time. A Merkle tree based mechanism is used to protect the security of the of-chain data (raw sensing data) with the condensed Data placed on the blockchain. The system was tested with the IOTA Shimmer test network, and the Ethereum test network. The second phase focuses on developing an Ethereum smart contract to manage access control for storing and retrieving condensed data on the blockchain. The smart contract introduces three levels of authorization (read, write, and admin) to regulate data access securely. Gas consumption optimization is achieved through a tag-based secure data-store mechanism embedded in the smart contract design. In the fnal phase, a deep learning model using Convolution Neural Networks (CNN) is introduced to detect vulnerabilities in smart contracts. Four input techniques—Word2Vec, FastText, Bag of Words (BoW), and TF-IDF—are compared for their efectiveness in identifying six types of vulnerabilities. TF-IDF emerges as the most efcient input technique, consistently achieving high detection rates (90% to 100%) across all vulnerability types. In particular, TF-IDF excels in detecting the Reentrancy vulnerability, achieving performance metrics of 96% to 97%. Word2Vec and FastText performed comparably with slight changes, however BoW consistently dropped behind (open full item for complete abstract)

Committee: Wenbing Zhao (Advisor); Timothy V Arndt (Committee Member); Hongkai Yu (Committee Member); Lili Dong (Committee Member); Sun S. Chung (Committee Member) Subjects: Computer Engineering; Computer Science

Master of Science, The Ohio State University, 2024, Electrical and Computer Engineering

Image denoising is a fundamental problem in image processing where a high fidelity image is recovered from a noise corrupted version. Denoising is fundamental because, from the Bayesian perspective denoisers are believed to also encode information about the prior probability distribution of images. This in turn, makes denoisers a widely applicable tool in many image inverse problems like compressive sensing, deblurring, in-painting, super-resolution, etc. As a result various algorithmic approaches for denoising have been studied in the past decades. However, data-driven denoising methods, which learn to denoise images from large image datasets using deep neural networks, have demonstrated far superior performance compared to the classical algorithmic methods while having much faster inference times. The data-driven methods can be broadly classified into two categories: blind and non-blind methods. While non-blind methods require knowledge of the noise level contained within the image, blind methods which require no such information are more practical. However, the performance of many recent state-of-the-art blind denoisers depend heavily on the noise levels used during training. In more recent work, ideas of inducing scale and normalization equivariance properties in denoisers have been explored in order to make denoisers more robust to changes in noise levels from training to test data. In our work we extend upon this idea, where we introduce a method to make any given denoiser normalization equivariant using a simple idea of instance normalization, which improves the noise level robustness of the denoiser by a significantly large margin with minimal change to the underlying architecture. In this thesis, we theoretically formulate our idea from the perspective of minimizers of the Wasserstein-1 distance between empirical distributions of training and test data, and propose a more practically feasible 2-pixel approximation that yi (open full item for complete abstract)

Committee: Philip Schniter (Advisor); Lee Potter (Committee Member) Subjects: Electrical Engineering

Master of Computing and Information Systems, Youngstown State University, 2024, Department of Computer Science and Information Systems

The field of coding education is rapidly evolving, with emerging technologies playing a pivotal role in transforming traditional learning methodologies. This thesis introduces Dr. Lego, an innovative framework designed to revolutionize the assessment and understanding of block-based coding through the integration of sophisticated deep learning models. Dr. Lego combines cutting-edge technologies such as MobileNetV3 (Howard, 2019), for visual recognition and BERT (Devlin et al., 2018), and XLNet (Yang et al., 2019) for natural language processing to offer a comprehensive approach to evaluating coding proficiency. The research methodology involves the meticulous curation of a diverse dataset comprising projects from the LEGO SPIKE app (LEGO Education, 2022), ensuring that the models are subjected to a broad range of coding scenarios. Leveraging the dynamic educational environment provided by the LEGO SPIKE app (LEGO Education, 2022), Dr. Lego empowers users to design and implement various coding projects, fostering hands-on learning experiences. This thesis delves into methodologies aimed at enhancing coding education by exploring model integration, data generation, and fine-tuning of pre-trained models. Dr. Lego not only evaluates coding proficiency but also provides cohesive and insightful feedback, enhancing the learning experience for users. The adaptability of the framework highlights its potential to shape the future of coding education, paving the way for a new era of interactive and engaging learning experiences.

