NOISE-OPTIMAL BINARY-SYNAPSE NEURAL NETWORKS

We examine the possibility of improving the performance of discrete-sy napse neural networks, functioning as content-addressable memories, by the inclusion of noise in their training procedure, and study the eff ects on the training itself. Pattern stability field distributions for optimized networks are illustrated for various levels of training noi se, including the noiseless, maximally stable, regime. We show that th e clipped Hebb rule is optimal in the high training noise limit, but t hat simulated annealing cannot be relied upon to identify a well defin ed optimal network for an arbitrary, finite, training-noise, in contra st to the case for continuous-synapse systems. Training by use of a co ntinuous-synapse network, whose synapses are subsequently clipped, is also addressed.