Four USC Students Receive CRA Outstanding Undergraduate Researcher Awards

Four USC students with diverse research interests spanning machine learning, practical AI applications, robotics, and data privacy have received national recognition from the Computing Research Association (CRA) for their exceptional contributions to computing research. 

Dutch Hansen was selected as a finalist for the prestigious CRA Outstanding Undergraduate Researcher Award, while Aryan Gulati, Hao Jiang, and Ryan Wang received honorable mentions for this award. 

The CRA Outstanding Undergraduate Researcher Award is one of the most prestigious recognitions for undergraduate computing research in North America. The award spotlights students who demonstrate exceptional potential in computing research fields.

Building fair and reliable AI systems

Dutch Hansen, a mathematics major, was recognized for his work in bridging the gap between the theoretical and practical aspects of reliable machine learning. In collaboration with his supervisor, Vatsal Sharan, an assistant professor of computer science, he investigated the fairness of machine learning algorithms as they make predictions about individuals.

Hansen’s work focuses on a concept called multicalibration, which describes the inherent uncertainty of machine learning predictions over subsets of a data distribution. Hansen’s findings were published at NeurIPS 2024 and hold potential connections to many areas, including causal inference and AI-assisted decision making.

Developing reliable machine learning models

Aryan Gulati, a computer science major with a minor in technology commercialization, earned an honorable mention for developing innovative out-of-distribution (OOD) detection techniques for text data. Mentored by Swabha Swayamdipta, a WiSE Gabilan Assistant Professor of Computer Science, and Antonio Ortega, a Dean’s Professor of Electrical and Computer Engineering, he tackled the critical task of identifying data that differs substantially from the training data used in machine learning models – an essential process to ensure model robustness and reliability. 

Presented at EMNLP 2024, Gulati’s research introduces an OOD detection technique that achieves comparable accuracy to existing approaches while being up to 11 times faster and consuming 97% less memory. As the former co-president of USC’s Center for AI in Society’s Student Branch (CAIS++), Gulati’s work reflects his commitment to developing machine learning solutions that benefit society. 

Training robots to handle complex tasks

Hao Jiang, who majors in applied computational mathematics and computer science, received an honorable mention for his research on ways to enable robots to perform complex manipulation tasks. Working in USC’s SLURM Lab under the guidance of Daniel Seita, an assistant professor of computer science, Jiang helped develop a learning-based approach that leverages reinforcement learning and simulation techniques to train robots for complex tasks. This approach successfully bridges the gap between simulation and real-world robot deployment. His work was recently published at the International Symposium of Robotics Research (ISRR) 2024. 

“It’s been fascinating to see how even small changes in a simulation’s design can have significant real-world impacts.” Hao Jiang. 

Safeguarding data creators’ privacy

Computer science major Ryan Wang also received an honorable mention for his work on data transparency in large language models. Collaborating with research advisor Robin Jia,  an assistant professor of computer science, and PhD mentor Johnny Wei, he developed a novel data watermarking technique that leverages injected randomness to enable statistical tests, providing guarantees for data creators to track whether their content has been used to train AI models. If a watermark is detected in a model, it serves as strong evidence that the corresponding data was included during the training process. 

Despite the computational challenges of validating his watermarks on large-scale models like ChatGPT, Wang found an innovative solution. “Since we couldn’t train these models ourselves, we instead leveraged the randomness already present in their training data — such as popular hash functions or GitHub commit IDs — as data watermarks. Using this approach, we demonstrated that our watermarks, when repeated at least 12 times in a 361-billion-token dataset, could be reliably detected by modern LLMs,” said Wang. 

Published on March 10th, 2025

Last updated on March 10th, 2025