2025 Summer Research Fellowship Projects
Ten projects are selected for the 2025 summer program. Projects are mentored by faculty from across all schools at Stevens. Information about each of the selected projects can be found below.
Investigating Honeypot Attacks
Faculty Mentor and Affiliation: Dr. Yasser Morgan, Department of Systems and Enterprises
Project Description: Honeypots capture non-signature based attacks. An important analysis of the origins of honeypot attacks would allow us to understand the socio-economic and political basis for attacks.
Desired Prerequisites: Knowledge of statistical and regression analysis
Can "Guided AI Use" Motivate Students to Learn Biology More Deeply?
Faculty Mentor and Affiliation: Dr. Woo Lee, Department of Chemistry and Chemical Biology
Project Description: Design and test novel learning activities, that integrate generative AI tools, to help students learn biology more deeply and at a high level. Visit here for examples of learning activities. https://woolee.substack.com/p/can-guided-ai-use-motivate-students
Desired Prerequisites: BIO181, BIO182, BIO291; CH115-118; Prompting skills; Statistical analysis skills
Generative AI for Computational Geometry
Faculty Mentor and Affiliation: Dr. Kishore Pochiraju, Department of Systems and Enterprises
Project Description: Generative design can be an effective tool for computational geometry and product engineering, enabling the automatic synthesis of designs satisfying user specifications. This research aims to develop geometric design tools that rely on Retrieval-augmented generation (RAG) to synthesize forms that satisfy multiple constraints. The tools address the need for automated, constraint-driven shape generation, particularly in applications with complex mechanical, manufacturing and geometric constraints. Using open-source research tools, students work on key techniques which may include 3D point cloud data segmentation, surface reconstruction, slicing, and solid generation driven by surface quality, smoothness, curvature consistency, and geometric fit requirements. The goal for the summer program will be to demonstrate geometry generation for a human wearable device such as a sports helmet.
Desired Prerequisites: Programming (Python); Computer Graphics; Interest in shape design
Harnessing the Power of Large Language Models for Portfolio Optimization
Faculty Mentor and Affiliation: Dr. Zonghao Yang, School of Business
Project Description: Information in financial markets exists in various modalities (e.g., numeric, text, images, audio) and is growing in volume. The ability to analyze information accurately and efficiently provides a competitive edge in investment. Recent advances in large language models (LLMs) have demonstrated exceptional capabilities in reasoning, extensive knowledge, and multi-modal processing. These models can analyze diverse information streams at speeds far exceeding human capabilities, suggesting their potential value for investment analysis. This project explores how to leverage LLMs' unique capabilities for portfolio optimization, addressing a significant gap between the models' demonstrated abilities and their practical applications in investment decision-making. Specifically, this project investigates the application of Large Language Models in portfolio optimization, leveraging their capabilities in processing extensive and diverse information sets across multiple modalities, including text (e.g., news articles, financial reports), images (e.g., product images, event slides), and audio (e.g., earnings calls).
Desired Prerequisites: Proficiency in Python programming is essential; Completion of at least one introductory-level course in machine learning is required; Students from diverse academic backgrounds are encouraged to apply, including those studying finance, information systems, financial engineering, computer science, and related fields.
AI Predictions of Molecular/Biomolecular Properties as Structural Fingerprints
Faculty Mentor and Affiliation: Dr. Yong Zhang, Department of Chemistry and Chemical Biology
Project Description: AI has been transforming the research field with two Nobel Prizes being recently awarded. However, AI tools in chemistry areas are mostly known for fast predictions but with significantly less accuracy than quantum chemistry predictions, especially for systems beyond normal organic molecules. Building on our previous strength in high accuracy quantum chemical predictions of many experimental spectroscopic properties such as NMR and IR parameters (structural fingerprints) which have successfully refined and determined 3D structures of important protein functional sites, we are developing new deep-learning (DL) based AI approaches to reach the accuracy toward the quantum chemistry level while keeping its speedy feature for some properties. In this summer, we will focus on certain spectroscopic properties that are useful molecular fingerprints for future structure determination of metalloproteins. Many undergraduate and graduate students of our lab have co-authored publications in prestigious and even world top scientific journals.
Desired Prerequisites: Python proficient with prior deep learning work experience
Identifying Physiological Patterns Associated with Improved Movement Performance
Faculty Mentor and Affiliation: Dr. Raviraj Nataraj, Department of Biomedical Engineering
Project Description: Our lab has developed platforms using virtual reality and instrumented wearables for motor (movement) rehabilitation of persons with brain or spinal cord injuries. We collected physiological data from skin-surface recordings of brain, muscle, and electrodermal activity during training with these platforms. We seek to identify trends within these data sets that can predict training conditions (e.g., task difficulty, level of automated assistance) that lead to improved performance. To meet this project goal, we propose having a student investigate the application of artificial neural networks and other data mining models as predictive filters to assess when persons are assuming mental and physical states that are more prime for optimal performance with our computerized interfaces. The impact of such an approach should advance rehabilitative technologies and support new pathways to adapt movement-driven interfaces (e.g., any VR training system) to optimize performance based on monitored physiological responses.
