2024 SURE Fellows

Aaron Bordeaux

Image

• Mentor: Professor Jason Nieh
• Home Institution: Princeton University
• Hometown: St. Michael's, AZ

Implementing Quantum Virtual Machines for Enhanced Resource Utilization in Quantum Computing

This paper introduces HyperQ, a novel system designed to enhance the efficiency and performance of quantum cloud computing by implementing quantum virtual machines (qVMs). We integrated HyperQ into the IBM Quantum Platform and evaluated its effectiveness from provider and user perspectives. Our evaluation focused on throughput, resource utilization, latency reduction, and fidelity maintenance improvements. Using a subset of QASMBench benchmarks, we demonstrated that HyperQ significantly increases throughput and utilization by up to six and seven times, respectively while reducing latency by a similar factor. Importantly, these improvements do not compromise the fidelity of quantum program execution. HyperQ's ability to perform both space and time scheduling of quantum workloads showcases its potential to effectively optimize near-term quantum computing resources. Our results indicate that HyperQ is a scalable and practical solution for enhancing quantum cloud services, providing substantial performance gains without requiring significant changes to existing quantum hardware or compilation tools.

Aaron Katznelson | Sia Lab

Image

• Mentor: Abigail Ayers; Cristia Victoriano; Sam Sia; Sia Labs; 
• Home Institution: Yeshiva University
• Hometown: Woodmere, NY

Optimizing Operation of Sample Preparation Device For Blood-Borne Pathogen Diagnostics

Blood sample preparation is a prerequisite in diagnostic testing for blood-borne diseases such as Hepatitis C Virus. However, current methods are too complex and expensive to use in a widespread testing application. The lab has previously developed a point-of-care friendly device that utilizes microfiltration and RNA binding magnetic microparticles to process small volumes of blood to be converted into an RNA sample that is suitable for diagnostic testing. This work explores optimizations to the device in addition to a secondary device that was designed to increase throughput and optimize user operation and ease of use.

Abidur Rahman | Lightwave Lab

Image

• Mentor: Prof Alex Meng; Lightwave Research Laboratory; 
• Home Institution: University of Michigan; Ann Arbor; 
• Hometown: Detroit, MI

Enhancing Distributed Computing with Optical Interconnects in Data Centers and High-Performance Computing (HPC) Systems

Optical photonics interconnects are emerging as a highly promising solution for distributed deep learning (DDL) due to their ability to address the critical bottlenecks in communication among distributed computing units (CUs). As deep learning models like OpenAI’s GPT-3 and GPT-4, Meta’s Llama 2 and Llama 3, and Google’s Gemini grow in size and complexity, the need for distributed training across multiple CUs has become essential. However, traditional electronic interconnects often struggle to provide the necessary bandwidth and low latency required for efficient communication in these distributed systems [1],[2]. Optical interconnects provide high-bandwidth, low- latency communication paths make silicon photonics an ideal solution for overcoming the communication bottlenecks in distributed deep learning. As a result, these technologies are poised to significantly improve the performance and scalability of deep learning models, enabling faster and more efficient training of increasingly complex neural networks.

Amy Wu | DNA Lab

Image

• Mentor: Dr. Vishal Misra; Distributed Network Analysis; (DNA) Lab
• Home Institution: University of Florida
• Hometown: Miami, Florida

DIVING DEEP: HOW DO LLMS LEARN ON THE FLY?

Large Language Models (LLMs) have revolutionized natural language processing, demonstrating remarkable capabilities in text generation and under-standing. One of the many phenomenons in LLMs is in-context learning, where task-specific responses are generated by providing the model with task-specific prompts [1]. This study investigates the behavior of LLMs through the lens of Bayesian learning, where the model constantly updates its token generating distribution based on new evid-ence. To further explore this topic, we propose a practical open-source tool for visualizing token probabilities during text generation, providing empirical insight into the models’ next token prediction process. This lets us analyze how the model is constantly learning in real-time.

   Andrea Patricia Sepulveda-Vargas | CEAL Lab

Image

• Mentor: Prof Brian Smith; CEAL Labs; 
• Home Institution: University of Puerto Rico
• Hometown: Puerto Rico, Mayaquez

Giving City Residents Greater Access to Local Quality-of-Life Data

Cities produce large amounts of data such as public service request data and local government meeting records. However, much of this data requires technical knowledge or a time investment from the user to access, making this data accessible only to a limited group of people. We believe there is an opportunity to make this information more accessible to everyday community residents. In this project, we explored how we could process publicly available data such as that from NYC Open Data and NYC Community Boards, to make it more transparent to residents. Our findings from Harlem, New York reveal patterns within the data that point to possible hyper-local issues. Future work points to user-facing systems that compile and expose these patterns in an accessible way, leading to greater community awareness and more consistent action from governments. To this end, we present high-fidelity prototypes for a mobile app that visualizes these community-specific concerns.

AnMei Dasbach-Prisk | Systems + Network Lab 

Image

• Mentor: Prof Asaf Cidon, Systems and Networking Lab 
• Home Institution: Cabrillo College
• Hometown: Santa Cruz, California 

Evaluating BEC Emails Generated by LLMs

Business Email Compromise (BEC) attacks pose a significant[1] and evolving threat. This research contributes to a larger project aimed at evaluating email scams and building a comprehensive training dataset to improve scam detection. Specifically, we explore the potential of Large Language Models (LLMs) to enhance the persuasiveness of spear phishing email templates[2]. We assess the effectiveness of different prompting techniques using G-Eval[3], a framework that evaluates LLM generated output provided custom criteria. Our findings overall demonstrate LLMs' ability to improve template quality.

