Concept formation and statistical learning in nonhomogeneous neural nets

The authors present an analysis of complex nonhomogeneous neural nets, an adaptive statistical learning algorithm, and the potential use of these types of systems to perform a general sensor fusion problem. First, an extension to the theory of statistical neurodynamics is introduced to include the analysis of complex nonhomogeneous neuron pools consisting of three subnets. Second, a statistical learning algorithm is developed based on the differential geometrical theory of statistical inference for the adaptive updating of the synaptic interconnection weights. The statistical learning algorithm is merged with the subnets of nonhomogeneous nets, and it is shown how these ensembles of nets can be applied to solve a general sensor fusion problem.