Generalization in the Hopfield model: numerical results

(Received 11 April 1991, accepted 16 April 1991)

Abstract
The generalization capability of the Hopfield model is studied numerically. There is a critical number of noisy examples that should be presented before the system grasps the pure patterns. Above , the generalization error falls off as a power of the presented examples. The critical number of examples increases exponentially with the noise level in the examples and linearly with the number of patterns to be learned.