Desired Prerequisites: Experience with signal processing and AI models for data mining and/or predictive modeling
Algorithm Bias in Targeting for Online Advertisement
Faculty Mentor and Affiliation: Dr. Pallavi Pal, School of Business
Project Description: There has been a recent debate on how online algorithms might create discriminatory results for certain groups in society. In this paper, we look at how online algorithms decide who gets to see an online ad and if the demographics of an individual change the probability of seeing a specific ad. This is especially important in employment, housing, and job search categories. In this paper, we want to analyze what precisely leads to the bias in the algorithm and suggest ways to undo the bias. Online ads are sold through an auction mechanism.
Desired Prerequisites: Proficiency in R/python; Econometric background preferred but not required
Towards Safer Mixture-of-Expert Models
Faculty Mentor and Affiliation: Dr. Zining Zhu, Department of Computer Science
Project Description: The mixture-of-experts (MoE) architecture has been adopted by many of the highest-performing large language models including GPT-4, Mixtral, and DeepSeek. MoE supports an unprecedented parameters scale, achieving promising performance with high efficiency. However, the quest towards understanding and controlling the mechanisms of MoE models has unique challenges: The MoE models are usually too large to run on a single GPU, and the novel components in MoE models, e.g., the router modules, the expert modules, and the multi-head latent attention modules. Both properties prevent people from applying existing mechanistic interpretability frameworks to understand and control the MoE models, introducing unique opaqueness to the model users and developers. To address these challenges, we propose to develop novel technologies to explain and control large MoE models on multiple GPUs. To start with, we propose to incorporate tensor parallelization to one widely-used explanation method (Logit Lens) and one controlling method (Steering Vector), assessing and intervening on the mechanisms of large MoE models on tasks with profound societal impacts: machine deception, hallucination, and privacy breach. Then, we will develop other explanation and steering technologies for large MoE models. We will leverage the insights learned from the interpretability analysis of MoE models to improve the models' safety behavior, therefore aligning these high-performing models to societal needs.
Desired Prerequisites: Familiarity with pytorch, transformers, and NLP
Accelerating RLHF-based LLM Alignment with Serverless Computing
Faculty Mentor and Affiliation: Dr. Hao Wang, Department of Electrical and Computer Engineering
Project Description: Reinforcement learning from human feedback (RLHF) aims to align large language models (LLMs) with human preferences to deliver higher-quality responses in specific fields, enhancing reasoning abilities and accuracy for tasks like coding and mathematics, as demonstrated by DeepSeek R1 and ChatGPT. However, current RLHF methods are computationally expensive and require significant resources, which are often underutilized due to varying workload demands. Serverless computing, known for its scalability, flexibility, and efficient resource usage, can dynamically allocate or release resources based on needs. This makes it well-suited to optimize RLHF processes by accelerating alignment and reducing costs through better resource efficiency. The goal of this project is to explore how serverless computing can accelerate RLHF-based LLM alignment by enabling more cost-effective and scalable RLHF while enhancing the response accuracy of LLMs.
Desired Prerequisites: Pythong programming (a must); Basic understanding of reinforcement learning and LLMs; Experience with the Ray framework and distributed computing (optional)
Investigating Quantitative Analysis Techniques For Verified Federated Learning Systems
Faculty Mentor and Affiliation: Dr. William Eiers, Department of Computer Science
Project Description: Federated Learning (FL) is a machine learning setting in which many clients (e.g., personal computers, IoT, mobile devices, organizations) collaboratively train a machine learning model under the guidance of a global server, where the training data is decentralized and no client shares training data with another client. The distributed nature of federated learning brings unique difficulties not seen in traditional machine learning systems. Moreover, the communication efficiency of the system directly impacts the accuracy and performance of the FL system. In this project, we aim to quantitatively assess the impact of data reduction solutions or data compression solutions used in FL on the robustness and accuracy of the learned global model. To do this, we will investigate and apply quantitative analysis techniques for machine learning and neural network models to federated learning systems. Additionally, this project will investigate quantitative information flow/information theory techniques for measuring the amount of information lost due to communication deficiencies in order to develop quantitative measures that can provide sound and quantitative results about the data reduction approaches used. By the end of this project, we aim to develop the groundwork for creating a quantitative model for modelling the impact of data loss on the accuracy and robustness of the learned model within FL systems.
Desired Prerequisites: Basic programming skills; Proficiency with python; Familiarity with neural networks/ML training techniques; Familiarity with distributed computing systems
