Oscar Madrid Padilla, University of California, Berkeley
Main Content
 When
 22 January 2019 from 3:30 PM to 4:30 PM
 Where
 201 Thomas Building
 Contact Name
 Lorey Burghard
 Add event to calendar

vCal
iCal
The fused lasso, also known as (anisotropic) total variation denoising, is widely used for piecewise constant signal estimation with respect to a given undirected graph. In this talk I will describe theory and methods for the fused lasso. Two classes of problems will be discussed: denoising on graphs, and nonparametric regression on general metric spaces. For the first of these tasks, I will provide a general upper bound on the mean squared error of the fused lasso that depends on the sample size and the total variation of the underlying signal. I will show that such upper bound is minimax when the graph is a tree of bounded degree, and I will present a surrogate estimator that attains the same upper bound and can be found in linear time. The second part of the talk will focus on extending the fused lasso to general nonparametric regression. The resulting approach, which we call the Knearest neighbors (KNN) fused lasso, involves (i) computing the KNN graph of the design points; and (ii) performing the fused lasso over this KNN graph. I will discuss several theoretical advantages over competing approaches: specifically, the estimator inherits local adaptivity from its connection to the fused lasso, and it inherits manifold adaptivity from its connection to the KNN approach. Finally, I will briefly mention some of my other research directions.