Annie Jaswal | Barmak Lab

Image

• Mentor: Prof Katayun Barmak, Barmak Lab 
• Home Institution: Colby College
• Hometown: Mamaroneck, New York 

Recrystallization Cycling and Quenching Study for the Synthesis and Characterization of Transition-Metal Dichalcogenides

Two-dimensional transition-metal dichalcogenides (TMDs) are a class of layered materials that have shown significant potential for applications in electronics and opto-electronics due to their unique electrical and optical properties when exfoliated down to molecular monolayers.1 High-purity materials as well as large crystal sizes are necessary for understanding and optimizing this potential.2 A two-step flux synthesis method has been developed to produce TMDs with significantly reduced concentrations of charged and isovalent point defects.3 However, substitutional defect densities remain at densities greater than 1011 cm-2 , and the relatively small sizes of the resulting crystals are not ideal as they hinder exfoliation yields and device fabric-ation. To address these issues, this effort uses a recrystallization cycling process to attempt to further reduce point defect densities and produce larger crystals. However, the recrystallization effort is found to introduce impurities, increasing the defect density, and have minimal effect on the crystal size.  Additionally, crystal growth dynamics are investigated through ex-situ crystal size measurements throughout the synthesis process via quenching and backlight imaging, maximizing the efficiency of the synthesis process.

Aroon Sankoh | Software Systems Lab 

Image

• Mentor: Prof Junfeng Yang, Software Systems Lab
• Home Institution: Washington University in St. Louis 
• Hometown: Boston, Mass 

Developing a Novel Audio Dataset for the Implementation of Robust Deepfake Detection Models

Misinformation is everywhere in today’s digital age. One straightforward and subtle way to spread it is through the use of deepfakes—videos or audio that are digitally altered to represent someone else. Given the advancements in deepfake technology and the approaching 2024 US Presidential election, it is more pressing than ever to develop robust deepfake detection models to combat the spread of misinformation. Our team has spent weeks developing a novel audio dataset to implement state-of-the-art deepfake detection models with. This dataset contains 1.2 million datapoints of spoofed (deepfake) audio from >15 distinct generation systems and an equal amount of bonafide (genuine) audio from multiple websites. After contributing to the dataset, I trained and tested the RawGAT-ST deepfake detection model on it. The results indicate that training deepfake detection models on a wide variety of deepfake detection generation sources will improve the models' accuracy on out-of-distribution data.

Ashveen Banga | Roar Lab

Image

• Mentor: Prof Sunil Agrawal, Roar Lab 
• Home Institution: Carnegie Mellon University
• Hometown: San Diego, California 

Enhancing Navigation Skills:

Immersive Virtual Reality Training for Alzheimer’s Disease Patients

Spatial navigation skills are crucial for daily living. Allocentric navigation is one of two main strategies used by people while navigating an environment; it involves relying on spatial memory of how key landmarks in an environment relate to each other [1]. Allocentric navigation is negatively impacted when hippocampus function is impaired, which occurs in Alzheimer’s disease [2]. Training allocentric navigation abilities thus could potentially improve everyday function and quality of life for people with Alzheimer’s disease.

Austin Gibson | Center for Theoretical Neuroscience 

Image

• Mentor: Ashok Litwin-Kumar, Jasmine Stone, Center for Theoretical Neuroscience, Zuckerman Institute, Columbia University 
• Home Institution: Tuskegee University
• Hometown: Opelika, Alabama 

Curriculum learning and experience replay in a model of a cognitive decision-making task.

Humans can learn and easily perform a wide variety of cognitive tasks, and recent research (1) establishes recurrent neural networks (RNNs) as a modeling technique in cognitive neuroscience. Both humans and networks can learn faster when trained on tasks in a specific order, a method termed ‘curriculum learning’ (3). However, when trained on multiple tasks sequentially, networks’ performance can decrease on previously learned tasks, a phenomenon termed ‘catastrophic forgetting’. Techniques such as replay can alleviate catastrophic forgetting. Here we investigate the effect of training order and replay on network performance of multiple cognitive tasks commonly studied in neuroscience.

Azalea Bailey  

Image

• Mentor: Zhou Yu
• Home Institution: University of California, Berkeley
• Hometown: Poughkeepsie, New York

Synthesizing Human Conversation Data with Large Language Models

Artificial Intelligence (AI) chatbots are revolutionizing various industries [1]. Companies are motivated to invest in chatbots to reduce labor costs and streamline their operations. However, training these chatbots can be challenging due to the need for historical conversation data, which is often not readily available and costly to obtain [2]. This paper explores two methods for synthesizing conversation data between a company’s AI chatbot and a Large Language Model (LLM) that mimics human dialogue [3]. The first method involves prompting an LLM to generate utterances based on the company’s AI chatbot’s user dialogue intent categorization tree [3]. The second method prompts an LLM to generate a user goal for the conversation to achieve based on a description of the company [3]. This study finds that LLMs can replicate human dialogue at a statistically significant level regarding conversation intent distribution and Bidirectional Encoder Representations from Transformers Score (BERTScore) [4].

Cristian Pena | Josz Lab

Image

• Mentor: Prof Cedric Josz, Josz Lab
• Home Institution: Florida Atlantic University 
• Hometown: Fort Lauderdale, Florida

Low-Rank Matrix Factorization for Moving Object Detection in Video Data

This study investigates the application of low-rank matrix factorization for moving object detection in video data. The primary goal is to enhance the extraction of moving objects from a sequence of video frames by approximating the data matrix representing the video as a rank-one matrix plus noise. This approach addresses the challenge of isolating moving elements within a video stream, contributing significantly to fields such as surveillance, traffic monitoring, and video analytics [1, 2].

