Defense Notices
All students and faculty are welcome to attend the final defense of EECS graduate students completing their M.S. or Ph.D. degrees. Defense notices for M.S./Ph.D. presentations for this year and several previous years are listed below in reverse chronological order.
Students who are nearing the completion of their M.S./Ph.D. research should schedule their final defenses through the EECS graduate office at least THREE WEEKS PRIOR to their presentation date so that there is time to complete the degree requirements check, and post the presentation announcement online.
Upcoming Defense Notices
Mahmudul Hasan
Assertion-Based Security Assessment of Hardware IP Protection MethodsWhen & Where:
Eaton Hall, Room 2001B
Committee Members:
Tamzidul Hoque, ChairEsam El-Araby
Sumaiya Shomaji
Abstract
Combinational and sequential locking methods are promising solutions for protecting hardware intellectual property (IP) from piracy, reverse engineering, and malicious modifications by locking the functionality of the IP based on a secret key. To improve their security, researchers are developing attack methods to extract the secret key.
While the attacks on combinational locking are mostly inapplicable for sequential designs without access to the scan chain, the limited applicable attacks are generally evaluated against the basic random insertion of key gates. On the other hand, attacks on sequential locking techniques suffer from scalability issues and evaluation of improperly locked designs. Finally, while most attacks provide an approximately correct key, they do not indicate which specific key bits are undetermined. This thesis proposes an oracle-guided attack that applies to both combinational and sequential locking without scan chain access. The attack applies light-weight design modifications that represent the oracle using a finite state machine and applies an assertion-based query of the unlocking key. We have analyzed the effectiveness of our attack against 46 sequential designs locked with various classes of combinational locking including random, strong, logic cone-based, and anti-SAT based. We further evaluated against a sequential locking technique using 46 designs with various key sequence lengths and widths. Finally, we expand our framework to identify undetermined key bits, enabling complementary attacks on the smaller remaining key space.
Masoud Ghazikor
Distributed Optimization and Control Algorithms for UAV Networks in Unlicensed Spectrum BandsWhen & Where:
Nichols Hall, Room 246 (Executive Conference Room)
Committee Members:
Morteza Hashemi, ChairVictor Frost
Prasad Kulkarni
Abstract
UAVs have emerged as a transformative technology for various applications, including emergency services, delivery, and video streaming. Among these, video streaming services in areas with limited physical infrastructure, such as disaster-affected areas, play a crucial role in public safety. UAVs can be rapidly deployed in search and rescue operations to efficiently cover large areas and provide live video feeds, enabling quick decision-making and resource allocation strategies. However, ensuring reliable and robust UAV communication in such scenarios is challenging, particularly in unlicensed spectrum bands, where interference from other nodes is a significant concern. To address this issue, developing a distributed transmission control and video streaming is essential to maintaining a high quality of service, especially for UAV networks that rely on delay-sensitive data.
In this MSc thesis, we study the problem of distributed transmission control and video streaming optimization for UAVs operating in unlicensed spectrum bands. We develop a cross-layer framework that jointly considers three inter-dependent factors: (i) in-band interference introduced by ground-aerial nodes at the physical layer, (ii) limited-size queues with delay-constrained packet arrival at the MAC layer, and (iii) video encoding rate at the application layer. This framework is designed to optimize the average throughput and PSNR by adjusting fading thresholds and video encoding rates for an integrated aerial-ground network in unlicensed spectrum bands. Using consensus-based distributed algorithm and coordinate descent optimization, we develop two algorithms: (i) Distributed Transmission Control (DTC) that dynamically adjusts fading thresholds to maximize the average throughput by mitigating trade-offs between low-SINR transmission errors and queue packet losses, and (ii) Joint Distributed Video Transmission and Encoder Control (JDVT-EC) that optimally balances packet loss probabilities and video distortions by jointly adjusting fading thresholds and video encoding rates. Through extensive numerical analysis, we demonstrate the efficacy of the proposed algorithms under various scenarios.
Ganesh Nurukurti
Customer Behavior Analytics and Recommendation System for E-CommerceWhen & Where:
Eaton Hall, Room 2001B
Committee Members:
David Johnson, ChairPrasad Kulkarni
Han Wang
Abstract
In the era of digital commerce, personalized recommendations are pivotal for enhancing user experience and boosting engagement. This project presents a comprehensive recommendation system integrated into an e-commerce web application, designed using Flask and powered by collaborative filtering via Singular Value Decomposition (SVD). The system intelligently predicts and personalizes product suggestions for users based on implicit feedback such as purchases, cart additions, and search behavior.
The foundation of the recommendation engine is built on user-item interaction data, derived from the Brazilian e-commerce Olist dataset. Ratings are simulated using weighted scores for purchases and cart additions, reflecting varying degrees of user intent. These interactions are transformed into a user-product matrix and decomposed using SVD, yielding latent user and product features. The model leverages these latent factors to predict user interest in unseen products, enabling precise and scalable recommendation generation.
To further enhance personalization, the system incorporates real-time user activity. Recent search history is stored in an SQLite database and used to prioritize recommendations that align with the user’s current interests. A diversity constraint is also applied to avoid redundancy, limiting the number of recommended products per category.
The web application supports robust user authentication, product exploration by category, cart management, and checkout simulations. It features a visually driven interface with dynamic visualizations for product insights and user interactions. The home page adapts to individual preferences, showing tailored product recommendations and enabling users to explore categories and details.
In summary, this project demonstrates the practical implementation of a hybrid recommendation strategy combining matrix factorization with contextual user behavior. It showcases the importance of latent factor modeling, data preprocessing, and user-centric design in delivering an intelligent retail experience.