Committee: Abdu Arslanyilmaz PhD (Advisor); Feng Yu PhD (Committee Member); Carrie Jackson EdD, BCBA (Committee Member) Subjects: Computer Science; Engineering; Information Systems; Robotics; Teaching

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

Distributed multi-agent systems are becoming increasingly crucial for robotics applications due to their potential for efficiency, resilience, scalability, and achieving complex tasks. Such applications include environmental mapping, search after natural disasters, and military systems. Hence, swarm robotics has been a focus of research for several years. However, controlling these systems in a distributed sense relying on local information is particularly challenging. Centralized methods using global information are often faced with the issue of scalability and significant computational requirements. All the information is sent to a central agent, which computes the actions of each agent. This is a bottleneck; such systems are not robust or scalable and often impractical. As the number of agents increases, decentralized control becomes a necessity. Each agent must decide its own action based on the local information from its neighbors. Traditional control-theoretic methods have been used in literature to develop such controllers. However, it is highly challenging to obtain an efficient decentralized controller that would enable individual agents to use local information to compute their control actions to achieve a global task. This dissertation uses data-driven Artificial Intelligence and Machine Learning techniques to develop decentralized control schemes for multi-agent systems. Such approaches have shown immense promise in solving complex problems such as facial recognition, anomaly detection, smart manufacturing, robotics, etc. This dissertation aims to develop a scalable and decentralized controller for large-scale heterogeneous robot swarms exhibiting segregative, aggregative, and navigation behaviors by learning in a supervised and unsupervised manner. This work presents two learning approaches: a) an imitation learning-based framework for segregation and aggregation tasks and b) a multi-agent deep graph reinforcement learning framework for cohesiv (open full item for complete abstract)

Committee: Manish Kumar Ph.D. (Committee Chair); David Thompson Ph.D. (Committee Member); Ali Minai Ph.D. (Committee Member); Xiaodong Jia Ph.D. (Committee Member) Subjects: Mechanical Engineering

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

Fingerprinting is a technique that identifies websites, software and devices by leveraging a group of information from users. An attacker can acquire users' secrets by only analyzing side-channel features from a system, such as network packet size and direction, power usage or CPU usage. In traditional fingerprinting attacks, a large amount of human effort is required as an attacker has to manually extract effective features for attacking purpose. This kind of attacking approach is easy to be defended as a defender can invalidate the attack by modifying the target features that are used in the attack. However, for AI-based fingerprinting, handcrafted feature is not necessary anymore. An attacker can train a machine learning classifier over raw data directly and achieve an impressive classification results. In this proposal, I propose to design effective and efficient defenses against deep neural network based fingerprinting attack. Firstly, I propose to improve the efficiency of existing defense against neural network based stream fingerprinting. Many defense algorithms have been proposed to defeat stream fingerprinting. However, most of those existing defense algorithms need extremely high bandwidth overhead in order to make the defense effective. In this dissertation, I leverage feature selection methods to analyze the feature space in stream fingerprinting. Instead of treating network packets equally when adding noises, we distinguish important packets using feature selection algorithms and add more noise to those important packets. Secondly, I propose to design an efficient defense against CPU cache based website fingerprinting. Recently, a new attack was proposed to monitor the cache occupancy of the Last Level Cache on a user's CPU. Although a defense was proposed, it is not effective when an attacker adapts the classifier with defended data. In this dissertation, I investigate the behavior of cache occupancy channel and reveal the reaso (open full item for complete abstract)

Committee: Boyang Wang Ph.D. (Committee Chair); Nirnimesh Ghose Ph.D. (Committee Member); Tingting Yu Ph.D. (Committee Member); Nan Niu Ph.D. (Committee Member); Wen-Ben Jone Ph.D. (Committee Member) Subjects: Computer Engineering