Danna A. Soriana | Adeleye Lab

Image

• Mentor: Prof Adeyemi Adeleye, Adeleye Lab
• Home Institution: Stanford University
• Hometown: Brooklyn, New York

Modifying Biochar for Improved Dissolved Metal Adsorption and Recovery

Biochar is a versatile, charcoal-like substance that is made through the partial combustion of biomass in the presence of limited oxygen. The material has a large specific surface area and a porous structure, rich in amino acids. These chemical and physical properties have made biochar an appealing candidate for a novel tool that sustainably removes environmental toxins. Our goal is to modify biochar to enhance its performance for cadmium ion removal from water, and determine the influence of solution concentration, pH, and treatment time on cadmium adsorption.

Francesca Hudnall-Saez  

Image

• Mentor: Jessica Z. Liu 
• Home Institution: San Francisco State University
• Hometown: Santa Cruz, California

Finite Element Analysis to Model the Effects of Osteoarthritis in the Knee

Osteoarthritis (OA) of the knee is a degenerative disease that begins with the breakdown of the joint cartilage [2]. The presence of OA in the knee can lead to damage of the cartilage tissue, tendons, synovium and bone, causing chronic pain, swelling, loss of motion, and may lead to excessive stress on the joint [1]. The subchondral bone (SB), calcified cartilage (CC), and articular cartilage (AC) provide the knee with the mechanical strength and support to withstand applied loads. The CC and AC that lies on top of the SB are composed of cells called chondrocytes, which secrete extracellular matrix containing proteins, collagen, and proteoglycans [4]. These proteoglycans are hydrophilic cells, retaining water molecules in CC and AC that give cartilage the mechanical properties necessacery to resist and distribute compressive forces [8]. Pathological processes that affect the joint morphology lead to changes in the biomechanics of the knee, and has been correlated to the presence of OA [1]. Understanding how OA affects the stresses, strains, and load distributions across the SB and CC interface of the knee is integral to engineering cartilage that is capable of replacing damaged tissue. The objective of this study is to use finite element analysis and modeling techniques to identify how the morphology of the knee, and the changes experienced with the presence of OA, affect the stress and strain distributions across the interface.

Geoffrey Enwere | Banta Lab

Image

• Mentor: Prof Scott Banta, Banta Lab
• Home Institution: Massachusetts Institute of Technology
• Hometown: Hempstead, New York

Protein-based Approaches for Biosensor Development.

Biosensors are critical tools for detecting and processing biological signals, with applications in environmental monitoring, medical diagnostics, and bioengineering. The current study investigates two distinct approaches for biosensor module development: protein-based and cell- based systems. This research focuses on engi- This research focuses on engineering bacterial two-component systems (TCS) and developing directed evolution platforms to enhance biosensor performance.neering bacterial two-component systems (TCS) and developing directed evolution platforms to enhance biosensor performance. The goal is to im- improve the specificity and efficiency of biosensors, particularly in detecting small molecules and facilitating rapid biological information processing.

Izabella Zmirska | Stem Cell Engineering Lab

Image

• Mentor: Dr. Kam Leong, Nanotherapeutics & Stem Cell Engineering Lab
• Home Institution: University of Florida
• Hometown: St. Augustine, Florida

Eudragit Coated Polyplexes for Non-viral Gene Delivery

Gene editing can precisely alter the genome, addressing various genetic disorders and providing a means of achieving chronic protein therapy. Non-viral vectors, such as polymer-based polyplexes, are promising gene delivery systems that exhibit lower cytotoxicity, immunogenicity, and mutagenesis compared to their viral counterparts [1]. Polyplexes offer an innovative method for oral gene delivery, potentially making gene therapeutics more convenient for frequent administration to patients. Chitosan grafted with branched polyethyleneimine (CS-g-bPEI) is an excellent candidate for these polyplexes due to the mucoadhesiveness of chitosan and the buffering capacity of branched polyethyleneimine at acidic endosomal pH. Furthermore, its cationic nature enhances uptake by gut epithelial cells [2].

Orally delivered polyplexes often have poor bioavailability in the colon, making them suboptimal for treating diseases like colon cancer or irritable bowel disease. To enhance the potential of CS-g-PEI polyplexes for targeted oral gene delivery to the colon, we have coated them with Eudragit, a commercial polymer that releases drugs to specific sites along the gastrointestinal tract based on pH. Specifically, we tested Eudragit S100 and Eudragit FS30D, which promote site-specific drug release in the colon due to their solubility at a pH of 7 [3]. By experimenting with different formulations and synthesis methods for making Eudragit-coated polyplexes, we successfully demonstrated their tunability in charge and size. We also assessed their gene editing efficiency and cytotoxicity through in vitro studies.

James Casey | Kelly Lab

Image

• Mentor: Dr. Shaina Kelly, Kelly Lab
• Home Institution: University of Southern California
• Hometown: Los Angeles, California

Thermal Conductivity of Synthetic Materials Across Varying Saturation States

In subsurface environments, geological formations retain moisture and fluids through infiltration across the Earth's crust, affecting processes such as geothermal energy extraction and gas storage, rendering the assessment of the thermal conductivity and fluid saturation relationship paramount. Synthetic materials with known compositions and porosities are standardized benchmarks to investigate these phenomena. They offer a systematic framework to understand the behavior of natural materials under diverse conditions. However, despite their utility, the characterization of these synthetics remains sparse, with chall-enges ranging from the accessibility of effective thermal conductivity measurements to existing method-ologies inadequately addressing the influence of saturation levels on thermal properties. This research addresses these gaps by characterizing the porosity, permeability, and thermal conductivity of micro- and nanoporous synthetic materials across different saturation states. These workflows are designed to enhance the accuracy and reliability of models applied to complex, heterogeneous natural materials.

