Deconvolution and Classification

• Activity Number: 217
• Type: Invited
• Date/Time: Monday, July 30, 2012 : 2:00 PM to 3:50 PM
• Sponsor: Section on Nonparametric Statistics
• Author(s): Raymond Carroll*+ and Aurore Delaigle and Peter Hall
• Companies: Texas A&M University and University of Melbourne and University of Melbourne
• Keywords: Classification ; Crossvalidation ; Deconvolution ; Measurement Error

Abstract

In a series of papers on Lidar data, magically good classification rates are claimed once data are deconvolved and a dimension reduction technique applied. The latter can certainly be useful, but it is not clear a priori that deconvolution is a good idea in this context. After all, deconvolution adds noise, and added noise leads to lower classification accuracy. While an alternative is to just use the observed data and ignore the convolution, we will present theoretical and numerical results that suggest a third method involving a carefully selected transformation other than deconvolution has better classification properties.