Doctor of Philosophy, The Ohio State University, 2023, Electrical and Computer Engineering

The advent of deep neural networks (DNNs) have brought exciting new possibilities for the realization of automated driving functions. These data-driven methods have been widely applied to various driving tasks, including end-to-end urban driving. However, the use of these methods beyond simulated tests remains limited due to two significant shortcomings: (i) the lack of model transparency and (ii) the difficulty of generalizing beyond the training distribution. This dissertation aims to investigate methods for addressing the transparency and safety mitigation of learning-based controllers, specifically deep reinforcement learning (DRL) methods, to enable safe and predictable driving. To enhance interpretability, an interpretable and causal state representation, coined the driving forces, is proposed. This representation captures the causal relationship between the state and the produced control action by leveraging force features to encode the influence of internal and external factors on the ego vehicle. By training a DRL agent on this representation within a highway driving environment, the ability of the driving forces to encode and interpret the state-action causalities was demonstrated. Furthermore, an alternative paradigm for online adaptation by modifying the formulation of the driving forces is proposed to mitigate the behavior of the ego vehicle. The results showed that the ego vehicle was successful in mitigating its behavior and following desired new and unseen behaviors, without requiring modification to the underlying DNN. To address model transparency in black-box DNN-based driving policies, a knowledge distillation framework that combines interpretable decision trees with rule learning algorithms is proposed. This framework learns decision rule sets that represent the decision boundaries of the original driving policy. The driving forces are utilized to abstract the original state representation and ensure the interpretability of the learned explanati (open full item for complete abstract)

Committee: Umit Ozguner (Advisor); Keith Redmill (Committee Member); Qadeer Ahmed (Committee Member); Gladys Mitchell (Committee Member) Subjects: Artificial Intelligence; Computer Engineering; Electrical Engineering; Transportation

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

In this thesis, I address the problem of computational tool recommendation to suggest during the construction of life science analysis workflows. The major motivation for such a system is to mitigate the time required by bioinformaticians in researching and selecting tools to complete an analysis. Constructing workflows is a time-consuming process that requires many careful decisions and extensive domain knowledge. The recent, rapid expansion of Bioinformatics research has led to new tools appearing daily that further perplexes the tool selection process. A great source of information for learning and constructing new analyses is to consult existing ones. A system that can learn latent connections between tools from existing workflows and use them to suggest downstream tools or tool sequences in a new workflow-in-progress should be highly valuable to researchers in the creation process. The Bioinformatics Tool Recommendation system (BTR) is proposed to accomplish this task. BTR is a deep learning architecture that makes use of emergent graph neural network technology to find the most relevant successive tools for an input workflow query. Workflow construction is framed as a session-based recommendation problem and relevant techniques are applied. The method leverages a novel approach in representing workflows as directed acyclic graphs, rather than linear tool sequences, that sees benefits in recommendation performance and logical function. An attention mechanism is used to highlight recent workflow context and drop low-relevance tools. Semantic tool descriptions are mined and incorporated using a domain-specific language processing approach. Experiments show a significant improvement over the closest-related previous work for the automatic evaluation metrics.

Committee: Tingting Yu Ph.D. (Committee Chair); Nan Niu Ph.D. (Committee Member); Jinze Liu PhD (Committee Member); Raj Bhatnagar Ph.D. (Committee Member) Subjects: Computer Science

