Joseph Simonis, distinguished data scientist at Edward Jones, will discuss fitting functions to data and how the principles behind physics-informed neural networks (PINNs) can be used on real-world problems in gravitational physics.
"Every physics and math major has fit a line to data. Fewer realize that the basic idea, choose parameters of a function, is essentially the foundation of modern machine learning, including neural networks. In this talk, I'll start with that familiar problem: fitting a line, then a curve, and eventually a function with thousands of parameters and no straightforward geometric picture at all. Along the way, we'll see why smoothness and the chain rule suddenly become important, and how we can quietly replace 'fit a function to data" with "fit a function to a differential equation.' That simple shift is the idea behind physics-informed neural networks (PINNs).
"I'll then show what happens when you try to use a PINN on a real, unsolved problem in gravitational physics. The basic idea is surprisingly simple; making it actually work is considerably less so. We'll look at some of the practical challenges and the trial and error involved in overcoming them, and I'll close with a broader question: how much theoretical elegance do we need when solving real problems? In physics we care deeply about getting the theory right. In industry, we also need approaches that actually get us answers."
Join us on Friday, Sept. 11, for this exciting presentation from Simonis. Lunch will be available in Hayes 216 from 11:45 a.m. to 12:15 p.m., and the presentation will begin in Hayes 211/213 at 12:10 p.m. We hope to see you there!