Srijanya Chetikaneni
Plant Disease Prediction Using Transfer LearningWhen & Where:
Eaton Hall, Room 2001B
Committee Members:
David Johnson, ChairPrasad Kulkarni
Han Wang
Abstract
Timely detection of plant diseases is critical to safeguarding crop yields and ensuring global food security. This project presents a deep learning-based image classification system to identify plant diseases using the publicly available PlantVillage dataset. The core objective was to evaluate and compare the performance of a custom-built Convolutional Neural Network (CNN) with two widely used transfer learning models—EfficientNetB0 and MobileNetV3Small.
All models were trained on augmented image data resized to 224×224 pixels, with preprocessing tailored to each architecture. The custom CNN used simple normalization, whereas EfficientNetB0 and MobileNetV3Small utilized their respective pre-processing methods to standardize the pretrained ImageNet domain inputs. To improve robustness, the training pipeline included data augmentation, class weighting, and early stopping.
Training was conducted using the Adam optimizer and categorical cross-entropy loss over 30 epochs, with performance assessed using accuracy, loss, and training time metrics. The results revealed that transfer learning models significantly outperformed the custom CNN. EfficientNetB0 achieved the highest accuracy, making it ideal for high-precision applications, while MobileNetV3Small offered a favorable balance between speed and accuracy, making it suitable for lightweight, real-time inference on edge devices.
This study validates the effectiveness of transfer learning for plant disease detection tasks and emphasizes the importance of model-specific preprocessing and training strategies. It provides a foundation for deploying intelligent plant health monitoring systems in practical agricultural environments.
Ahmet Soyyigit
Anytime Computing Techniques for LiDAR-based Perception In Cyber-Physical SystemsWhen & Where:
Nichols Hall, Room 250 (Gemini Room)
Committee Members:
Heechul Yun, ChairMichael Branicky
Prasad Kulkarni
Hongyang Sun
Shawn Keshmiri
Abstract
The pursuit of autonomy in cyber-physical systems (CPS) presents a challenging task of real-time interaction with the physical world, prompting extensive research in this domain. Recent advances in artificial intelligence (AI), particularly the introduction of deep neural networks (DNN), have significantly improved the autonomy of CPS, notably by boosting perception capabilities.
CPS perception aims to discern, classify, and track objects of interest in the operational environment, a task that is considerably challenging for computers in a three-dimensional (3D) space. For this task, the use of LiDAR sensors and processing their readings with DNNs has become popular because of their excellent performance However, in CPS such as self-driving cars and drones, object detection must be not only accurate but also timely, posing a challenge due to the high computational demand of LiDAR object detection DNNs. Satisfying this demand is particularly challenging for on-board computational platforms due to size, weight, and power constraints. Therefore, a trade-off between accuracy and latency must be made to ensure that both requirements are satisfied. Importantly, the required trade-off is operational environment dependent and should be weighted more on accuracy or latency dynamically at runtime. However, LiDAR object detection DNNs cannot dynamically reduce their execution time by compromising accuracy (i.e. anytime computing). Prior research aimed at anytime computing for object detection DNNs using camera images is not applicable to LiDAR-based detection due to architectural differences. This thesis addresses these challenges by proposing three novel techniques: Anytime-LiDAR, which enables early termination with reasonable accuracy; VALO (Versatile Anytime LiDAR Object Detection), which implements deadline-aware input data scheduling; and MURAL (Multi-Resolution Anytime Framework for LiDAR Object Detection), which introduces dynamic resolution scaling. Together, these innovations enable LiDAR-based object detection DNNs to make effective trade-offs between latency and accuracy under varying operational conditions, advancing the practical deployment of LiDAR object detection DNNs.
Rahul Purswani
Finetuning Llama on custom data for QA tasksWhen & Where:
Eaton Hall, Room 2001B
Committee Members:
David Johnson, ChairDrew Davidson
Prasad Kulkarni
Abstract
Fine-tuning large language models (LLMs) for domain-specific use cases, such as question answering, offers valuable insights into how their performance can be tailored to specialized information needs. In this project, we focused on the University of Kansas (KU) as our target domain. We began by scraping structured and unstructured content from official KU webpages, covering a wide array of student-facing topics including campus resources, academic policies, and support services. From this content, we generated a diverse set of question-answer pairs to form a high-quality training dataset. LLaMA 3.2 was then fine-tuned on this dataset to improve its ability to answer KU-specific queries with greater relevance and accuracy. Our evaluation revealed mixed results—while the fine-tuned model outperformed the base model on most domain-specific questions, the original model still had an edge in handling ambiguous or out-of-scope prompts. These findings highlight the strengths and limitations of domain-specific fine-tuning, and provide practical takeaways for customizing LLMs for real-world QA applications.
Rithvij Pasupuleti
A Machine Learning Framework for Identifying Bioinformatics Tools and Database Names in Scientific LiteratureWhen & Where:
LEEP2, Room 2133
Committee Members:
Cuncong Zhong, ChairDongjie Wang
Han Wang
Zijun Yao
Abstract
The absence of a single, comprehensive database or repository cataloging all bioinformatics databases and software creates a significant barrier for researchers aiming to construct computational workflows. These workflows, which often integrate 10–15 specialized tools for tasks such as sequence alignment, variant calling, functional annotation, and data visualization, require researchers to explore diverse scientific literature to identify relevant resources. This process demands substantial expertise to evaluate the suitability of each tool for specific biological analyses, alongside considerable time to understand their applicability, compatibility, and implementation within a cohesive pipeline. The lack of a central, updated source leads to inefficiencies and the risk of using outdated tools, which can affect research quality and reproducibility. Consequently, there is a critical need for an automated, accurate tool to identify bioinformatics databases and software mentions directly from scientific texts, streamlining workflow development and enhancing research productivity.
