CMStatistics 2016: Start Registration
View Submission - CMStatistics
B0661
Title: Fitting generalized multivariate Huber losses Authors:  Ami Wiesel - Hebrew University (Israel) [presenting]
Eli Peker - Hebrew University (Israel)
Abstract: A class of generalized multivariate Huber (GMH) loss functions is considered. Our goal is computationally efficient parameter estimation in linear models contaminated by non Gaussian yet correlated noise. We define a class of convex GMH loss functions with structured covariance matrices and flexible regularization parameters. This framework includes the classical weighted least squares and Hubers function as special cases. Next, we assume access to a secondary dataset of independent noise realizations, and we use this data to choose the best GMH function associated with the data. We use this fitted function to perform parameter estimation in a linear model by solving a simple convex optimization problem. We demonstrate the advantages of our proposed technique in heavy tailed correlated noise distributions.