Generalization and Approximation Capabilities of Multilayer Networks
Yoshikane Takahashi · Neural Computation · 1993
This paper develops a theory for constructing 3-layered networks. The theory allows one to specify a finite discrete set of training data and a network structure (minimum intermediate units, synaptic weights and biases) that generalizes and approximates any given continuous mapping between sets of contours on a plane within any given permissible error.