Juan Felipe Gomez

Welcome! I am a fifth year PhD student in physics at Harvard University advised by Flavio Calmon and funded by the Department of Energy Computational Science Graduate Fellowship.

In the summer of 2022 and 2023, I worked as a member of the Princeton Plasma Physics Lab and Center for Statistics and Machine Learning at Princeton, where I was advised by William Tang. Previously, I was a Caltech undergraduate in Physics and grew up in Atlanta, Georgia.

My PhD research is focused on privacy-preserving machine learning, motivated by the following claims:

• Machine learning datasets often contain sensitive information, making it crucial to address privacy concerns and prevent adversaries from exploiting model outputs to identify individuals.
• Rigorous and interpretable privacy definitions are needed to ensure privacy guarantees, high model accuracy, and foster public trust.

Differential privacy has emerged as a framework that directly addresses these challenges. As such, my research focuses on advancing the theory and practice of differential privacy.

There are many surprising connections between differential privacy and physics. I (along with some amazing coauthors!) explored some of these connections in: The saddle-point method in differential privacy and Schrödinger mechanisms: Optimal differential privacy mechanisms for small sensitivity.

I have also worked on the trustworthiness of large language models, mainly in the role randomness plays in high stakes downstream tasks such as toxicity detection. While at Princeton, I worked on large-scale distributed pre-training of a foundation-scale model for clean-energy fusion applications. We presented some of our preliminary results at Super Computing 2024. See our presentation here!