The bioNerDS system, a prior effort to address this challenge, uses a rule-based named entity recognition (NER) approach, achieving an F1 score of 63% on an evaluation set of 25 articles from BMC Bioinformatics and PLoS Computational Biology. By integrating the same set of features such as context patterns, word characteristics and dictionary matches into a machine learning model, we developed an approach using an XGBoost classifier. This model, carefully tuned to address the extreme class imbalance inherent in NER tasks through synthetic oversampling and refined via systematic hyperparameter optimization to balance precision and recall, excels at capturing complex linguistic patterns and non-linear relationships, ensuring robust generalization. It achieves an F1 score of 82% on the same evaluation set, significantly surpassing the baseline. By combining rule-based precision with machine learning adaptability, this approach enhances accuracy, reduces ambiguities, and provides a robust tool for large-scale bioinformatics resource identification, facilitating efficient workflow construction. Furthermore, this methodology holds potential for extension to other technological domains, enabling similar resource identification in fields like data science, artificial intelligence, or computational engineering.
Vishnu Chowdary Madhavarapu
Automated Weather Classification Using Transfer LearningWhen & Where:
Nichols Hall, Room 246 (Executive Conference Room)
Committee Members:
David Johnson, ChairPrasad Kulkarni
Dongjie Wang
Abstract
This project presents an automated weather classification system utilizing transfer learning with pre-trained convolutional neural networks (CNNs) such as VGG19, InceptionV3, and ResNet50. Designed to classify weather conditions—sunny, cloudy, rainy, and sunrise—from images, the system addresses the challenge of limited labeled data by applying data augmentation techniques like zoom, shear, and flip, expanding the dataset images. By fine-tuning the final layers of pre-trained models, the solution achieves high accuracy while significantly reducing training time. VGG19 was selected as the baseline model for its simplicity, strong feature extraction capabilities, and widespread applicability in transfer learning scenarios. The system was trained using the Adam optimizer and evaluated on key performance metrics including accuracy, precision, recall, and F1 score. To enhance user accessibility, a Flask-based web interface was developed, allowing real-time image uploads and instant weather classification. The results demonstrate that transfer learning, combined with robust data preprocessing and fine-tuning, can produce a lightweight and accurate weather classification tool. This project contributes toward scalable, real-time weather recognition systems that can integrate into IoT applications, smart agriculture, and environmental monitoring.
RokunuzJahan Rudro
Using Machine Learning to Classify Driver Behavior from Psychological Features: An Exploratory StudyWhen & Where:
Eaton Hall, Room 1A
Committee Members:
Sumaiya Shomaji, ChairDavid Johnson
Zijun Yao
Alexandra Kondyli
Abstract
Driver inattention and human error are the primary causes of traffic crashes. However, little is known about the relationship between driver aggressiveness and safety. Although several studies that group drivers into different classes based on their driving performance have been conducted, little has been done to explore how behavioral traits are linked to driver behavior. The study aims to link different driver profiles, assessed through psychological evaluations, with their likelihood of engaging in risky driving behaviors, as measured in a driving simulation experiment. By incorporating psychological factors into machine learning algorithms, our models were able to successfully relate self-reported decision-making and personality characteristics with actual driving actions. Our results hold promise toward refining existing models of driver behavior by understanding the psychological and behavioral characteristics that influence the risk of crashes.
Md Mashfiq Rizvee
Energy Optimization in Multitask Neural Networks through Layer SharingWhen & Where:
Eaton Hall, Room 2001B
Committee Members:
Sumaiya Shomaji, ChairTamzidul Hoque
Han Wang
Abstract
Artificial Intelligence (AI) is being widely used in diverse domains such as industrial automation, traffic control, precision agriculture, and smart cities for major heavy lifting in terms of data analysis and decision making. However, the AI life- cycle is a major source of greenhouse gas (GHG) emission leading to devastating environmental impact. This is due to expensive neural architecture searches, training of countless number of models per day across the world, in-field AI processing of data in billions of edge devices, and advanced security measures across the AI life cycle. Modern applications often involve multitasking, which involves performing a variety of analyzes on the same dataset. These tasks are usually executed on resource-limited edge devices, necessitating AI models that exhibit efficiency across various measures such as power consumption, frame rate, and model size. To address these challenges, we introduce a novel neural network architecture model that incorporates a layer sharing principle to optimize the power usage. We propose a novel neural architecture, Layer Shared Neural Networks that merges multiple similar AI/NN tasks together (with shared layers) towards creating a single AI/NN model with reduced energy requirements and carbon footprint. The experimental findings reveal competitive accuracy and reduced power consumption. The layer shared model significantly reduces power consumption by 50% during training and 59.10% during inference causing as much as an 84.64% and 87.10% decrease in CO2 emissions respectively.
Fairuz Shadmani Shishir
Parameter-Efficient Computational Drug Discovery using Deep LearningWhen & Where:
Eaton Hall, Room 2001B
Committee Members:
Sumaiya Shomaji, ChairTamzidul Hoque
Hongyang Sun
Abstract
The accurate prediction of small molecule binding affinity and toxicity remains a central challenge in drug discovery, with significant implications for reducing development costs, improving candidate prioritization, and enhancing safety profiles. Traditional computational approaches, such as molecular docking and quantitative structure-activity relationship (QSAR) models, often rely on handcrafted features and require extensive domain knowledge, which can limit scalability and generalization to novel chemical scaffolds. Recent advances in language models (LMs), particularly those adapted to chemical representations such as SMILES (Simplified Molecular Input Line Entry System), have opened new ways for learning data-driven molecular representations that capture complex structural and functional properties. However, achieving both high binding affinity and low toxicity through a resource-efficient computational pipeline is inherently difficult due to the multi-objective nature of the task. This study presents a novel dual-paradigm approach to critical challenges in drug discovery: predicting small molecules with high binding affinity and low cardiotoxicity profiles. For binding affinity prediction, we implement a specialized graph neural network (GNN) architecture that operates directly on molecular structures represented as graphs, where atoms serve as nodes and bonds as edges. This topology-aware approach enables the model to capture complex spatial arrangements and electronic interactions critical for protein-ligand binding. For toxicity prediction, we leverage chemical language models (CLMs) fine-tuned with Low-Rank Adaptation (LoRA), allowing efficient adaptation of large pre-trained models to specialized toxicological endpoints while maintaining the generalized chemical knowledge embedded in the base model. Our hybrid methodology demonstrates significant improvements over existing computational approaches, with the GNN component achieving an average area under the ROC curve (AUROC) of 0.92 on three protein targets and the LoRA-adapted CLM reaching (AUROC) of 0.90 with 60% reduction in parameter usage in predicting cardiotoxicity. This work establishes a powerful computational framework that accelerates drug discovery by enabling both higher binding affinity and low toxicity compounds with optimized efficacy and safety profiles.
