Computer Sciences Commons^™

Distxplore: Distribution-Guided Testing For Evaluating And Enhancing Deep Learning Systems, Longtian Wang, Xiaofei Xie, Xiaoning Du, Meng Tian, Qing Guo, Zheng Yang, Chao Shen

Research Collection School Of Computing and Information Systems

Deep learning (DL) models are trained on sampled data, where the distribution of training data differs from that of real-world data (i.e., the distribution shift), which reduces the model's robustness. Various testing techniques have been proposed, including distribution-unaware and distribution-aware methods. However, distribution-unaware testing lacks effectiveness by not explicitly considering the distribution of test cases and may generate redundant errors (within same distribution). Distribution-aware testing techniques primarily focus on generating test cases that follow the training distribution, missing out-of-distribution data that may also be valid and should be considered in the testing process. In this paper, we propose a novel …

(Un)Likelihood Training For Interpretable Embedding, Jiaxin Wu, Chong-Wah Ngo, Wing-Kwong Chan, Zhijian Hou

Research Collection School Of Computing and Information Systems

Cross-modal representation learning has become a new normal for bridging the semantic gap between text and visual data. Learning modality agnostic representations in a continuous latent space, however, is often treated as a black-box data-driven training process. It is well known that the effectiveness of representation learning depends heavily on the quality and scale of training data. For video representation learning, having a complete set of labels that annotate the full spectrum of video content for training is highly difficult, if not impossible. These issues, black-box training and dataset bias, make representation learning practically challenging to be deployed for video …

Deeparc: Modularizing Neural Networks For The Model Maintenance, Xiaoning Ren, Yun Lin, Yinxing Xue, Ruofan Liu, Jun Sun, Zhiyong Feng, Jinsong Dong

Research Collection School Of Computing and Information Systems

Neural networks are an emerging data-driven programming paradigm widely used in many areas. Unlike traditional software systems consisting of decomposable modules, a neural network is usually delivered as a monolithic package, raising challenges for some maintenance tasks such as model restructure and re-adaption. In this work, we propose DeepArc, a novel modularization method for neural networks, to reduce the cost of model maintenance tasks. Specifically, DeepArc decomposes a neural network into several consecutive modules, each of which encapsulates consecutive layers with similar semantics. The network modularization facilitates practical tasks such as refactoring the model to preserve existing features (e.g., model …

Uavs And Deep Neural Networks: An Alternative Approach To Monitoring Waterfowl At The Site Level, Zachary J. Loken

LSU Master's Theses

Understanding how waterfowl respond to habitat restoration and management activities is crucial for evaluating and refining conservation delivery programs. However, site-specific waterfowl monitoring is challenging, especially in heavily forested systems such as the Mississippi Alluvial Valley (MAV)—a primary wintering region for ducks in North America. I hypothesized that using uncrewed aerial vehicles (UAVs) coupled with deep learning-based methods for object detection would provide an efficient and effective means for surveying non-breeding waterfowl on difficult-to-access restored wetland sites. Accordingly, during the winters of 2021 and 2022, I surveyed wetland restoration easements in the MAV using a UAV equipped with a dual …

Faire: Repairing Fairness Of Neural Networks Via Neuron Condition Synthesis, Tianlin Li, Xiaofei Xie, Jian Wang, Qing Guo, Aishan Liu, Lei Ma, Yang Liu

Research Collection School Of Computing and Information Systems

Deep Neural Networks (DNNs) have achieved tremendous success in many applications, while it has been demonstrated that DNNs can exhibit some undesirable behaviors on concerns such as robustness, privacy, and other trustworthiness issues. Among them, fairness (i.e., non-discrimination) is one important property, especially when they are applied to some sensitive applications (e.g., finance and employment). However, DNNs easily learn spurious correlations between protected attributes (e.g., age, gender, race) and the classification task and develop discriminatory behaviors if the training data is imbalanced. Such discriminatory decisions in sensitive applications would introduce severe social impacts. To expose potential discrimination problems in DNNs …

Investigating Continual Learning Strategies In Neural Networks, Christopher Tam, Luiz Fernando Capretz

Electrical and Computer Engineering Publications

This paper explores the role of continual learning strategies when neural networks are confronted with learning tasks sequentially. We analyze the stability-plasticity dilemma with three factors in mind: the type of network architecture used, the continual learning scenario defined and the continual learning strategy implemented. Our results show that complementary learning systems and neural volume significantly contribute towards memory retrieval and consolidation in neural networks. Finally, we demonstrate how regularization strategies such as elastic weight consolidation are more well-suited for larger neural networks whereas rehearsal strategies such as gradient episodic memory are better suited for smaller neural networks.