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

Traditional drug discovery is costly and time-consuming. With the availability of large-scale molecular interaction networks, novel predictive modeling strategies have become vital to study the effect of drugs. Graphs are a powerful and flexible data structure in this regard. Biomedical graphs encompass the complex relationships between drugs, diseases, genes, and other micro/macroscopic effects of drugs. Hence, analyzing and modeling graphs can be valuable in identifying novel insights for drug discovery and its effects. Recently, deep learning research has made significant advances in image, speech, and natural language domains. The research in these fields has fostered progress in applying neural networks to graphs, referred to as graph neural networks (GNNs), for learning and identifying valuable hidden insights in graphs. While these GNNs are effective in learning representations, early research has focused primarily on optimizing GNNs for simple graph structures. Real-world graphs, however, tend to have complex characteristics such as heterogeneity, multi-modality, and combinatoriality. These complexities are particularly apparent in biomedical graphs, particularly in the areas of drug repurposing, virtual screening, and drug-drug interaction studies. This hinders the ability of GNNs to learn accurate representations and fully understand a drug's behavior within the human body. Furthermore, for most current methods, the interpretation of the inferred predictions has not been investigated in detail, leading to skepticism in their adoption, especially in biomedical and healthcare domains. The work in this thesis aims to enhance the capabilities of GNNs for complex networks by studying and generating hypotheses for drug discovery and drug-drug interaction studies in biological networks. To achieve this, GNNs have been investigated and improved with three specific aims. Aim 1 is to develop GNNs that take heterogeneous networks as input and use multi (open full item for complete abstract)

Committee: Anil Jegga DVM MRes (Committee Chair); Raj Bhatnagar Ph.D. (Committee Member); Mayur Sarangdhar PhD (Committee Member); Ali Minai Ph.D. (Committee Member); Yizong Cheng Ph.D. (Committee Member) Subjects: Computer Science

Doctor of Philosophy, The Ohio State University, 2023, Computer Science and Engineering

Attributed graphs, i.e. graphs with attributes associated with nodes, are popular data representations used to capture interactions between entities. In recent years, interest has been growing in developing data mining techniques on attributed graphs for different learning tasks such as node classification and clustering, link prediction and graph classification. Graph Neural Networks (GNN) are emerging as the state-of-the-art graph mining models. Within GNN, Graph Convolutional Networks (GCN) have been used with great success in various domains such as recommendation systems, social network analysis, AI-powered drug discovery etc. However, there are multiple challenges from different aspects in using GCN based approaches: 1) high time cost resulting from the frequent loading of data to GPUs during training; 2) limited learning ability resulting from over-smoothing issues inherent to GCN; 3) difficulties in scaling to large-scale graphs due to limited GPU memory; 4) restrictions on access to centralized training datasets when data sharing is prohibited due to privacy or commercial restrictions. This dissertation focuses on developing multiple efficient GNN based approaches for attributed graphs learning aimed at solving the aforementioned challenges. First, we present a general framework, in which multiple GCN methods (GraphSage, clusterGCN, GraphSaint) can be accelerated by reducing the frequency of data transfer to GPUs without noticeable degradation on learning ability. Second, to relieve both the over-smoothing and scalability issues of GCN, we describe our scalable deep clustering framework, Random-walk based Scalable Learning (RwSL) focusing on the node clustering task. Previous work like GCN based DGI, SDCN, DMoN or graph filtering based AGC, SSGC, AGE do not scale to large-scale graphs due to their use of non-scalable graph convolution operations. In contrast, RwSL can scale to graphs of arbitrary size by employing a parallelizable random-walk based algor (open full item for complete abstract)

Committee: Rajiv Ramnath (Advisor); Gagan Agrawal (Committee Member); Srinivasan Parthasarathy (Committee Member); Radu Teodorescu (Committee Member); Ruoming Jin (Committee Member) Subjects: Computer Engineering; Information Technology