Soma Pal
Truths about compiler optimization for state-of-the-art (SOTA) C/C++ compilersWhen & Where:
Eaton Hall, Room 2001B
Committee Members:
Prasad Kulkarni, ChairEsam El-Araby
Drew Davidson
Tamzidul Hoque
Jiang Yunfeng
Abstract
Compiler optimizations are critical for performance and have been extensively studied, especially for C/C++ language compilers. Our overall goal in this thesis is to investigate and compare the properties and behavior of optimization passes across multiple contemporary, state-of-the-art (SOTA) C/C++ compilers to understand if they adopt similar optimization implementation and orchestration strategies. Given the maturity of pre-existing knowledge in the field, it seems conceivable that different compiler teams will adopt consistent optimization passes, pipeline and application techniques. However, our preliminary results indicate that such expectation may be misguided. If so, then we will attempt to understand the differences, and study and quantify their impact on the performance of generated code.
In our first work, we study and compare the behavior of profile-guided optimizations (PGO) in two popular SOTA C/C++ compilers, GCC and Clang. This study reveals many interesting, and several counter-intuitive, properties about PGOs in C/C++ compilers. The behavior and benefits of PGOs also vary significantly across our selected compilers. We present our observations, along with plans to further explore these inconsistencies in this report. Likewise, we have also measured noticeable differences in the performance delivered by optimizations across our compilers. We propose to explore and understand these differences in this work. We present further details regarding our proposed directions and planned experiments in this report. We hope that this work will show and suggest opportunities for compilers to learn from each other and motivate researchers to find mechanisms to combine the benefits of multiple compilers to deliver higher overall program performance.
Nyamtulla Shaik
AI Vision to Care: A QuadView of Deep Learning for Detecting Harmful Stimming in AutismWhen & Where:
Eaton Hall, Room 2001B
Committee Members:
Sumaiya Shomaji, ChairBo Luo
Dongjie Wang
Abstract
Stimming refers to repetitive actions or behaviors used to regulate sensory input or express feelings. Children with developmental disorders like autism (ASD) frequently perform stimming. This includes arm flapping, head banging, finger flicking, spinning, etc. This is exhibited by 80-90% of children with Autism, which is seen in 1 among 36 children in the US. Head banging is one of these self-stimulatory habits that can be harmful. If these behaviors are automatically identified and notified using live video monitoring, parents and other caregivers can better watch over and assist children with ASD.
Classifying these actions is important to recognize harmful stimming, so this study focuses on developing a deep learning-based approach for stimming action recognition. We implemented and evaluated four models leveraging three deep learning architectures based on Convolutional Neural Networks (CNNs), Autoencoders, and Vision Transformers. For the first time in this area, we use skeletal joints extracted from video sequences. Previous works relied solely on raw RGB videos, vulnerable to lighting and environmental changes. This research explores Deep Learning based skeletal action recognition and data processing techniques for a small unstructured dataset that consists of 89 home recorded videos collected from publicly available sources like YouTube. Our robust data cleaning and pre-processing techniques helped the integration of skeletal data in stimming action recognition, which performed better than state-of-the-art with a classification accuracy of up to 87%
In addition to using traditional deep learning models like CNNs for action recognition, this study is among the first to apply data-hungry models like Vision Transformers (ViTs) and Autoencoders for stimming action recognition on the dataset. The results prove that using skeletal data reduces the processing time and significantly improves action recognition, promising a real-time approach for video monitoring applications. This research advances the development of automated systems that can assist caregivers in more efficiently tracking stimming activities.
Alexander Rodolfo Lara
Creating a Faradaic Efficiency Graph Dataset Using Machine LearningWhen & Where:
Eaton Hall, Room 2001B
Committee Members:
Zijun Yao, ChairSumaiya Shomaji
Kevin Leonard
Abstract
Just as the internet-of-things leverages machine learning over a vast amount of data produced by an innumerable number of sensors, the Internet of Catalysis program uses similar strategies with catalysis research. One application of the Internet of Catalysis strategy is treating research papers as datapoints, rich with text, figures, and tables. Prior research within the program focused on machine learning models applied strictly over text.
This project is the first step of the program in creating a machine learning model from the images of catalysis research papers. Specifically, this project creates a dataset of faradaic efficiency graphs using transfer learning from pretrained models. The project utilizes FasterRCNN_ResNet50_FPN, LayoutLMv3SequenceClassification, and computer vision techniques to recognize figures, extract all graphs, then classify the faradaic efficiency graphs.
Downstream of this project, researchers will create a graph reading model to integrate with large language models. This could potentially lead to a multimodal model capable of fully learning from images, tables, and texts of catalysis research papers. Such a model could then guide experimentation on reaction conditions, catalysts, and production.
Amin Shojaei
Scalable and Cooperative Multi-Agent Reinforcement Learning for Networked Cyber-Physical Systems: Applications in Smart GridsWhen & Where:
Nichols Hall, Room 246 (Executive Conference Room)
Committee Members:
Morteza Hashemi, ChairAlex Bardas
Prasad Kulkarni
Taejoon Kim
Shawn Keshmiri
Abstract
Significant advances in information and networking technologies have transformed Cyber-Physical Systems (CPS) into networked cyber-physical systems (NCPS). A noteworthy example of such systems is smart grid networks, which include distributed energy resources (DERs), renewable generation, and the widespread adoption of Electric Vehicles (EVs). Such complex NCPS require intelligent and autonomous control solutions. For example, the increasing number of EVs introduces significant sources of demand and user behavior uncertainty that can jeopardize grid stability during peak hours. Traditional model-based demand-supply controls fail to accurately model and capture the complex nature of smart grid systems in the presence of different uncertainties and as the system size grows. To address these challenges, data-driven approaches have emerged as an effective solution for informed decision-making, predictive modeling, and adaptive control to enhance the resiliency of NCPS in uncertain environments.