A System For The Detection Of Adversarial Attacks In Computer Vision Via Performance Metrics, Sarah Reynolds

Doctoral Dissertations and Master's Theses

Adversarial attacks, or attacks committed by an adversary to hijack a system, are prevalent in the deep learning tasks of computer vision and are one of the greatest threats to these models' safe and accurate use. These attacks force the trained model to misclassify an image, using pixel-level changes undetectable to the human eye. Various defenses against these attacks exist and are detailed in this work. The work of previous researchers has established that when adversarial attacks occur, different node patterns in a Deep Neural Network (DNN) are activated within the model. Additionally, it is known that CPU and GPU …

Effects Of Weight Initialization Methods On Ffn's, Ida K. Karem

The Cardinal Edge

Weight initialization is the method of determining starting values of weights in a neural network. The way this method is done can have massive effects on the network[2, 3, 6, 9] and can halt training if not handled properly. On the other hand, if initialization is chosen tactfully it can improve training and accuracy greatly. The initialization method usually called Normalized Xavier will be referred to as Nox in this paper to avoid confusion with the Xavier initialization method. This study analyzes five methods of weight initialization(Nox, He, Xavier, Plutonian, and Self-Root), two of them …

Dynamic Graph Enhanced Contrastive Learning For Chest X-Ray Report Generation, Mingjie Li, Bingqian Lin, Zicong Chen, Haokun Lin, Xiaodan Liang, Xiaojun Chang

Computer Vision Faculty Publications

Automatic radiology reporting has great clinical potential to relieve radiologists from heavy workloads and improve diagnosis interpretation. Recently, researchers have enhanced data-driven neural networks with medical knowledge graphs to eliminate the severe visual and textual bias in this task. The structures of such graphs are exploited by using the clinical dependencies formed by the disease topic tags via general knowledge and usually do not update during the training process. Consequently, the fixed graphs can not guarantee the most appropriate scope of knowledge and limit the effectiveness. To address the limitation, we propose a knowledge graph with Dynamic structure and nodes …

Physics-Guided Deep Learning For Solar Wind Modeling At L1 Point, Robert M. Johnson

All Graduate Theses and Dissertations, Spring 1920 to Summer 2023

Neural networks are adept at finding patterns that are too long and too small for humans to find in data. Usually, this power is used to generate predictions with greater accuracy than most alternative models. However, we can also use this power to understand more about the data we train these networks on. We do this by changing the data that the networks train on and the data they are tested on. This allows us to both control the maximum length of a pattern and to compare data between different groups, in our case, different solar cycles. This thesis is …

Neuro-Symbolic Representations For Information Retrieval, Laura Dietz, Hannah Bast, Shubham Chatterjee, Jeff Dalton, Jian Yun Nie, Rodrigo Nogueira

Computer Science Faculty Research & Creative Works

This tutorial will provide an overview of recent advances on neuro-symbolic approaches for information retrieval. A decade ago, knowledge graphs and semantic annotations technology led to active research on how to best leverage symbolic knowledge. At the same time, neural methods have demonstrated to be versatile and highly effective. From a neural network perspective, the same representation approach can service document ranking or knowledge graph reasoning. End-to-end training allows to optimize complex methods for downstream tasks. We are at the point where both the symbolic and the neural research advances are coalescing into neuro-symbolic approaches. The underlying research questions are …

Lidar Segmentation-Based Adversarial Attacks On Autonomous Vehicles, Blake Johnson

Scholarly Horizons: University of Minnesota, Morris Undergraduate Journal

Autonomous vehicles utilizing LiDAR-based 3D perception systems are susceptible to adversarial attacks. This paper focuses on a specific attack scenario that relies on the creation of adversarial point clusters with the intention of fooling the segmentation model utilized by LiDAR into misclassifying point cloud data. This can be translated into the real world with the placement of objects (such as road signs or cardboard) at these adversarial point cluster locations. These locations are generated through an optimization algorithm performed on said adversarial point clusters that are introduced by the attacker.

