How Can Machine-Learning Methods Help to Make Scientific Inferences?
Machine learning in the form of standard supervised classification algorithms is not all that useful or productive in the natural sciences, because the (effective) goals of these algorithms aren't very similar to the goals of scientific inquiry. However, the ML community has delivered great ideas and methods for building, fitting, and validating extremely flexible models. I argue that if we want to exploit the good things about ML but achieve truly scientific goals, we need to do two things: We need to augment or modify the (currently trivial) causal structure of the ML methods to represent our very strong domain-specific beliefs about how the data are generated. And we need to be careful to use ML methods only in the parts of our problems for which we don't care about the latent structure or parameters (that is, use them to model nuisances, not use them to do everything). I give examples from stellar astrophysics where adding ML components into larger causal models has created new scientific capabilities.
5:00pm - 5:30pm: Welcome Remarks
5:30pm - 6:30pm: Talk
Speaker: Professor David W. Hogg
After a PhD in Physics from Caltech and a few years at the Institute for Advanced Study, Hogg came to New York University in 2001, and was granted tenure there in 2007. His work at NYU has ranged from fundamental cosmological measurements to stellar dynamics to planet search and characterization. His work includes a significant engineering component. He spends a part of each year at the Max Planck Institute for Astronomy in Heidelberg, Germany, where he is a visiting member of the research staff, and a part of each week at the Flatiron Institute of the Simons Foundation, where he is a group leader.