Neural-network performance assessment in sonar applications

Abstract

The authors focus on passive sonar applications which involve analyzing data with unknown signals. A general set of signal events (which are classified by a human aural analysis) are used for network training. The primary objective of the application is to discriminate between target and nontarget event categories. A ground truth (GT) and classical decision theory are used in assessing various neural-network (NN) classifiers operating on the DARPA Phase 1 data set. Changes in classifier operating point are shown to vary results between classifier type. These results show the importance of identifying the objective of the NN application before performance assessment is made.<>