DISSERTATION DEFENSE Author : Jeremy Lewis Advisor : Dr. Jason O’Kane Date : Nov 16th , 2018 Time : 3:00 pm Place : Meeting room 2265, Innovation Center Abstract This thesis explores the problem of generating coverage paths---that is, paths that pass within some sensor footprint of every point in an environment--for mobile robots. It both considers models for which navigation is a solved problem but motions are constrained, as well for models in which navigation must be considered along with coverage planning because of the robot's unreliable sensing and movements. The motion contraint we adopt for the former is a common constraint, that of a Dubins vehicle. We extend previous work that solves this coverage problem as a traveling salesman problem (TSP) by introducing a practical heuristic algorithm to reduce runtime while maintaining near-optimal path length. Furthermore, we show that generating an optimal coverage path is NP-hard by reducing from the Exact Cover problem, which provides justification for our algorithm's conversion of Dubins coverage instances to TSP instances. Extensive experiments demonstrate that the algorithm does indeed produce path lengths comparable to optimal in significantly less time. In the second model, we consider the problem of coverage planning for a particular type of very simple mobile robot. The robot must be able to translate in a commanded direction (specified in a global reference frame), with bounded error on the motion direction, until reaching the environment boundary. The objective, for a given environment map, is to generate a sequence of motions that is guaranteed to cover as large a portion of that environment as possible, in spite of the severe limits on the robot's sensing and actuation abilities. We show how to model the knowledge available to this kind of robot about its own position within the environment, show how to compute the region whose coverage can be guaranteed for a given plan, and characterize regions whose coverage cannot be guaranteed by any plan. We also describe an algorithm that generates coverage plans for this robot, based on a search across a specially-constructed graph. Simulation results demonstrate the effectiveness of the approach.

THESIS DEFENSE Department of Computer Science and Engineering University of South Carolina Author : Fawad Kirmani Advisor : Dr. John Rose Date : Nov 14th , 2018 Time : 3:00 pm Place : Meeting room 2265, Innovation Center Abstract The goal of this work is to improve proteotypic peptide prediction with lower processing time and better efficiency. Proteotypic peptides are the peptides in protein sequence that can be confidently observed by mass-spectrometry based proteomics. One of the widely used method for identifying peptides is tandem mass spectrometry (MS/MS). The peptides that need to be identified are compared with the accurate mass and elution time (AMT) tag database. The AMT tag database helps in reducing the processing time and increases the accuracy of the identified peptides. Prediction of proteotypic peptides has seen a rapid improvement in recent years for AMT studies for peptides using amino acid properties like charge, code, solubility and hydropathy. We describe the improved version of a support vector machine (SVM) classifier that has achieved similar classification sensitivity, specificity and AUC on Yersinia Pestis, Saccharomyces cerevisiae and Bacillus subtilis str. 168 datasets as was described by Web-Robertson et al. [13] and Ahmed Alqurri [10]. The improved version of the SVM classifier uses the C++ SVM library instead of the MATLAB built in library. We describe how we achieved these similar results with much lesser processing time. Furthermore, we tested four machine learning classifiers on Yersinia Pestis, Saccharomyces cerevisiae and Bacillus subtilis str. 168 data. We performed feature selection from scratch, using four different algorithms to achieve better results from the different machine learning algorithms. Some of these classifiers gave similar or better results than the SVM classifiers with fewer features. We describe the results of these four classifiers with different feature sets.

DISSERTATION DEFENSE Department of Computer Science and Engineering University of South Carolina Author : Bing Feng Advisor : Dr. Jijun Tang Date : Nov 12th , 2018 Time : 12:00 pm Place : Meeting room 2267, Innovation Center Abstract Studies of bioinformatics develop methods and software tools to analyze biological data and provide insight of the mechanisms of biological processes. Machine learning techniques have been widely used by researchers for disease prediction, disease diagnosis, and bio-marker identification. Using machine learning algorithms to diagnose diseases has a couple of advantages. Besides solely relying on the doctors’ experiences and stereotyped formulas, researchers could use learning algorithms to analyze sophisticated, high-dimensional and multimodal biomedical data, and construct prediction/classification models to make decisions even when some information was incomplete, unknown, or contradictory. In this study, we first build an automated computational pipeline to reconstruct phylogenies and ancestral genomes for two high-resolution real yeast whole genome datasets. We further compare the results with recent studies and publications show that we reconstruct very accurate and robust phylogenies and ancestors. We also identify and analyze conserved syntenic blocks among reconstructed ancestral genomes and present yeast species. Next, we analyzed the metabolic level dataset obtained from the positive mass spectrometry of human blood samples. We applied machine learning algorithms and feature selection algorithms to construct diagnosis models of Chronic kidney diseases (CKD). We also identify the most critical metabolite features and study the correlations among the metabolite features and the developments of CKD stages. The selected metabolite features provided insights into CKD early stage diagnosis, pathophysiological mechanisms, CKD treatments and medicine development. Finally, we use deep learning techniques to build accurate Down Syndrome (DS) prediction/screening models based on the analysis of newly introduced Illumina human genome genotyping array. We proposed a bi-stream convolutional neural network (CNN) architecture with nine layers and two merged CNN models, which took two input chromosome SNP maps in combination. We evaluated and compared the performances of our CNN DS predictions models with conventional machine learning algorithms and single-stream CNN models. We visualized the feature maps and trained filter weights from intermediate layers of our trained CNN model. We further discussed the advantages of our method and the underlying reasons for their performances differences.