Jessica Xie  

Image

• Mentor: Parth Gami, Lab Ultrasound and Elasticity Imaging Lab (Prof Elisa Konofagou) 
• Home Institution: University of Missouri 
• Hometown: Columbia, Missouri

Towards At-Home Applications of Pulse Wave Imaging (PWI): Investigating Optimal Sampling Parameters for

Real-Time Pulse Wave Velocity (PWV) Estimation

Pulse wave imaging (PWI) is a noninvasive, high-frame-rate, ultrasound imaging technique that tracks the pulse wave-induced arterial wall motion. PWI has been demonstrated to provide estimates of central arterial mechanics (carotid, aorta), including pulse wave velocity (PWV) at end-diastole (PWVED) and end-systole (PWVES), compliance, and pulse pressure (PP)1-2 . Quantification of central arterial properties could improve the characterization, monitoring, and prediction of cardiovascular diseases like hypertension, diabetes, and Alzheimer’s. Recent advancements in piezoelectric micromachined ultrasound technology (PMUT) has enabled the development of wearable applications of ultrasound imaging3 , which could be utilized to translate PWI to at-home settings. While the current PWI technique provides reliable results, PWI requireS offline processing, manual intervention, ~15-18 minutes of run-time, and GPU-based computing. Since at-home applications of PWI will require real-time PWV and PP estimation, it is necessary to reduce the data size and computational complexity of PWI. This study focuses on investigating the optimal, minimum number of lateral positions required for accurate PWV estimation.

Jayla McCoy | Soft Tissue Labs  

Image

• Mentor: Prof Kristin Myers, Soft Tissue Lab 
• Home Institution: University of Kentucky
• Hometown: Cleveland, Ohio

Using Computational Mechanics to Improve Preterm Birth Risk Assessment

Preterm birth (PTB), defined as birth before 37 weeks of pregnancy, is the leading cause of childhood morbidity and mortality [1], affecting 1 in every 10 infants in the U.S. [2]. Shockingly, the global PTB rate has decreased by just 0.14% from 2010 to 2020 [2], highlighting very few advancements in the past decade. Up to 70-80% of all PTB cases occur unexplained or spontaneously (sPTB) [3], often resulting from premature rupture of membranes, premature uterine contractions, and premature cervical softening pathways that activate the remodeling of the cervix. Currently, there is no reliable method to predict sPTB. For routine check-ups, transvaginal ultrasound measurement of cervical length (TVCL) is the preferred method for sPTB risk stratification but has limited sensitivity [4]. A possible new and promising method is the Pregnolia aspiration device system [5], which can assess the stiffness of the cervix through a proxy parameter known as the aspirated cervical stiffness (aCS) index. Since changes in cervical stiffness correlate to different stages of cervical remodeling during pregnancy [6], this device has the potential to help better understand and quantify cervical remodeling to predict sPTB. By combining the aCS index with patient specific measurements obtained from ultrasound through in-silico computational mechanics tools such as finite element (FE) modeling, we can obtain further insights to the deformation and stress fields in the cervix and aim to further improve this method for sPTB prediction.

Joyce Gill | PC Enabled Abilities Lab

Image

• Mentor: Prof Brian A. Smith, Computer Enabled Abilities Lab 
• Home Institution: Grinnell College
• Hometown: Seoul, South Korea

Leveraging Augmented Reality to Preserve Community History

As gentrification continues to rapidly transform many communities, the intangible heritage of a community’s mem-ories and lived experiences are at risk of being lost, leading to the gradual erosion of a community’s history and cultural heritage. In this context, Augmented Reality (AR) has emerged as a powerful medium for integ-rating physical and digital realms, offering novel possibilities for preserving community history. This paper explores the following questions: 1. How can AR be effectively utilized to preserve com-munity history? 2. What are the key UI/UX design considerations for developing a community-centered AR application? This study details a case study focused on Harlem, New York where we conducted formative studies to inform our design process and prototyped Community AR, an interactive AR platform designed to preserve community history while enhancing civic engagement. We concluded by conducting a pilot study, through which we gained key insight into AR’s potential as a tool to bridge the gap between physical spaces with their historical narratives.

Julianne Hyvert 

Image

• Mentor: Prof Sinisa Vulelic, Vukelic Group
• Home Institution: Northwestern University 
• Hometown: Buffalo, New York

Synthesizing Human Conversation Data with Large Language Models

Myopia, commonly referred to as near-sightedness, is a common refractive error caused by light being focused anterior to the retina, leading to blurred vision. If left uncorrected, high myopia leaves patients at increased risk of developing other ocular pathologies, including glaucoma, choroidal neovascularization, and cataracts [1,2]. Currently, 28.3% of the global popul-ation is affected by myopia, and projections indicate that nearly half of the world's population will be myopic by 2050 [3]. Refractive surgeries, such as laser-assisted in situ keratomileusis (LASIK), are popular options for patients seeking permanent vision correction. While effective in improving visual acuity, these invasive procedures can lead to postoperative complications, including epithelial ingrowth, diffuse lamellar keratitis, and corneal flap dislocation [4]. An alternative treatment involves soaking the cornea in the photosensitizer riboflavin and exposing it to ultraviolet-A (UVA) light to induce the formation of collagen crosslinks and reduce myopic refractive error [5]. However, this procedure has primarily demonstrated effectiveness in patients with low myopia [6]. Combining LASIK with riboflavin and UVA-assisted crosslinking has been shown to significantly reduce myopic regression compared to LASIK alone, demonstrating the potential for collagen crosslinking to provide greater long-term efficacy and stability in the treatment of myopia [7].

