Hierarchical associative networks

The authors study a modified Hopfield model of associative memory with a learning rule proposed by Personnaz et al (1985) for the special case of ultrametrically correlated patterns. The formula for the synaptic strength tells the 'teacher' how much stress to put on details compared to averages at each level. It is still a local rule if one assumes that the ultrametric correlation structure of the patterns is given a priori. The result is also given in terms of the Parisi function q(x). In a special limit they get the same result as Parga and Virasora (1986).