Neural discriminating analysis on preprocessed data

A neural discriminating analysis is developed in order to reduce significantly the dimension of input data spaces in pattern recognition problems. It searches for the restricted set of orthogonal directions in the input data space that classifies events with maximum discrimination efficiency. Improvements on the performance of the resulting compact discriminator are shown to be obtained when the discriminating analysis is performed on an image space that results from the application of a preprocessing map on the original input data space. As a case study, a particle discriminator for high-energy experimental physics is successfully developed, achieving efficiencies above 97%. The implementation of the method in a multiprocessor environment based on digital signal processor (DSP) technology is also addressed for online operation.