Kamaula Rowe  

Image

• Mentor: Assistant Prof Tanvir Ahmed Khan  
• Home Institution: Princeton University
• Hometown: Queens, New York

FusionBench: Analyzing Kernel Fusion in Vision Models

From data centers to mobile devices, Machine Learning (ML) drives a variety of use cases including computer vision1 , natural language processing2 , and speech recognition3 . ML benchmarks are crucial for evaluating the performance of ML models, software, and hardware. Established benchmarks from TorchBench4 and HuggingFace5 evaluate ML tasks but do not focus on kernel fusion optimizations. Kernel fusion combines consecutive operators into a single kernel, improving efficiency by minimizing memory access overhead and enhancing computational performance6 . Despite its significance, to the best of our knowledge, there is a lack of benchmark suites dedicated to this technique, creating a gap in comparison and innovation. Reliable metrics are needed to understand, optimize, and automate kernel fusion, particularly for tensor operations like convolution. Without these benchmarks, the potential advantages and effects of kernel fusion remain underexplored, hindering progress in optimizing ML model performance and automation in ML compilers. In this work, we propose FusionBench, a novel benchmark suite to study the effectiveness of kernel fusion in improving performance of various ML workloads.

Katie Cox | Hess Lab

Image

• Mentor: Henry Hess, Hess Lab
• Home Institution: Dartmouth College
• Hometown: McLean, Virginia

Analyzing the GATTA-PAINT Nanoruler using DNA PAINT

DNA Point Accumulation for Imaging of Nanoscale Topography, or DNA PAINT 1-3 , is a method used to precisely localize single DNA molecules on a surface by imaging short fluorescent oligonucleotides in solution that bind to them. An adaptation of this technique is used in a potential method of protein sequencing. This method involves connecting peptides of interest to DNA molecules on the surface and detecting the binding of short fluorescent oligonucleotides conjugated to organic receptors for amino acid side chains in solution. Since this method yields imaging data of considerable complexity, we required a testbed to validate our analysis pipeline. To do this, data from the GATTA-PAINT Nanoruler kit by GATTA-quant Inc., which utilizes the ATTO-655 red fluorophore to visualize the binding of complementary oligonucleotides in solution to oligonucleotides on the surface, was compared to sample data.

Kay Russi | Correa Lab

Image

• Mentor: Dr. Santiago Correa, Correa Lab 
• Home Institution: Clemsen University
• Hometown: Charleston, South Carolina

Evaluating lipid-based delivery of dsDNA agonists for activation of cGAS-STING pathway

In eukaryotes, DNA is primarily found in the nucleus and mitochondria. The presence of cytosolic DNA is typically a sign of internal dysregulation or the presence of pathogens. [1] The cyclic GMP-AMP synthase (cGAS)-stimulator of interferon genes (STING) pathway is responsible for detecting cytosolic DNA and eliciting an innate immune response to protect the cell. [2] Activation of this pathway plays an important role in cancer immunotherapies. [3] To activate this pathway, exogenous double-stranded DNA (dsDNA) must be delivered into the cytosol of a cell to stimulate cGAS. When cGAS detects dsDNA, cGAMP is produced, which activates STING. In response, type I interferons are produced, eliciting an immune response, such as dendritic cell maturation, which can lead to more adaptive immune cell activity, and overcoming an immunosuppressive tumor environment. [4] This study looks to use lipid-based nanoparticles (LBNPs) encapsulated with a dsDNA agonist to safely and effectively stimulate the cGAS- STING pathway.

Kedar Krishnan | Cellular Engineering Lab

Image

• Mentor: Neeraj Sakhrani, Cellular Engineering Lab
• Home Institution: Johns Hopkins University
• Hometown: Basil, Switzerland

Development of a Blood-Joint Transwell System to Investigate the Effect of High Blood Glucose

Exposure in a Diabetic Osteoarthritis Model

Osteoarthritis (OA) is a degenerative joint disease characterized by cartilage deterioration and synovial inflammation.1 Previous studies have shown that type 2 diabetes mellitus (DM) and its associated hyperglycemia may increase the progression and incidence of OA.2 The connection between both diseases has historically been attributed to increasing age and joint loading due to obesity.3 However, the underlying pathophysiological mechanisms implicated in DM and OA have not been thoroughly investigated due to the associated comorbidity involved with treating this patient population.4 Furthermore, the complex interactions between blood vessels, end-othelial cells, synovium, and articular cartilage necessitates the development of in vitro models that recapitulate physiological conditions of the joint space. In this study, we develop a novel blood-joint transwell system consisting of human umbilical vein endothelial cells (hUVEC), fibroblast-like synoviocytes (FLS), and articular chondrocytes (ACs) (Fig. 1). The transwell system allows for cross-talk between the cell types either across the transwell membrane or via paracrine factors released into shared media. As such, the system can recapitulate signaling within the joint and model the effects of DM-induced hyperglycemia in the bloodstream on OA-associated cartilage degradation.

