Return to Colloquia & Seminar listing
n-TangentProp: A Quasilinear Algorithm for Taking Derivatives of Neural Networks
Student-Run Research Seminar| Speaker: | Kyle Chickering, UC Davis |
| Related Webpage: | math.ucdavis.edu/~krc |
| Location: | 3106 MSB |
| Start time: | Wed, Jan 29 2025, 12:10PM |
Physics-informed neural networks are a numerical method which uses deep learning to compute the solution to differential equations. After enjoying a two-year period of feverish interest, recent work has critcized the field for not producing any substantial results, encouraging computationally inefficiency, and lacking theoretical justifictaion.
In this talk we discuss the basics of PINNs, some of the reasons for their inefficiency, and the author's recent work on partially addressing these inefficiencies.