Parallel and Distributed Systems for Constructive Neural Network Learning

A construct ive learning algori thm dynamical ly creates a problem-specific neural ne twork architecture rather t h a n learning o n a pre-specified architecture. W e propose a parallel vers ion of our recently presented construct ive neural ne twork learning algori thm. Parallelization provides a computat ional speedup by a fact o r of O ( t ) where t is t h e number of training ezamples. Distributed and parallel implementa t ions u n d e r p4 using a ne twork of Workstat ions and a Touchstone D E L T A are examined. Ezper imenta l results indicate tha t algori thm parallelization m a y result n o t only in improved computat ional t i m e , but also in better predict ion quality.