Fast Teaching of Boltzmann Machines with Local Inhibition

Abstract

Local clusters of lateral inhibition are modelled softly by supplementing the objective function (rather than by strict competition) for the Input_to_Output Boltzmann machine. This frustrates unwanted complexity of the induced internal representation of data. Furthermore, incremental teaching (shaping) incorporates new data with minimal retraining of previously learned data. Consequent learning rates are well over an order of magnitude better than the standard models, although maximum storage capacity is marginally reduced.