A connectionist approach to learning discrete-valued patterns

Recently, there has been much interest in connectionist learning as alternative to symbolic learning methods. The purpose of the paper is two fold: 1) it considers briefly the fundamental differences between the symbolic and connectionist approaches to learning by relating them to a common mathematical framework; and 2) a model which can be categorized as both symbolic and connectionist is described. We describe how a connectionist network, which can be regarded as minimal is constructed. Nodes are created dynamically through a transformation process aimed at preserving a formal lattice structure. Every node formed is created in order to capture a specific correlation between a group of input/output values. The node is assigned a strength related to the prominence of the correlation. Dominant nodes associated with general (short) patterns enable the system to recognize patterns even though they contain noise. The inference and transformation is achieved through marker propagation and is based on node strengths.