Fall 2023

AI Ethics for Smart Health through Smart Watches
Lots of personal data can be obtained from wearable devices such as smart watches. This data can be used to improve health, for example to learn to detect health problems and to check whether people adhere to doctor’s exercise recommendations. This project will conduct a thorough study of the ethical issues in using AI systems in this domain, with recommendations of how AI systems for smart health should be designed with ethical considerations in mind.
AI/ML assisted fault detection in foundry processed devices
Highly accurate fault detection in foundry produced microelectronics is crucial to ensuring quality of devices that leave the foundry. However, current defect detection flows are human-centric, which produces a bottleneck. The objective of this project is to leverage recent advances in AI/ML to develop automated techniques that can 1) identify manufacturing defects in microelectronics using imagery collected at the foundry, and 2) determine whether the identified defect will impact the performance of the manufactured component.
Analyzing Open Source Software Ecosystems
Open source runs a lot of the world’s critical software systems, but there is much that’s unknown in how maintainers, developers and other parts of the software ecosystem function. Help us analyze a large corpus of open source data — both source code and patch conversations — to better understand them! We’ll study things like rise to influence, authorship styles, malware analysis, topic modeling and social network analysis!
Application of AI, ML and NLP in understanding and preventing a serious aviation safety problem in the US - Runway Safety
This project, which is co-advised by Dr. Yolanda Gil, will use AI/ML/NLP to understand root-causes of one of the serious aviation safety problem in the US - runway incursions. The Aviation Safety Reporting Systems, which is administered by the NASA and is an untapped treasure trove of textual data, will be used for this project.
Assessing the California Public Sector Job Market
Public sector institutions at local, state, and federal levels are facing an unprecedented hiring crisis in competition for new talent. Yet there is no systematic understanding of the needs and openings across these levels of government to inform stakeholders such as universities, community colleges, and high schools on the current and emerging hiring trends in what constitutes approximately 15-20% of the entire labor market. In this project, students will develop algorithms that continuously scrap relevant job sites used by these governments to assess both developed and emerging hiring trends by aptitudes, professions, entry-levels, mobility, location, and other important attributes. In so doing, the project will inform researchers in public policy, public administration, political science, and labor economics as well as practitioners in government and associated stakeholders.
Automated question type coding of forensic interviews
Question type coding is used in research on forensic interviewing to distinguish between best practice open-ended questions, and closed-ended and leading questions that interviewers are trained to avoid. Most research teams in the field rely on a time-consuming and labor-intensive method of question type coding whereby a researcher codes every question in the interview, and a second researcher codes a subset to demonstrate inter-rater reliability. We are currently working with a graduate of the Masters in Computer Science program at USC on a project exploring automated question type coding of forensic interviews with victims of child abuse. In collaboration with the student, we have trained a large language model (RoBERTa) to distinguish between question types based on a rudimentary classification system. In the next stage of the project, we are aiming to finetune the model and use zero shot and few shot prompting to make distinctions for which there is limited manually-coded data.
Bad Writing is "Fine": Tuning an LLM to Suggest Improvements

Prototype an approach to fine-tune a large language model (LLM) to help diagnose areas to improve a specific writing product. For example, scientific papers require consistent language but in creative writing variety matters. Proposed steps are:

  1. Writing Product: Coordinate with project mentors to choose a common and important writing product, such as a position paper or an academic conference. Identify/gather a rubric and a corpus.
  2. Inject Bad Writing: For each element of the rubric, develop prompts for generative AI to decrease the quality of writing based on the rubric (i.e., make it worse). This will form a training data set of the good example and version worse on certain characteristics.
  3. Fine Tune: Students will be expected to attempt to fine tune an LLM (e.g., LLAMA 2) based on this synthetically generated data
  4. Evaluate: Research if tuning suggests better domain-specific areas to improve.

This project aligns with ongoing work with the USC Generative AI Center.

Build a multilingual decipherment system
We will build a working system that can decipher a letter substitution cipher into 14 languages and beyond, based on https://aclanthology.org/2021.acl-long.561/ then apply it to languages it has never seen
Building a Platform for NFL Data Insights
Open source sports data such as the nflverse has lead to a massive increase in public sports analytics. But it’s still hard to process, subset, visualize and analyze this data. This project will build a general-purpose analysis platform and dashboard, similar to what many teams use internally. Using the nflfastr data, this platform will allow interested individuals to select the play parameters they’re interested in, and will provide relevant analysis, visualization and insight. Ideally, we’ll set up the dashboard on the internet, and open source the project, allowing others to expand the available datasets, analyses and visualizations.
Does Municipal Broadband Deliver as Promised? An examination of broadband pricing and household adoption in areas served by muni networks.
Broadband networks owned and/or operated by local governments (“muni networks”) are increasingly seen as a key tool to close the digital divide in Internet availability and adoption. There is however only anecdotal evidence about whether muni networks deliver on the promise of more affordable broadband in communities of little interest to traditional ISPs - typically disadvantaged communities. Taking advantage of the greater level of resolution in the new FCC broadband availability maps, this project will examine broadband pricing and adoption at the address level in areas served by muni networks, using a matched sample of comparable areas as a reference point. The goal of the project is to empirically assert whether muni networks are delivering on the promise of more affordable services, and whether this results in more household adoption than expected. The project is a component of an ongoing collaboration with digital equity advocacy organizations.