As a powerful data-driven approach, Multi-Agent Reinforcement Learning (MARL) enables agents to learn and adapt in dynamic and uncertain environments. However, MARL techniques introduce complexities related to communication, coordination, and synchronization among agents. In this PhD research, we investigate autonomous control for smart grid decision networks using MARL. First, we examine the issue of imperfect state information, which frequently arises due to the inherent uncertainties and limitations in observing the system state.
Second, we focus on the cooperative behavior of agents in distributed MARL frameworks, particularly under the central training with decentralized execution (CTDE) paradigm. We provide theoretical results and variance analysis for stochastic and deterministic cooperative MARL algorithms, including Multi-Agent Deep Deterministic Policy Gradient (MADDPG), Multi-Agent Proximal Policy Optimization (MAPPO), and Dueling MAPPO. These analyses highlight how coordinated learning can improve system-wide decision-making in uncertain and dynamic environments like EV networks.
Third, we address the scalability challenge in large-scale NCPS by introducing a hierarchical MARL framework based on a cluster-based architecture. This framework organizes agents into coordinated subgroups, improving scalability while preserving local coordination. We conduct a detailed variance analysis of this approach to demonstrate its effectiveness in reducing communication overhead and learning complexity. This analysis establishes a theoretical foundation for scalable and efficient control in large-scale smart grid applications.
Asrith Gudivada
Custom CNN for Object State Classification in Robotic CookingWhen & Where:
Nichols Hall, Room 246 (Executive Conference Room)
Committee Members:
David Johnson, ChairPrasad Kulkarni
Dongjie Wang
Abstract
This project presents the development of a custom Convolutional Neural Network (CNN) designed to classify object states—such as sliced, diced, or peeled—in robotic cooking environments. Recognizing fine-grained object states is critical for context-aware manipulation yet remains a challenging task due to the visual similarity between states and the limited availability of cooking-specific datasets. To address these challenges, we built a lightweight, non-pretrained CNN trained on a curated dataset of 11 object states. Starting with a baseline architecture, we progressively enhanced the model using data augmentation, optimized dropout, batch normalization, Inception modules, and residual connections. These improvements led to a performance increase from ~45% to ~52% test accuracy. The final model demonstrates improved generalization and training stability, showcasing the effectiveness of combining classical and advanced deep learning techniques. This work contributes toward real-time state recognition for autonomous robotic cooking systems, with implications for assistive technologies in domestic and elder care settings.
Tanvir Hossain
Gamified Learning of Computing Hardware Fundamentals Using FPGA-Based PlatformWhen & Where:
Nichols Hall, Room 250 (Gemini Room)
Committee Members:
Tamzidul Hoque, ChairEsam El-Araby
Sumaiya Shomaji
Abstract
The growing dependence on electronic systems in consumer and mission critical domains requires engineers who understand the inner workings of digital hardware. Yet many students bypass hardware electives, viewing them as abstract, mathematics heavy, and less attractive than software courses. Escalating workforce shortages in the semiconductor industry and the recent global chip‑supply crisis highlight the urgent need for graduates who can bridge hardware knowledge gaps across engineering sectors. In this thesis, I have developed FPGA‑based games, embedded in inclusive curricular modules, which can make hardware concepts accessible while fostering interest, self‑efficacy, and positive outcome expectations in hardware engineering. A design‑based research methodology guided three implementation cycles: a pilot with seven diverse high‑school learners, a multiweek residential summer camp with high‑school students, and a fifteen‑week multidisciplinary elective enrolling early undergraduate engineering students. The learning experiences targeted binary arithmetic, combinational and sequential logic, state‑machine design, and hardware‑software co‑design. Learners also moved through the full digital‑design flow, HDL coding, functional simulation, synthesis, place‑and‑route, and on‑board verification. In addition, learners explored timing analysis, register‑transfer‑level abstractions, and simple processor datapaths to connect low‑level circuits with system‑level behavior. Mixed‑method evidence was gathered through pre‑ and post‑content quizzes, validated surveys of self‑efficacy and outcome expectations, focus groups, classroom observations, and gameplay analytics. Paired‑sample statistics showed reliable gains in hardware‑concept mastery, self‑efficacy, and outcome expectations. This work contributes a replicable framework for translating foundational hardware topics into modular, game‑based learning activities, empirical evidence of their effectiveness across secondary and early‑college contexts, and design principles for educators who seek to integrate equitable, hands‑on hardware experiences into existing curricula.
Hara Madhav Talasila
Radiometric Calibration of Radar Depth Sounder Data ProductsWhen & Where:
Nichols Hall, Room 317 (Richard K. Moore Conference Room)
Committee Members:
Carl Leuschen, ChairPatrick McCormick
James Stiles
Jilu Li
Leigh Stearns
Abstract
Although the Center for Remote Sensing of Ice Sheets (CReSIS) performs several radar calibration steps to produce Operation IceBridge (OIB) radar depth sounder data products, these datasets are not radiometrically calibrated and the swath array processing uses ideal (rather than measured [calibrated]) steering vectors. Any errors in the steering vectors, which describe the response of the radar as a function of arrival angle, will lead to errors in positioning and backscatter that subsequently affect estimates of basal conditions, ice thickness, and radar attenuation. Scientific applications that estimate physical characteristics of surface and subsurface targets from the backscatter are limited with the current data because it is not absolutely calibrated. Moreover, changes in instrument hardware and processing methods for OIB over the last decade affect the quality of inter-seasonal comparisons. Recent methods which interpret basal conditions and calculate radar attenuation using CReSIS OIB 2D radar depth sounder echograms are forced to use relative scattering power, rather than absolute methods.
