Learning regular and irregular examples separately

Oka and Yoshida (1992) proposed GLLL, a hybrid neural network architecture-of a global and a local learning module, and demonstrated its high accuracy and efficiency. In this paper the authors analyze learning in GLLL directing their attention to separation of examples between the two modules. The authors´ findings are: 1) regular and irregular examples are distinguished and learned separately by the two modules; 2) learning progresses in three stages, and overgeneralization occurs in the second stage; 3) outputs for highly unusual inputs are produced as if they are members of regular examples. These findings fit qualitatively human learning data reported by Marcus et al. (1990) and Pinker (1991).