PhD, University of Cincinnati, 2022, Arts and Sciences: Mathematical Sciences

Expensive computer models (simulators) are frequently used to simulate the behavior of a complex system in many scientific fields because an explicit experiment is very expensive or dangerous to conduct. Usually, only a limited number of computer runs are available due to limited sources. Therefore, one desires to use the available runs to construct an inexpensive statistical model, an emulator. Then the constructed statistical model can be used as a surrogate for the computer model. Building an emulator for high dimensional outputs with the existing standard method, the Gaussian process model, can be computationally infeasible because it has a cubic computational complexity that scales with the total number of observations. Also, it is common to impose restrictions on the covariance matrix of the Gaussian process model to keep computations tractable. This work constructs a flexible emulator based on a deep neural network (DNN) with feedforward multilayer perceptrons (MLP). High dimensional outputs and limited runs can pose considerable challenges to DNN in learning a complex computer model's behavior. To overcome this challenge, we take advantage of the computer model's spatial structure to engineer features at each spatial location and then make the training of DNN feasible. Also, to improve the predictive performance and avoid overfitting, we adopt a data augmentation technique into our method. Finally, we apply our approach using data from the UVic ESCM model and the PSU3D-ICE model to demonstrate good predictive performance and compare it with an existing state-of-art emulation method.

Committee: Won Chang Ph.D. (Committee Member); Xia Wang Ph.D. (Committee Member); Emily Kang Ph.D. (Committee Member) Subjects: Statistics

MS, University of Cincinnati, 2022, Education, Criminal Justice, and Human Services: Information Technology-Distance Learning

Among other cancers worldwide, lung cancer is the leading cause of death. The lives that we lose every year to lung cancer are more than combined of those lost to pancreatic, breast, and prostate cancer. However, lung cancer receives the least amount of research funds for each life lost to cancer each year. Lung cancer receives $3,580 per lost life, pancreatic cancer receives $4796 per lost life, prostate cancer receives $8116 per lost life, and breast cancer receives $19050 per lost life. The survival rate for lung cancer patients is very low compared to other cancer patients. If doctors diagnose a patient with stage I lung cancer, the survival rate will be 55%, which means that the patient will most likely survive cancer for five or more years. However, the survival rate will drop to 5% if the patient is diagnosed with stage IV lung cancer. Diagnosing cancer at an early stage gives doctors more time for their treatment plan, increasing the survival rate or even becoming cancer-free. In this thesis, we aim to develop a deep learning model that will help doctors predict and diagnose lung cancer early to save more lives. This thesis proposes a 2D CNN architecture, using IQ-OTH/NCCD - Lung Cancer Dataset in Kaggle. The dataset consists of 1097 CT scan images, which include three classes, normal cases, malignant cases, and benign cases. The experiment shows that the model has achieved high performance with 99.45% accuracy, and 1.75% loss. The weighted average is 99% and 99% for the macro average. The proposed model can be a particularly useful tool to support radiologists' decisions in predicting and classifying lung cancer.

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

BA, Oberlin College, 2022, Computer Science

In open multiagent systems, multiple agents work together or compete to reach the goal while members of the group change over time. For example, intelligent robots that are collaborating to put out wildfires may run out of suppressants and have to leave the place to recharge; the rest of the robots may need to change their behaviors accordingly to better control the fires. Thus, openness requires agents not only to predict the behaviors of others, but also the presence of other agents. We present a deep reinforcement learning method that adapts the proximal policy optimization algorithm to learn the optimal actions of an agent in open multiagent environments. We demonstrate how openness can be incorporated into state-of-the-art reinforcement learning algorithms. Simulations of wildfire suppression problems show that our approach enables the agents to learn the legal actions.

Committee: Adam Eck (Advisor) Subjects: Artificial Intelligence; Computer Science

PhD, University of Cincinnati, 2022, Engineering and Applied Science: Computer Science and Engineering

Identification of the mechanism of action (MoA) of a small molecule which causes pharmacological effects on cellular networks governing gene expression levels is an important field of study for the purpose of drug development and repurposing. While gene expression can be used for the prediction of small molecule MoA using traditional machine learning algorithms, these algorithms do not consider the underlying complexity of cellular level biological networks driving gene expression. In particular, capturing predictive features from the polarity of interaction in cell signaling networks where nodes in the network either activate or inhibit other nodes is still a challenging problem for the prediction of drug MoA. We propose an ensemble deep learning meta-algorithm for predicting small molecule MoA from gene expression data using unsigned and signed graph convolutional networks (GCN). We developed a GCN algorithm to extract features from signed networks and combined predictive probabilities with that of an unsigned GCN using stacking. Our ensemble methodology improves the overall predictive capabilities significantly when compared to unsigned or signed GCN.

