LEARNING GRAY-TONED PATTERNS IN NEURAL NETWORKS

The problem of learning multi-state patterns in neural networks is investigated. An analysis of the space of couplings (Gardner approach) yields the distribution of local fields, the critical storage capacity alpha-c and the minimum number of errors for an overloaded network. For noisy local fields the classification error is minimized if the local fields of the patterns are allowed to lie in intervals of finite width. A fast converging, adaptive learning algorithm is presented, which finds the coupling matrix of optimal stability.