Adaptive Subspace Self-Organizing Map and Its Applications in Face Recognition

Recently Kohonen proposed the Adaptive Subspace Self-Organizing Map (ASSOM) for extracting subspace detectors from the input data. In the ASSOM, all subspaces represented by the neurons are constrained to intersect the origin in the feature space. As a result, it cannot compensate for the mean present in the data set. In this paper we propose affined subspaces for constructing a set of linear manifolds. This gives rise to a modified ASSOM known as the Adaptive Manifold Self-Organizing Map (AMSOM). In some cases, AMSOM performs many orders of magnitude better than ASSOM. We apply AMSOM to face recognition. Since some face images may share a manifold due to similarities present in the images, we use a multi-layer neural network to divide the manifold into sub-areas each of which corresponds to a single class, e.g., a face class for Smith. Our experiment results show that this approach performs better than those obtained using the standard Principal Component Analysis (PCA) method.