As an active target calibration is not possible for past field seasons, a method that uses natural targets will be developed. Unsaturated natural target returns from smooth sea-ice leads or lakes are imaged in many datasets and have known scattering responses. The proposed method forms a system of linear equations with the recorded scattering signatures from these known targets, scattering signatures from crossing flight paths, and the radiometric correction terms. A least squares solution to optimize the radiometric correction terms is calculated, which minimizes the error function representing the mismatch in expected and measured scattering. The new correction terms will be used to correct the remaining mission data. The radar depth sounder data from all OIB campaigns can be reprocessed to produce absolutely calibrated echograms for the Arctic and Antarctic. A software simulator will be developed to study calibration errors and verify the calibration software. The software for processing natural targets and crossovers will be made available in CReSIS’s open-source polar radar software toolbox. The OIB data will be reprocessed with new calibration terms, providing to the data user community a complete set of radiometrically calibrated radar echograms for the CReSIS OIB radar depth sounder for the first time.
Past Defense Notices
SUSHIL BHARATI
Vision Based Adaptive Obstacle Detection, Robust Tracking and 3D Reconstruction for Autonomous Unmanned Aerial VehiclesWhen & Where:
246 Nichols Hall
Committee Members:
Richard Wang, ChairBo Luo
Suzanne Shontz
Abstract
Vision-based autonomous navigation of UAVs in real-time is a very challenging problem, which requires obstacle detection, tracking, and depth estimation. Although the problems of obstacle detection and tracking along with 3D reconstruction have been extensively studied in computer vision field, it is still a big challenge for real applications like UAV navigation. The thesis intends to address these issues in terms of robustness and efficiency. First, a vision-based fast and robust obstacle detection and tracking approach is proposed by integrating a salient object detection strategy within a kernelized correlation filter (KCF) framework. To increase its performance, an adaptive obstacle detection technique is proposed to refine the location and boundary of the object when the confidence value of the tracker drops below a predefined threshold. In addition, a reliable post-processing technique is implemented for an accurate obstacle localization. Second, we propose an efficient approach to detect the outliers present in noisy image pairs for the robust fundamental matrix estimation, which is a fundamental step for depth estimation in obstacle avoidance. Given a noisy stereo image pair obtained from the mounted stereo cameras and initial point correspondences between them, we propose to utilize reprojection residual error and 3-sigma principle together with robust statistic based Qn estimator (RES-Q) to efficiently detect the outliers and accurately estimate the fundamental matrix. The proposed approaches have been extensively evaluated through quantitative and qualitative evaluations on a number of challenging datasets. The experiments demonstrate that the proposed detection and tracking technique significantly outperforms the state-of-the-art methods in terms of tracking speed and accuracy, and the proposed RES-Q algorithm is found to be more robust than other classical outlier detection algorithms under both symmetric and asymmetric random noise assumptions.
MOHSEN ALEENEJAD
New Modulation Methods and Control Strategies for Power Electronics InvertersWhen & Where:
1 Eaton Hall
Committee Members:
Reza Ahmadi, ChairGlenn Prescott
Alessandro Salandrino
Jim Stiles
Huazhen Fang*
Abstract
The DC to AC power Converters (so-called Inverters) are widely used in industrial applications. The multilevel inverters are becoming increasingly popular in industrial apparatus aimed at medium to high power conversion applications. In comparison to the conventional inverters, they feature superior characteristics such as lower total harmonic distortion (THD), higher efficiency, and lower switching voltage stress. Nevertheless, the superior characteristics come at the price of a more complex topology with an increased number of power electronic switches. The increased number of power electronics switches results in more complicated control strategies for the inverter. Moreover, as the number of power electronic switches increases, the chances of fault occurrence of the switches increases, and thus the inverter’s reliability decreases. Due to the extreme monetary ramifications of the interruption of operation in commercial and industrial applications, high reliability for power inverters utilized in these sectors is critical. As a result, developing simple control strategies for normal and fault-tolerant operation of multilevel inverters has always been an interesting topic for researchers in related areas. The purpose of this dissertation is to develop new control and fault-tolerant strategies for the multilevel power inverter. For the normal operation of the inverter, a new high switching frequency technique is developed. The proposed method extends the utilization of the dc link voltage while minimizing the dv/dt of the switches. In the event of a fault, the line voltages of the faulty inverters are unbalanced and cannot be applied to the three phase loads. For the faulty condition of the inverter, three novel fault-tolerant techniques are developed. The proposed fault-tolerant strategies generate balanced line voltages without bypassing any healthy and operative inverter element, makes better use of the inverter capacity and generates higher output voltage. These strategies exploit the advantages of the Selective Harmonic Elimination (SHE) and Space Vector Modulation (SVM) methods in conjunction with a slightly modified Fundamental Phase Shift Compensation (FPSC) technique to generate balanced voltages and manipulate voltage harmonics at the same time. The proposed strategies are applicable to several classes of multilevel inverters with three or more voltage levels.
XIAOLI LI
Constructivism LearningWhen & Where:
246 Nichols Hall
Committee Members:
Luke Huan, ChairVictor Frost
Bo Luo
Richard Wang
Alfred Ho*
Abstract
Aiming to achieve the learning capabilities possessed by intelligent beings, especially human, researchers in machine learning field have the long-standing tradition of borrowing ideas from human learning, such as reinforcement learning, active learning, and curriculum learning. Motivated by a philosophical theory called "constructivism", in this work, we propose a new machine learning paradigm, constructivism learning. The constructivism theory has had wide-ranging impact on various human learning theories about how human acquire knowledge. To adapt this human learning theory to the context of machine learning, we first studied how to improve leaning performance by exploring inductive bias or prior knowledge from multiple learning tasks with multiple data sources, that is multi-task multi-view learning, both in offline and lifelong setting. Then we formalized a Bayesian nonparametric approach using sequential Dirichlet Process Mixture Models to support constructivism learning. To further exploit constructivism learning, we also developed a constructivism deep learning method utilizing Uniform Process Mixture Models.
