Representing taxonomical hierarchy of knowledge by structured Boltzmann machine
Taku Okuno, Yukinori Kakazu · 2002
The Boltzmann machine for content-addressable memory is structured to explicitly deal with taxonomical hierarchy of learned concepts embedded in its weights. It is realized by iterating three processes: extracting a subnetwork which represents an abstract concept, replacing it with a unit, and generating new layer by connecting it with the subnetwork. By this architecture, constraints on such hierarchy embedded in knowledge can be utilized to process knowledge to some extent. Its effectiveness is demonstrated by computer simulations.>