- Location: 4327 Stevenson Center TN
- Contact: Andreas Berlind
- Email: email@example.com
- Phone: 6469246688
- Website: https://as.vanderbilt.edu/physics/events/colloquium/
- Audience: Free and Open to the Public
Speaker: Professor of Physics and Data Science, New York University
We have spent more than a century building elaborate (computational) physical models, some of which are extremely successful. These include (for example) QCD, cosmological structure formation, and Earth climate. Despite their successes, these physical models disagree with the data in structured and repeatable ways, and we have enormous amounts of data to discover, test, and measure these disagreements. In many cases, we can combine good physical knowledge with good data to build a predictive model that is more powerful than any model built with either used on its own. Here I demonstrate these general points with the specific example of red-giant stars, where our data-driven model is now delivering more precise measurements of detailed stellar chemical compositions (the products of stellar nucleosynthesis) than any purely physical model. I will give comments on statistics and criticisms of machine learning that are relevant to many scientific domains.