Performance of Classifier Architectures With the RNADS Feature Space
Abstract
To evaluate the efficiency of the remote netted acoustic/seismic sensor array (RNADS) for classification, we must investigate the performance of various classification algorithms. Currently, the U.S. Army Research Laboratory (ARL) is developing an acoustic/seismic target classifier using a backpropagation neural network (BPNN) algorithm. Various techniques for extracting features have been evaluated to improve the confidence level and probability of correct identification (ID). For any given feature space, the BPNN creates complex boundaries in the hyperspace occupied by the feature vectors and only one hidden layer is required to create hyperplanes as decision boundaries; this, however, may not be the ideal classifier. Alternately, nonparametric and parametric classifier architectures are being investigated, since it is the mutual relationship between features and classifiers that allows the maximum recognition performance. Intuitively, we expect the BPNN to perform well, based on results from k-means analysis techniques. Using a hierarchical k-means analysis tool, we determined that only a few feature data clusters exist for each class. These 'feature pockets' may comprise about 40 percent of the training data in some instances and, in fact, have been suggested to be useful in a minimum distance classifier or beneficial in learning vector quantization. The nonparametric classifier architectures make no assumption about the statistics of the feature space distribution and instead, rely on the data to estimate classification parameters. They have advantages when the features are created using nonlinear processes with highly non-Gaussian statistics and allow flexibility in the tradeoff of computation, memory, training, and testing.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jan 01, 1999
- Accession Number
- ADA390190
Entities
People
- Mark Wellman
Organizations
- United States Army Research Laboratory