Constructively learning a near-minimal neural network architecture

Rather than iteratively manually examining a variety of pre-specified architectures, a constructive learning algorithm dynamically creates a problem-specific neural network architecture. Here we present an revised version of our parallel constructive neural network learning algorithm which constructs such an architecture. The three steps of searching for points on separating hyperplanes, determining separating hyperplanes from separating points and selecting separating hyperplanes generate a near-minimal architecture. As expected, experimental results indicate improved network generalization.