Abstract:
This thesis presents a deep learning approach for classifying all seven crystal systems from scanning transmission electron microscopy (STEM) images, addressing challenges caused by variable material orientations. Using a large and diverse dataset of 6,670 simulated HAADF-STEM images with randomly oriented structures, the model incorporates a novel Fractal Average Pooling (FAP) mechanism that captures structural complexity by computing the fractal dimension of feature maps. The proposed method achieves high accuracy across multiple datasets, demonstrating that fractal-based features significantly enhance classification reliability and open new avenues in materials image analysis.

Research Adviser: Aggelos Katsaggelos

STEM Image Classification Using Fractal-Based Neural Network

Abstract:
This study presents a deep learning framework that predicts prostate cancer molecular subtypes (PAM50 and PSC) directly from H&E-stained whole-slide images using a patch-wise MIL approach with a MobileNetV2 backbone and a novel three-branch attention mechanism. Trained on 641 WSIs, the model showed promising accuracy, revealed subtype-specific tissue patterns, and demonstrated potential as a cost-effective, non-invasive alternative to genomic assays for guiding personalized therapy.

Research Adviser: Lee Cooper, Department of Electrical & Computer Engineering

Deep Learning-Based Framework for Predicting Prostate Cancer Molecular Subtypes

Abstract:
This thesis explores the application of diffusion policy to the shepherding problem, extending from a single-dog model to a two-dog system. Experimental results show how different flock distributions and herd sizes affect success rate and completion time.

Research Adviser: Randy Freeman, Department of Electrical and Computer Engineering

Diffusion Policy for Shepherding: Extending from One To Two Shepherds

Abstract:
This thesis introduces a multimodal wearable sensing system for accurate and privacy-conscious detection of eating behavior, essential for dietary monitoring and health interventions. It uses an upward-facing thermal infrared camera to detect hand-to-mouth gestures and activates a forward-facing RGB camera only when necessary, preserving user privacy. The system, built on compact hardware with an ESP32-S3 microcontroller and an efficient on-device machine learning model, achieves 98% gesture recognition accuracy while remaining energy-efficient and scalable.

Research Adviser: Nabil Alshurafa

A Wearable Platform for Privacy-Aware Eating Detection via Multimodal Sensing

Abstract:
This thesis compares multi-agent strategies for autonomous exploration of Martian caves, focusing on a novel Deep Reinforcement Learning (RL) method versus a state-of-the-art algorithmic approach (UDZRSR). The RL method, designed with custom reward functions and phased strategies, outperforms the algorithm in larger cave environments due to its adaptability, efficient energy use, and dynamic role allocation. The study concludes that combining algorithmic precision with machine learning flexibility offers the most effective solution for planetary exploration and related Earth-based applications.

Research Adviser: Qi Zhu

Deep Reinforcement Learning Versus Algorithmic Approaches in Multi-Agent Mars Cave Exploration

Abstract:
This paper explores the deployment of Federated Learning (FL) on a 5G-NR Standalone (SA) testbed using resource-constrained IoT devices (Raspberry Pis) and a central server with Software Defined Radio and O-RAN software. The authors implement FL using the Flower framework and develop a custom instrumentation tool to analyze performance across 5G, WiFi, and Ethernet networks. Experimental results show that 5G uplink latency significantly impacts FL convergence, being 33.3× slower than Ethernet and 17.8× slower than WiFi, and it worsens the straggler effect during training.

Research Adviser: Igor Kadota

Federated Learning over 5G, WiFi, and Ethernet: Measurements and Evaluation

Abstract:
This thesis presents a hardware-efficient approach for deploying quantized neural networks on resource-constrained FPGAs, specifically the Zynq XC7Z020, by using piecewise-linear approximations to replace sigmoid activations. This technique reduces memory usage and enables shared compute resources between dense and activation layers, optimizing performance in memory-limited scenarios. Despite some trade-offs in scheduling and timing, the implementation achieves timing closure and maintains competitive accuracy, making it suitable for real-time edge applications.

Research Adviser: Seda Ogrenci

A Linearization-Based Approach to Cross-Layer Resource Sharing in Neural Network FPGA Deployment

Abstract:
This dissertation details the design and validation of a compact, wireless near-infrared spectroscopy (NIRS) device for continuous monitoring of tissue oxygen saturation (StO₂) in pediatric patients, addressing the limitations of traditional bulky, wired systems. The device integrates low-power components, dual-wavelength LED illumination, and optimized hardware to achieve high accuracy, extended battery life, and clinical-grade performance. Bench-top and clinical trials confirmed its reliability and feasibility, positioning it as a promising solution for non-invasive, long-term pediatric monitoring in both hospital and home settings.

