Training noise adaptation in attractor neural networks
K. Y. Michael Wong, David C. Sherrington · Journal of Physics A Mathematical and General · 1990
The authors consider synaptic neural networks which minimise the output error of the stored patterns when the input patterns are ensembles of their noisy versions with overlap m t with the clean patterns. When m t is infinitesimally less than 1, the network automatically attains maximal stability, confirming the usefulness of training noises in enhancing memory associativity. When m t drops below 1, the field distribution has two bands for large m t , and one continuous band for small m t . Errorless retrieval is impossible for training noises of the order N 0 . With the increase in training noise, the retrieval overlap deteriorates, although memory associativity does increase for sufficiently low storage.