Abstract
A Hopfield-type neural network that can store ultrametrically organised patterns with a finite magnetisation, or bias, is studied. Both the patterns and their ancestors are remembered in one network. A homogeneous field (or threshold) is included in the model. In zero field the (replica symmetric) mean-field equations map onto those of the Hopfield model, implying that transition temperatures, storage capacity, etc. are the same. By varying appropriately the field and/or a constant added to all the synaptic strengths, it is possible to climb up and down the hierarchical tree or to focus on a certain level in the hierarchy and thereby increase the capacity slightly.

This publication has 12 references indexed in Scilit: