Basins of attraction in a neural network model trained with external fields

Abstract

A neural network model with optimal connections trained with ensembles of external, discrete, noisy fields is studied. Allowing for non-zero errors in the storage, novel behaviour is observed, which is reflected in the model's retrieval map. Improvements in the model's content addressability is determined by comparing the maximum storage level at which there is a near 100% basin of attraction. The case presented here has the external field applied during training, during retrieval, and during both with statistically equal parameters. In all three the content addressability is improved over the external field network, with the equal training and retrieval fields case having the largest improvement. However, the apparent domination of the retrieval over the training field perhaps suggests this simple equality is unlikely to be the optimal relationship.