Research Adviser: John Rogers

A compact, wireless Low-power functional Near Infrared Spectroscopy Device for continuous StO2 monitoring in Pediatric Application

Abstract:
This thesis addresses the security risks posed by large language models (LLMs) generating code that uses newly discovered, vulnerable APIs due to outdated training data. To mitigate this, the proposed tool, APILOT, maintains a real-time dataset of outdated APIs and guides LLMs to avoid them during code generation. Evaluations show that APILOT reduces outdated API usage by an average of 89.42% across multiple LLMs, with minimal impact on performance.

Research Adviser: David Zaretsky

APILOT: Securing LLM-Generated Code by Avoiding Outdated API Usage

Abstract:
This thesis presents VIR-V, a custom RISC-V Rocket coprocessor designed to accelerate VCODE, an intermediate vector dataflow representation derived from the NESL language, which expresses nested data-parallelism. VIR-V efficiently maps these operations to hardware for parallel execution, supporting key operations like arithmetic, scans, reductions, and permutations. Implemented in Chisel and evaluated with FireSim, VIR-V achieves a 1.89× speedup over baseline CPU performance, demonstrating its effectiveness and compatibility within the RISC-V ecosystem, with future potential for enhancements.

Research Adviser: Peter Dinda

VIR-V: A RISC-V RoCC Accelerator for VCODE Computing

Abstract:
This report investigates the use of Kolmogorov-Arnold Networks (KANs) for speech bandwidth extension (BWE), aiming to reconstruct high-frequency components in narrowband speech to enhance quality. A novel architecture integrates KANs into a Transformer-augmented U-Net (TUNet), combining attention-based global context modeling with learnable activation functions for improved spectral reconstruction. Although the method shows promise in capturing complex patterns, it underperforms the baseline due to computational and training challenges, highlighting areas for future optimization.

Research Adviser: Stephen Xia

Analysis for application of KAN for Bandwidth Extension

Abstract:
This study presents a modular framework for efficient curation of large-scale pathology image datasets using content-based image retrieval (CBIR), a novel weighted similarity metric, and a closed-loop refinement process. A triplet-loss-based feature extractor and multi-feature similarity fusion enable effective retrieval across diverse datasets, with task-specific models performing best within domains and foundation models excelling in rare or mimic-rich cases. Applied to a real-world funisitis dataset, the system demonstrates that iterative retrieval with minimal manual input can significantly enrich relevant samples, supporting scalable pathology research and rare disease discovery.

Research Adviser: Jeffrey Goldstein

Closed-Loop CBIR for Scalable Pathology Image Curation: A Case Study on Umbilical Cord Funisitis

Abstract:
Accurate classification of autonomous vehicle (AV) behaviors is essential for improving system performance and safety, and this study introduces LLM-MLFFN, a novel framework that enhances classification using multimodal data and large language model (LLM)-based semantic reasoning. The framework integrates statistical and behavioral features with LLM-generated semantic descriptions, fusing them through a dual-channel attention network to achieve high prediction accuracy. Tested on the Waymo dataset, LLM-MLFFN reached 94% accuracy, outperforming existing models and demonstrating the value of combining numerical and semantic data for robust AV behavior analysis.

Research Adviser: David Zarestky

Large Language Model-Enhanced Multi-Level Feature Fusion Network for Autonomous Driving Behavior Classification

Abstract:
Episodic Reinforcement Learning tasks are often hindered by sparse rewards and hidden trap states that lead to failure without clear feedback. To tackle this, the proposed Time-Weighted Contrastive Reward Learning (TW-CRL) framework uses both successful and failed demonstrations, incorporating temporal context to learn a dense reward function that highlights critical success and failure states. Experiments on navigation and robotic manipulation tasks show that TW-CRL improves learning efficiency and robustness, outperforming existing IRL methods.

Research Adviser: Stephen Xia

Time-Weighted Contrastive Reward Learning for Efficient Inverse Reinforcement Learning

Abstract:
As climate change increases drought stress, understanding CAM photosynthesis—known for its water-use efficiency—is vital, yet often overlooked in biology education due to a lack of accessible tools. This thesis introduces PhytoBits, a low-cost bio-electronic system that uses implanted electrodes and microcontrollers to detect CAM activity in living plants, enabling hands-on learning. By capturing bio-electrical signals tied to CAM’s acid accumulation cycles and integrating them into educational modules, the system makes complex plant physiology both observable and teachable in classroom settings.