Kevin Han | Robotics Lab

Image

• Mentor: Prof Brian Plancher, Accessible and Acelerated Robotics Lab
• Home Institution: University of Texas at Austin
• Hometown: Houston, Texas

A UNIVERSAL FRAMEWORK FOR ENSEMBLED DEEP FEDERATED REINFORCEMENT

LEARNING AND APPLICATIONS TO MICRO UAVS

Federated Reinforcement Learning (FRL) enables multiple agents with identical state and action spaces in independent and varied environments to collaboratively learn an optimal policy. This approach is beneficial in scenarios where agent privacy is crucial, such as in energy grids or medicine. Using PyTorch, I develop a deep FRL framework capable of supporting and ensembling any RL algorithm, and enhance two novel momentum-based algorithms, FEDSVRPG-M and FEDHAPG-M[1], alongside other state-of-the-art (SotA) RL algorithms, to train crazyflie drones. The novelty in these momentum-based algorithms are shown with their guaranteed convergence to a stationary point of the average performance function, despite environment heterogeneity. However, this is under the assumption that they are the only local algorithms.

For the other local algorithms, I use Proximal Policy Optimization (PPO) [2], Soft Actor-Critic (SAC) [3], and Twin-Delayed Deep Deterministic Policy Gradient (TD3) [4]. Each of these algorithms have their independent strengths. PPO is more of an on-policy algorithm that is stable and provides reliable performance but can be sample inefficient and sensitive to hyperparameters. SAC and TD3 are off-policy algorithms. SAC offers excellent exploration and sample efficiency but is computationally demanding. TD3 is robust against overestimation bias and sample efficient but can be slower to converge. Combined with deep learning, my ensemble method aims to speed up learning by leveraging the strengths of each other SotA sub-algorithm without losing too much of the convergence benefits of momentum-based FRL. I also explore the benefits of aggregation in value function estimation to determine if critics benefit from FRL too.

Furthermore, on the simulation side, I modify the gympybullet-drones[5] platform to include domain randomizations for wind and mass conditions, enhancing sim2real transfer. Using this FRL platform, I train crazyflie drones for various tasks and plan to incorporate layer freezing and LQR-based supervised learning for subtasks like hovering to advance robot learning..

Leo Yang | Samuel Sia Lab

Image

• Mentor: Terry Chern, Samuel Sia Lab
• Home Institution: Reed College
• Hometown: Wuhan, China

Upper Arm Length Prediction

Wearable medical devices are revolutionizing medical care.

However, the accuracy of these measure-ments can be compromised by body movements, particularly when the device is mounted on the arm. While the upper arm when compared to the lower arm is less susceptible to known covariates; depending on the orientation of the upper arm, these artifacts may still affect biomedical data acquisition, necessitating a method to account and negate these effects. To address this issue, it is important to determine the distance of device placement from the patient’s shoulder for an upper arm based device. By colocalizing a motion sensing device with where the measurement is taken, we can gather motion data to assist in this compensation. This study proposes an algorithm based on classical mechanics to estimate this distance by using the acceleration data, thereby enhancing the reliability of biomedical measurements in various arm orientations.

Liao Zhu | WiMNet Lab

Image

• Mentor: Gil Sussman, WiMNet Lab
• Home Institution: University of Iowa
• Hometown: Iowa City, Iowa

Characterizing 5G Interference Effects on Weather Sensors for Enhanced Spectrum Sharing in Urban Environments

The deployment of 5G wireless networks, while promising faster communication and greater connectivity, poses a significant challenge to the accuracy of weather forecasts due to potential radiation leakage into frequency bands used by weather sensors [1]. Two main focuses this summer dealt with utilizing a mobile 28 GHz IBM Phased Array Antenna Module (PAAM) to measure the effects of its transmission on a radiometer (an instrument that senses electromagnetic radiation in the atmosphere) and attempting to recreate a neural network capable of accurately predicting atmospheric data despite the presence of 5G interference. Our aim is to characterize how the interference affects the radiometer to feed into other technology we have for spectrum sharing, ultimately contributing to improved weather forecasting in dense urban environments.

Lucas Sha | Zhou Yu Lab

Image

• Mentor: Zhou Yu Lab
• Home Institution: Brown University
• Hometown: Boston, Massachusetts

Building Proficiency-Adaptive Chatbots for English Language Learners

Chatbots powered by LLMs are being used by English learners to practice their language skills in natural, conversational contexts. However, state of the art bots often generate outputs requiring an already high degree of proficiency in English to understand, making it difficult for less proficient English learners to converse with them.

In this project, we explore techniques to make chatbots automatically adapt their outputs in response to the detected proficiency level of the user conversing with them.

Marcos Javith | Correa Lab

Image

• Mentor: Prof. Santiago Correa, Correa Lab 
• Home Institution: University of Florida
• Hometown: Orlando, Florida

Cell-Derived Materials as Hydrogel Components

 
Extracellular vesicles (EVs) are nanocarriers composed of a mixture of lipids, surface proteins, and membrane proteins that equip them with the necessary markers to reach their targets for cell-to-cell communication, making them attractive for targeted thera-peutic delivery. [1] Additionally, protein nanocages are able to mimic the structure of viruses and are highly engineerable for antigen identities and combinations, making them very attractive for vaccine development and overall strengthening of the immune system (2). However, encapsulating cargo into these nanocarriers without compromising the cargo’s activity and the nanocarrier’s integrity is a challenging process. Studies have shown successful expression of cell-derived EVs and protein nanocages in Escherichia coli (E. coli). [2, 3] In this process, designed membrane proteins are encapsulated in EVs and secreted. Similarly, proteins self-assemble into icosahedral virus-like protein nanocages. Direct expression of antigen-containing nanocarriers in bacteria eliminates the need for additional engineering steps that could compromise material integrity. This study focuses on optimizing the expression, isolation, and characterization of these nanocarriers and integrating them into modified hyalu-ronic acid-based hydrogels. Loading into hydrogels reduces immunogenicity by limiting exposure to non-target tissues, ensures a sustained release rate, and provides an injectable medium for these nanocarriers, collectively enhancing therapeutic effectiveness. (4)

Rashfigur Rahman | DNA Lab

Image

• Mentor: Prof. Vishal Misra, Distributed Networks Analysis (DNA) Lab 
• Home Institution: Boston University
• Hometown: Somerset, Massachusetts

DIVING DEEP: HOW DO LLMS LEARN ON THE FLY?

