Physical Sciences and Mathematics Commons^™

Optimization And Application Of Graph Neural Networks, Shuo Zhang

Dissertations, Theses, and Capstone Projects

Graph Neural Networks (GNNs) are widely recognized for their potential in learning from graph-structured data and solving complex problems. However, optimal performance and applicability of GNNs have been an open-ended challenge. This dissertation presents a series of substantial advances addressing this problem. First, we investigate attention-based GNNs, revealing a critical shortcoming: their ignorance of cardinality information that impacts their discriminative power. To rectify this, we propose Cardinality Preserved Attention (CPA) models that can be applied to any attention-based GNNs, which exhibit a marked improvement in performance. Next, we introduce the Directional Node Pair (DNP) descriptor and the Robust Molecular Graph …

Out-Of-Distribution Generalization Of Deep Learning To Illuminate Dark Protein Functional Space, Tian Cai

Dissertations, Theses, and Capstone Projects

Dark protein illumination is a fundamental challenge in drug discovery where majority human proteins are understudied, i.e. with only known protein sequence but no known small molecule binder. It's a major road block to enable drug discovery paradigm shift from single-targeted which looks to identify a single target and design drug to regulate the single target to multi-targeted in a Systems Pharmacology perspective. Diseases such as Alzheimer's and Opioid-Use-Disorder plaguing millions of patients call for effective multi-targeted approach involving dark proteins. Using limited protein data to predict dark protein property requires deep learning systems with OOD generalization capacity. Out-of-Distribution (OOD) …

Learn Biologically Meaningful Representation With Transfer Learning, Di He

Dissertations, Theses, and Capstone Projects

Machine learning has made significant contributions to bioinformatics and computational biol­ogy. In particular, supervised learning approaches have been widely used in solving problems such as bio­marker identification, drug response prediction, and so on. However, because of the limited availability of comprehensively labeled and clean data, constructing predictive models in super­ vised settings is not always desirable or possible, especially when using data­hunger, red­hot learning paradigms such as deep learning methods. Hence, there are urgent needs to develop new approaches that could leverage more readily available unlabeled data in driving successful machine learning ap­ plications in this area.

In my dissertation, …

Speech Enhancement Using Speech Synthesis Techniques, Soumi Maiti

Dissertations, Theses, and Capstone Projects

Traditional speech enhancement systems reduce noise by modifying the noisy signal to make it more like a clean signal, which suffers from two problems: under-suppression of noise and over-suppression of speech. These problems create distortions in enhanced speech and hurt the quality of the enhanced signal. We propose to utilize speech synthesis techniques for a higher quality speech enhancement system. Synthesizing clean speech based on the noisy signal could produce outputs that are both noise-free and high quality. We first show that we can replace the noisy speech with its clean resynthesis from a previously recorded clean speech dictionary from …

Robust Neural Machine Translation, Abdul Rafae Khan

Dissertations, Theses, and Capstone Projects

This thesis aims for general robust Neural Machine Translation (NMT) that is agnostic to the test domain. NMT has achieved high quality on benchmarks with closed datasets such as WMT and NIST but can fail when the translation input contains noise due to, for example, mismatched domains or spelling errors. The standard solution is to apply domain adaptation or data augmentation to build a domain-dependent system. However, in real life, the input noise varies in a wide range of domains and types, which is unknown in the training phase. This thesis introduces five general approaches to improve NMT accuracy and …

Deep Learning Based Medical Image Analysis With Limited Data, Jiaxing Tan

Dissertations, Theses, and Capstone Projects

Deep Learning Methods have shown its great effort in the area of Computer Vision. However, when solving the problems of medical imaging, deep learning’s power is confined by limited data available. We present a series of novel methodologies for solving medical imaging analysis problems with limited Computed tomography (CT) scans available. Our method, based on deep learning, with different strategies, including using Generative Adversar- ial Networks, two-stage training, infusing the expert knowledge, voting based or converting to other space, solves the data set limitation issue for the cur- rent medical imaging problems, specifically cancer detection and diagnosis, and shows very …

Towards Improving Accuracy And Interpretability Of Deep Learning Based On Satellite Image Classification, Yamile Patino Vargas

Dissertations and Theses

ABSTRACT

The study of satellite images provides a way to monitor changes in the surface of the Earth and the atmosphere. Convolutional Neural Networks (CNN) have shown accurate results in solving practical problems in multiple fields. Some of the more recognized fields using CNNs are satellite imagery processing, medicine, communication, transportation, and computer vision. Despite the success of CNNs, there remains a need to explain the network predictions further and understand what the network is determining as valuable information.

There are several frameworks and methodologies developed to explain how CNNs predict outputs and what their internal representations are [1, 4, …

Multimodal Sensing And Data Processing For Speaker And Emotion Recognition Using Deep Learning Models With Audio, Video And Biomedical Sensors, Farnaz Abtahi

Dissertations, Theses, and Capstone Projects

The focus of the thesis is on Deep Learning methods and their applications on multimodal data, with a potential to explore the associations between modalities and replace missing and corrupt ones if necessary. We have chosen two important real-world applications that need to deal with multimodal data: 1) Speaker recognition and identification; 2) Facial expression recognition and emotion detection.

The first part of our work assesses the effectiveness of speech-related sensory data modalities and their combinations in speaker recognition using deep learning models. First, the role of electromyography (EMG) is highlighted as a unique biometric sensor in improving audio-visual speaker …