In this paper we study the conditions necessary for an autoassociative neural network to store structured patterns built from a predefined set of smaller configurations, which we can treat as words composed by letters. First, we show that no second-order noniterative local learning rule allows an efficient storage of words (but interneural couplings of an order at least equal to the number of letters in a word are neccessary). Besides, for the case of three-letter words, we show that any second-order coupling leads to frustration when two letters are presented. Then, we derive some properties of a neural network model based on a generalized high-order Hebb’s rule and coupled subnets, and we show that it solves efficiently the problem of storage and retrieval of words.
