Nonlinear data compression and representation by combining self-organizing map and subspace rule

Neural network learning algorithms combining Kohonen´s self-organizing map and Oja´s principal component type learning rule are studied for data compression and estimation of the tangent spaces of the feature manifold. The approach can also be thought as a combination of vector quantization and transform coding. The algorithms are derived from certain optimization criteria leading to the local analysis of the data. A novel application for clustering and classification of overlapping classes is presented. Simulations justify the performance of the algorithms