Knowledge representation in a multilayered Hopfield network

A scheme is presented for representing knowledge in a multilayered Hopfield network. Each layer of this network is composed of units (representing microfeatures) and pairwise constraints among them. Each layer uses a Hopfield algorithm to find coalitions of units corresponding to local minima in an energy landscape. Each distinct coalition is assumed to correspond to some interesting cluster of microfeatures that are mutually reinforcing. A unique grandmother unit at the next layer is associated with every emerging coalition and serves as both a label and an abstraction for that coalition. Such grandmother units are in turn microfeatures for the subsequent layer and are also configured in a Hopfield network with pairwise constraints. This network can exhibit reconstruction, focus of attention, and switching of attention by using a second set of feedback layers with downward connections.<>