Abstract
When dealing with the curse of dimensionality (small sample size with many dimensions), feature subset selection is an important preprocessing strategy. This issue is particularly germane to the discrimination of class-labeled high-dimensional biomedical spectra as is often acquired from magnetic resonance and infrared spectrometers. A technique is presented that stochastically selects feature subsets with varying cardinality for discrimination by probabilistic neural networks. The results are benchmarked against two classifiers using the entire feature set both with and without feature averaging. The new technique had significantly fewer misclassifications than either of the benchmarks.