MOHANAD AL-IBADI
Array Processing Techniques for Ice-Sheet Bottom TrackingWhen & Where:
317 Nichols Hall
Committee Members:
Shannon Blunt, ChairJohn Paden
Eric Perrins
Jim Stiles
Huazhen Fang*
Abstract
In airborne multichannel radar sounder signal processing, the collected data are most naturally represented in a cylindrical coordinate system: along-track, range, and elevation angle. The data are generally processed in each of these dimensions sequentially to focus or resolve the data in the corresponding dimension such that a 3D image of the scene can be formulated. Pulse-compression is used to process the data along the range dimension, synthetic aperture radar (SAR) processing is used to process the data in the along-track dimension, and array-processing techniques are used for the elevation angle dimension. After the first two steps, the 3D scene is resolved into toroids with constant along-track and constant range that are centered on the flight path. The targets lying in a particular toroid need to be resolved by estimating their respective elevation angles.
In the proposed work, we focus on the array processing step, where several direction of arrival (DoA) estimation methods will be used to resolve the targets in the elevation-angle dimension, such as MUltiple Signal Classification (MUSIC) and maximum-likelihood estimation (MLE). A tracker is then used on the output of the DoA estimation to track the ice-bottom interface. We propose to use the tree re-weighted message passing algorithm or Kalman filtering, based on the array-processing technique, to track the ice-bottom. The outcome of this is a digital elevation model (DEM) of the ice-bottom. While most published work assumes a narrowband model for the array, we will use a wideband model and focus on issues related to wideband arrays. Along these lines, we propose a theoretical study to evaluate the performance of the radar products based on the array characteristics using different array-processing techniques, such as wideband MLE and focusing-matrices methods. In addition, we will investigate tracking targets using a sparse array composed of three sub-arrays, each separated by a large multiwavelength baseline. Specifically, we propose to develop and investigate the performance of a Kalman tracking solution to this wideband sparse array problem when applied to data collected by the CReSIS radar sounder.
QIAOZHI WANG
Towards the Understanding of Private Content -- Content-based Privacy Assessment and Protection in Social NetworksWhen & Where:
2001B Eaton Hall
Committee Members:
Bo Luo, ChairFengjun Li
Richard Wang
Heechul Yun
Prajna Dhar*
Abstract
In the 2016 presidential election, social networks showed their great power as a “modern form of communication”. With the increasing popularity of social networks, privacy concerns arise. For example, it has been shown that microblogs are revealed to audiences that are significantly larger than users' perceptions. Moreover, when users are emotional, they may post messages with sensitive content and later regret doing so. As a result, users become very vulnerable – private or sensitive information may be accidentally disclosed, even in tweets about trivial daily activities.
Unfortunately, existing research projects on data privacy, such as the k-anonymity and differential privacy mechanisms, mostly focus on protecting individual’s identity from being discovered in large data sets. We argue that the key component of privacy protection in social networks is protecting sensitive content, i.e. privacy as having the ability to control dissemination of information. The overall objectives of the proposed research are: to understand the sensitive content of social network posts, to facilitate content-based protection of private information, and to identify different types of sensitive information. In particular, we propose a user-centered, quantitative measure of privacy based on textual content, and a customized privacy protection mechanism for social networks.
We consider private tweet identification and classification as dual-problems. We propose to develop an algorithm to identify all types of private messages, and, more importantly, automatically score the sensitiveness of private message. We first collect the opinions from a diverse group of users w.r.t. sensitiveness of private information through Amazon Mechanical Turk, and analyze the discrepancies between users' privacy expectations and actual information disclosure. We then develop a computational method to generate the context-free privacy score, which is the “consensus” privacy score for average users. Meanwhile, classification of private tweets is necessary for customized privacy protection. We have made the first attempt to understand different types of private information, and to automatically classify sensitive tweets into 13 pre-defined topic categories. In proposed research, we will further include personal attitudes, topic preferences, and social context into the scoring mechanism, to generate a personalized, context-aware privacy score, which will be utilized in a comprehensive privacy protection mechanism.
STEVE HAENCHEN
A Model to Identify Insider Threats Using Growing Hierarchical Self-Organizing Map of Electronic Media IndicatorsWhen & Where:
1 Eaton Hall
Committee Members:
Hossein Saiedian, ChairArvin Agah
Prasad Kulkarni
Bo Luo
Reza Barati
Abstract
Fraud from insiders costs an estimated $3.7 trillion annually. Current fraud prevention and detection methods that include analyzing network logs, computer events, emails, and behavioral characteristics have not been successful in reducing the losses. The proposed Occupational Fraud Prevention and Detection Model uses existing data from the field of digital forensics along with text clustering algorithms, machine learning, and a growing hierarchical self-organizing map model to predict insider threats based on computer usage behavioral characteristics.
The proposed research leverages research results from information security, software engineering, data science and information retrieval, context searching, search patterns, and machine learning to build and employ a database server and workstations to support 50+ terabytes of data representing entire hard drives from work computers. Forensic software FTK and EnCase are used to generate disk images and test extraction results. Primary research tools are built using modern programming languages. The research data is derived from disk images obtained from actual investigations when fraud was asserted and other disk images when fraud was not asserted.
The research methodology includes building a data extraction tool that is a disk level reader to store the disk, partition, and operating system data in a relational database. An analysis tool is also created to convert the data into information representing usage patterns including summarization, normalization, and redundancy removal. We build a normalizing tool that uses machine learning to adjust the baselines for company, department, and job deviations. A prediction component is developed to derive insider threat scores reflecting the anomalies from the adjusted baseline. The resulting product will allow identification of the computer users most likely to commit fraud so investigators can focus their limited resources on the suspects.