You are invited to join Women in Computing November event on Tuesday, Nov. 6. Pizza will be provided and everyone - all genders and majors is welcome! Topic: Professional Development When: Tuesday, November 6th, start from 7 pm Where: Room 2277, IBM Innovation Center/Horizon 2 (the building next to Strom Thurmond Fitness Center that has the IBM logo on the side). Main agenda: Internship preparation, Resume reviews, and LinkedIn headshots

DISSERTATION DEFENSE Author : Zhonghao Liu Advisor : Dr. Jianjun Hu Date : Oct. 29th , 2018 Time : 9:00 am Place : Meeting room 2267, Innovation Center Abstract Machine learning and deep learning methods have been increasingly applied to solve challenging and important bioinformatics problems such as protein structure pre- diction, disease gene identification, and drug discovery. However, the performances of existing machine learning based predictive models are still not satisfactory. The question of how to exploit the specific properties of bioinformatics data and couple them with the unique capabilities of the learning algorithms remains elusive. In this dissertation, we propose advanced machine learning and deep learning algorithms to address two important problems: mislocation-related cancer gene identification and major histocompatibility complex-peptide binding affinity prediction. Our first contribution proposes a kernel-based logistic regression algorithm for identifying potential mislocation-related genes among known cancer genes. Our algorithm takes protein-protein interaction networks, gene expression data, and subcellular location gene ontology data as input, which is particularly lightweight comparing with existing methods. The experiment results demonstrate that our proposed pipeline has a good capability to identify mislocation-related cancer genes. Our second contribution addresses the modeling and prediction of human leukocyte antigen (HLA) peptide binding of human immune system. We present an allele-specific convolutional neural network model with one-hot encoding. With extensive evaluation over the standard IEDB datasets, it is shown that the performance of our model is better than all existing prediction models. To achieve further improvement, we propose a novel pan-specific model on peptide-HLA class I binding affinities prediction, which allows us to exploit all the training samples of different HLA alleles. Our sequence-based pan model is currently the only algorithm not using pseudo sequence encoding — a dominant structure-based encoding method in this area. The benchmark studies show that our method could achieve state-of-the-art performance. Our proposed model could be integrated into existing ensemble methods to improve their overall prediction capabilities on highly diverse MHC alleles. Finally, we present a LSTM-CNN deep learning model with attention mechanism for peptide-HLA class II binding affinities and binding cores prediction. Our model achieved very good performance and outperformed existing methods on half of tested alleles. With the help of attention mechanism, our model could directly output the peptide binding core based on attention weight without any additional post- or pre- processing.

ACM is hosting a 24 hour Code-A-Thon on Friday, October 26th in Swearingen 1D11 at 7PM (and also online) . The Code-A-Thon is open to all majors and all skill levels. If you can’t make the opening event, compete online here: http://www.hackerrank.com/usc-acm-fall-2018-145-division http://www.hackerrank.com/usc-acm-fall-2018-146-division http://www.hackerrank.com/usc-acm-fall-2018-240-division http://www.hackerrank.com/usc-acm-fall-2018-350-division Pick the division of the highest CS course you are enrolled in or have taken (out of 145, 146, 240, and 350). If you haven’t taken any of these CS classes, take the 145 division. If you are a CS graduate student, or have already taken 350, compete in the 350 division. You can pick a lower division, but you won’t be able to compete for prizes. Prizes are:

• 32GB flash drive & wireless keyboard
• Raspberry Pi 3 B+
• 100,000 mAh mobile battery

Let me know if you have any questions, James Coman President, ACM University of South Carolina | Class of 2019 ACM Friendface Page