Large Language Models (LLMs) have revolutionized natural language processing, demonstrating remarkable capabilities in text generation and understanding. One of the many phenomenons in LLMs is in-context  learning, where task-specific responses are generated by providing the model with task-specific prompts [1]. This study investigates the behavior of LLMs through the lens of Bayesian learning, where the model  constantly updates its token generating distribution based on new evidence.To further explore this topic, we propose a practical open-source tool for visualizing token probabilities during text generation, providing empirical insight into the models’ next token prediction process. This lets us analyze how the model is constantly learning in real-time.

Ryley McGovern | BERL Lab

Image

• Mentor: Prof. Bianca Howard, BERL Lab 
• Home Institution: Hamilton College
• Hometown: North Bergen, Indiana

Optimizing Energy Burden to Incorporate into Building Decarbonization Pathways to Retrofit Existing Buildings

Buildings account for two thirds of New York City’s greenhouse gas emissions which calls for the necessity of transforming the urban building stock to deliver on climate change goals.1 Whilst climate change policies seek to address climate change mitigation, they also seek to address additional societal challenges such as alleviating energy burden.2 Energy burden, which is the proportion of annual  household income that is spent on energy utility costs2, remains to be a persistent issue amongst low-income households. Low-income households spend a higher percentage of their income on electricity and gas bills than any other income group5. To begin to address this issue, energy burden was incorporated as a metric into an optim-ization to determine how building decarbonization pathways changes the “optimal” way to retrofit existing buildings.

Sara Gomez | ARiSE Lab

Image

• Mentor: Prof. Baishakhi Ray, ARiSE Lab 
• Home Institution:  Vanderbilt University
• Hometown: East Greenwich, Rhode Island

PrimeVUL was previously introduced as a dataset for training and evaluating large language models 

PrimeVUL was previously introduced as a dataset for training and evaluating large language models (LLMs) for vulnerability detection (VD), but research revealed the considerable gap between capabilities and practical requirements for deploying code LLMs in security roles. This project aims to enhance the detection and fixing of security vulnerabilities in open-source codebases through stub testing. By focusing on a sample of vulnerabilities from the TensorFlow codebase, this research implements stub tests to recreate vulnerabilities and validate fixes. Key findings highlight the effectiveness of stub testing in improving software reliability, proving the value of automated test generation for training LLMs in VD as stub tests provide modality of information, allowing for dynamic vulnerability tracing during development. When implemented, automated stub test generation ensures detected vulnerabilities are addressed, while data augmentation techniques simulate edge cases likely to cause vulnerabilities, both of which provide comprehensive detection and fixing solutions to enhance software robustness and reliability.

Shirly Gottlieb | Simunovic Lab

Image

• Mentor: Dr. Mijo Simunovic, Simunovic Lab 
• Home Institution: University of Delaware
• Hometown: Windmere, New York

Direct Fibroblast Reprograming with CRISPRa

This project aims to evaluate the use of CRISPR activation (CRISPRa) in con-verting fibroblast cells to endodermal cells which can then develop into spherical organoids of liver, intestine, and colon progenitors[1][2]. Whole gene insertion and reprogramming with lentivirus has been used before but is time-consuming and expensive [3]. CRISPRa is a powerful gene expression tool that allows one to overexpress specific genes through tailored single guide RNAs (sgRNAs) [4]. SgRNAs direct the CRISPR system to precise genomic locations, which facilitates the reprogramming of human stem cells. Endoderm cells, which generate tissues and cells like those found in the embryo and liver, are invaluable for studying disease mechanisms, understanding embryonic development and facilitating cell differentiation. This approach holds promise for advancing drug discovery and developing novel therapeutic strategies. My work focuses on the early stages of this project: creating sgRNA plasmids for the transcription factors required for liver organoid creation: FOXA3, HNF4a, CDX2 and GATA6 [5].

Shruti Roy | ROAR Lab

Image

• Mentor: Prof. Sunil Agrawal, ROAR Lab 
• Home Institution: Princeton University
• Hometown: Bridgewater, New Jersey

Development and Analysis of Multi-Dimensional Head-Trunk-Pelvis Models During Seated Reaching Tasks

Seated reaching tasks are incredibly common and crucial for ease and quality of everyday life. These tasks are even more important for individuals recovering from stroke or spinal cord injuries, which often limits patients to upper body movement initiated from seated positions, e.g. wheelchairs.

In order to develop support devices and physical therapy for spinal cord injury (SCI) patients, detailed movement and coordination data is needed to obtain insight into healthy movement patterns of the pelvis, trunk, and head during seated reaching tasks. Coordination data is especially useful for identifying segment
dependencies and guiding training protocols during physical therapy.

