CS+: Computer Science Projects Beyond the Classroom

Lead: Rong Ge

Description: "Recently there has been some works in how simple transformers can extract and store knowledge in their parameters (see e.g., https://arxiv.org/abs/2412.06538 and the references). In general, understanding how transformers store and recall knowledge is a key step in understanding how large language models work. In this project we hope to run some experiments on small scale transformers to form hypothesis towards understanding where knowledge is stored, how efficient can the storage be and more explicitly what is the mechanism of knowledge recall."

Goals/Deliverables: We will start by reading papers on this direction to get up to speed (this is an extremely active area and there might be more results between now and summer). Then we will design several experiments to see how transformers store knowledge in different settings, and form hypothesis. We will then design experiments to verify or refute the hypothesis and/or analyze the mechanisms theoretically (if it seems feasible). The final product is most likely a report although if everything goes smoothly it has the potential to turn into a paper.
 

Student Background/Prerequisites: 

1. Familiarity with deep learning and pytorch, basic math courses such as linear algebra, multivariate calculus and probability.
2. Prior experience as a UTA providing help in any course will be useful but not mandatory, as is rich first-hand help-seeking experiences as a student. 
    

Lead: Danfeng Zhang
 

Description: Trusted Execution Environments (TEEs) such as Intel SGX and ARM TrustZone provide hardware-isolated regions for running sensitive computations. While TEEs strengthen system security, they remain vulnerable to side-channel attacks: attacks that exploit information leaked through timing, memory access patterns, cache behavior, or other unintended channels.

In this project, students will investigate how side-channel vulnerabilities arise in TEEs and explore practical techniques for detecting, demonstrating, or mitigating such attacks. Depending on interest, work may involve building small proof-of-concept attacks, analyzing existing TEE implementations, designing defenses, or evaluating performance-security trade-offs. 

 Goals/Deliverables: 

1.  Learn how TEEs work and why they are vulnerable to side channel attacks.
2. Learn how to model system behavior, analyze system security, and mitigate vulnerabilities considering performance-security trade-offs.
    

Student Background/Prerequisites: This project is ideal for students with a solid background in operating systems and systems security.

 

Lead: Alberto Bartesaghi

Description: Cryogenic electron microscopes -or cryo-EM for short- allow researchers to peer at the microscopic shape of cellular proteins like never before. These machines blast proteins with a 300,000-volt beam of electrons so that highly sensitive detectors underneath can tease out their shapes based on the interaction that occurs. Being able to “see” proteins -life’s crucial building materials- can help determine how they work. Recognizing protein structure and function is essential for scientists trying to design better drugs to tackle some the world’s most devastating diseases, including HIV, cancer, COVID-19 and Alzheimer’s disease. A 300,000-volt electron beam is, however, extremely damaging to the proteins it is trying to image. To help protect the samples in the machine, researchers cryogenically freeze them to help maintain their integrity and use very low electron doses to prevent structural damage which results in extremely noisy images.

An emerging modality of cryo-EM called cryo-electron tomography (cryo-ET) uses computerized tomography principles to provide an accurate representation of the 3D molecular architecture of entire cells. The mining of the rich information contained in the native cellular environment is hindered by the crowded nature of cells populated by many different molecular species. The accurate detection of individual molecules in 3D is a critical step towards allowing the visualization of these molecular machines at high-resolution [1]. Motivated by recent advances in deep neural network approaches for molecular pattern mining [2], this project seeks to improve these methods to detect the position of challenging macromolecules within 3D images of frozen hydrated cells with the ultimate goal of understanding cellular function and disease at the molecular level.

[1] Liu, HF., Zhou, Y., Huang, Q. et al. In situ structure determination of conformationally flexible targets using nextPYP. Nature Protocols (2025). https://doi.org/10.1038/s41596-025-01218-9.

[2] Huang, Q., Zhou, Y. & Bartesaghi, A. MiLoPYP: self-supervised molecular pattern mining and particle localization in situ. Nature Methods 21, 1863–1872 (2024). https://doi.org/10.1038/s41592-024-02403-6."

 Goals/Deliverables: 

1. As part of this project, students will write computer code that will take as input 3D volumes of cells and automatically detect the location of multiple molecular species so they can later be extracted and used for high-resolution 3D visualization. Students will carry out the development in a dedicated high-performance computing (HPC) environment and at the end of the project will incorporate their code to the web-based application nextPYP (https://www.nextpyp.app). Depending on the progress they make, a research paper describing their approach and presenting results obtained on real datasets will be produced.
    

Student Background/Prerequisites: Skills required include Python programming (including PyTorch/TensorFlow frameworks), experience with image analysis, data science, or computer vision. Familiarity with HPC cluster environments is desirable.
 

Lead: Siqi Liu

We conduct research in theoretical computer science, focusing on the properties of random error-correcting codes. Error-correcting codes play a central role in data transmission and communication protocols by introducing carefully designed redundancy that allows errors incurred during transmission to be detected and corrected.

Polynomial codes are a well-studied family of error-correcting codes with many desirable properties, such as local correctability, list decodability, and local testability. In this project, we aim to understand whether these nice properties are preserved when random modifications are applied to these codes.
 

Goals/Deliverables: Students will become familiar with state-of-the-art research on error correcting codes, gain an understanding of several applications of these codes, and produce a written report or research paper on the topic.
 

Student Background/Prerequisites: A strong foundation in discrete math, probability, and computational complexity is highly desirable.
 

Lead: Xiaowei Yang

Description: Network topology information is useful in many ways. For instance, it enables users and network operators debug network issues such as routing inefficiencies and supports operators in planning future network expansions. However, topology discovery remains an open challenge in networking research. Recent advances in AI/ML can help address this open challenge. Many ISPs, including cloud providers such as AWS, Azure, and Google Cloud, publish schematic maps of their networks online.  In this project, you will investigate how to apply AI/ML techniques to discover network topology maps by crawling the public Internet and detecting the existence of such maps.
 

Goals/Deliverables: A large corpus of publicly available network topology maps.
 

Student Background/Prerequisites: Basic programming and scripting skills and understand how to apply AI/ML to identify network maps.