An Analysis Of Text-Based Machine Learning Models For Vulnerability Detection, Kollin Ryne Napier

Theses and Dissertations

With an increase in complexity of software, developers rely more on reuse and dependencies in their source code via code snippets. As a result, it is becoming harder to identify and mitigate vulnerabilities. Although traditional analysis tools are still utilized, machine learning models are being adopted to expand efforts and combat such threats. Given the possibilities towards usage of such models, research in this area has introduced various approaches which vary in usability and prediction. In generalizing models to a more natural language approach, researchers have opted to train models on source code to identify existing and potential vulnerabilities. Exploratory …

Improving The Performance Of Automatic Target Recognition Using Data Augmentation., Rolando Holmes

Electronic Theses and Dissertations

Automatic Target Recognition (ATR) is a task that aims to recognize targets based on data obtained from visual sensors such as Infrared cameras. Machine learning, in particular Deep Learning, has been the dominate approach to solve this task. However, in order to achieve high performance, Deep Learning requires a large amount of data that may not be available, which poses an issue when developing robust models. In this thesis, we propose the use of data augmentation from an automatic detector as an alternative source to supplement the need for more training data. We use a two stage approach to develop …

Integration Of Neural Network And Distance Relay To Improve The Fault Localization On Transmission Lines, Linh Tran

Turkish Journal of Electrical Engineering and Computer Sciences

Power transmission lines are integral and very important components of power systems. Because of the length of these lines and the complexity of the power grids, the lines may encounter various incidents such as lightning strike, shortage, and breakage. When an incident or a fault occurs, a fast process of identification, localization, and isolation of the fault is desired. An accurate fault localization would have a great impact in reducing the restoration time of the system. One of the most popular solutions for fault detection and localization is the distance relays using the impedance-based algorithms. However, these relays are still …

Nviz: Unraveling Neural Networks Through Visualization, Kevin Hoffman

Mathematics, Computer Science & Statistics Presentations

The growing utility of artificial intelligence (AI) is attributed to the development of neural networks. These networks are a class of models that make predictions based on previously observed data. While the inferential power of neural networks is great, the ability to explain their results is difficult because the underlying model is automatically generated. The AI community commonly refers to neural networks as black boxes because the patterns they learn from the data are not easily understood. This project aims to improve the visibility of patterns that neural networks identify in data. Through an interactive web application, NVIZ affords the …

Loss Scaling And Step Size In Deep Learning Optimizatio, Nora Alosily

Dissertations

Deep learning training consumes ever-increasing time and resources, and that is
due to the complexity of the model, the number of updates taken to reach good
results, and both the amount and dimensionality of the data. In this dissertation,
we will focus on making the process of training more efficient by focusing on the
step size to reduce the number of computations for parameters in each update.
We achieved our objective in two new ways: we use loss scaling as a proxy for
the learning rate, and we use learnable layer-wise optimizers. Although our work
is perhaps not the first …

Convolutional Neural Networks Analysis Reveals Three Possible Sources Of Bronze Age Writings Between Greece And India, Shruti Daggumati, Peter Z. Revesz

School of Computing: Faculty Publications

This paper analyzes the relationships among eight ancient scripts from between Greece and India. We used convolutional neural networks combined with support vector machines to give a numerical rating of the similarity between pairs of signs (one sign from each of two different scripts). Two scripts that had a one-to-one matching of their signs were determined to be related. The result of the analysis is the finding of the following three groups, which are listed in chronological order: (1) Sumerian pictograms, the Indus Valley script, and the proto-Elamite script; (2) Cretan hieroglyphs and Linear B; and (3) the Phoenician, Greek, …

Applications Of Generative Adversarial Networks In Single Image Datasets, Dylan E. Cramer

Scholarly Horizons: University of Minnesota, Morris Undergraduate Journal

One of the main difficulties faced in most generative machine learning models is how much data is required to train it, especially when collecting a large dataset is not feasible. Recently there have been breakthroughs in tackling this issue in SinGAN, with its researchers being able to train a Generative Adversarial Network (GAN) on just a single image with a model that can perform many novel tasks, such as image harmonization. ConSinGAN is a model that builds upon this work by concurrently training several stages in a sequential multi-stage manner while retaining the ability to perform those novel tasks.

