About the maximum information and maximum likelihood principles in neural networks

Neural networks with radial basis functions are considered., and the Shanno n information in their output concerning input. The role of information-pre serving input transformations is discussed when the network is specified by the maximum information principle and by the maximum likelihood principle. A transformation is found which simplifies the input structure in the sens e that it minimizes the entropy in the class of all information-preserving transformations. Such transformation need not be unique - under some assump tions it may be any minimal sufficient statistics.