Talks

Preference Learning & LLMs

I wrote a blog-esque document on training LLMs for a lab presentation. Per its audience, it devotes a lot of energy towards motivating RLHF and DPO from a statistical perspective, via Bradley-Terry.

ML Interpretability

Prof. Giles Hooker and I co-wrote two papers quantifying and reducing uncertainty in SHAP and other feature importance scores. These slides overview both papers. I have presented this work at the following venues:

• Center for Human-Compatible AI
• World Conference on Explainable Artificial Intelligence
• Apple, Data Analytics and Quality group (Hardware Technologies)

Epidemic Severity Rates

Many important epidemiological metrics, such as the case-fatality rate, relate two time series to one another. These “severity rates” are typicaly estimated with a ratio of aggregate counts, for example deaths today divided by cases L days ago. While seemingly reasonable, we show these ratio estimators have stark failure modes. We derive bias expressions and propose robust ML-based alternatives.

I have presented on the bias and estimation of severity rates to a wide range of audiences.

• Public health: CDC, California Department of Public Health, Santa Clara DPH
• Academic: Delphi Group, UCSF MINDSCAPE, Tibshirani & Hooker labs