Procrustes: A Feature Set Reduction Technique
Abstract
This report explores the effectiveness of a new method for feature reduction and interpretation called Procrustes ordering. The investigation is performed using real data from eleven acoustic signal classes; hold-out studies are used to establish confidence in the conclusions reached. A significance test of the Procrustes angles based on a feature generation model is proposed. Additionally, an experimental statistical methodology is introduced to evaluate varying feature orderings derived from multiple trials using the same data set. Procrustes ordering is used in conjunction with a new variation of Fisher's method called smoothed Fisher. The variation is obtained by using a recently developed maximum likelihood trained probabilistic neural network, called Streit's Probabilistic Neural Network (SPNN), to provide smoothed estimates of the parameters defining the Fisher projection space. The results show that, on the given data set, Procrustes ordering used in conjunction with smoothed Fisher is an excellent method for feature reduction and interpretation. In addition, it is shown that Procrustes ordering is suitable for in situ application because it is fast and easy to compute on serial computers. Procrustes ordering, Probabilistic neural network
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 30, 1994
- Accession Number
- ADA282552
Entities
People
- Roy L. Streit
- Stephen G. Greineder
- Tod Luginbuhl
Organizations
- Naval Undersea Warfare Center