Trend Filtering, From Univariate to Graphs, and Old and to New: Neyman Seminar

Seminar: Neyman Seminar | February 24 | 4-5 p.m. | 1011 Evans Hall

 Ryan Tibshirani, Carnegie Mellon University

 Department of Statistics

This talk is centered around trend filtering, a relatively recent method for nonparametric regresson based on penalizing the L1 norm of discrete derivatives. We will discuss some of the unique features of this method that "make it work", and briefly cover extensions to additive models and graphs. We will finish by discussing connections to what are an old topic in numerical analysis---discrete splines---and then give some new results related to these connections with potentially exciting implications for bridging the discrete-time and continuous-time views for smoothing, interpolation, and representation.

 Berkeley, CA 94720, 5106422781