Verrels Eugeneo | CEAL Lab

Image

• Mentor: Prof. Brian A. Smith, CEAL Lab 
• Home Institution: Pomona College
• Hometown: Truth or Consequences

Fostering More Equivalent Access to Images for Blind and Low Vision Users

Blind and low-vision (BLV) individuals primarily access images through alternative text (”alt-text”) and image descriptions. However, alt-text can often be inaccurate or insufficiently detailed, failing to convey the full complexity of images. These complexities include subtle details such as colors, object sizes, positions, and depth. Providing BLV users with a way of understanding these intricacies can offer them with a richer, more nuanced understanding of images, which could be particularly valuable within educational contexts. In this work, we conducted a formative interview with a low-vision student about their experiences with images within educational settings. From this interview, we discovered that conveying the texture of materials is crucial to gain a better understanding of a digitized image. We then developed an audio-based prototype that communicates material texture information to BLV users interacting with images on touchscreen devices. We hope that this work can significantly enhance education within STEM fields, particularly in disciplines where understanding material properties and spatial relationships are essential, such as biology, geology, and engineering.

Vibhaakshayaa Satish Kumar

Image

• Mentor: Prof. Ioannis (John) Kymissis, 
• Home Institution: University of Washington
• Hometown: Bothell, Washington

Transient State Analysis of MicroLEDs

MicroLEDs are commonly applied in display technologies, but there exists other applications such as in Visible Light Communication (VLC), imaging, and distance detection. For VLC, the microLED must quickly switch from on and off states to modulate at high frequencies. Similarly, imaging and distance detection need the microLED to operate at high frequency and luminance. Electrical and optical charact-eristics such as rise and fall times (transition between 10% and 90% max voltage), and the cutoff frequency (signal drops below 50% its original voltage while LED is driven at a constant voltage) are useful to evaluate an LEDs performance. This project explores developing a method to capture these characteristics using an Indium Gallium Nitride (InGaN) microLED. The intention is to identify LEDs with reduced rise and fall times to improve data transfer capabilities and maintain at optimal luminance.

Vidushi Jindal | A2R Lab

Image

• Mentor: Dr. Brian Plancher, A2R Lab 
• Home Institution: Rutgers University
• Hometown: Edison, New Jersey

Optimizing Edge Robotics with YOLO, SORT, and TinyMPC for Enhanced Object Tracking and Control

Our research focuses on integrating YOLO for object detection, SORT for object tracking, and TinyMPC for model predictive control to enhance the efficiency and accuracy of robotic systems. Using the CrazyFlie 2.1, a small and low-cost robot equipped with the AI Deck and GAP8 IoT processor, we aim to improve real-time obstacle avoidance and dynamic trajectory tracking. This integration allows for better handling of complex trajectories and actuator constraints, promoting the use of affordable technology in robotics.

William Otoo-Mensah 

Image

• Mentor: Christine P. Hendon
• Home Institution: University at Buffalo
• Hometown: Long Island, New York

22 swine hearts were used to create the dataset

 In order to prevent complications, pulmonary vein isolation, or PVI, is a crucial part of radiofrequency ablation (RFA) therapy for atrial fibrillation. PVI relies on accurately identifying the junction between the pulmonary vein and the left atrium. The precision and dependability of conventional techniques like optical coherence tomography (OCT) and electroanatomic mapping are constrained. Based on near-infrared spectroscopy (NIRS) data, this work offers a unique method for improved pulmonary vein recognition using deep learning, more especially 1D convolutional neural networks (CNNs). 22 swine hearts were used to create the dataset, containing near-infrared spectra with reflectance measurements at 1024 wavelengths between 500 and 1100 nanometers. Three categories have been assigned to the data: normal myocardium (0) , lesions (1) and pulmonary veins (3), RFA lesions, and. Splitting the data by heart and remapping the labels (from label 3 to 2), restructuring the features for CNN input, and standardizing the data with the "StandardScaler" are important phases in the data processing process.

Emanuel Mendiola-Ortiz 

Image

• Mentor: Junfeng Yang
• Home Institution: Penn State University
• Hometown: Kennet Square, Pennsylvania

Developing Robust Audio Deepfake Detector

The proliferation of deepfake technology has raised significant concerns about the authenticity of audio content, necessitating robust detection mechanisms. Our research focuses on generating high-quality audio deepfakes and developing effective detection models. With the hopes it will help counterattack adversaries using it for bad intentions. This study aims to develop a real-world deepfake audio detector.

Nigel Cole 

Image

• Mentor: Dr. Treena Arinzeh
• Home Institution: Southern University A + M
• Hometown: Dallas, Texas

The regulatory effect of GAG-mimetics on activated human osteoarthritic chondrocytes

Osteoarthritis (OA) is a degenerative joint disease characterized by the progressive breakdown of articular cartilage. As OA progresses, it causes pain, stiffness, and loss of joint function, affecting 528 million people, or approximately 7% of the global population [1]. Due to the limited regenerative ability of articular cartilage, damage to this tissue can result in long-term functional impairment and pain. Current treatments for OA primarily focus on mitigating pain symptoms, and there is no drug capable of effectively halting the progression of the disease [2]. Tissue engineering is a biomedical engineering discipline that creates biological substitutes to replace diseased or damaged human tissue, e.g. cartilage tissue. Glycosaminoglycans (GAGs) are the main component in the cartilage tissue, essential for providing homeostasis in cartilage tissue [3]. Herein, we used a mimetic of sulfated GAGs derived from cellulose to examine its effect on OA chondrocytes. Electrospinning technique was used to produce fibrous scaffolds containing the GAG mimetic. The scaffolds were seeded with OA chondrocytes and cultured in the presence of pro-inflammatory molecules.