DISSERTATION DEFENSE Department of Computer Science and Engineering University of South Carolina Author : Chao Chen Advisor : Dr. Gabriel Terejanu Date : Oct. 15th , 2018 Time : 2:30 pm Place : Meeting room 2267, Innovation Center Abstract Normal neural networks trained with gradient descent and back-propagation have received great success in various applications. On one hand, point estimation of the network weights is prone to over-fitting problems and lacks important uncertainty information associated with the estimation. On the other hand, exact Bayesian neural network methods are intractable and non-applicable for real-world applications. To date, approximate methods have been actively under development for Bayesian neural networks, including but not limited to, stochastic variational methods, Monte Carlo dropouts, and expectation propagation. Though these methods are applicable for current large networks, there are limits of these approaches with either under estimation or over-estimation of uncertainty. Extended Kalman filters (EKFs) and unscented Kalman filters (UKFs), which are widely used in data assimilation community, adopt a different perspective of inferring the parameters. Nevertheless, EKFs are incapable of dealing with highly non-linearity, while UKFs are inapplicable for large network architectures. Ensemble Kalman filters (EnKFs) serve as great methodology in atmosphere and oceanology disciplines targeting extremely high-dimensional, non-Gaussian, and nonlinear state-space models. So far, there is little work that applies EnKFs to estimate the parameters of deep neural networks. By considering neural network as a nonlinear function, we augment the network prediction with parameters as new states and adapt the state-space model to update the parameters. In the first work, we describe the ensemble Kalman filter, two proposed algorithms for training both fully-connected and Long Short-term Memory (LSTM) networks, and experiment it with a synthetic dataset, 10 UCI datasets, and a natural language dataset for different regression tasks. To further evaluate the effectiveness of the proposed training scheme, we trained a deep LSTM network with the proposed algorithm, and applied it on five real-world sub-event detection tasks. With a formalization of the sub-event detection task, we develop an outlier detection framework and take advantage of the Bayesian Long Short-term Memory (LSTM) network to capture the important and interesting moments within an event. In the last work, we develop a framework for student knowledge estimation using Bayesian network. By constructing student models with Bayesian network, we can infer the new state of knowledge on each concept given a student. With a novel parameter estimate algorithm, the model can also indicate misconception on each question. Furthermore, we develop a predictive validation metric with expected data likelihood of the student model to evaluate the design of questions.

THESIS DEFENSE Department of Computer Science and Engineering University of South Carolina Author : Lacie Renee Stiffler Advisor : Dr. Bakos Date : October 10th, 2018 Time : 10:00 am Place : 2267, Storey Innovation Center Abstract Artificial neural networks are an effective machine learning technique for a variety of data sets and domains, but exploiting the inherent parallelism in neural networks requires specialized hardware. Typically, computing the output of each neuron requires many multiplications, evaluation of a transcendental activation function, and transfer of its output to a large number of other neurons. These restrictions become more expensive when internal values are represented with increasingly higher data precision. A spiking neural network eliminates the limitations of typical rate-based neural networks by reducing neuron output and synapse weights to one-bit values, eliminating hardware multipliers, and simplifying the activation function. However, a spiking neural network requires a larger number of neurons than what is needed in a comparable rate-based network. In order to determine if the benefits of spiking neural networks outweigh the costs, we designed the smallest spiking neural network and rate-based artificial neural network that achieved 90\% or comparable testing accuracy on the MNIST data set. After estimating the FPGA storage requirements for synapse values of each network, we concluded rate-based neural networks need significantly fewer bits than spiking neural networks.

THESIS DEFENSE Department of Computer Science and Engineering Author : George Akhvlediani Advisor : Dr. Buell Date : September 28th Time : 10:00 am Place : 2265, Storey Innovation Center Abstract This research examines conflicts accompanying the proliferation of computer technology and, more specifically, constellations of dependency in the always expanding volume of software, platforms, and the firms/individuals using them. We identify a pervasive phenomenon of “divarication” in the growing variety of progressively specialized systems and system roles. As software systems enter new thresholds of sophistication, they effectively aggregate many distinct components and protocols. Consequently, we are confronted with a diverse ecology of stratified and thereby incompatible software systems. Software inherits the limitations and potential flaws of its constituent parts, but unlike physical machinery, it isn’t readily disassembled in instances of failure. The individuals using these systems have no means to dissect and analyze their tools, and thus are necessarily dependent on developer assistance. We assert that divarication is a consequence of interfacing, and particularly in the way computer interfaces operate as the sole point of contact between a user and a software system. Taking Charles S. Peirce’s three types of sign (the icon, index, and symbol) into special consideration, we observe that computer interfaces seldom employ iconic representation. In other words, these interfaces do not reflect the interior logic that drives them; they bear no resemblance to their referent(s). Merely “using” software doesn’t promise any insight into how that software works. We argue that this circumstance makes divarication inevitable. Opaque elements are assembled together into opaque wholes, and so the magnitude of this problem will likely scale with increasing software sophistication. As the thesis title indicates, we bring Peirce’s notion of “iconicity” into accompaniment with “interfacing”, forming an abstract paradigm in response to divarication. We intend to infuse a software platform with a recurrent protocol of iconicity, to develop a platform that allows at least partial disassembly and examination of the programs it facilitates. We composed a diagrammatic design scheme; a blueprint for software platforms that might emulate “interfacing iconicity”. We developed a prototype platform, implementing this structural logic. This initial prototype is a rudimentary HTML rendering platform, one that articulates the relationship between plain-text code, its Document Object Model (DOM) representation, and the rendered “page” itself. Currently, this prototype is a useful analog for our argument. Since it offers a distinct perspective on the connections between text markup and its systemic interpretation, it may also have educational utility. However, it is not yet a fully realized implementation of our design paradigm, and at this stage a conclusion on whether the latter genuinely addresses divarication would be premature.