From thresholding dimension reduction to informative component extraction

Abstract

Generally, pattern recognition systems is designed with a relatively small amount of training data on parameter estimates; moreover, during test, finite sample size of testing data might bring trouble for the expected bias and variance of the models. Plug-in test statistics suffer from large estimation errors, often causing the performance to degrade as the measurement vector dimension increases. Thresholding dimensionality reduction method is briefly introduced first. An extension of this idea as informative component extraction is discussed for recognition system, especially in biometrics. A novel nominal model as the population distribution is introduced to reduce the dimension. Two different kind of benefits are obtained from this method are discussed. The modified test statistic is evaluated with a set of processed physical signals. Authentication testing for the exponential distribution is examined first. Special attention is paid to a high dimension Gaussian model with unknown mean and variances. Moreover, the performance is examined with different sample size.