Effectiveness Of A Timing Side Channel For Deriving Neural Network Depth, Matthew P. Weeks

Theses and Dissertations

From facial recognition on cell phones to vehicle traffic modeling for city planning, integrating ML models can be an expensive investment in resources. Protecting that investment is difficult, as information about the model and how it was built can be leaked through multiple channels, such as timing and memory access. In this thesis, one method of extracting data through a timing side-channel is examined across multiple hardware and software configurations to determine its reliability for general use. While attempting to determine the layer count of a target model solely from its inference time, the research determined that it is not …

Learning And Understanding User Interface Semantics From Heterogeneous Networks With Multimodal And Positional Attributes, Meng Kiat Gary Ang, Ee-Peng Lim

Research Collection School Of Computing and Information Systems

User interfaces (UI) of desktop, web, and mobile applications involve a hierarchy of objects (e.g., applications, screens, view class, and other types of design objects) with multimodal (e.g., textual and visual) and positional (e.g., spatial location, sequence order, and hierarchy level) attributes. We can therefore represent a set of application UIs as a heterogeneous network with multimodal and positional attributes. Such a network not only represents how users understand the visual layout of UIs but also influences how users would interact with applications through these UIs. To model the UI semantics well for different UI annotation, search, and evaluation tasks, …

Emotion Classification Of Indonesian Tweets Using Bidirectional Lstm, Aaron K. Glenn, Phillip M. Lacasse, Bruce A. Cox

Faculty Publications

Emotion classification can be a powerful tool to derive narratives from social media data. Traditional machine learning models that perform emotion classification on Indonesian Twitter data exist but rely on closed-source features. Recurrent neural networks can meet or exceed the performance of state-of-the-art traditional machine learning techniques using exclusively open-source data and models. Specifically, these results show that recurrent neural network variants can produce more than an 8% gain in accuracy in comparison with logistic regression and SVM techniques and a 15% gain over random forest when using FastText embeddings. This research found a statistical significance in the performance of …

Continual Learning-Based Optimal Output Tracking Of Nonlinear Discrete-Time Systems With Constraints: Application To Safe Cargo Transfer, Behzad Farzanegan, S. (Sarangapani) Jagannathan

Electrical and Computer Engineering Faculty Research & Creative Works

This Paper Addresses a Novel Lifelong Learning (LL)-Based Optimal Output Tracking Control of Uncertain Non-Linear Affine Discrete-Time Systems (DT) with State Constraints. First, to Deal with Optimal Tracking and Reduce the Steady State Error, a Novel Augmented System, Including Tracking Error and its Integral Value and Desired Trajectory, is Proposed. to Guarantee Safety, an Asymmetric Barrier Function (BF) is Incorporated into the Utility Function to Keep the Tracking Error in a Safe Region. Then, an Adaptive Neural Network (NN) Observer is Employed to Estimate the State Vector and the Control Input Matrix of the Uncertain Nonlinear System. Next, an NN-Based …

Adversarial Training Of Deep Neural Networks, Anabetsy Termini

CCAC Theses and Dissertations

Deep neural networks used for image classification are highly susceptible to adversarial attacks. The de facto method to increase adversarial robustness is to train neural networks with a mixture of adversarial images and unperturbed images. However, this method leads to robust overfitting, where the network primarily learns to recognize one specific type of attack used to generate the images while remaining vulnerable to others after training. In this dissertation, we performed a rigorous study to understand whether combinations of state of the art data augmentation methods with Stochastic Weight Averaging improve adversarial robustness and diminish adversarial overfitting across a wide …

More Human Than Human: Llm-Generated Narratives Outperform Human-Llm Interleaved Narratives, Zoie Zhao, Sophie Song, Bridget Duah, Jamie C. Macbeth, Scott Carter, Monica Van, Nayeli Bravo, Matthew Klenk, Katherine Sieck, Alexandre Filipowicz

