A Recursive Procedure for Updating Quadratic Forms and Its Application to Feature Selection.

摘要

A survey of the problem of the selection of measurements is presented. The Mahalanobis distance is introduced as a selection criterion, and approached from an algebraic point of view. A new method for updating the inverse of a positive definite matrix is derived, and applied to the selection problem. Two interactive procedures are developed and used in experiments with the classic problem of Fisher, with high-resoluton radar images, and with aircraft photographs. Moments are used as features in the last two experiments. (Author)