Towards an information-theoretic approach to kernel-based topographic map formation

A new information-theoretic learning scheme is introduced for kernel-based topographic map formation. The kernel parameters are adjusted so as to maximize the differential entropies of the kernel outputs and, at the same time, to minimize the mutual information between these outputs. The learning scheme is based on infomax learning supplemented with a cooperative/competitive stage to achieve topographically-organized maps. As a potential application, we consider density estimation.