Stochastic orthogonal and nonorthogonal subspace basis pursuit

Component analysis, or basis methods, provide a lower-dimensional representation of a given data set for compression, compaction, or discrimination. Stochastic basis pursuit addresses the problem of finding an optimal basis, either orthogonal or nonorthogonal, for improved pattern discrimination for pattern recognition applications. In this paper, the results of experiments performed with two stochastic optimization techniques as applied to the optimal basis problem are reported. The cost function is a quadratic discriminant function. Testing is done using three publicly available databases and ten-fold cross-validation. Empirical results demonstrate a twelve to fifteen percent average performance improvement over previous results.