Research Adviser: Nivedita Arora

Phytobits: A Bioelectronic Sensor and Microcontroller-Based Curriculum for Cam Photosynthesis Education

Abstract:
This study investigated the relationship between periventricular and deep white matter hyperintensities (WMH) and cognitive function in 81 elderly individuals using automated segmentation and cerebrovascular measurements. Deep WMH volumes were more strongly linked to cognitive decline than periventricular WMH, and both subtypes showed moderate negative correlations with regional cerebral blood flow, particularly in the frontal and insular regions. The findings highlight that increased WMH burden reflects broader cerebrovascular compromise and arterial stiffness, supporting the need for subregional WMH analysis in improving diagnostic precision for cognitive impairment.

Research Adviser: Lirong Yan

Deep and Periventricular White Matter Hyperintensity Segmentation and Their Relationship with Cognitive Decline and Vessel Hemodynamics

Abstract:
Cell segmentation is difficult due to complex cell shapes, overlapping boundaries, and the limited availability of expert-labeled data. To overcome these challenges, this study introduces a Multiscale Dilated Fusion Attention (MDFA) module within a UNet framework, which uses dilated convolutions and multiscale feature fusion to better capture diverse spatial information. The proposed method significantly improves segmentation accuracy, especially in low-data and low-training scenarios, outperforming baseline models.

Research Adviser: David Zaretsky

Cell Image Segmentation System Based on Improved UNet

Abstract:
This study evaluates three communication strategies—Full Communication, Bernoulli, and Smart Communication—in a multi-agent patrol system under partial observability, using MAPPO with Graph Neural Networks. Results show that Smart Communication outperforms others by achieving the lowest average idleness with minimal communication, making it ideal for real-world use, while Full Communication performs well but is less efficient due to high communication overhead. The Bernoulli model struggled with instability and higher idleness, highlighting the importance of intelligent, adaptive communication in multi-agent coordination.

Research Adviser: Qi Zhu

Selective Communication Strategies for Multi-Agent Patrolling Systems

Abstract:
Superconducting qubits are strong candidates for scalable quantum computing, but simulating their energy spectra is computationally expensive due to repeated diagonalization of large Hamiltonians. This work adapts Eigenvector Continuation (EVC), a reduced-basis method, to accelerate these calculations for fluxonium and 0-π qubits, introducing monolevel and multilevel schemes and using PCA to address numerical instabilities. Results show exponential error convergence and up to 30× speed-up for sparse 0-π qubit Hamiltonians, significantly reducing runtime while maintaining high accuracy.

Research Adviser: Jens Koch

Eigenvector Continuation for Superconducting Qubits

Abstract:
This paper introduces a variational deep learning framework for face recognition under atmospheric turbulence, combining a Variational Autoencoder (VAE) with an ArcFace-based backbone to enhance robustness against blur and noise. Two models are proposed: VAE-FR, which implicitly models degradation, and JDVAE-FR, which explicitly predicts spatial blur and noise patterns for improved feature extraction, while keeping inference efficient by discarding decoders post-training. Experiments on the IJB-C benchmark show significant performance gains under challenging conditions, with JDVAE-FR achieving an AUC of 99.6932% in blur+noise scenarios and 99.7317% in noise-only settings, demonstrating strong resilience and improved discriminability.

Research Adviser: Aggelos Katsaggelos

Variational Deep Atmospheric Turbulence Restoration and Correction for Face Recognition with InsightFace

Abstract:
This project uses a gem5-based RISC-V vector architecture simulator to study how varying scalar core configurations influence performance. Testing across scalar and vectorized benchmarks reveals that scalar core design significantly affects execution time and instruction throughput, even in vector-intensive workloads. The results emphasize the critical role of scalar cores in vector systems and showcase the importance of architectural configurability in simulation research.

Research Adviser: Russ Joseph

Vector Processor Performance Model

Abstract:
This research presents a fine-tuning strategy for AV-HuBERT to enhance audio-visual speech recognition performance in challenging drone noise environments. By replacing the original Transformer encoder with a Conformer and adding Compression and Recovery (CAR) modules, the model better captures dependencies and improves noise resilience. Experiments using the LRS2 dataset and custom drone noise recordings show that the modified AV-HuBERT outperforms the original, validating the effectiveness of the architectural changes.