Our primarily plan to evaluate and validate our research results is via empirical study, statistical evaluation and benchmarking with tests of precision and recall from a second set of disk images.
JAMIE ROBINSON
Code Cache Management in Managed Language VMs to Reduce Memory Consumption for Embedded SystemsWhen & Where:
129 Nichols Hall
Committee Members:
Prasad Kulkarni, ChairBo Luo
Heechul Yun
Abstract
The compiled native code generated by a just-in-time (JIT) compiler in managed language virtual machines (VM) is placed in a region of memory called the code cache. Code cache management (CCM) in a VM is responsible to find and evict methods from the code cache to maintain execution correctness and manage program performance for a given code cache size or memory budget. Effective CCM can also boost program speed by enabling more aggressive JIT compilation, powerful optimizations, and improved hardware instruction cache and I-TLB performance.
Though important, CCM is an overlooked component in VMs. We find that the default CCM policies in Oracle’s production-grade HotSpot VM perform poorlyeven at modest memory pressure. We develop a detailed simulation-based framework to model and evaluate the potential efficiency of many different CCM policies in a controlled and realistic, but VM-independent environment. We make the encouraging discovery that effective CCM policies can sustain high program performance even for very small cache sizes.
Our simulation study provides the rationale and motivation to improve CCM strategies in existing VMs. We implement and study the properties of several CCM policies in HotSpot. We find that in spite of working within the bounds of the HotSpot VM’s current CCM sub-system, our best CCM policy implementation in HotSpot improves program performance over the default CCM algorithm by 39%, 41%, 55%, and 50% with code cache sizes that are 90%, 75%, 50%, and 25% of the desired cache size, on average.
AIME DE BERNER
Application of Machine Learning Techniques to the Diagnosis of Vision DisordersWhen & Where:
2001B Eaton Hall
Committee Members:
Arvin Agah, ChairNicole Beckage
Jerzy Grzymala-Busse
Abstract
In the age of data collection and as we search for knowledge, over time numerous techniques have been developed and used to capture, manipulate, and to process data to acquire the hidden correlations, relations, patterns, and mappings that one may not be able to see. Computers as machines with the help of improved algorithms have proven to provide Artificial Intelligence (AI) by applying models to predict outcomes within an acceptable margin of error. Through performance metrics applied using Data Mining and Machine Learning models to predict human vision disorders, we are able to see promising models. AI techniques used in this work include an improved version of C.45 called C.48, Neuro-Networks, K-Nearest-Neighbor, Random Forest, Support Vector Machines, AdaBoost, among many. The best predictive models were determined that could be applied to the diagnosis of vision disorders, focusing on Strabismus, the need for patient referral to a specialist.
HAO XUE
Understanding Information Credibility in Social NetworksWhen & Where:
246 Nichols Hall
Committee Members:
Fengjun Li, ChairLuke Huan
Prasad Kulkarni
Bo Luo
Hyunjin Seo
Abstract
With the advancement of Internet, increasing portions of people's social and communicative activities now take place in the digital world. The growth and popularity of online social networks have tremendously facilitate the online interaction and information exchange. More people now rely online information for news, opinions, and social networking. As the representative of online social-collaborative platforms, online review systems has enabled people to share information effectively and efficiently. A large volume of user generated content is produced daily, which allows people to make reasonable judgments about the quality of service or product of an unknown provider. However, the freedom and ease of of publishing information online has made these systems no longer the sources of reliable information. Not only does biased and misleading information exist, financial incentives drive individual and professional spammers to insert deceptive reviews to manipulate review rating and content. What's worse, advanced Artificial Intelligence has made it possible to generate realistic-looking reviews automatically. In this proposal, we present our work of measuring the credibility of information in online review systems. We first propose to utilize the social relationships and rating deviations to assist the computation of trustworthiness of users. Secondly, we propose a content-based trust propagation framework by extracting the opinions expressed in review content. The opinion extraction approach we used was a supervised-learning based methods, which has flexibility limitations. Thus, we propose a enhanced framework that not only automates the opinion mining process, but also integrates social relationships with review content. Finally, we propose our study of the credibility of machine-generated reviews.
MOHAMMADREZA HAJIARBABI
A Face Detection and Recognition System for Color Images using Neural Networks with Boosting and Deep LearningWhen & Where:
2001B Eaton Hall
Committee Members:
Arvin Agah, ChairPrasad Kulkarni
Bo Luo
Richard Wang
Sara Wilson*
Abstract
A face detection and recognition system is a biometric identification mechanism which compared to other methods is shown to be more important both theoretically and practically. In principle, the biometric identification methods use a wide range of techniques such as machine learning, computer vision, image processing, pattern recognition and neural networks. A face recognition system consists of two main components, face detection and recognition.
In this dissertation a face detection and recognition system using color images with multiple faces is designed, implemented, and evaluated. In color images, the information of skin color is used in order to distinguish between the skin pixels and non-skin pixels, dividing the image into several components. Neural networks and deep learning methods has been used in order to detect skin pixels in the image. In order to improve system performance, bootstrapping and parallel neural networks with voting have been used. Deep learning has been used as another method for skin detection and compared to other methods. Experiments have shown that in the case of skin detection, deep learning and neural networks methods produce better results in terms of precision and recall compared to the other methods in this field.
The step after skin detection is to decide which of these components belong to human face. A template based method has been modified in order to detect the faces. The designed algorithm also succeeds if there are more than one face in the component. A rule based method has been designed in order to detect the eyes and lips in the detected components. After detecting the location of eyes and lips in the component, the face can be detected.
After face detection, the faces which were detected in the previous step are to be recognized. Appearance based methods used in this work are one of the most important methods in face recognition due to the robustness of the algorithms to head rotation in the images, noise, low quality images, and other challenges. Different appearance based methods have been designed, implemented and tested. Canonical correlation analysis has been used in order to increase the recognition rate.