Computer Science: Faculty Publications

Narrative story generation has gained emerging interest in the field of large language models. The present paper aims to compare stories generated by an LLM only (non-interleaved) with those generated by interleaving human-generated and LLM-generated text (interleaved). The study’s hypothesis is that interleaved stories would perform better than non-interleaved stories. To verify this hypothesis, we conducted two tests with roughly 500 participants each. Participants were asked to rate stories of each type, including an overall score or preference and four facets—logical soundness, plausibility, understandability, and novelty. Our findings indicate that interleaved stories were in fact less preferred than non-interleaved stories. …

Artificial Intelligence-Enabled Exploratory Cyber-Physical Safety Analyzer Framework For Civilian Urban Air Mobility, Md. Shirajum Munir, Sumit Howlader Dipro, Kamrul Hasan, Tariqul Islam, Sachin Shetty

VMASC Publications

Urban air mobility (UAM) has become a potential candidate for civilization for serving smart citizens, such as through delivery, surveillance, and air taxis. However, safety concerns have grown since commercial UAM uses a publicly available communication infrastructure that enhances the risk of jamming and spoofing attacks to steal or crash crafts in UAM. To protect commercial UAM from cyberattacks and theft, this work proposes an artificial intelligence (AI)-enabled exploratory cyber-physical safety analyzer framework. The proposed framework devises supervised learning-based AI schemes such as decision tree, random forests, logistic regression, K-nearest neighbors (KNN), and long short-term memory (LSTM) for predicting and …

Sign Language Recognition Using A Hybrid Machine Learning Model, Peeyusha Shivayogi

Master's Projects

Sign Language is a visual language used by millions of people around the world. American Sign Language (ASL) is one of the most popular sign languages and the third most popular language in the United States. Automatic recognition of ASL signs can help bridge the communication gap between deaf and hearing individuals. In this project, we explore the use of deep learning models for ASL sign recognition, using the MNIST dataset as a benchmark. We preprocessed the data by reshaping the images to the input layer size of the models and normalized the pixel values. We evaluated five popular deep-learning …

Machine Learning-Based Anomaly Detection In Cloud Virtual Machine Resource Usage, Tarun Mourya Satveli

Master's Projects

Anomaly detection is an important activity in cloud computing systems because it aids in the identification of odd behaviours or actions that may result in software glitch, security breaches, and performance difficulties. Detecting aberrant resource utilization trends in virtual machines is a typical application of anomaly detection in cloud computing (VMs). Currently, the most serious cyber threat is distributed denial-of-service attacks. The afflicted server's resources and internet traffic resources, such as bandwidth and buffer size, are slowed down by restricting the server's capacity to give resources to legitimate customers.

To recognize attacks and common occurrences, machine learning techniques such as …

Leveraging Tweets For Rapid Disaster Response Using Bert-Bilstm-Cnn Model, Satya Pranavi Manthena

Master's Projects

Digital networking sites such as Twitter give a global platform for users to discuss and express their own experiences with others. People frequently use social media to share their daily experiences, local news, and activities with others. Many rescue services and agencies frequently monitor this sort of data to identify crises and limit the danger of loss of life. During a natural catastrophe, many tweets are made in reference to the tragedy, making it a hot topic on Twitter. Tweets containing natural disaster phrases but do not discuss the event itself are not informational and should be labeled as non-disaster …

Dfhic: A Dilated Full Convolution Model To Enhance The Resolution Of Hi-C Data, Bin Wang, Kun Liu, Yaohang Li, Jianxin Wang

Computer Science Faculty Publications

Motivation: Hi-C technology has been the most widely used chromosome conformation capture(3C) experiment that measures the frequency of all paired interactions in the entire genome, which is a powerful tool for studying the 3D structure of the genome. The fineness of the constructed genome structure depends on the resolution of Hi-C data. However, due to the fact that high-resolution Hi-C data require deep sequencing and thus high experimental cost, most available Hi-C data are in low-resolution. Hence, it is essential to enhance the quality of Hi-C data by developing the effective computational methods.

Results: In this work, we propose …