Research Adviser: Stephen Xia

Finetuning Audio-visual model in noisy environments

Abstract:
This thesis compares structured and unstructured approaches for predicting ICU length of stay (LOS) within the first 24 hours using the MIMIC-IV dataset. Structured data (demographics, vitals, labs, ICD codes) fed into classical models like XGBoost achieved the highest performance (AUROC = 0.805, R² = 0.352) with low compute cost, while transformer-based models using free-text notes (e.g., ClinicalBERT) achieved similar accuracy but required significantly more training time. The study concludes that both modalities provide strong early predictive signals, and the choice between them should consider factors like computational resources, integration ease, and clinical workflow compatibility.

Research Adviser: Zachary Wood-Doughty

Early ICU Length-of-Stay Prediction on MIMIC-IV: A Dual Approach with Clinical Features and Textual Notes

Abstract:
This study introduces a deep learning framework that predicts prostate cancer molecular subtypes (PAM50 and PSC) directly from H\&E-stained whole-slide images, achieving AUCs of 0.774 and 0.724, respectively. The results highlight image-based AI as a cost-effective, non-invasive alternative to molecular testing, with potential clinical utility in guiding personalized therapy.

Research Adviser: Lee Cooper

A Deep learning approach for Prostate Molecular Subtype Prediction

Abstract:
This thesis explores a series of plasmonic waveguide designs aimed at achieving strong optical confinement while mitigating intrinsic ohmic losses. Starting with a basic Ag–InP planar structure, the study introduces an InGaAs gain layer to enhance vertical confinement and mode–gain overlap, followed by an etched InP ridge design for improved lateral confinement. The final hybrid Ag–InGaAs–InP ridge waveguide, optimized with a 300 nm InGaAs layer, demonstrates low loss, high confinement, and positive net gain, offering a viable solution for compact and efficient plasmonic integrated circuits.

Research Adviser: Seng-Tiong Ho

Optimization of Plasmonic Waveguide Structure with Optical Gain for Realization of Lossless Plasmonic Integrated Circuits

Abstract:
This thesis introduces a low-cost, wireless sensor system designed to monitor CAM photosynthesis in real time, addressing limitations of traditional bulky and expensive tools. The system uses biocompatible electrodes, low-power electronics, and the LoRa communication protocol to enable scalable, long-term data collection in field conditions. Developed with the Chicago Botanic Garden, it also integrates anomaly detection and tinyML to support researchers with real-time insights and data management.

Research Adviser: Nivedita Arora

A Novel Electrochemical IoT Sensor for Cheap, Scalable, and Real-Time Plant Metabolism Monitoring of CAM Plants

Abstract:
This thesis explores advanced cerebrovascular image analysis by combining MRI with deep learning, introducing the 4DST U-Net—a spatiotemporal convolutional model—for accurate segmentation of cerebral vessels in 4D MRA, achieving strong generalizability across diverse datasets. It also investigates the relationship between arterial damping capacity, perivascular space (PVS) enlargement, and cognitive function, revealing that reduced damping correlates with increased PVS burden and potential cognitive decline. Together, these findings establish a multi-scale framework that merges technical innovation with physiological insight, offering promising tools for early diagnosis and personalized treatment of neurovascular and cognitive disorders.

Research Adviser: Lirong Yan

Advanced Neurovascular Image Processing and Analysis Using Deep Learning: Non-Contrast Enhanced 4DMRA Segmentation and Quantitative Assessment of Perivascular Spaces

Abstract:
This study presents a cascade enhancement framework for cine cardiac MRI (CMR) that improves segmentation accuracy by progressively refining image quality across three domains: contrast, sharpness, and inhomogeneity. Unlike prior methods, it uses high-quality references from within the dataset itself and enhances images before feeding them into a 2D U-Net for biventricular segmentation. Evaluations on two public datasets (ACDC and M&M) showed improved image quality, segmentation performance, and consistency in functional measurements, demonstrating the framework's potential to enable more reliable automated CMR analysis.

Research Adviser: Mohammed Elbaz

Cascade Disentangled Quality Enhancement for Improved Biventricular Segmentation in Cine CMR

Abstract:
This thesis presents a finite element simulation study of neuromorphic devices, focusing on memtransistors and organic electrochemical transistors (OECTs), using COMSOL Multiphysics to model key electrostatic and transport mechanisms. Simulations showed that back-gated memtransistors provide superior gate control through enhanced modulation of the Schottky barrier, while drift-diffusion modeling of OECTs accurately captured experimental trends, including high on/off ratios and distinct p/n-type behavior. The developed models offer a validated framework for optimizing low-power, scalable neuromorphic devices, supporting future advances in bioelectronics and brain-inspired computing.