Committee: Mario Medvedovic Ph.D. (Committee Member); Gowtham Atluri Ph.D. (Committee Member); Ali Minai Ph.D. (Committee Member); Jaroslaw Meller Ph.D. (Committee Member); Raj Bhatnagar Ph.D. (Committee Member) Subjects: Bioinformatics

Doctor of Philosophy (PhD), Ohio University, 2022, Electrical Engineering & Computer Science (Engineering and Technology)

Speech separation separates the speech of interest from background noise (speech enhancement) or interfering speech (speaker separation). While the human auditory system has extraordinary speech separation capabilities, designing artificial models with similar functions has proven to be very challenging. Recently, waveform deep neural network (DNN) has become the dominant approach for speech separation with great success. Improving speech quality and intelligibility is a primary goal for the speech separation tasks. Integrating human speech elements into waveform DNNs has proven to be a simple yet effective strategy to boost objective performance (including speech quality and intelligibility) of speech separation models. In this dissertation, three solutions are proposed to integrate human speech elements into waveform speech separation solutions in an effective manner. First, we propose a knowledge-assisted framework to integrate pretrained self-supervised speech representations to boost the performance of speech enhancement networks. To enhance the output intelligibility, we design auxiliary perceptual loss functions that rely on speech representations pretrained on large datasets, to ensure the denoised network outputs sound like clean human speeches. Our second solution is for speaker separation, where we design a speaker-conditioned model that adopts a pretrained speaker identification model to generate speaker embeddings with rich speech information. Our third solution takes a different approach to improve speaker separation solutions. To suppress information of non-target speakers in auxiliary-loss based solutions, we introduce a loss function that can maximize the distance between speech representations of separated speeches and speeches of clean non-target speakers. In this dissertation, we also address a practical issue in frame-based DNN SE solution: frame stitching, where the input context to be observed in a network is often limited, resulting (open full item for complete abstract)

Committee: Jundong Liu (Advisor); Razvan Bunescu (Committee Member); Li Xu (Committee Member); Avinash Karanth (Committee Member); Martin J. Mohlenkamp (Committee Member); Jeffrey Dill (Committee Member) Subjects: Computer Science

Doctor of Philosophy, The Ohio State University, 2021, Electrical and Computer Engineering

The fields of computer vision and machine learning have made enormous strides in developing models which solve tasks only humans have been capable of solving. However, the models constructed to solve these tasks came at an enormous price in terms of computational resources and data collection. Motivated by the sustainability of continually developing models from scratch to tackle every additional task humans can solve, researchers are interested in efficiently constructing new models for developing solutions to new tasks. The sub-fields of machine learning devoted to this line of research go by many names. Such names include multi-task learning, transfer learning, and few-shot learning. All of these frameworks use the same assumption that knowledge should be shared across models to solve a set of tasks. We define knowledge as the set of conditions used to construct a model that solves a given task. By shared knowledge, we are referring to conditions that are consistently used to construct a set of models which solve a set of tasks. In this work, we address two sets of tasks posed in the fields of computer vision and machine learning. While solving each of these sets of tasks, we show how each of our methods exhibits a novel implementation of shared knowledge leading to many implications for future work in developing systems that further emulate the abilities of human beings. The first set of tasks fall within the sub-field of action analysis, specifically the recognition of intent. Instead of a data-driven approach, we construct a hand-crafted model to infer between intentional/non-intentional movement using common knowledge concepts known by humans. These knowledge concepts are ultimately used to construct an unsupervised method to infer between intentional and non-intentional movement across levels of abstraction. By layers of abstraction we mean that the model needed to solve the most abstract instances of intent recognition, is useful in developing models whi (open full item for complete abstract)

Committee: Aleix Martinez (Advisor); Abhishek Gupta (Committee Member); Yingbin Liang (Committee Member) Subjects: Artificial Intelligence; Computer Engineering; Computer Science