Dystal: a self-organizing ANN with pattern independent training time

Abstract

As the difficulty of problems increase, artificial neural networks (ANNs) that use nonlinear optimization suffer from degraded execution speed, particularly with respect to learning time. Dystal is an ANN which does not suffer this degradation. Dystal is an ANN based on properties of associative learning found in biological neural networks. To verify these theoretical properties of Dystal, the authors implement Dystal on MasPar, a massively parallel machine. They show that the execution time is independent of both the separability of the patterns and the number of output units, and that the training time is linear with the number of patterns in the training data set. That is, the number of iterations through the training set to achieve learning is small and independent of pattern content or training set size. >