Research Adviser: Mark Hersam

Finite Element Simulation of Neuromorphic Devices

Abstract:
This study introduces an integrated visible-light OCT and fluorescence SLO system for in vivo validation of retinal ganglion cell (RGC) axon imaging, addressing the limitations of postmortem flat-mount validation methods. Using transgenic Eno2-YFP mice, the system demonstrated highly consistent imaging between vis-OCT fibergraphy (vis-OCTF) and SLO, with a Pearson correlation coefficient of 0.991 across measured axon bundle widths. These results highlight the potential of this multimodal system for non-invasive, longitudinal studies of RGC damage in optic neuropathies.

Research Adviser: Hao Zhang

Integrated visible-light optical coherence tomography and fluorescence scanning laser ophthalmoscopy

Abstract:
This thesis explores the integration of machine learning into human-centered sensing systems to enhance intelligence, usability, and robustness in real-world applications. It introduces three key systems: DomAIn, a natural language-driven smart home automation platform; DUal-NET, a transformer-based speech enhancement model for bone-conduction microphones; and SoleSense, a shoe-based wearable for outdoor sound event detection using a compact transformer-convolution model. Together, these systems showcase how domain-aware design and modern ML architectures can make contextual computing more accessible, effective, and deployable in everyday environments.

Research Adviser: Stephen Xia

Human-Centered Sensing Across Wearables, Audio, and Smart Environments: A Multi-Modal, Model-Driven Approach

Abstract:
This thesis explores the use of large language models (LLMs) to decode unvoiced electromyography (EMG) signals into text without relying on paired audio, making it more practical for individuals who cannot produce vocal speech. A novel EMG adaptor is introduced to map EMG features into the LLM’s input space, achieving a 0.49 word error rate and outperforming specialized models by nearly 20%, even with just six minutes of training data. This work marks an important step toward enabling LLMs to understand silent speech through articulatory biosignals like surface EMG.

Research Adviser: Qi Zhu

Multi-modality and Large Language Model

Abstract:
This thesis addresses decentralized spectrum sharing in wireless networks, where multiple source-destination pairs must independently learn optimal transmission strategies without coordination or knowledge of others. It introduces Fair Share RL (FSRL), a reinforcement learning-based approach that incorporates state augmentation, risk-aware architecture, and fairness-oriented rewards to maximize throughput while promoting fairness. Simulations show that FSRL significantly outperforms a standard baseline, achieving up to 89% greater fairness in challenging conditions and 48.1% improvement on average.

Research Adviser: Igor Kadota

Fair Dynamic Spectrum Access via Fully Decentralized Multi-Agent Reinforcement Learning

Abstract:
This thesis tackles the challenge of optimizing Age of Information (AoI) in real-world wireless networks, which is critical for applications requiring fresh, time-sensitive data such as autonomous vehicles and IoT systems. It addresses two practical issues: the base station’s limited knowledge of true AoI due to unreliable communication, and the impact of highly variable update sizes. The proposed solutions include an MMSE-based AoI estimator, randomized and Max-Weight scheduling policies, and a Lyapunov-based Age-Debt policy, all of which demonstrate strong analytical and simulated performance in minimizing AoI under realistic constraints.

Research Adviser: Igor Kadota

Optimizing Age-of-Information in Real-World Networks

Abstract:
This thesis addresses the limitations of current AI hardware by exploring neuromorphic computing, which mimics the brain's efficiency through emerging devices like memtransistors made from 2D materials such as MoS₂. It highlights the potential of multi-terminal memtransistors for integrating memory and processing, while identifying key challenges, including non-scalable synthesis methods, high operating voltages, and the unexplored influence of contact metals on device performance. The work focuses on improving MoS₂ synthesis, optimizing memtransistor fabrication, and engineering contact materials to advance neuromorphic computing and bridge materials science with computer engineering and neuroscience.

Research Adviser: Mark Hersam

Nanomaterial Synthesis and Memtransistors Fabrication for Next-generation Microelectronics

Abstract:
This thesis develops two innovative wireless and flexible bioelectronic systems addressing critical clinical needs with minimally invasive and non-contact approaches. The first is a battery-free TAVI pacemaker that uses metasurface-assisted wireless power transfer for safe, deep-tissue cardiac stimulation without an internal power source. The second is a non-invasive breast milk volume monitoring platform employing flexible antennas and a novel algorithm, demonstrating potential for wearable integration and real-time wireless sensing.

Research Adviser: John Rogers

Design and Evaluation of Wireless and Flexible Bioelectronic Systems: From Implantable Pacemakers to Non-Invasive Milk Volume Monitoring