Supervised Competitive Learning with Backpropagation Network and Fuzzy Logic

Abstract: "SCL assembles a set of learning modules into a supervised learning system to address the stability-plasticity dilemma. Each learning module acts as a similarity detector for a prototype, and includes prototype resetting (akin to that of ART) to respond to new prototypes. Here (Part I) we report SCL results using backpropagation networks as the learning modules. We used two features extractors: about 30 energy-based features, and a combination of energy-based and graphical features (about 60). SCL recognized 98% (energy) and 99% (energy/graphical) of test digits, and 91% (energy) and 96% (energy/graphical) of test letters