Using neural networks for dynamic reduction of the features space dimensionality

The paper presents a modification of the Principal Components Analysis (PCA) for dimensionality reduction. The peculiarity of the method is that linear subspaces are removed dynamically via special criterion, which utilizes variances of data projections onto eigenvectors´ estimates, whose coordinates are defined approximately. The latter gives rise to the method´s name: pseudo-PCA. Two algorithms based upon this method are proposed for dimensionality reduction. The first algorithm applies idea of the Generalized Hebbian Algorithm, while the second algorithm uses evolutionary approach for neural networks training. The algorithms are tested and studied on classification problems and obtained results show applicability and efficiency of the algorithms for dimensionality reduction even though linear subspaces for removal are defined approximately.